See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/274056442

Intelligent Management Systems for Energy Efficiency in Buildings: A Survey

Article in ACM Computing Surveys · July 2014

DOI: 10.1145/2611779

CITATIONS READS

103 1,105

5 authors, including:

Alessandra De Paola Marco Ortolani

University of Palermo University of Palermo
61 PUBLICATIONS 1,146 CITATIONS 75 PUBLICATIONS 1,031 CITATIONS

SEE PROFILE SEE PROFILE

Giuseppe Lo Re Giuseppe Anastasi

University of Palermo University of Pisa
178 PUBLICATIONS 2,665 CITATIONS 208 PUBLICATIONS 8,853 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Giuseppe Lo Re on 25 May 2015.

The user has requested enhancement of the downloaded file.

Intelligent Management Systems for Energy Efficiency
in Buildings: A Survey
Article

Accepted version

A. De Paola, M. Ortolani, G. Lo Re, G. Anastasi, S.K. Das

ACM Computing Surveys

It is advisable to refer to the publisher’s version if you intend to cite

from the work.

Publisher: ACM

NDS LAB - Networking and Distributed Systems

http://www.dicgim.unipa.it/networks/
 A
Intelligent Management Systems for Energy Efficiency in Buildings:
A Survey
ALESSANDRA DE PAOLA, University of Palermo, Italy
MARCO ORTOLANI, University of Palermo, Italy
GIUSEPPE LO RE, University of Palermo, Italy
GIUSEPPE ANASTASI, University of Pisa, Italy
SAJAL K. DAS, Missouri University of Science and Technology, USA

In recent years, reduction of energy consumption in buildings has increasingly gained interest among re-

T
searchers mainly due to practical reasons, such as economic advantages and long-term environmental sus-
tainability. Many solutions have been proposed in the literature to address this important issue from com-
plementary perspectives, which are often hard to capture in a comprehensive manner. This survey article
aims at providing a structured and unifying treatment of the existing literature on intelligent energy man-
agement systems in buildings, with a distinct focus on available architectures and methodology supporting
a vision transcending the well-established smart home vision, in favor of the novel Ambient Intelligence
paradigm. Our exposition will cover the main architectural components of such systems, beginning with the
basic sensory infrastructure, moving on to the data processing engine where energy saving strategies may
be enacted, to the user interaction interface subsystem, and finally to the actuation infrastructure necessary
AF
to transfer the planned modifications to the environment. For each component we will analyze di↵erent
solutions, and we will provide qualitative comparisons, also highlighting the impact that a single design
choice can have on the rest of the system.
Categories and Subject Descriptors: A.1 [General Literature]: Introductory and Survey; I.2 [Artificial
Intelligence]: General; J.0 [Computer Applications]: Applications and Expert Systems
General Terms: Design, Algorithms, Measurement
Additional Key Words and Phrases: Building Management Systems, Energy Saving, Ambient Intelligence
ACM Reference Format:
A. De Paola, M. Ortolani, G. Lo Re, G. Anastasi, S. K. Das, 2013. Intelligent Management Systems for
Energy Efficiency in Buildings: a Survey ACM Comput. Surv. V, N, Article A (January YYYY), 35 pages.
DOI:http://dx.doi.org/10.1145/0000000.0000000

1. INTRODUCTION
Technological advancements stimulate novel products and services, which however in-
evitably result into intensive resource (e.g., energy) consumption. At the same time, global
R
awareness about their costs in terms of energy footprint is arising for the sake of environment
protection. In fact, current rates of worldwide energy utilization are no longer a↵ordable,
and therefore an increasing number of governments are promoting policies for sustainable
development and clever use of global energy resources. Their ultimate aim is a significant
reduction of the overall polluting emissions, and the adoption of suitable strategies for re-
ducing unnecessary energy wastes. Merely limiting the use of novel services would however
pose an unacceptable burden on the end user. Hence, rather than cutting services, the re-
search in the field of energy efficiency must focus on the optimization of resource usage yet
providing an adequate level of comfort for the users.
D

A proper characterization of energy consumption in an environment is necessary in order

to identify the main causes of wastes. A relevant fraction of worldwide energy consump-
tion is tightly related to indoor systems for residential, commercial, public, and industrial
premises; it has been estimated that residential and commercial buildings account today
for about 20% of the world’s total energy consumption. In this domain, the energy bill due
to environmental control – including heating, ventilation and air conditioning (HVAC) – is
dominating, especially in the developed countries [U.S. Energy Information Administration
2010]. Residential appliances consume about 30% of the total electricity consumption and

ACM Computing Surveys, Vol. V, No. N, Article A, Publication date: January YYYY.
A:2 A. De Paola et al.

produce 12% of all energy-related CO2 emissions; for instance, about 54% of the energy
consumption in US residential buildings is due to HVAC systems, and about 6% to artificial
lighting, while in commercial buildings HVAC and artificial lighting systems account for
40% and 15% of energy consumption, respectively [Perez-Lombard et al. 2008].
A wide variety of systems and methodologies have thus been proposed in the literature to
address the issue of reducing energy consumption in residential and commercial buildings.
These proposals are based on di↵erent yet complementary perspectives, and often take an
inter-disciplinary approach which makes it hard to obtain a comprehensive view of the state
of the art in the energy management of buildings.
The lack of a structured and unifying view over the available approaches and methodolo-
gies to be adopted during the design of such energy-aware systems was the main trigger for

T
undertaking the research underlying this survey. We specifically focused on the underlying
architectures and methodologies, as well as on the necessary techniques that go beyond the
well-established smart home paradigm, thus progressing toward intelligent Building Man-
agement Systems (BMSs), in accordance with the Ambient Intelligence (AmI) vision. The
ideal application scenario for AmI considers the user as the focus of a pervasive environment
augmented with sensors and actuators, where an intelligent system monitors environmen-
tal conditions and takes the proper actions to satisfy user requirements [Remagnino and
Foresti 2005]. AmI systems are characterized by a low intrusiveness, by the capability to
AF
adapt themselves to the users’ behavior and to anticipate their requirements. In the specific
context of a BMS for energy saving, this visionary goal becomes even more complex due to
the presence of contrasting goals, i.e., satisfaction of user requirements and minimization of
energy consumption.
Throughout this survey, we will identify the main components constituting a BMS;
namely, a sensory infrastructure for monitoring energy consumption and environmental
features, a data processing engine for processing sensory data and performing energy sav-
ing strategies, a user interaction interface subsystem, and an actuation infrastructure for
modifying the environmental state. For each component we will analyze di↵erent solutions
presented in the literature. Whenever possible, we will provide qualitative comparisons of
various approaches with respect to their specific features. We will also highlight the impact
that a single design choice can have on the rest of the system. To qualitatively evaluate
di↵erent BMS we will identify a set of relevant characteristics. In this assessment the end
users have a relevant role; besides being a↵ected by too strict energy-saving policies, users
might be hassled by other structural features, such as a set of invasive devices, or by algo-
rithmic features, such as learning methods that force them to have a continuous interaction.
R
In general we will refer to these aspects as the “user comfort”, and we will emphasize the
characteristics of di↵erent solutions in terms of scalability and complexity of the proposed
architecture, intrusiveness of the deployed sensory and actuating devices, and the resulting
impact of technology on user comfort.
Although some e↵ort has already been made in this domain [Cook and Das 2007; Froehlich
et al. 2011; Lu et al. 2010], there still exist significant research challenges. The focus on users
imposes great commitment on reducing intrusiveness of the deployed equipment, and pushes
toward the development of intelligent algorithms which do not require user-driven training.
This entails reducing the amount of a priori information that needs to be provided by in-
D

stallers, as well as limiting preliminary o↵-line training and minimizing explicit interactions
with users. Advances in this direction will allow for systems requiring minimal deployment
e↵ort, thanks to the lack of need for manual configuration, except for basic structural in-
formation. With such capability of self-configuration in place, it will be possible to devise
BMSs able to self-adapt to previously unseen scenarios; the unpredictability of the envi-
ronment may be due to variations in the performance of the actuators, modifications in
the habits or preferences of users, or to changes in climate. In a visionary perspective, it

ACM Computing Surveys, Vol. V, No. N, Article A, Publication date: January YYYY.
Intelligent Management Systems for Energy Efficiency in Buildings: A Survey A:3

is conceivable to think of self-organizing architectures able to modify their own software,

given a simple high-level description, with the aim of developing truly autonomic systems.
In order to carry out the research underlying this survey, we considered papers appeared
on journals and conference proceedings published by the most relevant research associations
and publishers, in the last decade, with a specific focus on the past four years, besides a few
less recent works which can be considered as milestones in their field. The works included
in the survey were selected on the basis of their relevance, scientific soundness and of the
presence of significant results for the field.
Before starting with the discussion of each of the mentioned issues, Section 2 introduces
the main approaches to energy saving in buildings, while Section 3 states the requirements
of a BMS and describes a reference architecture. Then, Section 4 discusses a number of

T
possible architectures proposed in the literature for practical BMSs. Section 5 provides an
overview of the di↵erent approaches to energy monitoring, of the available sensory devices for
measuring energy consumptions, and of energy models proposed in the literature. Section 6
focuses on technologies and methodologies for premise occupancy detection and learning
the user preferences. Section 7 surveys intelligent techniques, such as user profiling, pattern
detection and pattern prediction, that are instrumental to energy saving.
AF
2. GENERAL APPROACHES TO ENERGY EFFICIENCY
Four general approaches have been identified in the literature for reducing electrical energy
consumptions in buildings [Corucci et al. 2011], namely user awareness about energy con-
sumptions, reduction of standby consumptions, scheduling of flexible tasks, and adaptive
control of electrical equipments. We will briefly discuss each of them in the following.
Energy Consumption Awareness. The simplest approach to energy efficiency is to
provide appropriate feedback to the users about energy consumptions so as to increase
their awareness and encourage eco-friendly behaviors. User awareness has been leveraged in
many commercial and prototype systems such as Google PowerMeter [PowerMeter 2011],
Microsoft Hohm [Microsoft Hohm 2011], Berkeley Energy Dashboard [Pulse Energy Inc.
2013], AlertMe [AlertMe 2013], Cambridge Sensor Kit (CSK) for Energy [Taherian et al.
2010] and E2Home or Energy-Efficient Home [Ghidini and Das 2012]. Providing simple
feedback can valuably influence the user behavior [Darby 2006]. However, to reduce costs,
these systems typically provide only aggregate measures of energy consumption. Hence, they
do not allow to identify the specific device or behavior causing the highest energy waste.
Although user awareness is the basic approach to energy efficiency, its e↵ectiveness is quite
R
limited. Experimental studies carried out in a real building [Jiang et al. 2009b] have shown
that the sole provision of feedback is not sufficient to ensure significant energy savings in
the long term.
Reducing Standby Consumption. Another simple approach consists in eliminating or
drastically reducing energy wastes due to electrical appliances left in standby mode. Despite
its apparent simplicity, such an approach can produce significant energy savings. It has been
estimated that most consumer electronics (such as TVs, set-top boxes, hi-fi equipments) and
office devices (e.g., printers, IP phones) consume more energy in standby mode than in active
mode, as they remain in standby for very long times [International Energy Agency 2003].
D

The standby mode can be detected by monitoring the energy consumption of the specific
device. This requires a metering infrastructure which, of course, should have a very low
energy consumption [Jiang et al. 2009a]. Once the standby mode has been detected, the
device can be switched o↵. To this end, di↵erent strategies can be used to trade o↵ energy
saving for user satisfaction. The easiest way is to let the user decide about when to switch o↵
a device that entered the standby mode [Corucci et al. 2011]. A more sophisticated approach
consists of taking into account information related to the user presence, or in learning their
behavior.

ACM Computing Surveys, Vol. V, No. N, Article A, Publication date: January YYYY.
 A:4 A. De Paola et al.

Activity Scheduling. The widespread adoption of smart technology in many electrical

appliances enables the scheduling of their activity plans for energy optimization. In case
of constraints on the energy peak demand, or in the presence of time-dependent fares, ad
hoc strategies can be implemented for determining the optimal scheduling of energy-hungry
tasks that do not require user interaction (e.g., washing machine, dishwasher). The BMSs
proposed in [Corucci et al. 2011] and [The AIM Consortium 2008] allow the user to specify
the exact time period when a certain task is to be executed by a specific appliance. Such a
policy makes sense only when energy fares vary over time, but their variations are known a
priori.
Adaptive control. A significant fraction of energy is wasted due to an expensive use of
HVAC and artificial lighting systems, thus adaptive control on such systems is essential for

T
e↵ective energy management in buildings. The enacted policies should not negatively a↵ect
the comfort perceived by the user, otherwise the reaction would be an immediate rejection of
any automatic control, thus discarding the possibility of energy saving. The use of intelligent
techniques for user-presence detection and prediction is advised to adaptively tune the
activation time of electrical equipments, especially for those whose latency in bringing the
environment into the desired conditions is non negligible (see Section 7.1). Techniques for
learning user preferences may also be extremely useful for adaptively managing electrical
appliances, as they help to avoid overestimating user needs and just take into account their
AF
actual requirements.
3. THE REFERENCE BUILDING MANAGEMENT SYSTEM
To be e↵ective, the previous general approaches need to be implemented in an automated
BMS capable of enforcing an intelligent utilization of electrical appliances – with respect to
user preferences – so as to reduce electrical energy consumptions in the buildings without
negatively a↵ecting user comfort.
The focus on energy-awareness imposes several functional requirements, related to:
— sensing the environmental conditions (e.g., temperature, light intensity, etc.);
— monitoring energy consumption;
— modifying the environmental conditions;
— interacting with users, in order to send them notifications, and to gather feedbacks and
commands from them;
— detecting context (e.g., user presence, actions performed by the user);
— predicting the context;
— learning user habits and preferences;
R
— learning the energy consumption of appliances;
— learning the e↵ects of the actuators on the environment state;
— planning the optimal sequence of actions leading to energy saving while satisfying the user
requirements, according to system goals.
Moreover, it is highly desirable that the following non-functional requirements are also
fulfilled:
— low intrusiveness of the interaction with the user;
— low intrusiveness of physical devices and infrastructure;
— scalability with respect to the number of devices, areas, and occupants;
D

— extensibility after addition of new devices, thanks to proper abstractions;

— ease of deployment;
— software modularity;
— interoperability, both with respect to physical devices and with respect to other software
systems.
All the requirements listed above should find correspondence in the choice for the architec-
ture of the entire system. Event though the complexity of any significant BMS discourages

ACM Computing Surveys, Vol. V, No. N, Article A, Publication date: January YYYY.
 Intelligent Management Systems for Energy Efficiency in Buildings: A Survey A:5

Users
SaaS Abstraction

User Interaction Interface

Energy Plane
User Information Processor

Processing
Engine
Knowledge Manager

Raw Data Processor

Middleware

Sensory Infrastructure Actuation

T
(Energy and Ambient) Infrastructure

Environment

Fig. 1. Main components of the reference Building Management System for energy efficiency.

the use of a single architectural paradigm to capture all of its essential aspects, some general
considerations apply. For instance, most complex functionalities require the use of artificial
intelligent techniques, whose implementation needs to preserve unitarity of reasoning, which
AF
is not easily obtainable in a fully distributed environment. On the other hand, such systems
also rely on a close connection with the surrounding environment, which directly translates
into the need for a pervasive physical infrastructure.
According to this general scheme a BMS should comprise the following components (see
Figure 1), designed as suggested:
— Sensory and Actuation Infrastructure: constitutes the connection to the real world;
the sensor devices will comprise energy/power meters for measuring energy consumptions,
and sensors for acquiring environmental data (e.g., temperature, light intensity, etc.) and
context information (e.g., user presence), whereas the actuation infrastructure will consist
of all the physical devices in the building that can influence the state of the environment
(e.g., HVAC or artificial lighting systems);
— Middleware: connects the lower distributed infrastructure with the centralized process-
ing modules, dealing with the extreme heterogeneity of the devices nowadays available at
the physical level. It should be easily extensible with respect to the adoption of new de-
vices; an e↵ective approach is to design these modules in a component-based fashion. The
R
sensing and actuating devices in this case may be programmed as individual specialized
software components, exporting a common interface allowing for their aggregation into
more complex modules, to be used by the upper layers.
— Processing Engine: is constituted by specialized components implementing advanced
functionalities, such as targeting the energy consumption of appliances, and the e↵ects of
the actuators on the environment, learning user preferences, and recognizing their current
activities. A di↵erent architectural paradigm may be needed to tackle the inherent com-
plexity of the intelligent core, and to ensure its modularity. The approach suggested here
is to group the related software components into logical levels according to the provided
D

functionalities, following a 3-tier model mirroring the increasing level of abstraction dur-
ing data processing from the environment up to the user. Those components will likely
benefit from a centralized implementation.
— User Interaction Interface: provides interaction with the end users in order to send
them notifications to stimulate appropriate behaviors, and to gather feedbacks and com-
mands from them. A paradigm shift is necessary at this level; a fully distributed imple-
mentation is probably the wisest choice, and a smoother user experience can be provided
by developing the applications in the context of a SaaS (Software as a Service) infrastruc-

ACM Computing Surveys, Vol. V, No. N, Article A, Publication date: January YYYY.
A:6 A. De Paola et al.

ture. Besides favoring scalability, an immediate advantage for users would be that they
would only need very thin clients to access the system, so that the interaction may be
very natural and the overall system would result minimally intrusive.
Besides these components, the BMS should include software modules for energy aware-
ness. Unlike the other components, those modules should be spread across all layers for
better efficiency, as indicated by the Energy Plan in Figure 1.

4. BMS ARCHITECTURES
After introducing the general architecture of a BMS for energy efficiency, and briefly de-
scribed the main functionalities it should implement, we now survey a number of architec-
tural solutions proposed in the literature, and we analyze and compare them from di↵erent

T
viewpoints, such as architectural model (e.g., centralized vs. distributed), internal organiza-
tion (e.g., single layer vs. multi-layer), networking protocols, ability to support heterogeneity
in sensing technologies, and so on. Moreover, we compare di↵erent solutions with respect
to such software quality attributes, as modularity, extensibility and interoperability.

4.1. Plain Support for Energy Awareness

The first considered solution is a monitoring system based on Web-enabled Power Out-
AF
lets [Weiss and Guinard 2010]. Since the system is only intended to stimulate user awareness
to energy consumption, there is no actuation infrastructure. A web-based user interaction
interface is responsible for sending appropriate notification messages to the user. Each ap-
pliance is connected through a Power Outlet, i.e., a power meter that measures the energy
consumption of the appliance and sends the acquired information to a Gateway, using a
standard communication protocol (e.g., Bluetooth or ZigBee). By providing a RESTful
Application Programming Interface (API) [Richardson and Ruby 2007], the Gateway seam-
lessly integrates the smart power outlets into the Web. This allows users to easily access
their energy consumption through a Web browser. At the same time, it opens the system
to application developers. Such an approach would appear overly simplistic with respect to
the ideal BMS; complete focus on energy monitoring does not allow to relate consumption
to the current environmental state, nor does it allow to automatically control actuators. A
very fine-grained energy monitoring by unintrusive devices would, on the other hand, be
advisable for the realization of an ideal BMS, possibly based on a more complex architecture.
In the previous solution, the integration of power outlets with the World Wide Web is
mediated through an intermediate gateway. A further evolution consists of a direct integra-
R
tion of power meters, and possibly any other smart device, by exploiting the Web-of-Things
(WoT) paradigm. The latter is the extension of the well-known Internet of Things (IoT)
paradigm to the Web [Guinard et al. 2011]. Following the WoT approach, any smart object
(e.g., power meter, sensor, actuator) hosts a tiny web server. Hence, it can be fully inte-
grated in the Web by reusing and adapting technologies and patterns commonly used for
traditional Web content. An application framework for a smart home following the WoT
paradigms has been proposed in HomeWeb [Kamilaris et al. 2011]; this solution is charac-
terized by some degree of modularity since it is based on a web-service approach.
The solutions discussed so far rely on a centralized architecture and are able to support
D

heterogeneous embedded devices, thus providing a basic support for interoperability and
extensibility, even if these potential characteristics are not fully exploited.

4.2. Integration of Actuators and Environmental Sensors

A centralized architecture is also implemented by the iPower system [Yeh et al. 2009]
(see Figure 2). In iPower, a central server interacts with heterogeneous sensory and actu-
ator devices. Specifically, a Wireless Sensor Network (WSN) [Benini et al. 2006] is used to
monitor environmental conditions and to measure energy consumptions, while actuation is

ACM Computing Surveys, Vol. V, No. N, Article A, Publication date: January YYYY.
Intelligent Management Systems for Energy Efficiency in Buildings: A Survey A:7

Interaction with the user

BMS Architecture
Web user interface Smartphone

via Internet User

Central
Database Legend
Server
X10 controller
Serial Connection
Ethernet
X10 actuators
Power Line
Base station
Wireless Environmental
sensor
Energy sensor
Ambient Power
Measurements Consumption Actions

Environment Appliances

Fig. 2. iPower Architecture.

performed by X10 [X10 2013] devices connected to the server via Power Line Communica-
tion (PLC). Since wireless sensors have a limited transmission range they may not be able
to communicate directly with the server. Hence, to extend the system coverage, sensing de-
vices send their data to a local base station. Base stations are then connected to the server
through an Ethernet high-speed LAN. To manage heterogeneity with a sufficient degree of
abstraction, iPower relies on a multi-layer architecture. A Service Layer is defined, within
the central server, to provide an abstraction over the physical layer (as defined by the inter-
faces of the Open Service Gateway initiative (OSGi) platform [Gu et al. 2004]) according
to a service-based paradigm. On top of the Service Layer lies the intelligent system logic,
based on a rule-based reasoning engine. Rules are defined by the system administrator by
means of a high-level language and translated into service requests for the actuators. iPower
paves the way for an interoperable, modular and extensible solution. The iPower solution,
despite the adoption of slightly more intrusive sensors and actuators, allows to monitor en-
vironmental quantities, besides energy consumption; moreover, a hierarchical organization
vouches for medium scalability. However, it is our belief that a greater e↵ort is necessary in
terms of scalability, also with respect to the software components devoted to reasoning. The
rule-based engine guarantees a coherent source of reasoning, albeit a reactive one, and does
not support prediction. Finally, the actuating infrastructure appears too simple to enact
automatic control of actuators, and merely allows for tuning their supply power.
A centralized approach, similar to that used in iPower, is also considered by GreenBuild-
ing [Corucci et al. 2011]. Unlike iPower, GreenBuilding uses an unstructured (i.e., single-
tier) architecture and combines the energy monitoring and control functionalities into a
single infrastructure (i.e., power meters are also actuators). In addition, sensing devices for
environmental monitoring can be fully integrated in the same unique wireless infrastruc-
ture. A similar solution is also proposed in [Wen and Agogino 2008], where a prototype of
a wireless actuation module is presented that can be fully integrated within the monitoring
WSN. Using a single (wireless) infrastructure for monitoring and control lessens the burden
of technology integration. On the other hand, it reduces the flexibility in deciding the gran-
ularity of the monitoring/control process. As for iPower, this architectural solution aims
at the right direction, but does not appear fully adequate yet, due to the simple actuating
system and the lack of explicit support for intelligent reasoning.

4.3. Hierarchical Architectures for Improved Scalability

A more complex architecture capable of providing advanced support to heterogeneous sen-
sory and actuator infrastructures, is used in the Sensor9k system [De Paola et al. 2012].

ACM Computing Surveys, Vol. V, No. N, Article A, Publication date: January YYYY.
A:8 A. De Paola et al.

Central AmI Server

AmI AmI AmI Legend
Applications Applications Applications
Gateway
AmI Interface
Base station
BMS Architecture
Energy User
User

Direct
access
Consumption Presence Environmental
Profiler sensor
Manager Detector
Energy sensor
Physical Abstraction Interface
Workstation
Relay board
Wireless Serial Connection
RFId Reader

Ethernet Actuator

Wired Sensor
Power
Line

Ambient Users Power

Measurements Actions
Feedback Consumption

Environment User Appliances

Fig. 3. Sensor9k Architecture.

As shown in Figure 3, the system architecture is organized according to a three-tier model.

The Physical layer includes all the sensory and actuation devices, the Middleware layer is
composed of a set of building blocks for implementing basic services, and, finally, the Appli-
cation layer hosts the control logic and consists of various AmI applications. The inclusion
of a Physical Abstraction Interface ensures support against the heterogeneity of physical de-
vices, as it takes care of exporting higher-level abstractions identifying the basic monitored
units. Such system aims to address scalability with respect to the number of monitored
areas, which is typically the major limitation of centralized solutions, through a hierarchy
of gateways implementing di↵erent middleware functionalities. The sensory and actuating
infrastructure of such architecture presents a degree of complexity suitable for implement-
ing advanced energy saving strategies through adaptive control of actuators and user action
monitoring. On the other hand, the high dependence on the available appliances requires
ad-hoc design and development of sensors on actuators, thus limiting the generality with
respect to the adoption across di↵erent scenarios.
The idea of a hierarchical architecture with gateways interconnecting di↵erent technolo-
gies is also proposed in [Capone et al. 2009], in the context of the AIM project [The AIM
Consortium 2008]. The main goal is the construction of a bridge between a smart home and
the smart power grid in order to control the energy consumption of appliances. As shown
in Figure 4, the proposed architecture is a two-level hierarchy. At the topmost level lies
the AIM Gateway, whose task is the coordination of a set of Energy Management Devices
(EMDs) [Tompros et al. 2008; Tompros et al. 2009], each of which manages a number of
appliances. EMDs implement the actual control logic that includes power monitoring and
power control. These functionalities exploit the energy profiles of appliances, i.e., the as-
sociation between predicted energy consumption and operating mode of the appliance, as
outlined in Section 5.3. EMDs represent the element, within the AIM Project architecture,
allowing the abstraction from a specific physical layer; they may be implemented di↵erently
depending on the home network, but still expose the same communication API to the AIM
Gateway. The latter takes care of providing a unified interface for inter-EMD communi-
cation, both internally (i.e., for the system users) and externally (i.e., towards the energy
provider). Most of the energy management functionalities, such as appliances control and
user profiling, are hosted by the Gateway. By exploiting EMDs, the Gateway can learn the
operating mode of appliances and take actions to modify it, by deactivating the appliance

ACM Computing Surveys, Vol. V, No. N, Article A, Publication date: January YYYY.
Intelligent Management Systems for Energy Efficiency in Buildings: A Survey A:9

Interaction with the user

Home Communication Equipment

Wi-fi, Ethernet, … User

AIM Residential Gateway Legend

Energy consumption Intelligent power EMD (Energy

Monitoring & Control Consumption Metering Management
Device)
Device Virtual Environment
Workstation
Identity-based Security & Personalized Smartphone
User Privileged Service
Wireless
management Access Creation
router
BMS Architecture

IP Connectivity Washing
machine
Wi-fi, Ethernet, … Refrigerator

TV
KNX, Power line, Zigbee, ….
Home Cinemas
White Good A/V Equipment Sensors and actuators
embedded in appliances

Smart Appliances

Fig. 4. AIM Reference Architecture.

or by turning its operating mode to a less consuming one. AIM Gateways are implemented
with ESTIA gateways [The ESTIA Consortium 2008]. This specific technology was chosen
since these devices are based on the open services execution framework of OSGi [Lee et al.
2003; Gu et al. 2004]. Among the presented architectures, this is the one that allows for
the greatest scalability, extensibility, modularity and interoperability, due to its hierarchi-
cal architecture and also its partially distributed control logic with respect to high-level
functionalities. These considerations would locate AIM close to the ideal reference BMS.

4.4. Comparison among Di↵erent Architectural Solutions

The main features of the architectures described so far are summarized in Table I. Those
architectures may be regarded as representative of the three main types of approaches de-
scribed in this section: Web-enabled Power Outlet represents the simplest possible solution,
as it provides only a basic support for energy monitoring; iPower stands for solutions also

Table I. Comparison between di↵erent architectures.

Web-enabled iPower Sensor9k AIM
Power Outlets Architecture
Ambient Sensor Technolo- None WSN WSN, RFID, WSN, RFID
gies User action
sensors
Energy-aware Technologies Power meters Wireless Root/Wireless Energy Man-
power meters power meters agement De-
and actuators vices (EMD)
Architecture Model One-tier Multi-tier Multi-tier Multi-tier
Support for Heterogeneity None OSGi OpenGIS- OSGi
based
Control Logic Deployment Centralized Centralized Centralized Distributed
Interoperability Low Medium High High
Scalability Low Medium High High
Extensibility Low Medium Medium High

ACM Computing Surveys, Vol. V, No. N, Article A, Publication date: January YYYY.
A:10 A. De Paola et al.

Controlling every actuator /

Predictive behavior
Action Capability / Strategy Complexity
Ideal BMS

Sensor9k
AIM

iPower

WSN for Light GreenBuilding

Pure monitoring / "
Energy awareness

Web power outlets

Zero-tier Multi-tiers
Hierarchical Layering

Fig. 5. Comparison among di↵erent architectures with respect to two qualitative dimensions.
The assessed architectures are: Sensor9K [De Paola et al. 2012], AIM [Capone et al. 2009], iPower [Yeh
et al. 2009], GreenBuilding [Corucci et al. 2011], WSN for Light [Wen and Agogino 2008], and Web power
outlets [Weiss and Guinard 2010].
integrating di↵erent actuators and sensors, thus moving toward an interoperable, modular
and extensible BMS; Sensor9k and AIM Architecture represent hierarchical solutions that
try to meet the requirements we identified for an ideal BMS; the respective peculiarities in
their distributed approaches justify a deeper analysis of both proposals.
An interesting contribution to the definition of an appropriate level of abstraction for
heterogeneous devices is presented in BOSS [Dawson-Haggerty et al. 2013]; although this
work does not define a particular type of architecture, it proposes a distributed operating
system to manage heterogeneous devices in a BMS. For this purpose BOSS includes a
Presentation Layer Hardware, as an extension of sMAP [Dawson-Haggerty et al. 2010] and
a Hardware Abstraction Layer that allows developers to interact with devices via semantic
queries. In addition to the above architectures, and other similar ones not considered here
for the sake of space, a number of smart-home solutions have been proposed in the literature
that fall into the broad field of Smart Spaces. They are typically general-purpose solutions,
and do not specifically consider the goal of energy saving [Roy et al. 2007; Cook and Das
2004; Dawson-Haggerty et al. 2010; Helal et al. 2005]. Even if most of them could be extended
to address energy efficiency too, they are beyond the scope of this survey. A detailed review
of solutions presented in the literature for smart environments and ambient intelligence
systems is reported in [Cook and Schmitter-Edgecombe 2009; Cook and Das 2007; Sadri
2011]. Finally, an overview of wired and wireless communication technologies for building
automation can be found in [Qiu and Deconinck 2011] and [Gomez and Paradells 2010],
respectively.
As a final consideration, we can state that a requirement for an e↵ective BMS for res-
idential control is the presence of a rich sensory and actuating infrastructure, with good
scalability. Figure 5 compares di↵erent approaches with respect to two qualitative dimen-
sions, namely the hierarchical layering of their BMS architectures (and their complexity,
which influences scalability) versus the support for advanced actuation capabilities and
the resulting possibility of performing complex strategies of energy saving. According to
this qualitative analysis, the ideal architectural choice would fall close to the solution pro-
posed in [Capone et al. 2009]. A solution providing no support for actuation, but just pure
monitoring, only enables very simple energy saving strategies aiming at stimulating energy
awareness in users, but entirely delegating to the users the choice about modifying their

ACM Computing Surveys, Vol. V, No. N, Article A, Publication date: January YYYY.
Intelligent Management Systems for Energy Efficiency in Buildings: A Survey A:11

Systems for Energy Monitoring

Indirect Monitoring Direct Monitoring Hybrid Monitoring

Coarse-graned Medium-graned Fine-graned

Energy Monitoring Energy Monitoring Energy Monitoring

Fig. 6. Taxonomy of Energy Monitoring Systems.

T
habits. On the other hand, a varied actuating system, allowing for individual control of ac-
tuators via specific signals, enables complex strategies. Intelligent systems may exploit such
complexity, and plan the optimal sequence of actions that can satisfy users and minimize
energy consumptions.

5. MEASURING ENERGY CONSUMPTION

A precise and timely knowledge of energy consumption is an essential requirement for
enforcing any energy saving strategy. Hence, measuring energy consumption is the basic
AF
functionality of BMSs targeted to energy efficiency. The monitoring of energy consumption
in buildings indirectly provides information about the user habits and context [Beckel et al.
2013], in terms of the occupancy levels and of the kind of activities carried on by the
inhabitants [Kim et al. 2009b]. Thus, a relevant issue to be addressed during the design
phase of a monitoring system is how to monitor energy consumption with the required
precision and granularity while preserving the user privacy.

5.1. Systems for Energy Monitoring

Systems for energy monitoring can be classified according to di↵erent criteria, e.g., the type
of sensors they use, or the spatial granularity used for collecting data. With respect to sen-
sors, it is possible to distinguish between direct, indirect, and hybrid monitoring systems.
Direct monitoring systems use electricity sensors for directly measuring energy consump-
tion, while indirect systems infer energy consumption by measuring other quantities such as
temperature and/or noise. Finally, hybrid systems rely on both approaches. Direct monitor-
ing systems can be further classified into fine-grained, medium-grained and coarse-grained
systems, depending on the level of spatial granularity they use in collecting data about
R
electrical energy consumption. The taxonomy is graphically summarized in Figure 6.
5.1.1. Indirect monitoring. As expected, indirect monitoring systems are so called because
they do not use electricity sensors for measuring the energy consumption of appliances.
Instead, they indirectly infer information about energy consumption by measuring other
physical quantities that are somewhat related to energy consumption. This approach lever-
ages the fact that appliances typically a↵ect other observable environmental variables, such
as temperature, ambient noise, vibrations or electromagnetic field. Specifically, data pro-
vided by sensors are combined with a consumption model of the appliance in order to
D

obtain an estimate of its energy consumption. An indirect monitoring system is proposed

in [Schoofs et al. 2010], where a wireless sensor network is used to measure physical quan-
tities such as noise, temperature and vibrations. Each appliance is identified by a specific
pattern of its sensory measurements. For instance, switching on a kettle is associated with
temperature rising, a variation in vibration, and ambient noise. However, the paper does
not specify how the system is provided with the association between sensory patterns and
specific operating appliance; additionally, the simplicity of this approach limits its applica-
bility to feedback-based systems. Given the use of signature-based models for environmental

ACM Computing Surveys, Vol. V, No. N, Article A, Publication date: January YYYY.
A:12 A. De Paola et al.

measurements, this solution could be viable in centralized intelligence architectures, using

a distributed sensor infrastructure.
Whenever a model for appliance energy consumption is available, any system able of au-
tomatically detecting appliances could be used for performing indirect energy monitoring.
Those systems include the approach proposed in [Gupta et al. 2010], which exploits informa-
tion coming from the energy distribution network, other than explicit energy consumption.
The proposed approach analyzes high frequency electro-magnetic interferences generated
by the electronic devices powered through a switch mode power supply (SMPS) (used in
fluorescent lighting and in many electronic devices). Due to the limited applicability to a
specific class of actuators, such technology should be just regarded as complementary to the
energy monitoring system. For instance, this approach could be suitable for fully centralized

T
architectures, where the pervasiveness of sensory devices is minimal.
With reference to the ideal BMS, indirect energy monitoring systems are not suitable since
their use would require building models for actuators which, especially when environmental
measurements are involved, would have to be done in situ, thus being invasive for users,
not well generalizable, and consequently slowing down the deployment of the entire BMS.

5.1.2. Direct monitoring. Unlike indirect systems, a direct monitoring system measures en-
ergy consumption through ad hoc electricity sensors, typically referred to as power meters.
AF
The granularity used for direct energy monitoring spans from a single point of metering to
the monitoring of individual appliances. The rationale for using only a single power meter
is keeping intrusiveness at a very low level. These coarse-grained systems are referred to
as NILM (Non-Intrusive Load Monitoring) systems, or NALM (Non-intrusive Application
Load Monitoring) systems if the focus is on individual appliances. On the opposite end, fine-
grained systems allow to monitor individual appliances with a high precision but require
the deployment of a large number of power meters. Obviously, the granularity of monitoring
a↵ects the approach to the artificial reasoning carried on the collected sensory data and,
indirectly, also the possible energy-saving policies than can be used. NALM systems are
well suited for centralized architectures, with limited pervasiveness of sensory devices.
The NALM approach has been initially introduced by Hart [Hart 1992], who proposed a
system for measuring current and voltage at the root of the energy distribution network,
which is typically organized as a distribution tree. Variations in collected measurements,
after pre-processing, are compared to consumption profiles for the various appliances in
order to infer their activation or de-activation. Hart’s work has been seminal for a number
of subsequent works in the field of energy monitoring. Several approaches proposed in the
R
literature are based on the processing of measurements collected by a single point of mea-
surement [Laughman et al. 2003], and on the use of complex algorithms, such as Genetic
Algorithms [Baranski and Voss 2004] or Support Vector Machines [Patel et al. 2007] in
order to decompose the measurement into its components. However, some authors ques-
tion the e↵ectiveness of such disaggregation techniques in environments like office rooms,
where many loads are based on switched power supplies [Jiang et al. 2009b]. A survey of
disaggregation techniques for sensing energy consumption is presented in [Froehlich et al.
2011].
The alternative approach to a single point of sensing consists of monitoring energy con-
D

sumption at a finer grain. Brought to its ideal extreme, this approach would require a
detailed knowledge of every branch of the power distribution network, which, of course, is
not feasible in practice. Works presented in the literature only attempt to come close to this
ideal goal. The authors of [Jiang et al. 2009b] explore several practical techniques for ap-
proximately disaggregating the load tree using a relatively sparse set of power meters. The
possibility of relying on a fine-grain monitoring system is extremely advantageous for an
ideal BMS, as it allows to get information about consumption of specific appliances. Such
detailed monitoring, not available in the approaches mentioned so far, is useful to avoid

ACM Computing Surveys, Vol. V, No. N, Article A, Publication date: January YYYY.
Intelligent Management Systems for Energy Efficiency in Buildings: A Survey A:13

using consumption models for those appliances, thus eliminating the initial training phase
with its costs in terms of user discomfort. This solution is along the same line of heavily
decentralized architectures, such as those described in [Yeh et al. 2009; Weiss and Guinard
2010; Corucci et al. 2011].
Within the broad spectrum of granularity, there exists an intermediate position between
NILM systems and systems targeting each device individually. [Marchiori et al. 2011]
contains a proposal about measuring energy consumption only for those branches of the
energy distribution tree where some particular devices are connected. With respect to a
fine-grained approach, this method requires installation of fewer monitoring devices, while
compared to a NILM system, it allows to monitor the behavior of low consumption devices,
whose fingerprints would otherwise be overshadowed by high-powered devices. In particular,

T
this can be obtained by powering the latter class of devices on a dedicated circuit. Within
one specific branch, it is however necessary to use data analysis algorithms allowing for
a disaggregation of partial data. A similar approach is adopted in [Agarwal et al. 2009]
for monitoring energy consumption of buildings in a university campus. In such medium-
grained monitoring solutions, disaggregation techniques allow to obtain comparable results
to fine-grained, direct monitoring systems with fewer monitoring devices, and hence possibly
lower costs, but higher discomfort for users, due to the initial training phase for probabilistic
models.
AF5.1.3. Hybrid monitoring. Finally, a hybrid approach to monitoring, including both direct
and indirect parts, involves using both specific sensors for energy measurement (typically
in a single power meter at the root of the distribution tree), and indirect sensors for rec-
ognizing the operating status of appliances. An example of such a complex approach may
be found in [Kim et al. 2009a], where the authors propose a monitoring system based on
WSNs with magnetic, light and noise sensors, and including a power meter for monitoring
the overall energy consumption. The authors propose an automated calibration method for
learning the combination of appliances that best fits the collected sensory data and the
global consumption. The calibration method integrates two types of models. Specifically, a
model of the influence of magnetic field, depending on two a priori unknown calibration
parameters, is used for more complex appliances with many operating modes. On the con-
trary, appliances with fewer operating modes only require models associating the relative
consumption to each specific mode, which is estimated via the noise and light sensors. The
main disadvantage of this work is that the calibration is to be performed in situ and cannot
be carried out before the deployment since many unpredictable external factors may influ-
ence the measured environmental variables. It is worth pointing out that hybrid systems
R
are typically characterized by a coarse-grained direct monitoring of energy, with a single
sensor at the root of the energy distribution tree. This is usually coupled with a fine-grained
indirect monitoring.
5.1.4. Comparison of di↵erent energy monitoring systems. We believe that an ideal BMS that
is able to provide accurate description for actuator consumption without demanding ex-
cessively intrusive deployment, naturally calls for fine-grained direct monitoring. However,
when deployment costs are prohibitive, it is possible to reduce the number of used de-
vices and to rely on a disaggregation technique, starting from the branches of the energy
D

distribution tree.
Figure 7 reports a comparison of di↵erent energy monitoring systems together with some
of the previously discussed architectural solutions according to two qualitative dimensions,
namely the overall intrusiveness experienced by users, and the details on attainable moni-
toring. Values along the first dimension were attributed to assess both the intrusiveness of
deployed devices and the discomfort perceived by the users during the training phase, while
the second dimension is tightly related to the position of the assessed solutions within the
taxonomy depicted in Figure 6. Note that, as regards the sensory infrastructure, costs get

ACM Computing Surveys, Vol. V, No. N, Article A, Publication date: January YYYY.
A:14 A. De Paola et al.

fine grain
DIRECT
FINE‐GRAINED
Ideal BMS
High-Fidelity WSN

Web power outlets

Grain of energy monitoring
DIRECT
MEDIUM‐GRAINED WSN monitoring

Circuit monitoring

INDIRECT

ElectriSense
DIRECT
coarse grain

COARSE‐GRAINED
iPower

NALM approaches

intrusive non intrusive

Overall intrusiveness

Fig. 7. Comparison of di↵erent systems for energy monitoring, with respect to two qualitative dimensions.
The assessed energy monitoring systems are: High-Fidelity WSN [Jiang et al. 2009b], Web power out-
lets [Weiss and Guinard 2010], WSN monitoring [Schoofs et al. 2010], Circuit monitoring [Marchiori et al.
2011], ElectriSense [Gupta et al. 2010], iPower [Yeh et al. 2009].

higher as the systems get closer to the ideal one. When it is important to keep installation
costs below a given threshold, it will be necessary to trade part of the functionalities of the
final BMS for cost.

5.2. Devices for Energy Sensing

Besides the di↵erent approaches to energy monitoring, it is also necessary to consider the
technology to be used for creating the sensory infrastructure. In this section we will mainly
focus on the available technologies for energy sensing. However, we will not consider sensors
for environment monitoring, as they are beyond the scope of this survey, although they
are exploited into indirect and hybrid monitoring systems. A wide selection of sensory
technologies for energy sensing is currently available o↵-the-shelf. The choice of a given
technology directly a↵ects the complexity of the architecture supporting the monitoring
system and providing the integration with the rest of the BMS.
Coarse-grained direct energy monitoring systems exploit devices installed at the root
of the power distribution network; this represents an extremely simple and inexpensive
solution. Even though many devices have been designed to implement the “single point of
sensing” approach, it is still advisable to carefully assess the impact of the technology on
pre-existing premises, as well as its ease of deployment. In [Ruzzelli et al. 2010] the RECAP
(RECognition of electrical Appliances and Profiling in real-time) system is proposed for the
identification of energy fingerprints of appliances, which relies on a single wireless energy-
monitoring sensor clipped to the main electrical unit. The use of wireless communication
links avoids the need for the deployment of a communication infrastructure from scratch.
As regards the integration in the BMS architecture, NALM systems are more suited to
approaches with few pervasive devices and oriented to centralization.
Devices for monitoring energy consumption in medium-grained and fine-grained systems
are similar to the previously described ones, the only di↵erence being the density of deployed
devices. The monitoring system proposed in [Jiang et al. 2009b] exploits a sensory system
consisting of a network of heterogeneous wireless sensors, made of AC meters and light

ACM Computing Surveys, Vol. V, No. N, Article A, Publication date: January YYYY.
Intelligent Management Systems for Energy Efficiency in Buildings: A Survey A:15

sensors. The energy sensing nodes allow to collect active, reactive, and apparent power
measurements [Jiang et al. 2009a]; each node implements the IPv6/6LoWPAN stack, and
the wireless sensor networks is connected to other TCP/IP networks via a router. Other
solutions for AC power metering through a sensor network have also been presented in
the literature. In [Lifton et al. 2007] the “Plug” network is described, which is composed
of nodes fulfilling all the functional requirements of a normal power strip, and equipped
with an antenna and a CPU. An alternative solution is to use integrated sensor/actuator
platforms for energy consumption monitoring, through commercially available devices such
as WiSensys [WiSensys 2011], as suggested by the authors of [Corucci et al. 2011]. Currently
such devices are still expensive, so it might be convenient to allow for coarser granularity
of monitoring, by coupling a single energy sensor to a group of devices.

T
Monitoring and efficiently managing energy consumption of the sensing infrastructure
itself would deserve a separate discussion. This issue is extremely important in case of
sensor nodes powered by batteries with a limited energy budget, as in typical WSNs for
environmental and context monitoring. However, this topic is beyond the scope of this survey
(the reader can refer to [Anastasi et al. 2009] for a detailed overview on power management
in WSNs).

5.3. Models of Energy Consumption in Buildings

AF
Energy consumption models of individual appliances represent an alternative tool for energy
monitoring as they allow to estimate the overall energy consumption of buildings simply
relying on the knowledge about the status of each appliance (i.e., without requiring any
specific sensing infrastructure). For some devices, the corresponding energy consumption
in di↵erent operating modes can be retrieved from their technical specifications, such as
the Code of Conduct (CoC) edited by the European Commission, which, however, do not
cover the entire set of available appliances. A few years ago, in the context of the AIM
project [The AIM Consortium 2008], an experimental estimation of energy consumption
through some universal measurement methodologies was proposed [Foglar and S. 2008].
However, this approach is viable only when it is not possible to adopt the above mentioned
energy monitoring approaches, whose use could only be deferred to a later time. Recall
that the mere employment of energy consumption models carries with it the burden of a
preliminary training phase, which is likely to be done in situ, with consequent discomfort for
users; furthermore, they often require the support of a sensory infrastructure for recognizing
the activation state of actuators, just as in the indirect energy monitoring systems.
The specific methods to use depend on the particular class of appliances under consider-
R
ation. To this end, the following three classes are distinguished [Tompros et al. 2009]:
– Appliances whose instantaneous energy consumption is steady. This class includes ap-
pliances, such as lighting systems, whose energy consumption is approximately constant
during the entire operating cycle.
– Appliances whose instantaneous energy consumption is predictable. Such appliances are
characterized by a time-varying energy consumption, for which a predictive model can be
built. For instance, washing machines, dishwasher, and refrigerators fall in this class.
– Appliances whose instantaneous consumption is unpredictable. This class includes appli-
ances, such as HVAC systems, whose behavior is highly a↵ected by external factors (e.g.,
D

ambient temperature).
Appliances belonging to the first two classes can be easily modeled through a synthetic
profile, whereas those in the third class require an ad hoc infrastructure for measuring
energy consumptions.
Consumption models of single appliances can be used for run-time energy monitoring in
order to obtain an estimate of the current energy consumption of the building, or to decide
possible actions to be undertaken. They can also be used to steer the profiling process for
appliances currently in use, based only on the aggregated data about consumptions, as

ACM Computing Surveys, Vol. V, No. N, Article A, Publication date: January YYYY.
A:16 A. De Paola et al.

in [Prudenzi 2002], where consumption models and operating modes of appliances are used
to train a neural network in order to recognize the appliances currently in use. Exclusive
support for a posteriori analysis of appliance usage patterns makes such approach unfit for
use in a BMS which requires real-time knowledge about the state of actuators.
It is worth pointing out that the recognition of currently-active appliances only based on
the overall energy consumption is often a difficult task, due to the existence of appliances
with nearly identical power consumptions. The importance of simultaneously considering
both active and reactive power is highlighted in [Ruzzelli et al. 2010] which distinguishes the
di↵erent nature of various appliances. Specifically, an appliance can be classified as resistive,
inductive, or capacitive. Generally, appliances consume active power to carry on their tasks,
however, reactive power is also consumed due to the presence of inductors and/or capacitors

T
in their circuit. Therefore, a Unique Appliance Signature is proposed for combining several
pieces of information that can be collected from the electricity distribution network, such
as the active power and the power factor. The use of use active power, phase shift, current
crest factor and current signal harmonics is suggested in [Englert et al. 2013] so as to classify
appliances..
The problem of recognizing appliances in use starting from aggregate measurements is
also addressed in [Ducange et al. 2012], which exploits finite state machines based on fuzzy
transitions (FSMFT) and a novel disaggregation algorithm. FSMFTs are used to coarsely
AF
model how each type of appliance works. The disaggregation algorithm exploits a database
of FSMFTs for hypothesizing possible configurations of active appliances at each mean-
ingful variation of active and reactive aggregate powers. This approach is di↵erent from
previous NALM approaches because it exploits explicit construction of a model for energy
consumption of actuators; however both in terms of discomfort for users and monitoring
detail, it su↵ers from the same limitations as NALM systems.
The energy consumption models discussed so far may also be used to tune the run-time
behavior of the system. Furthermore, energy consumption models can provide valuable
information during design phase of a BMS. For this purpose, it is also possible to build
simulation tools for modeling the energy consumption of an entire building [Crawley et al.
2001; Ellis and Torcellini 2005].
6. ENVIRONMENT AND CONTEXT SENSING
BMSs are typically characterized by a wide set of functionalities, in addition to simple
energy monitoring, that allow to enable sophisticated energy-saving policies and automated
control of appliances. They include the ability to detect or predict the presence of occupants
R
within the monitored areas, as well as the ability to observe user actions in order to learn
their behavior. The practical integration of such functionalities into a BMS depends on the
availability of the corresponding enabling technologies. Moreover, integrating heterogeneous
sensors for environmental and context monitoring within the BMS sensor infrastructure
requires the availability of a software architecture that can abstract from low-level devices,
such as in [Yeh et al. 2009] and [Capone et al. 2009]. This section provides an overview of
the additional technologies needed for the development of advanced, intelligent BMSs.
6.1. Technologies for Occupancy Detection
D

The presence of users or their activity may be detected by di↵erent technologies, from
simpler ones (e.g., for motion detection or access monitoring) to more complex ones, e.g.,
Radio-Frequency Identification (RFID), laser scan or Global Positioning System (GPS).
In [Lu and Fu 2009] highly complex devices are developed for detecting user location in an
indoor environment. The employed devices fall into the category called Ambient Intelligence
Compliant Object (AICO), which includes apparently ordinary household objects pervad-
ing the environment, enriched with advanced functionalities such as transparent human-
environment interaction monitoring. The device specifically designed for user location de-

ACM Computing Surveys, Vol. V, No. N, Article A, Publication date: January YYYY.
Intelligent Management Systems for Energy Efficiency in Buildings: A Survey A:17

Table II. Technologies for presence and activity detection.

Sensor Description and References P R C I
Camera Indoor or outdoor deployment; requires ad-hoc software for user Vari- High Med. High
detection from their silhouettes. [Koile et al. 2003] [Kientz et al. able
2008] [Erickson et al. 2011] [Khalili et al. 2010]
Motion Based on passive infrared (PIR) technology for detecting peo- Low Low Low Low
Sensor ple/things by analysing the flow between source and sensor. [Lu
et al. 2010] [Gao and Whitehouse 2009] [Cook 2010] [Lu and Fu
2009] [Tapia et al. 2004] [Wilson and Atkeson 2005] [Mozer 1998]
[Agarwal et al. 2010] [Milenkovic and Amft 2013]
Motion Based on pyroelectric sensors, which detect human presence ex- Low Med. Med. Low
Sensor ploiting body heat. [Kobayashi et al. 2011]
Door Magnetic reed switches installed on doors/windows. [Lu et al. High Low Low Low

T
Sensor 2010] [Gao and Whitehouse 2009] [Cook 2010] [Tapia et al. 2004]
[Wilson and Atkeson 2005] [De Paola et al. 2012] [Kobayashi et al.
2011]
Floor Piezoelectric sensors placed under floor tiles for detecting the pres- High High High High
Sensor sure of a user stepping by. [Lu and Fu 2009] [Kidd et al. 1999]
Power Current and voltage sensors installed between an electric appli- High Med. Med. Low
Sensor ance and a power source for measuring current flow and voltage.
[Lu and Fu 2009] [Tapia et al. 2004] [Milenkovic and Amft 2013]
Object RFID sensors installed on each relevant object to interact with High High Med. High
AF
Id

Sound
Sensor
users wearing an RFID-detecting glove in their proximity. [Phili-
pose et al. 2004]
Indoor low-quality microphones may detect ambient noise, with Low Low Low Low
no recording of intelligible audio traces. [De Paola et al. 2012]
Precision (P):
Relevance (R):
Cost (C):
Intrusiveness (I):
accuracy of sensory measurements w.r.t. the monitored environmental quantity;
correlation between the monitored feature and the user presence or activity;
cost for designing/buying/deploying the sensors;
experienced intrusiveness for the user w.r.t. installing/using the sensors.

tection is called floor-AICO and consists of a floor tile with an embedded piezoelectric pad
for sensing the pressure caused by a user stepping onto the tile; the floor-AICO is connected
to a wireless sensor node which collects the sensory information and forwards it to a central
server. The complexity of this kind of device may limit its deployment only to specific areas.
Hence, even though the sensor precision is fully satisfactory, the overall system precision,
and its cost, may greatly vary depending on the chosen deployment density.
Table II, partially drawn from [Lu and Fu 2009], lists some of the available technologies for
R
user presence and activity detection. The choice of the most suitable one is not immediate, as
it is necessary to weigh up many factors, such as costs, expected performance, intrusiveness,
and privacy. The last two aspects, in particular, may prove critical for the acceptance of
the proposed BMS by end users. The issue of privacy safeguard when using video and
audio sensors has been discussed in [Lu and Fu 2009], [Campbell et al. 2002], [Tapia et al.
2004]. Privacy issues are also inherent when dealing with activity detection, regardless of
the employed sensory technology. As mentioned in Section 5, similar questions are raised
in [Kim et al. 2009b] in the context of energy monitoring.
Several solutions also deal with tradeo↵ between cost and performance: Infrared (IR)
D

motion sensors, for instance, are quite inexpensive and often deployed in households for
surveillance purposes; however, they usually convey inaccurate information. On the other
hand, door sensors are more accurate as regards the open/close status, but such information
is only partially correlated to the user presence.
As Table II shows, energy monitoring can be used to detect on-going activities, but
the obtained information is particularly useful for those activities that involve usage of
appliances with definite energy profile, as described in many existing works. As the cost of
the individual detector is not negligible, it might make sense to monitor only those electric

ACM Computing Surveys, Vol. V, No. N, Article A, Publication date: January YYYY.
A:18 A. De Paola et al.

Very precise /
Very relevant
Precision / Relevance AICO Ideal BMS

RFId Camera

information
fusion
Power Sensor

Motion Sensors (Heat)

Very Imprecise /

Motion Sensors (PIR)

Non relevant

Door Sensors Sound Sensors

intrusive non intrusive

Overall intrusiveness

Fig. 8. Comparison of di↵erent technologies for context sensing, with respect to the precision and relevance
of resulting information compared to the overall intrusiveness for the user. An information fusion process
could improve the quality of the obtained information while maintaining a low intrusiveness.

appliances whose usage appears to be correlated with specific activities or, alternatively, to
adopt activity analysis methods based on aggregated data about energy consumption.
The selection of the underlying technology needs a preliminary and accurate estimation of
the obtainable performance in terms of costs. The advantage of using more expensive devices
is that they generally provide more relevant information, while cheaper devices typically
produce noisy data, which additionally may be just marginally correlated with the observed
phenomenon (e.g., user presence or activity). On the other hand, the availability of cheaper
devices allows for the development of a truly pervasive sensory system, with a broader
coverage of the area of interest. We believe that performing multi-sensor fusion with plenty
of data, even if partially noisy, is the best choice since it allows for keeping costs low without
interfering excessively with user daily lives. While aiming at a low level of intrusiveness,
an optimal choice for the BMS could be the adoption of a wide set of motion and door
sensors such that the data coming from inexpensive devices are integrated with those coming
from other pre-installed devices, such as noise or power sensors. A graphical comparison
among di↵erent technologies for context monitoring is shown in Figure 8, which intuitively
highlights how information fusion may provide valuable support with low intrusiveness. A
detailed review of various information fusion methods for sensor networks is presented in
[Nakamura et al. 2007].

6.2. Technologies for Learning User Preferences

When user preferences are learned through explicit feedback, the system needs to include
human-computer interfaces in order to collect the end user opinion about the various en-
vironmental conditions. A trivial solution is to use touchscreens through which users may
communicate their preferences, possibly via a web-based interface; analogous solutions may
also be devised for personal mobile devices.
When considering implicit feedback, it may be necessary to add further sensing devices
to the BMS, thus performing discrete monitoring of their interactions with the actuators. A
very e↵ective solution may allow such interactions only through digital interfaces, as part
of the BMS itself. This is the choice adopted by iDorm [Hagras et al. 2007; Holmes et al.
2002], where the user can interact with the HVAC and lighting system exclusively via a

ACM Computing Surveys, Vol. V, No. N, Article A, Publication date: January YYYY.
Intelligent Management Systems for Energy Efficiency in Buildings: A Survey A:19

PC-based interface, a portable personal digital assistant (PDA) interface, a mobile phone
interface and a touchscreen interface. Such setting makes action monitoring extremely easy,
as the necessary information is readily available at the BMS. An analogous approach is used
in [Kolokotsa et al. 2005], where users can control actuators only through an ad hoc panel
that later sends all settings to the BMS. User-given settings are also stored into a smart
card in order to explicitly code the user preferences.
Albeit efficient, the option of forcing all user-actuator interactions to happen through
the BMS, thus forbidding any direct interaction, gets rid of the traditional control tools
(switches, remotes) which may make the whole system scarcely attractive to less experienced
users, such as elderly people. On the other hand, maintaining traditional ways of interaction
reduces the impact on the consolidated user habits, at the cost of a higher burden in terms

T
of the technology to be developed and installed, as well as of the overall BMS architectural
complexity. This approach has been adopted, for instance, in [De Paola et al. 2012], which
allows users to interact with actuators in the traditional way, but also includes specifically-
devised sensors for capturing the signals originated by electrical switches, remote controls,
and so on. In the scenario considered in [Vainio et al. 2008], a feedback is obtained whenever
the actuator state change was not caused by a command from the BMS; a similar approach
is also adopted in [Khalili et al. 2010]. In the Neural Network House [Mozer 1998] users
triggering the actuators generate implicit feedback, even though it is unclear whether this
AF
is realized by imposing the use of a BMS interface or via the addition of ad hoc sensors on
the actuators.
Besides being a source of implicit feedback, user actions also produce changes in the
environment state; it is thus necessary to handle the possible clashes between controls
generated by the users and by the BMS itself. The authors of [Vainio et al. 2008], for
instance, forcibly leave the actuator control to the end user in the 15 minutes following
any user interaction, thus avoiding an immediate overriding by the system, in case the user
preferences had not yet been perfectly assimilated.
7. INTELLIGENT SUPPORT TECHNIQUES
The energy savings policies enacted by BMSs may vary greatly and, depending on the com-
plexity of the adopted strategies, might require the use of artificial intelligence techniques.
Many works presented in the literature focus on the design of various intelligent functional-
ities, such as user profiling, predicting the occupancy status of the monitored premises, or
detecting the activity patterns of users.
This section describes possible intelligent functionalities to be added to a BMS for energy
R
efficiency, while focusing on the main AI approaches used for their actual implementation.
The same technique can be used for di↵erent purposes; likewise, di↵erent techniques may
prove useful for reaching one goal. This section first discusses the desired goals for designing a
BMS starting from its functional requirements. The correspondence between the underlying
techniques and their purpose is summarized in Table III.
7.1. Occupancy/Activity Detection and Prediction
Contextual information is fundamental in systems designed for energy saving in buildings.
The most relevant information is related to the presence of users in the areas of interest
D

and the activities carried on. Details about user presence may be used to switch the system
governing household into a low consumption regime, whenever users are absent. Such sys-
tems must be extremely reliable and reactive, and need to timely detect user arrival into
the monitored site in order to avoid the perception of an unacceptable comfort reduction,
or useless energy waste, thanks to the correct detection of when areas remain unoccupied.
The detection of on-going activities represents an evolution from user detection systems;
presence detection may indeed by regarded as a specific case of activity detection where the
“state” associated with a user may only assume two values: “absent” or “present”. Activity

ACM Computing Surveys, Vol. V, No. N, Article A, Publication date: January YYYY.
A:20 A. De Paola et al.

Table III. Matching between advanced functionalities for BMS support and the corresponding AI techniques.
AI Techniques Functionalities References
Rules Occupancy/Activity Detection [Mozer 1998] [Agarwal et al. 2010]
[Vazquez and Kastner 2011]
Data Mining Occupancy/Activity Prediction
[Das et al. 2002]
Occupancy/Activity Prediction [Mozer 1998]
Neural Networks
Learning User Preference [Choi et al. 2005]
Occupancy/Activity Detection [Thanayankizil et al. 2012]
[Lu and Fu 2009]
Bayesian Networks Learning User Preference [Kushwaha et al. 2004] [Chen et al.
2006] [Chen et al. 2009] [Hasan et al.
2009] [Lin and Fu 2007]
Hidden Markov Occupancy/Activity Detection [Lu et al. 2010] [Cook 2010] [De Paola

T
Models et al. 2012] [Duong et al. 2005]
[Milenkovic and Amft 2013]
Data Mining Occupancy/Activity Detection and [Rashidi et al. 2011]
Prediction
Fuzzy Logic Learning User Preference [Hagras et al. 2007] [Vainio et al. 2008]
[Kolokotsa et al. 2005]
Reinforcement Learning User Preference [Mozer 1998] [De Paola et al. 2012]
Learning [Khalili et al. 2010]
AF
detection requires finer detail both at the sensory level and at the inference level. If the
BMS includes a module for detecting/predicting activities carried on by users, it is possible
to adjust the actuators with respect to each activity in order to closely match the user needs
or to reduce, albeit partially, the overall energy consumption.
Most of the solutions reported in the literature are suited for integration into architectures
with centralized control logic, and present a varying degree of intrusiveness, depending on
the employed sensor devices, and on the discomfort perceived by users during the learning
process. A qualitative assessment of some of those works is presented in Figure 9, where
di↵erent systems are evaluated with respect to two qualitative dimensions. The first criterion
is the overall intrusiveness due to physical devices and to the user discomfort during the
learning phase; the second criterion is the complexity of the adopted energy saving strategy
enabled by an adequate set of actuators. There exist several approaches that adopt a slightly
intrusive sensory infrastructure composed of motion sensors, door sensors or audio sensors.
The complexity of the software infrastructure then makes up for it in terms of a wider set
of energy saving strategies.
R
7.1.1. Implementation Approaches. Although the desired goal may be merely occupancy de-
tection of the monitored areas, the data provided by simple sensors, such as motion or
door sensors, carry extremely relevant information with a sufficient degree of reliability. For
example, this approach has been proposed in [Agarwal et al. 2010], which suggests the use
of a simple rule-based approach for detecting user presence from motion sensors and door
sensors. As illustrated in Figure 9, such solution is still too naive to support predictive
energy saving, and only allows reactive behavior.
Besides detecting user presence, it is possible to implement a system for predicting it,
typically based on user behavior patterns expressed in a statistical form or through a set
D

of rules. The easiest way to do this consists of exploiting past sensory data to create an
environment occupational profile to be used as a static prediction model. Such a model can
be used to infer optimal configurations for the environmental control system (typically, for
HVAC systems), but such configurations do not vary over time and do not adapt to changes,
albeit minimal. A sample static environment occupation model can be found in [Gao and
Whitehouse 2009], as shown in Figure 10. In [Vazquez and Kastner 2011] the construction
of a statistical occupational profile is suggested by way of clustering techniques, used to
extract patterns from large amounts of data. The authors proposed to gather sensory data

ACM Computing Surveys, Vol. V, No. N, Article A, Publication date: January YYYY.
Intelligent Management Systems for Energy Efficiency in Buildings: A Survey A:21

Controlling every actuator /

Predictive behavior
Action Capability / Strategy Complexity
DYNAMIC OCCUPANCY/ NN House Ideal BMS
ACTIVITY PREDICTION

Observe AICO Smart Therm Softgreen

STATIC OCCUPANCY PREDICTION

SP Therm Occup Predict

Pure monitoring / "
Energy awareness

USER’S PRESENCE
OD BMS
DETECTION

intrusive non intrusive

Overall intrusiveness

Fig. 9. Comparison of di↵erent BMS exploiting the “occupancy detection” intelligent functionality.
The assessed BMS are: NNHouse [Mozer 1998], Observe [Erickson et al. 2011], AICO [Lu and Fu 2009],
Smart Therm [Lu et al. 2010], Softgreen [Thanayankizil et al. 2012], SP Therm [Gao and Whitehouse 2009],
Occup Predict [Vazquez and Kastner 2011], OD BMS [Agarwal et al. 2010].

from a simple set of door sensors, and to perform clustering on all daily 24-dimension pro-
files. Representative occupation profiles are the centroids of identified clusters. The authors
compare several clustering methods, such as fuzzy c-means, where the membership of inputs
to cluster is not strict but smoothed by a degree of membership. Figure 11 shows the result
of a fuzzy c-means performed on a set of one-dimensional data. Even though such solutions
enable predictive energy saving, they are not sufficient to let the prediction mechanism
adapt to changes in user behavior, which is why this approach is positioned far from the
ideal BMS.
Predictive models may also be realized by using neural networks, which allow inference
of an unknown function starting from a training dataset. In Neural Network House [Mozer
1998], neural networks are used, together with rules for occupancy detection, to predict
the binary occupational state of the monitored sites. Input data for the neural network is
provided by sensory readings coming from (binary) motion sensors. Despite being one of
the first works on this topic, this represents a good solution towards the ideal BMS logic.
Specifically, it uses a relatively intrusive sensor infrastructure, made up of sound sensors,

Unconditioned
Conditioned Conditioned
Time
Time Time
Miss Miss
Time Time

Time Occupancy Occupancy

0:00 Toff Tleave Tarrive Ton 23:59

Fig. 10. A typical schedule used by the Self-Programming Thermostat [Gao and Whitehouse 2009], on the
basis of a static environment occupation model.

ACM Computing Surveys, Vol. V, No. N, Article A, Publication date: January YYYY.
A:22 A. De Paola et al.

•• • ••• •• •• • ••• • • ••• • •• • •

Fig. 11. Possible resuts of fuzzy c-means clustering on one-dimensional data.

T
motion detectors and door and window status sensors; nevertheless it shows advanced ac-
tuating capabilities and adopts a complex energy saving strategy. Such features result in a
positive assessment which is represented by a position close to the ideal BMS as in Figure 9.
As this work was proposed a long while ago, it does not include an infrastructure for direct
energy monitoring, nor a scalable modular architecture.
An approach capable of dealing with the intrinsic uncertainty present in sensory readings
AF
and their partial correlation with the environmental features to be inferred is provided by
Bayesian (or Belief) Networks. An overview of several approaches using Bayesian Networks
(BN) for building occupancy detection is presented in [Dodier et al. 2006]. A straightforward
model for inferring the current occupancy of a site, or the on-going activities, involves the
use of statistical correlation of the instantaneous sensory information with the state of
interest [Thanayankizil et al. 2012]. In [Lu and Fu 2009] an augmented variant with a
multiple enhanced Bayesian Network (BN) is proposed that detects interleaved activities.
The proposed model makes explicit use of localization information obtained through the
smart floor for inferring the belief about a single activity. Moreover, the various sensory
readings are ranked according to a usefulness index, depending on the correlation of the
specific type of sensor with the activity to be inferred. In this way, the weight of a single
sensory reading into the data fusion process is tightly related to its relevance. This kind
of Bayesian network does not take into account past history of user behavior nor previous
sensory measurements. The above two solutions enable predictive energy saving strategies;
the first one [Thanayankizil et al. 2012] makes use of unintrusive information sources (such
as ID badges, WiFi signals, online calendars, device activity status, which are likely already
R
present in an office environment), which justifies its assessment in our figure. On the other
hand, the solution proposed in [Lu and Fu 2009] appears very intrusive, due to the massive
use of AICOs, which might be perceived by the users as unnatural objects, requiring heavy
modifications to pre-existing environments.
The most complete approach is certainly based on the use of a predictive model for site
occupancy and on its refinement on the basis of current sensory readings, so as to obtain
a dynamically evolving model according to actual environmental conditions. It is easily
noticeable how detection and prediction of user presence is, at this level, just a special case
of the more general problem of activity detection and prediction.
D

In order to include information about past states, one of the most common approaches
is the probabilistic one, through the use of Bayesian networks or, more specifically, Hidden
Markov Models (HMM). The easiest HMM is the traditional scheme, characterized by a
state variable influencing the value of a set of variables for sensory evidence, where the
probability to be in a given state only depends on the previous state. An example of this
scheme, adopted by [Lu et al. 2010], is illustrated in Figure 12. The proposed model uses
simple sensors to detect motion and doors status, deployed in the various environments of the
house and in front of the main entrance. The obtained information may be merged into the

ACM Computing Surveys, Vol. V, No. N, Article A, Publication date: January YYYY.
Intelligent Management Systems for Energy Efficiency in Buildings: A Survey A:23

Fig. 12. A simple Hidden Markov Model; the current activity of a user (yt ) a↵ects a set of sensory readings
(xt ); the transition from an activity to another one is represented by a probabilistic state transition model.

typical user behavior pattern through a HMM estimating the probability distribution of the
occupancy status of the house (Away, Active and Sleep). Together with [Thanayankizil et al.

T
2012] and [Mozer 1998], it represents the proposal more closely resembling the ideal BMS,
even though its focus is exclusively on the HVAC system, disregarding artificial lighting.
A similar model has also been proposed in [Cook 2010] and [De Paola et al. 2011]; in the
latter case, the HMM takes into account both the room occupancy level, and the estimated
number of the occupants.
Many modifications to HMMs with respect to the classical approach have been presented
in the literature. In [Duong et al. 2005], a Hidden Semi-Markov Model is presented which
explicitly accounts for the possible duration of the activities. Another variant for human
AF
activity detection considers hierarchical HMMs, which capture the natural complexity of
human behavior [Nguyen et al. 2005]. In [Erickson et al. 2011] is proposed the use of models
based on Markov Chains accepting images gathered from a set of cameras as input. Even
though the enabled strategies present sufficient complexity, their intrusiveness is consider-
ably high due to the use of cameras, which pose serious threats regarding user privacy.
Some proposals use HMMs jointly with data mining algorithms for predicting the most
likely sequence of actions. This is the approach followed in [Rashidi et al. 2011], which
proposed an activity discovering method based on data mining techniques for identifying
the most frequent sensory event sequences, which are presumably associated with actions
repeated over time. Such sequences provide the input for a set of multiple HMMs allowing
to detect the most likely sequence of actions. The employed data mining techniques rely
on a previous work [Das et al. 2002] which adopted a sequence matching approach for
detecting potential correspondences between the current sensory event sequence and the
system history in order to predict the most likely future actions.
As a final consideration, let us point out that the design of an ideal BMS must enable
predictive energy saving, with sufficient complexity, but low intrusiveness. To this end we
R
suggest the adoption of sound sensors, motion detectors and sensors for monitoring the
status of doors and windows. Moreover, the target sensor infrastructure will allow energy
monitoring and capturing the status of actuators in order to monitor user actions thus infer-
ring its preferences. Such sensory infrastructure will also be exploited to gather information
about user presence. Among the discussed methods, one of the most interesting proposals is
an HMM-based algorithm that predicts user behavior and copes with intrinsic uncertainty
in data.
7.1.2. Integration with Energy Saving Policies. Information about user presence is generally
D

exploited to actively modify the state of actuators. However, a few works are present in the
literature where such information is used only to provide users with contextual feedback.
For example, the authors of [Kobayashi et al. 2011] use it to trigger notifications activated
by simple rules, such as “if user presence is not detected, and lights are on, then send a
notifications to the user”. Control of the actuators is then delegated to the user, who may
interact with the system via mobile devices.
The simplest systems using information about user presence for environmental condition
controls are the reactive systems. Those immediately react to some specific sensory stimulus

ACM Computing Surveys, Vol. V, No. N, Article A, Publication date: January YYYY.
A:24 A. De Paola et al.

without relying on any model of the external world, or on any higher-level reasoning form.
An example is provided by lighting systems activated by motion sensors, or, within HVAC
systems, by the so-called “reactive thermostats” exploiting motion sensors, door sensors
or card key access systems. Such systems are often cause of a decay in user comfort since
it is possible that the energy saving mode is activated even when users are still present
within the site. Furthermore, a low energy-saving rate is obtained in the long run due to
the user habit of conservatively tuning the system in order to avoid an excessive reduction
of their own comfort. The conservative mode may correspond to a wide temporal tolerance
of inactivity before the energy saving mode is triggered, or to setting large setback values
to be used in the absence of users, so that no excessive discomfort is experienced when
the user presence is undetected. A preliminary analysis of energy consumption associated

T
with reactive thermostats can be found in [Gao and Whitehouse 2009]. However, it is shown
in [Agarwal et al. 2010] that under some conditions, an approach relying on them is sufficient
to obtain energy saving of about 10% – 15%. Anyway, reactive systems may be used as a
comparison baseline during experimental evaluation.
In case predictive models for site occupancy are adopted, slightly more complex policies
for energy saving can be implemented. A static predictive model is adopted in [Kastner
et al. 2010], which opted for a very straightforward strategy. An artificial model is used to
simulate the behavior of actuators in order to compute the necessary time to reach the pre-
AF
defined goal, given the present and the desired temperature conditions. Such information
about latency is used to reach the desired conditions at the time when the user is expected
to occupy the room.
The authors of [Gao and Whitehouse 2009] propose the Self-Programming Thermostat,
a system exploiting a static model of environment occupation to automatically create a
scheduling scheme for the high- and low-consumption modes of the HVAC system. The task
of defining the trade-o↵ between the expected comfort and the required energy saving is left
to the user, who needs to indicate (i) the maximum tolerance over the time interval during
which the user is present but the system (mistakenly) remains in the low-consumption
mode, and (ii) the temperatures associated with the two operating modes. The employed
sensors just detect motion and door status. More precisely, the goal is to determine the
schedule minimizing the “miss time”, on average, with respect to past statistics. As shown
in Figure 10, the “miss time” is defined as the duration of the interval when rooms are
occupied and the system is still in low-consumption mode. The scheduling efficiency is
defined as the reduction in the conditioning time with respect to the baseline schedule.
As already mentioned, the use of an accurate predictive model, able to quickly respond
R
to the changes in room occupancy state allows to use more “aggressive” settings for low-
consumption mode (e.g., a very low temperature) thus obtaining greater energy saving as
compared to reactive thermostats. In [Lu et al. 2010] a Smart Thermostat is proposed to
exploit the information about user presence, obtained via a HMM. Occupancy patterns
are used in the context of a hybrid approach aimed at minimizing the long-term expected
energy usage. The system tries to infer the optimal schedule for using the actuators so that
the area is heated whenever users are present, but not for unnecessarily longer periods, thus
balancing the expected costs of preheating too early and preheating too late. The HMM
is used to deactivate the HVAC system when the user presence is not deemed likely any
D

longer and when it is necessary to timely react to the user unforeseen re-entry in order
to recreate comfortable temperature conditions. The system also uses a statistical profile
of average exit and re-entry times, computed based on the past data, in order to pre-heat
the environment with the goal of minimizing the waste of energy in the long run without
sacrificing the user’s well-being. The predictive occupancy model for users and the timely
detection of their arrival may be exploited by two kinds of actuators: a low-cost system with
higher response time used to keep the environmental conditions consistent with the expected
occupancy patterns, and a more expensive one with respect to energy consumption, but with

ACM Computing Surveys, Vol. V, No. N, Article A, Publication date: January YYYY.
Intelligent Management Systems for Energy Efficiency in Buildings: A Survey A:25

Preferences Learning

Manual Input Automatic Learning

Static Learning Dynamic Learning

Explicit Feedback Implicit Feedback Explicit Feedback Implicit Feedback

Fig. 13. Taxonomy of systems for managing user preferences.

T
lower latencies, which may be used to bring the environment back to the desired state, after
unforeseen changes in the occupancy status.
A predictive model is used in [Erickson et al. 2011] for room occupancy in order to
tune ventilation and temperature setting of the HVAC system. For temperature control,
the authors propose a simple algorithm which activates the actuators to bring the room
temperature at the desired value only if the probability for that room to be occupied over-
comes a predefined threshold. For ventilation, the intensity is set proportionally to the
AF
expected number of occupants. An analogously simple model is proposed in [Thanayankizil
et al. 2012], where artificial lighting is tuned on the basis of a Bayesian model for room
occupancy. The proposed system adopts a lazy strategy for switching o↵ the lights if no
occupant is detected within a given time interval; a fast trigger is performed if user presence
is detected.

7.2. Learning User Preferences

Taken to its extreme, the optimal energy saving strategy consists of minimizing consumption
by switching o↵ all environmental control systems. Such a radical strategy is clearly unac-
ceptable as it does not take into account the secondary goal for any BMS, i.e., maximization
of the inhabitants’ well-being. The latter criterion steers environmental control policies, so
that trying to reach both goals concurrently allows us to find solutions representing an
acceptable compromise.
An ideal BMS would address the task of learning the mapping between user activities in
all time slots and the preferred environmental conditions. Such information may be used
for instance by the control system to infer the minimum acceptable comfort conditions that
R
also result in energy saving.
7.2.1. Implementation Approaches for Managing User Preferences. Systems for managing user
preferences can be classified according to the taxonomy presented in Figure 13.
User preferences may be manually defined by the system administrator, as in [Gao and
Whitehouse 2009] where the users specify the minimum and maximum temperature set-
points, or they may be automatically learned by the system. In the latter case, a more
realistic model for user preferences can be constructed, which more closely matches the true
user requirements and is free from potential evaluation errors by the system administrator.
D

Learning may be static or dynamic; in the first static case, the system is trained o↵-line,
even on real data, before the system is actually functioning and the user preference profile
does not change over time. An example which belongs to such class is a system that records
the interactions between the user and the HVAC system only for a training period, and
then computes the average temperature preferred by the user. A dynamic learning system
allows to modify the user profile as the system acquires up-to-date information. It is thus
able to adapt to preference modifications due to seasonal, mood or health changes of the
involved users. The previous example may be turned into a dynamic one, by extending the

ACM Computing Surveys, Vol. V, No. N, Article A, Publication date: January YYYY.
A:26 A. De Paola et al.

recording of user-HVAC interactions also during the on-line functioning, and by a periodic
computing of the preferred temperature through a mobile average. In order to carry on
the preference learning, two di↵erent kinds of feedback may be required from the users:
explicit or implicit. Explicit feedback is obtained when a user voluntarily ties a judgement
to a given environmental condition, for instance by using an interface (e.g., a touchscreen)
installed in the monitored area. In contrast, implicit feedback is obtained when the BMS is
able to perform non-intrusive user monitoring, by observing their interactions with the ac-
tuator or face expressions, and interprets the gathered information by associating it with a
hypothetical appreciation degree of current environmental conditions. For instance, a BMS
enriched with sensors on actuators and with the capability of perceiving the user presence
could reason as follows: if users interacts with the actuators, their preferences are expressed

T
by the chosen settings, while if users are present but do not interact with the actuators,
this fact implicitly signals that they accept the current environmental conditions. Explicit
feedback is more challenging to obtain, since it is unreasonable to force a user to express
their opinion about environmental conditions with excessive frequency, especially in the
case of actual deployments; moreover, such systems are indeed more invasive and might not
be well tolerated. Implicit feedback is perceived as more discreet, since the user may even
ignore to be monitored, but it generally requires installation of additional ad hoc devices
that have to be integrated with the rest of the BMS. Information coming from explicit and
AF
implicit feedback can be used both for static and dynamic learning. The works proposed
in [Fernández-Montes et al. 2009] and [Kushwaha et al. 2004] exploit statical learning of user
preferences. The former adopts explicit feedback obtained through a questionnaire compiled
by users about preferred lighting conditions; the latter adopts implicit feedback obtained
by recording the sequence of tasks performed by the user during a training phase and then
builds a Bayesian Network through a Case-Base Reasoning for coding the user preferred
sequence of tasks. A more common solution is to adopt explicit feedback within a dynamic
learning engine [Boton-Fernandez and Lozano-Tello 2011; Chen et al. 2006; Chen et al.
2009]. As an example, the authors of [Boton-Fernandez and Lozano-Tello 2011] propose
a system capable of recognizing activities performed by the user and which dynamically
learns frequent patterns in order to define a set of rules; the user is required to validate the
proposed rules and their acceptance or rejection is intended as an explicit feedback for the
learning engine. The prevalent approach for performing dynamic learning of user preferences
is to exploit implicit feedback [Hagras et al. 2007; Kolokotsa et al. 2005; Vainio et al. 2008;
Mozer 1998; De Paola et al. 2012; Khalili et al. 2010; Hasan et al. 2009; Lin and Fu 2007;
Choi et al. 2005]. An example is a BMS that evaluates the interaction of the users with the
R
actuators, as previously described, for obtaining an instantaneous evaluation of the adopted
policy and a learning mechanism based on a moving average for implementing a dynamic
behavior.
In our opinion, this solution is the one more closely matching the ideal requirements. Using
implicit feedbacks is very suitable in order to minimize user discomfort; dynamic learning
lets the system avoid the o↵-line learning phase, while on the other hand it is possible to
obtain autonomous adaptability to new scenarios. In order to support this functionality, the
sensory infrastructure needs to include appropriate sensors for detecting user interactions
with the actuators.
D

7.2.2. Implementation Approaches for Coding User Preferences. Methods for learning user pref-
erences may be divided into two great classes: tacit coding of preferences, by learning the
rules to be used to control actuators, and exhaustive coding, by associating the preferred
environmental conditions with every possible context (see Figure 14).
Tacit Preferences Coding. This category comprises all those approaches aiming at
the realization of “controllers”, with varying degrees of intelligence, whose rules reflect the
behavior expected on part of the user. A very simple approach has been proposed in [Choi

ACM Computing Surveys, Vol. V, No. N, Article A, Publication date: January YYYY.
Intelligent Management Systems for Energy Efficiency in Buildings: A Survey A:27

Preferences Coding

Tacit Coding Exhaustive Coding

Neural Network Rule Probabilistic

Preferred Value Preferred Range
Learning Learning Approach

Reinforcement
Fuzzy Rules SWRL Rules
Learning

Fig. 14. Taxonomy of systems for representing user preferences.

T
et al. 2005], where a neural network is trained to reproduce the association between context
information and the services selected by the users. Although the learning system is on-line,
hence dynamic, this method does not allow to consider imprecisions in the gathered data.
A di↵erent approach consists in a tacit coding by learning the set of rules to be performed
by the BMS, in order to satisfy user preferences. This subclass includes approaches based
on fuzzy controllers that adapt their rules based on the received feedback. Fuzzy controllers
AF
are largely employed in home automation, since they realize quite robust systems even in
the presence of uncertain or imprecise data. The use of a fuzzy system is suggested in [Doc-
tor et al. 2005; Hagras et al. 2007] which exploits three-dimensional membership functions
explicitly including a footprint of uncertainty. The use of a fuzzy controller exploiting con-
textual information has also been proposed in [Vainio et al. 2008] which exploits additional
information, such as user preference and the time of day, besides information related to
environmental conditions, e.g., lighting. In both approaches, the set of rules learned during
o↵-line training are successively tuned during the on-line usage, on the basis of implicit
feedback obtained from the users.
Reinforcement learning (RL) [Sutton and Barto 1998] is another technique suitable for
dynamically learning the rules that better match user preferences, which are thus indirectly
modeled. Each action, performed in a specific situation, is associated with a quality value
which is dynamically determined as a function of the rewards produced by the reaction
of the environment. When RL is used to learn the user preference, a negative reward is
typically associated with the last performed action if the user, operating on the actuators,
overwrote the setting proposed by the system; this approach is used in the Neural Net-
R
work House [Mozer 1998]. In [Khalili et al. 2010] the Hierarchical Reinforcement Learning
(HRL) is adopted to understand user preferences. This technique aims to address the slow
convergence shown by traditional RL systems. Some works in the literature propose the
adoption of ontologies for representing the rules for controlling the actuators, thus pro-
viding a tacit coding of user preference. IntelliDomo [Boton-Fernandez and Lozano-Tello
2011], for instance, expresses control rules as SWRL (Semantic Web Rule Languages). The
system exploits data mining techniques to discover frequent and periodic patterns in the
user behavior. Once those are found, they are coded into rules. Learning is dynamic and
on-line, and the user can trigger a change in the rules in any instant, by providing feedback
D

about the system performance.

In order to learn user preferences, it may be extremely useful to adopt a probabilistic
approach when considering potential uncertainty in the collected data, as well as possible
slight, sudden and irregular modifications in user behavior. In this case, Bayesian networks
(BN) represent an invaluable tool as they allow to express probabilistic connections among
di↵erent features of the world. A BN is used in [Kushwaha et al. 2004] to identify the se-
quence of actions to perform on the actuators in order to carefully imitate the user past
behavior. Learning is carried on by recording user actions, which represent implicit feed-

ACM Computing Surveys, Vol. V, No. N, Article A, Publication date: January YYYY.
A:28 A. De Paola et al.

back for the system, and by performing a Case-Based Reasoning (CBR). Relationships
between actions are represented via conditional probabilities tables. Learning is automatic,
but static, since no on-line adjustment is performed over the user profile. In [Chen et al.
2006] BNs are used for coding user preferences by a semi-supervised learning approach ex-
ploiting both labeled and unlabeled data. Labeled data is obtained from explicit interaction
with the users, who are explicitly interviewed whenever the system perform environmental
control actions that are not validated by the users themselves; such data is used to learn
and modify the structure of the BN. Unlabeled data is used during the system normal func-
tioning and records the association between the environmental conditions and the actuators
state; such data is used to update the conditional probability tables. A similar approach is
adopted in [Chen et al. 2009] and in [Yeh et al. 2011]. The works in [Hasan et al. 2009; Lin

T
and Fu 2007] also make use of BNs to code multiple user preferences. Such a task is partic-
ularly challenging because more than one user interacts with the same appliances, making
it necessary to recognize which user performed which action, and also because users inter-
act with and influence each other. Both approaches exploit implicit feedback to build the
preference model, using data gathered from sensors deployed in the environment; moreover,
both approaches rely on a two-tier system, where the lower tier uses a BN to represent the
preferences of an individual user, and the upper tier uses an additional BN to coordinate
the underlying one.
AFExhaustive Preferences Coding. A complementary category of approaches considers
the exhaustive coding of the desired environmental conditions in a given context, as opposed
to coding them implicitly by learning control rules. The exhaustive coding of user preferences
may point out a single preferred value (e.g., temperature, lighting) [Gao and Whitehouse
2009], or a preferred range per given physical quantity [Kolokotsa et al. 2005; Fernández-
Montes et al. 2009; Wen and Agogino 2008; Pan et al. 2008].
The Self-Programming Thermostat [Gao and Whitehouse 2009] uses exhaustive coding
for user preferences in terms of both the desired temperature and delay tolerance before
reaching the optimal conditions. In this case, preferences about environmental conditions
are coded via a single constraining value, while tolerance is expressed as a function of the
transition delay.
A very simple coding for users preferences regarding artificial lighting is proposed
in [Fernández-Montes et al. 2009]. Here the system restricts itself to learning the threshold
for lighting below which the user would perceive insufficient illumination. Learning is based
on explicit feedback obtained by periodically interviewing users with question forms about
the perceived lighting level quality. This training phase is carried out from the beginning of
R
the deployment and is not adjusted during normal functioning.
The same category also includes the work proposed in [Kolokotsa et al. 2005], where the
rules of a fuzzy system are modified according to the user preferences, exhaustively coded
into a smart card used for authentication and for interacting with the BMS. In this case,
preferences do not represent a tight constraint; rather they are coded as tolerance intervals
regarding the considered phenomena. Learning aims to identify which actions over the many
available actuators permits a reduction on the energy consumption, while allowing to reach
the requested environmental conditions.
The adoption of an exhaustive coding for user preferences is not as common in the litera-
D

ture, especially because it introduces an additional layer for knowledge representation inside
the BMS, thus adding complexity and increasing the possibility of errors. The tacit repre-
sentation of preferences within the system forces knowledge representation to be functional
to the adopted processing model (e.g., utility function for reinforcement learning, condi-
tional probability tables for Bayesian networks). Such representation could not be suitable
for interaction with the user, but it allows us to easily learn what the system needs and
does so in the most useful form.

ACM Computing Surveys, Vol. V, No. N, Article A, Publication date: January YYYY.
Intelligent Management Systems for Energy Efficiency in Buildings: A Survey A:29

7.2.3. Integration with Energy Saving Policies. An intelligent module for learning user prefer-
ences can be exploited in order to automate the control of the system actuators according to
the designed energy saving strategy. Many works for automatic learning of user preferences
consider user satisfaction as their only goal, disregarding energy saving issues altogether.
This is the case, for instance, in [Hagras et al. 2007] and [Vainio et al. 2008] where the
fuzzy controller is designed to learn the set of rules allowing the BMS to behave exactly as
the users would, and to adapt to their needs. Also, in [Boton-Fernandez and Lozano-Tello
2011] [Kushwaha et al. 2004; Chen et al. 2006; Hasan et al. 2009; Lin and Fu 2007], SWRL
rules or Bayesian networks are used only to select the action presumably preferable for the
user. Clearly, this direction is not much of interest in order to design a BMS for energy
saving. When user preferences are considered for energy saving purposes, three main classes

T
of approaches may be followed:
– Single objective function with a constraint: this approach considers matching user prefer-
ences as a tight constraint and energy saving as an objective function to be maximized [Gao
and Whitehouse 2009; Fernández-Montes et al. 2009; Wen and Agogino 2008; Pan et al.
2008];
– Single objective function: this approach considers both user well-being and energy saving
as part of the same objective function [Mozer 1998; Khalili et al. 2010; Singhvi et al. 2005];
– Multi-objective optimization: this approach consists of considering two separate functions,
AF
adopting a multi-objective optimization method [De Paola et al. 2012].
The Self-Programming Thermostat [Gao and Whitehouse 2009] falls into the first class.
The user poses a tight constraint over the desired temperature and the system cannot modify
that value; the system tunes the actuators so as to choose the optimal time to reach that set
point with the goal of saving energy whenever the user is not present in the area (see Figure
10 on page 21). An even simpler approach is proposed in [Fernández-Montes et al. 2009],
where, starting from the minimum acceptable value of lighting for the users, it is suggested
to tune the actuators so as to keep the lighting level just above that threshold. Also the
scheme in [Wen and Agogino 2008] proposes to combine energy saving and user preferences,
by considering the former criterion as an objective function to maximize (actually, energy
consumption is minimized) while the latter as a constraint to satisfy. In particular, the
problem is formulated as a linear optimization on the lighting value for the actuators to be
set on. The goal is to minimize lighting (assumed to be proportional to energy consumption)
constraining the lighting value to fall within a pre-fixed range.
In [Pan et al. 2008; Yeh et al. 2010] a control system is presented for artificial lighting
capable of respecting the constraints posed by user preferences (meant as desired lighting
R
range) and which attempts to minimize energy consumption. The system is devised for a
multi-user scenario; if the constraint combination does not allow any admissible solution,
the preference ranges are iteratively relaxed. Also in this case, the problem is formulated
in terms of linear programming, if user satisfaction is considered as a binary variable, or in
terms of sequential quadratic programming, if user satisfaction is considered as a continuous
variable, expressed as a Gaussian centered on the preference value.
In the Neural Network House project [Mozer 1998], the user discomfort and energy con-
sumption are regarded as two terms contributing to the same objective function to be
minimized. To this end, both quantities are expressed in the same measurement unit, i.e.,
D

in terms of currency. Action selection is performed in order to minimize the single objective
function, which includes a dynamically modified term in order to learn user preferences.
A similar approach is also adopted in [Khalili et al. 2010], which attempts to minimize
a convex function depending on the energy cost and user perceived utility. The adoption
of one only objective function is proposed in [Singhvi et al. 2005], which expresses it as a
linear combination of the user satisfaction function and the cost utility function (which is
inversely proportional to the energy saving). Moreover, a predictive model for user presence

ACM Computing Surveys, Vol. V, No. N, Article A, Publication date: January YYYY.
A:30 A. De Paola et al.

is embedded into their general model, by taking user preferences into account only if the
probability that the user is in the controlled area is non negligible.
Finally, another approach exists that considers user preferences and energy consumption
as two incommensurable quantities; thereby a multi-objective approach may consequently
be adopted. This choice is made for instance in [De Paola et al. 2012], where the user
preferences and the model of energy consumption are provided as input to a multi-objective
optimization system comparing various solutions by assessing their Pareto dominance. This
way, a potentially optimal set of solutions is selected, and the action to be performed is
selected therein according to a prefixed heuristic. In the reported case study, the solution
representing the median of the dominant front is chosen.

8. CONCLUSION AND CURRENT CHALLENGES

T
The significance of using energy saving strategies in building management has now been
fully recognized both from industry and the academic world. In this survey article we have
analyzed the technological, architectural, and algorithmic aspects that contribute to the
design of an energy-aware building management system (BMS). Our work has pointed out
that the guidelines for designing a BMS stem from the chosen policy for energy saving.
Depending on its complexity, and on the possibility for future expansions, the designer will
have to select the sensory and communication technology to deploy in the building, as well
AF
as the whole system architecture and the software modules providing intelligent support.
Despite the research e↵orts, many open issues remain to be addressed, also with respect
to potential industrial exploitation of BMSs. This aspect must not be disregarded as it is
likely to a↵ect the di↵usion and the actual impact of BMSs on global energy saving.
The first issue to be considered in order to encourage a commercial di↵usion of BMSs is
an accurate evaluation of the Return of Investment (ROI). Indeed, even disregarding the
costs of software design and development, the mere deployment of the required hardware
(sensors, actuators, communication infrastructure) has a non negligible cost. To the best of
our knowledge, in the literature there is no proposal about a simple and e↵ective tool for
estimating the yearly energy saving due to the adoption of a specific BMS, and consequently
the number of years required to recover the investment. Without such evaluation it is difficult
to envision a wide distribution among families and small institutions.
Other important issues concern the design of e↵ective BMSs characterized by an easy
management. One of the goals yet to be met is the definition of straightforward, semi-
automated configuration procedures, thus allowing for easy porting to home configurations
or small work environments without the need for the presence of specialized operators.
R
The difficulty lies both in the physical deployment of the sensing technology, which might
nevertheless require the intervention of technicians, but also in the association of meta-
information to sensors. This is required, for instance, in order to figure out which sensors are
deployed in which area, or which energy-sensing device is installed close to which appliance.
It would be advisable to transfer the know-how of Autonomic Computing into BMSs, and
more generally into Ambient Intelligence, in order to make the systems aware of their
physical structure, in terms of constituting components.
An analogous challenge must be addressed at a higher level, i.e., when considering the in-
telligent modules supporting the BMS. As shown, the majority of the intelligent approaches
D

supporting advanced energy saving policies requires a learning phase in order to let the sys-
tem acquire the necessary preliminary knowledge to carry its own activities. For instance,
a Bayesian network-based system needs to learn the conditional probability tables, and a
fuzzy system needs to learn its own rules, so the designer is often at a crossroads. It might
be assumed that such knowledge will be coded a priori into the BMS by some domain
expert, or through test scenarios analysis, and later re-used in di↵erent deployments, with
no relevant impact on the precision of inference (although such hypothesis does not appear
very reasonable). Otherwise, we need to accept that a non negligible technical as well as

ACM Computing Surveys, Vol. V, No. N, Article A, Publication date: January YYYY.
Intelligent Management Systems for Energy Efficiency in Buildings: A Survey A:31

theoretical gap is still present between the creation of BMS prototypes, and their practical
applicability at a large scale. Such gap is indeed represented by the lack of a semi-automated
mechanism for adapting the described intelligent systems to new scenarios. It is probably
possible to get to a compromise consisting of coding a priori part of the necessary knowl-
edge (regarding, for instance, the type of considered environment, or the generic connection
between the type of environment and the activities carried on therein), and subsequently
proceed to finely and adaptively tuning for new scenarios, based on the collected data or
due to limited contribution by the end-user.
Finally, a very important topic, which we could not examine in depth within this survey,
for the sake of brevity, is related to the use of techniques of intelligent planning for a
completely automated energy savings policy. The discussed examples of advanced policies

T
for actuator control just consider specific issues and typically act almost reactively based on
some pre-coded behavior. To the best of our knowledge, current literature does not report
works addressing the issue of the design of a comprehensive system, making full use of
intelligent techniques in order to become completely autonomous in controlling all aspects
of building management. Such a system should be able to correctly infer the environmental
state, to learn the needs and preferences of its inhabitants, and to predict the optimal
sequence of actions to carry on to reach its energy saving goals while respecting the user
requirements. The main difficulty lies in the substantial computational cost of traditional
AF
intelligent techniques for planning, especially in the context of complex scenarios, such as
BMSs which require planning over time. In our opinion, the direction to follow in this case
might be to identify a trade-o↵ between long-term planning systems, and reactive ones,
whose task would be to modify the long-term plans in order to address the unavoidable
environmental fluctuations, and the variations in user behavior.
ACKNOWLEDGMENTS
The authors are grateful to the anonymous reviewers for insightful comments and constructive suggestions
that helped us improve the quality of the manuscript significantly.
The work of S.K. Das is partially supported by the following NSF grants: 26-1004-54 titled the E2Home
(Energy Efficient Home) project, IIS-1064460 and CNS-1150192.
The work of A. De Paola, M. Ortolani, G. Lo Re and G. Anastasi is partially supported by the PO
FESR 2007/2013 grant G73F11000130004 funding the SmartBuildings project and by the PON R&C grant
MI01 00091 funding the SeNSori Project.
REFERENCES
Y. Agarwal, B. Balaji, R. Gupta, J. Lyles, M. Wei, and T. Weng. 2010. Occupancy-driven energy manage-
ment for smart building automation. In Proc. of the 2nd ACM Work. on Embedded Sensing Syst. for
R
Energy-Efficiency in Building (BuildSys ’10). 1–6.
Y. Agarwal, T. Weng, and R.K. Gupta. 2009. The energy dashboard: improving the visibility of energy
consumption at a campus-wide scale. In Proc. of the 1st ACM Work. on Embedded Sensing Syst. for
Energy-Efficiency in Buildings (BuildSys ’09). 55–60.
AlertMe. 2013. (2013). http://www.alertme.com/
G. Anastasi, M. Conti, M. Di Francesco, and A. Passarella. 2009. Energy conservation in wireless sensor
networks: A survey. Ad Hoc Networks 7, 3 (2009), 537–568.
M. Baranski and J. Voss. 2004. Genetic algorithm for pattern detection in NIALM systems. In Proc. of the
2004 IEEE Int. Conf. on Syst., Man, Cybern., Vol. 4. 3462–3468.
C. Beckel, L. Sadamori, and S. Santini. 2013. Automatic socio-economic classification of households using
D

electricity consumption data. In Proc. of the 4th Int. Conf. on Future energy Systems. 75–86.
L. Benini, E. Farella, and C. Guiducci. 2006. Wireless sensor networks: Enabling technology for ambient
intelligence. Microelectronics journal 37, 12 (2006), 1639–1649.
V. Boton-Fernandez and A. Lozano-Tello. 2011. Learning Algorithm for Human Activity Detection in Smart
Environments. In Proc. of the 2011 IEEE/WIC/ACM Int. Conf.s on Web Intelligence and Intelligent
Agent Technology (WI-IAT), Vol. 3. 45–48.
R. Campbell, J. Al-Muhtadi, P. Naldurg, G. Sampemane, and M.D. Mickunas. 2002. Towards security and
privacy for pervasive computing. In Proc. of the 2002 Mext-NSF-JSPS Int. Conf. on Software security:
theories and Syst. (ISSS’02). 1–15.

ACM Computing Surveys, Vol. V, No. N, Article A, Publication date: January YYYY.
A:32 A. De Paola et al.

A. Capone, M. Barros, H. Hrasnica, and S. Tompros. 2009. A New Architecture for Reduction of Energy
Consumption of Home Appliances. In Proc. of the Europ. Conf. of the Czech Presidency of the Council
of the EU “Towards eEnvironment”. 1–8.
Y.H. Chen, C.H. Lu, K.C. Hsu, L.C. Fu, Y.J. Yeh, and L.C. Kuo. 2009. Preference model assisted activity
recognition learning in a smart home environment. In Proc. of the 2009 IEEE/RSJ Int. Conf. on
Intelligent Robots and Syst. (IROS 2009). 4657–4662.
Z.Y. Chen, C.L. Wu, and L.C. Fu. 2006. Using semi-supervised learning to build bayesian network for
personal preference modeling in home environment. In Proc. of the 2006 IEEE Int. Conf. on Syst.,
Man, Cybern. (SMC’06), Vol. 5. 3816–3821.
J. Choi, D. Shin, and D. Shin. 2005. Research and implementation of the context-aware middleware for
controlling home appliances. IEEE Trans. on Consumer Electronics 51, 1 (2005), 301–306.
D.J. Cook. 2010. Learning Setting-Generalized Activity Models for Smart Spaces. IEEE Intell. Syst. 27, 1
(2010), 32–38.

T
D.J. Cook and S.K. Das. 2004. Smart environments: Technology, protocols and applications.
D.J. Cook and S.K. Das. 2007. How smart are our environments? An updated look at the state of the art.
Pervasive and Mobile Computing 3, 2 (2007), 53–73.
D.J. Cook and M. Schmitter-Edgecombe. 2009. Assessing the Quality of Activities in a Smart Environment.
Methods of informaion in medicine 48, 5 (2009), 480–485.
F. Corucci, G. Anastasi, and F. Marcelloni. 2011. A WSN-based Testbed for Energy Efficiency in Buildings.
In Proc. of the 16th IEEE Symp. on Computers and Commun. (ISCC’11). 990–993.
D.B. Crawley, L.K. Lawrie, F.C. Winkelmann, W.F. Buhl, Y.J. Huang, C.O. Pedersen, R.K. Strand, R.J.
AF Liesen, D.E. Fisher, M.J. Witte, and others. 2001. EnergyPlus: creating a new-generation building
energy simulation program. Energy and Buildings 33, 4 (2001), 319–331.
S. Darby. 2006. The e↵ectiveness of feedback on energy consumption. Technical Report, Environmental
Change Institute, University of Oxford. (2006). http://www.eci.ox.ac.uk/research/energy/downloads/
smart-metering-report.pdf
S.K. Das, D.J. Cook, A. Battacharya, E.O. III Heierman, and T.Y. Lin. 2002. The role of prediction
algorithms in the MavHome smart home architecture. IEEE Wireless Commun. 9, 6 (2002), 77–84.
S. Dawson-Haggerty, X. Jiang, G. Tolle, J. Ortiz, and D. Culler. 2010. sMAP: a simple measurement and
actuation profile for physical information. In Proc. of the 8th ACM Conf. on Embedded Networked
Sensor Syst. 197–210.
S. Dawson-Haggerty, A. Krioukov, J. Taneja, S. Karandikar, G. Fierro, N. Kitaev, and D. Culler. 2013.
BOSS: Building Operating System Services. In Proc. of the 10th USENIX Symp. on Networked Syst.
Design and Implementation (NSDI). 443–458.
S. Dawson-Haggerty, J. Ortiz, X. Jiang, J. Hsu, S. Shankar, and D. Culler. 2010. Enabling green building
applications. In Proc. of the 6th Workshop on Hot Topics in Embedded Networked Sensors. 1–5.
A. De Paola, S. Gaglio, G. Lo Re, and M. Ortolani. 2011. Multi-sensor fusion through adaptive Bayesian
networks. In AI*IA 2011: Artificial Intelligence Around Man and Beyond. Lecture Notes in Computer
Science, Vol. 6934. 360–371.
R
A. De Paola, S. Gaglio, G. Lo Re, and M. Ortolani. 2012. Sensor9k: A testbed for designing and experiment-
ing with WSN-based ambient intelligence applications. Pervasive and Mobile Computing 8, 3 (2012),
448–466.
F. Doctor, H. Hagras, and V. Callaghan. 2005. A type-2 fuzzy embedded agent to realise ambient intelligence
in ubiquitous computing environments. Information Sciences 171, 4 (2005), 309–334.
R.H. Dodier, G.P. Henze, D.K. Tiller, and X. Guo. 2006. Building occupancy detection through sensor belief
networks. Energy and buildings 38, 9 (2006), 1033–1043.
P. Ducange, F. Marcelloni, and D. Marinari. 2012. An Algorithm based on Finite State Machines with
Fuzzy Transitions for Non-Intrusive Load Disaggregation. In Proc. of the IFIP/IEEE Int. Conf. on
Sustainable Internet and ICT for Sustainability. 1–6.
T.V. Duong, H.H. Bui, D.Q. Phung, and S. Venkatesh. 2005. Activity recognition and abnormality detection
D

with the switching hidden semi-markov model. In Proc. of the 2005 IEEE Computer Society Conf. on
Computer Vision and Pattern Recognition (CVPR 2005), Vol. 1. 838–845.
P.G. Ellis and P.A. Torcellini. 2005. Simulating tall buildings using EnergyPlus. In Proc. of the 9th Int.
IBPSA Conf. 279–286.
F. Englert, T. Schmitt, S. Kößler, A. Reinhardt, and R. Steinmetz. 2013. How to auto-configure your smart
home?: high-resolution power measurements to the rescue. In Proc. of the 4th Int. Conf. on Future
energy Syst. (e-Energy ’13). 215–224.

ACM Computing Surveys, Vol. V, No. N, Article A, Publication date: January YYYY.
Intelligent Management Systems for Energy Efficiency in Buildings: A Survey A:33

V.L. Erickson, M.A. Carreira-Perpinan, and A.E. Cerpa. 2011. OBSERVE: Occupancy-based system for effi-
cient reduction of HVAC energy. In Proc. of the 10th ACM/IEEE Int. Conf. on Information Processing
in Sensor Networks (IPSN 2011). 258–269.
A. Fernández-Montes, L. Gonzalez-Abril, J.A. Ortega, and F.V. Morente. 2009. A study on saving energy
in artificial lighting by making smart use of wireless sensor networks and actuators. IEEE Network 23,
6 (2009), 16–20.
A. Foglar and Plosz. S. 2008. Appliance Profiles Specification. AIM Consortium, Deliverable 2.3. (2008).
http://www.ict-aim.eu/fileadmin/user files/deliverables/AIM-D2-3v1-0.pdf
J. Froehlich, E. Larson, S. Gupta, G. Cohn, M. Reynolds, and S. Patel. 2011. Disaggregated End-Use Energy
Sensing for the Smart Grid. IEEE Pervasive Computing 10, 1 (2011), 28–39.
G. Gao and K. Whitehouse. 2009. The self-programming thermostat: optimizing setback schedules based
on home occupancy patterns. In Proc. of the 1st ACM Work. on Embedded Sensing Syst. for Energy-
Efficiency in Buildings. 67–72.

T
G. Ghidini and S.K. Das. 2012. Improving home energy efficiency with E2Home: A Web-based application
for integrated electricity consumption and contextual information visualization. In Proc. of the IEEE
3rd Int. Conf. on Smart Grid Commun. (SmartGridComm). 471–475.
C. Gomez and J. Paradells. 2010. Wireless home automation networks: A survey of architectures and
technologies. IEEE Commun. Magazine 48, 6 (2010), 92–101.
T. Gu, H.K. Pung, and D.Q. Zhang. 2004. Toward an OSGi-based Infrastructure for Context-Aware Appli-
cations. IEEE Pervasive Computing 3, 4 (2004), 66–74.
D. Guinard, V. Trifa, F. Mattern, and E. Wilde. 2011. From the Internet of Things to the Web of Things:
AF Resource-Oriented Architecture and Best Practices. In Architecting the Internet of Things. 97–129.
S. Gupta, M.S. Reynolds, and S.N. Patel. 2010. ElectriSense: Single-point sensing using EMI for electrical
event detection and classification in the home. In Proc. of the 12th ACM Int. Conf. on Ubiquitous
computing. 139–148.
H. Hagras, F. Doctor, V. Callaghan, and A. Lopez. 2007. An incremental adaptive life long learning approach
for type-2 fuzzy embedded agents in ambient intelligent environments. IEEE Trans. on Fuzzy Syst. 15,
1 (2007), 41–55.
G.W. Hart. 1992. Nonintrusive appliance load monitoring. Proc. of the IEEE 80, 12 (1992), 1870–1891.
M.K. Hasan, K.A.P. Ngoc, Y.K. Lee, and S. Lee. 2009. Preference learning on an OSGi based home gateway.
IEEE Trans. on Consumer Electronics 55, 3 (2009), 1322–1329.
S. Helal, W. Mann, H. El-Zabadani, J. King, Y. Kaddoura, and E. Jansen. 2005. The Gator Tech Smart
House: A programmable pervasive space. Computer 38, 3 (2005), 50–60.
A. Holmes, H. Duman, and A. Pounds-Cornish. 2002. The iDorm: Gateway to heterogeneous networking
environments. In Proc. of the Int. ITEA Work. on Virtual Home Environments. 1–8.
International Energy Agency. 2003. Cool Appliance - Policy Strategies for Energy Efficient Homes. IEA
Publications, Paris, France. 62.168.68.98/StandardsLabels/downloads/01.pdf
X. Jiang, S. Dawson-Haggerty, P. Dutta, and D. Culler. 2009a. Design and implementation of a high-fidelity
ac metering network. In Proc. of the 2009 Int. Conf. on information Processing in Sensor Networks.
R
253–264.
X. Jiang, M. Van Ly, J. Taneja, P. Dutta, and D. Culler. 2009b. Experiences with a high-fidelity wireless
building energy auditing network. In Proc. of the 7th ACM Conf. on Embedded Networked Sensor Syst.
113–126.
A. Kamilaris, V. Trifa, and A. Pitsillides. 2011. HomeWeb: An application framework for Web-based smart
homes. In Proc. of the 18th Int. Conf. on Telecommunications (ICT 2011). 134–139.
W. Kastner, M.J. Kofler, and C. Reinisch. 2010. Using AI to realize energy efficient yet comfortable smart
homes. In Proc. of the 2010 8th IEEE International Work. on Factory Communication Syst. (WFCS).
169–172.
A.H. Khalili, C. Wu, and H. Aghajan. 2010. Hierarchical preference learning for light control from user feed-
back. In Proc. of the 2010 IEEE Computer Society Conf. on Computer Vision and Pattern Recognition
D

Work. (CVPRW). 56–62.

C. Kidd, R. Orr, G. Abowd, C. Atkeson, I. Essa, B. MacIntyre, E. Mynatt, T. Starner, and W. Newstetter.
1999. The aware home: A living laboratory for ubiquitous computing research. In Cooperative buildings.
Integrating information, organizations, and architecture. Lecture Notes in Computer Science, Vol. 1670.
191–198.
J.A. Kientz, S.N. Patel, B. Jones, E. Price, E.D. Mynatt, and G.D. Abowd. 2008. The Georgia Tech Aware
Home. In Proc. of the SIGCHI Conf. on Human factors in computing Syst. 3675–3680.

ACM Computing Surveys, Vol. V, No. N, Article A, Publication date: January YYYY.
A:34 A. De Paola et al.

Y. Kim, T. Schmid, Z.M. Charbiwala, and M.B. Srivastava. 2009a. ViridiScope: design and implementation
of a fine grained power monitoring system for homes. In Proc. of the 11th Int. Conf. on Ubiquitous
Computing. 245–254.
Y. Kim, T. Schmid, M.B. Srivastava, and Y. Wang. 2009b. Challenges in resource monitoring for residential
spaces. In Proc. of the 1st ACM Work. on Embedded Sensing Syst. for Energy-Efficiency in Buildings.
1–6.
K. Kobayashi, M. Tsukahara, A. Tokumasu, K. Okuyama, K. Saitou, and Y. Nakauchi. 2011. Ambient
intelligence for energy conservation. In Proc. of the 2011 IEEE/SICE Int. Symp. on System Integration
(SII). 375–380.
K. Koile, K. Tollmar, D. Demirdjian, H. Shrobe, and T. Darrell. 2003. Activity zones for context-aware
computing. In Proc. of the 5th Int. Conf. on Ubiquitous Computing (UbiComp 2003). 90–106.
D. Kolokotsa, K. Niachou, V. Geros, K. Kalaitzakis, G.S. Stavrakakis, and M. Santamouris. 2005. Imple-
mentation of an integrated indoor environment and energy management system. Energy and Buildings

T
37, 1 (2005), 93–99.
N. Kushwaha, M. Kim, D.Y. Kim, and W.D. Cho. 2004. An intelligent agent for ubiquitous computing
environments: smart home UT-AGENT. In Proc. of the 2nd IEEE Work. on Software Technologies for
Future Embedded and Ubiquitous Syst.. 157–159.
C. Laughman, K. Lee, R. Cox, S. Shaw, S. Leeb, L. Norford, and P. Armstrong. 2003. Power signature
analysis. Power and Energy Magazine, IEEE 1, 2 (2003), 56–63.
C. Lee, D. Nordstedt, and S. Helal. 2003. Enabling smart spaces with OSGi. IEEE Pervasive Computing
2, 3 (2003), 89–94.
AF
J. Lifton, M. Feldmeier, Y. Ono, C. Lewis, and J.A. Paradiso. 2007. A platform for ubiquitous sensor
deployment in occupational and domestic environments. In Proc. of the 6th Int. Conf. on Information
processing in sensor networks. 119–127.
Z.H. Lin and L.C. Fu. 2007. Multi-user preference model and service provision in a smart home environment.
In Proc. of the 2007 IEEE Int. Conf. on Automation Science and Engineering (CASE 2007). 759–764.
C.H. Lu and L.C. Fu. 2009. Robust location-aware activity recognition using wireless sensor network in an
attentive home. IEEE Trans. on Automation Science and Engineering 6, 4 (2009), 598–609.
J. Lu, T. Sookoor, V. Srinivasan, G. Gao, B. Holben, J. Stankovic, E. Field, and K. Whitehouse. 2010. The
smart thermostat: using occupancy sensors to save energy in homes. In Proc. of the 8th ACM Conf. on
Embedded Networked Sensor Syst. (SenSys ’10). 211–224.
A. Marchiori, D. Hakkarinen, Q. Han, and L. Earle. 2011. Circuit-Level Load Monitoring for Household
Energy Management. IEEE Pervasive Computing 10, 1 (2011), 40–48.
Microsoft Hohm. 2011. (2011). http://www.microsoft.com/environment/
M. Milenkovic and O. Amft. 2013. An opportunistic activity-sensing approach to save energy in office
buildings. In Proc. of the 4th Int. Conf. on Future energy Syst. 247–258.
M.C. Mozer. 1998. The Neural Network House: An Environment hat Adapts to its Inhabitants. In Proc. of
the Intelligent Environments AAAI Spring Symp. 110–114.
E.F. Nakamura, A.A.F. Loureiro, and A.C. Frery. 2007. Information fusion for wireless sensor networks:
R
Methods, models, and classifications. ACM Comput. Surveys 39, 3 (2007), 9:1–9:55.
N.T. Nguyen, D.Q. Phung, S. Venkatesh, and H. Bui. 2005. Learning and detecting activities from movement
trajectories using the hierarchical hidden Markov model. In Proc. of the 2005 IEEE Computer Society
Conf. on Computer Vision and Pattern Recognition (CVPR 2005), Vol. 2. 955–960.
M.S. Pan, L.W. Yeh, Y.A. Chen, Y.H. Lin, and Y.C. Tseng. 2008. A WSN-based intelligent light control
system considering user activities and profiles. IEEE Sensors Journal 8, 10 (2008), 1710–1721.
S.N. Patel, T. Robertson, J.A. Kientz, M.S. Reynolds, and G.D. Abowd. 2007. At the flick of a switch:
Detecting and classifying unique electrical events on the residential power line. In Proc. of the 9th Int.
Conf. on Ubiquitous computing. 271–288.
L. Perez-Lombard, J. Ortiz, and C. Pout. 2008. A review on buildings energy consumption information.
Energy and Buildings 40, 3 (2008), 394–398.
D

M. Philipose, K.P. Fishkin, M. Perkowitz, D.J. Patterson, D. Fox, H. Kautz, and D. Hahnel. 2004. Inferring
activities from interactions with objects. Pervasive Computing, IEEE 3, 4 (2004), 50–57.
Google PowerMeter. 2011. (2011). http://www.google.com/powermeter
A. Prudenzi. 2002. A neuron nets based procedure for identifying domestic appliances pattern-of-use from
energy recordings at meter panel. In Proc. of the 2002 IEEE Power Engineering Society Winter Meet-
ing, Vol. 2. 941–946.
Pulse Energy Inc. 2013. Berkeley Energy Dashboard. (2013). https://us.pulseenergy.com/UniCalBerkeley/
dashboard

ACM Computing Surveys, Vol. V, No. N, Article A, Publication date: January YYYY.
Intelligent Management Systems for Energy Efficiency in Buildings: A Survey A:35

Z. Qiu and G. Deconinck. 2011. Smart Meter’s feedback and the potential for energy savings in household
sector: A survey. In Proc. of the 2011 IEEE Int. Conf. on Networking, Sensing and Control. 281–286.
P. Rashidi, D.J. Cook, L. Holder, and M. Schmitter-Edgecombe. 2011. Discovering activities to recognize
and track in a smart environment. IEEE Trans. on Knowledge and Data Engineering 23, 4 (2011),
527–539.
P. Remagnino and G.L. Foresti. 2005. Ambient Intelligence: A New Multidisciplinary Paradigm. IEEE
Trans. on Syst., Man, Cybern., A, Syst. ,Humans 35, 1 (2005), 1–6.
L. Richardson and S. Ruby. 2007. RESTful web services.
A. Roy, S.K. Das, and K. Basu. 2007. A Predictive Framework for Location-Aware Resource Management
in Smart Homes. IEEE Trans. on Mobile Computing 6, 11 (2007), 1270–1283.
A.G. Ruzzelli, C. Nicolas, A. Schoofs, and G.M.P. O’Hare. 2010. Real-time recognition and profiling of
appliances through a single electricity sensor. In Proc. of 2010 7th Annual IEEE Commun. Society
Conf. on Sensor Mesh and Ad Hoc Commun. and Networks (SECON). 1–9.

T
F. Sadri. 2011. Ambient intelligence: A survey. ACM Comput. Surveys 43, 4 (2011), 36:1–36:66.
A. Schoofs, A.G. Ruzzelli, and G.M.P. O’Hare. 2010. Appliance Activity Monitoring Using Wireless Sensors.
In Proc. of the 9th ACM/IEEE Int. Conf. on Information Processing in Sensor Networks. 434–435.
V. Singhvi, A. Krause, C. Guestrin, J.H. Garrett Jr, and H.S. Matthews. 2005. Intelligent light control using
sensor networks. In Proc. of the 3rd Int. Conf. on Embedded networked sensor Syst. 218–229.
R.S. Sutton and A.G. Barto. 1998. Reinforcement learning: An introduction.
S. Taherian, M. Pias, G. Coulouris, and J. Crowcroft. 2010. Profiling energy use in households and office
spaces. In Proc. of the 1st Int. Conf. on Energy-Efficient Computing and Networking. 21–30.
AF
E. Tapia, S. Intille, and K. Larson. 2004. Activity recognition in the home using simple and ubiquitous
sensors. In Pervasive Computing. Lecture Notes in Computer Science, Vol. 3001. 158–175.
L.V. Thanayankizil, S.K. Ghai, D. Chakraborty, and D.P. Seetharam. 2012. Softgreen: Towards energy
management of green office buildings with soft sensors. In Proc. of the 2012 4th Int. Conf. on Com-
munication Systems and Networks (COMSNETS). 1–6.
The AIM Consortium. 2008. AIM - A novel architecture for modelling, virtualising and managing the energy
consumption of household appliances. (2008). http://www.ict-aim.eu/
The ESTIA Consortium. 2008. Enhanced Networked Architecture for Personalised Provision of AV Con-
tent within the Home Environment. (2008). http://www.gorenjegroup.com/en/filelib/gorenje group/
eu projects/eu project estia.pdf
S. Tompros, N. Mouratidis, M. Caragiozidis, H. Hrasnica, and A. Gavras. 2008. A pervasive network ar-
chitecture featuring intelligent energy management of households. In Proc. of the 1st Int. Conf. on
PErvasive Technologies Related to Assistive Environments. 1–6.
S. Tompros, N. Mouratidis, M. Draaijer, A. Foglar, and H. Hrasnica. 2009. Enabling Applicability of Energy
Saving Applications on the Appliances of the Home Environment. IEEE Network 23, 6 (2009), 8–15.
U.S. Energy Information Administration. 2010. International Energy Outlook 2010 - Highlights. Report
DOE/EIA-0484(2010). (2010). http://www.eia.doe.gov/oiaf/ieo/highlights.html
A.M. Vainio, M. Valtonen, and J. Vanhala. 2008. Proactive fuzzy control and adaptation methods for smart
R
homes. IEEE Intell. Syst. 23, 2 (2008), 42–49.
F.I. Vazquez and W. Kastner. 2011. Clustering methods for occupancy prediction in smart home control.
In Proc. of the 2011 IEEE Int. Symp. on Industrial Electronics (ISIE). 1321–1328.
M. Weiss and D. Guinard. 2010. Increasing energy awareness through web-enabled power outlets. In Proc.
of the 9th Int. Conf. on Mobile and Ubiquitous Multimedia. 20:1–20:10.
Y.J. Wen and A.M. Agogino. 2008. Wireless networked lighting systems for optimizing energy savings and
user satisfaction. In Proc. of the 2008. IEEE Wireless Hive Networks Conf. (WHNC 2008). 1–7.
D. Wilson and C. Atkeson. 2005. Simultaneous tracking and activity recognition (STAR) using many anony-
mous, binary sensors. In Pervasive computing. Lecture Notes in Computer Science, Vol. 3468. 329–334.
WiSensys. 2011. (2011). http://www.wisensys.com
D

X10. 2013. (2013). http://www.x10.com

H.W. Yeh, C.H. Lu, Y.C. Huang, T.H. Yang, and L.C. Fu. 2011. Cloud-Enabled Adaptive Activity-Aware
Energy-Saving System in a Dynamic Environment. In Proc. of the 2011 IEEE Ninth Int. Conf. on
Dependable, Autonomic and Secure Computing (DASC). 690–696.
L.W. Yeh, C.Y. Lu, C.W. Kou, Y.C. Tseng, and C.W. Yi. 2010. Autonomous light control by wireless sensor
and actuator networks. IEEE Sensors Journal 10, 6 (2010), 1029–1041.
L.W. Yeh, Y.C. Wang, and Y.C. Tseng. 2009. iPower: an energy conservation system for intelligent buildings
by wireless sensor networks. Int. J. of Sensor Networks 5, 1 (2009), 1–10.

ACM Computing Surveys, Vol. V, No. N, Article A, Publication date: January YYYY.
 A
Intelligent Management Systems for Energy Efficiency in Buildings:
A Survey – Appendix
ALESSANDRA DE PAOLA, University of Palermo, Italy
MARCO ORTOLANI, University of Palermo, Italy
GIUSEPPE LO RE, University of Palermo, Italy
GIUSEPPE ANASTASI, University of Pisa, Italy
SAJAL K. DAS, Missouri University of Science and Technology, USA

1. EVALUATION APPROACHES

T
In order to complete our survey on methodologies for BMS design and evaluation, we will
examine potential ways for researchers to conduct the experimental assessment of their
proposals. Namely, many other factors besides the main goal of energy saving must be con-
sidered, e.g., user comfort and system accuracy. During experimental evaluation, it may be
useful to compare one’s own approach with an ideal theoretical system, which would show
the desired behavior depending on the ground truth for the considered environmental quan-
tities (temperature, lighting, room occupancy). For instance the performance of a system
AF
for lighting control (including both a reactive and a predictive component) with a purely
reactive baseline algorithm is compared in [Lu et al. 2010a], using the same sensors and
actuators, and with an optimal algorithm capable of estimating the exact lighting level at
any time of the day. The optimal algorithm does not introduce delays, always satisfies the
user, and provides the theoretical upper bound on energy saving and user comfort.
Additional meaningful comparisons can be made about the performance of currently
available commercial systems, or the state of the art from research. The main difficulty
in comparing di↵erent approaches is the strong dependence from the BMS physical infras-
tructure and from the real trend of environmental quantities, whereas it would be highly
desirable to devise a comparative evaluation preliminarily to the physical deployment. A
solution to this problem could be provided by the adoption of simulation tools. As regards
to the HVAC system control, a comparison can be performed with the clues provided by
the EnergyPlus simulator [Crawley et al. 2001], which allows to reproduce indoor heat
flows starting from the structural model of the building, and from the actuator model and
settings. The advantages arising from the adoption of Energy Plus derive mostly from its
supply of predefined packages with models of actuators and buildings which may be easily
composed in the simulation so as to allow researchers to focus on the policies and algorithms
R
for energy saving.
The remainder of this section describes the criteria to adopt for assessing BMS perfor-
mance, and specifically user comfort and energy saving. In particular, the subsection dealing
with energy saving assessment will present many of the simulation techniques reported in
the literature.

1.1. User Comfort

One of the most relevant features a BMS must possess is the ability to guarantee user
D

comfort. Such characteristic allows critical assessment of the achieved energy saving by
discriminating between the contribution due to the user action and the one due to the
e↵ectiveness of the energy policies; the metric to be used, however, largely depends on the
examined approach.
Even though the literature contains several explicit models to measure user comfort [Pfaf-
ferott et al. 2007; Schumann et al. 2010], such models are generally too complex and depend
on too many variables to be of practical interest. An example is the thermal comfort index
presented in [Olesen and Parsons 2002], which is computed as a function of air temperature,

ACM Computing Surveys, Vol. V, No. N, Article A, Publication date: January YYYY.
A:2 A. De Paola et al.

radiance, humidity, air speed, occupants’ clothing and activity. Due to their intrinsic com-
plexity, such indices are seldom adopted, as explained in [Erickson et al. 2011]. Analogous
indices exist for evaluating the adequacy of the ventilation system, such as the standard pro-
posed by ASHRAE [Ashrae Standards 2013], which assesses ventilation speed with respect
to the number of people present in the room and to the a↵ected area.
At the opposite end of the spectrum of approaches for user comfort evaluation are all
methods relying on subjective assessment by the users themselves. The easiest way, but also
the least e↵ective and most intrusive one, consists of asking them to fill in forms to express
their opinion about the system performance. Such an approach may become extremely dull
for users as well as for evaluators, since they need to devise such forms, propose them to
users and process the answers. Nevertheless, it has been adopted in [Kolokotsa et al. 2005;

T
Fernández-Montes et al. 2009] since it does not require developing any specific hardware or
software for automated collection and processing of user feedback.
If the proposed system requires user preferences (e.g., desired temperature) to be static
and known a priori, a good indicator may be the time interval when the system is not
meeting the requirement, as in the case when the system keeps the low consumption mode
while the area is occupied by users. Such a metric has been proposed for instance in [Lu
et al. 2010b], where the “miss time” is measured as the time during which the home is
occupied but the temperature has not reached the desired level. The authors compare the
AF
obtained performance with the theoretical optimum, whose miss time would always be zero,
indicating an instantaneous response time. This comparison is useful for verifying the impact
on user comfort, as well as to provide an upper bound for energy consumption. In this case,
the goal is to obtain a tradeo↵ comfort and consumption, by minimizing the latter while
regarding the former as a constraint.
The lighting control system in [Lu et al. 2010a] uses, as a discomfort metric, the root
mean square error of the actual light intensity at the desk over the (known a priori) user
desirable setpoint, and an analogous choice was also made in [Khalili et al. 2010]. Similarly,
the schemes in [Pan et al. 2008; Yeh et al. 2010] measure the user comfort via an explicit
preference model, either in terms of preference interval, as proposed in [Wen and Agogino
2008], or in terms of a Gaussian centered on the preference value and considering two
indices. In the first case, a “GAP” index is used, representing the di↵erence between the
provided lighting and the requested one; if lighting falls within the desired interval, the gap
is 0, otherwise it is computed as the interval range. In the second case, the comfort index
is directly provided by the value of the Gaussian function. Clearly, an explicit model for
the user preferences does not require direct interaction with them to check whether their
R
preferences have been satisfied or not. However, it is dubious if this assessment method can
account for the variability and unpredictability of real-life scenarios.
If the system was designed to include the necessary devices to monitor user actions, it
may use such information to obtain a measure of the comfort perceived by users in a given
environmental condition. Whenever the users interact with the actuators to modify the
settings imposed by the BMS, they are in fact implicitly conveying a negative assessment
about the chosen management policy. Additionally, the farther the user setting is from the
system one, the more negative the opinion of the user. Such kind of assessment is reported
in the literature mainly for systems which attempt to dynamically learn user preferences.
D

The authors of [Doctor et al. 2005], for instance, indirectly use this metric; since they
propose a fuzzy system whose rules are dynamically adapted based on user feedback, the
authors consider the cumulative number of rule adaptations as a metric and observe its
trend to assess the user satisfaction over time. Finally, in [Hasan et al. 2009] is considered
as discomfort indicator the fact that users overwrite the system setting for more than three
times for a specific service.

ACM Computing Surveys, Vol. V, No. N, Article A, Publication date: January YYYY.
Intelligent Management Systems for Energy Efficiency in Buildings: A Survey – Appendix A:3

1.2. Energy Saving

Besides user comfort, the other important metric to take into account is energy saving,
typically expressed with respect to a baseline system. The comparison can be carried out in
many ways, e.g., by measuring the actual energy consumption through an energy monitoring
system, by performing a theoretical assessment, or by simulating the consumption through
specific software tools.
An actual comparison via monitoring does not require the baseline system to be managed
by a BMS. It is sufficient that the considered scenario includes a tool for energy measure-
ment (e.g., the sensor for aggregated measures at the root of the distribution tree). Such
a comparison can provide a ground truth during assessment of a new proposal. However,
exact monitoring in certain comparison scenarios is impractical, so a straightforward, yet

T
relatively simple approach is to hypothesize an energy consumption model to be used to as-
sess the hypothetical performance of the proposed system. To this aim, some of the models
proposed in the literature could be adopted, although sometimes they may become exces-
sively complex and simplified models are used instead. This mainly occurs with artificial
lighting systems; for instance, in [Wen and Agogino 2008] and [Pan et al. 2008] the lighting
measure is assumed to be directly proportional to the energy consumption. This assumption
greatly simplifies experimental assessment, but it must be noted that it may prove to be
unrealistic in the presence of artificial lighting sources realized with heterogeneous tech-
AF
nologies, or in spaces where artificial lighting is combined with natural one. Furthermore,
the presence of HVAC systems makes the computation of the theoretical consumption more
difficult.
For all the above considerations, software simulation is the most widely used solution
for the assessment of di↵erent approaches in diverse environmental conditions. The most
well-known and widespread tool for simulation of energy consumption in buildings is En-
ergyPlus [Crawley et al. 2001], which is based on thermal modeling of the entire building.
It has been developed by the U.S. Department of Energy and allows for assessing the ef-
fect and the energy consumption of di↵erent HVAC and lighting systems, under di↵erent
operating modes and external environmental conditions. The used energy model is highly
detailed and requires a precise description of the physical characteristics of the building
(such as building materials, facing direction of walls, floors, roofs, windows, and doors) as
well as of all the installed actuators, from HVAC to lighting systems. It is also possible
to provide information about the occupational pattern of users in their rooms in order to
assess their wellness. EnergyPlus allows either to make use of “templates” for modeling
the energy consumption of di↵erent devices, or to explicitly provide the specifications to
R
be coded into the simulator. The simulator computes the outdoor climate trends based
on known models and on past sensing obtained from the most common weather stations.
Such information is then used to employ some well-known thermal-transfer equations in
order to compute the instantaneous indoor environmental conditions, as well as the global
energy consumption. The main advantage of using such a “global” simulator, as opposed
to considering the energy consumption models of di↵erent devices individually, is that it
also takes into account the tight coupling of loads, system and plant, as well as the pos-
sibility to model the availability of renewable energy sources. Moreover, it is possible to
obtain simulation models for the energy consumption in buildings where the impact of the
D

environmental factors on the consumption itself is non negligible and not trivial to capture,
as in the case of skyscrapers, for which environmental conditions (e.g., temperature, wind
speed and direction) may significantly vary between di↵erent floors [Ellis and Torcellini
2005]. Furthermore, some of the modules added in recent releases allow to model also the
environmental emissions due to energy use. The work in [Lu et al. 2010b] is a meaningful
example of EnergyPlus adoption; the Smart Thermostat is tested over several environmen-
tal conditions, which could only be realized in practice with a large-scale experimentation

ACM Computing Surveys, Vol. V, No. N, Article A, Publication date: January YYYY.
A:4 A. De Paola et al.

that would require relevant funding. The validation of the simulated model was compared
against the sensory measurements obtained through a set of devices actually deployed in a
residential building. The use of this simulator allowed to test the proposed system under
di↵erent building conditions and climate zones. The idea of a preliminary simulation to
drive the design of eco-friendly buildings, by an automated multivariate optimization tool
was exploited in [Ellis et al. 2006] and [Lee and Braun 2006] which use EnergyPlus to assess
di↵erent approaches to actuator control.
ATPlus is another proposal for a simulation software devised to model heat flows in
buildings considering the presence of multiple HVAC appliances and taking into account
the specifics of the weather model to adopt. This simulator was used in [Kastner et al. 2010]
to test a control strategy exploiting neural networks in order to establish the best moment

T
to activate the actuators.
The use of simulators to assess energy saving performance is reported in the literature
mainly with regard to HVAC systems, even though some works exist which exploit such
tools for artificial lighting. For example, in [Hammad and Abu-Hijleh 2010] is adopted the
Integrated environmental Solutions - Virtual Environments (IES-VE) [Integrated environ-
mental Solutions - Virtual Environments (IES-VE) 2013] commercial package to assess the
e↵ectiveness of various configurations of the proposed system with dynamic external lou-
vers. An overview focusing on the usability of some commercially or academically available
AF
simulators for the assessment of the energy performance of buildings is contained in [Attia
et al. 2009].
Besides choosing the simulation platform in order to estimate the parameters of interest, it
is also necessary to select some similar systems to be used as a comparison. Such comparison
may help assessing the actual e↵ectiveness of the employed techniques and may provide
justification for the adoption of a more complex and more e↵ective systems instead of a
cheaper one. For instance, if the proposal regards a technological innovation for some of the
actuators, it may be more meaningful to compare it against the most common commercial
devices. The performance of Smart Thermostat [Lu et al. 2010b] is compared against the
devices actually deployed in a residential building. In [Hammad and Abu-Hijleh 2010], a
comparison is made between their dynamic external louvers (able to modify the amount of
lighting filtering through the windows depending on the outdoor conditions, on the indoor
perceived lighting, and on the user feedback) with other lighting systems generally present
in office buildings. If, on the other hand, the focus is not on the individual device but on the
overall comparison strategy, it may be convenient to compare the proposed management
policy with a theoretical one, used on the same set of appliances. In [Kastner et al. 2010] the
R
control strategy for an HVAC system is compared with another strategy, typically adopted
in actual deployments, which forces the activation of the HVAC system to occur 45 minutes
earlier to the occupants’ arrival, regardless of the delay actually required by those appliances
before reaching the desired temperature. The performance of the lighting control system
is compared in [Lu et al. 2010a] exploiting natural sources as much as possible in order
to reach the desired lighting level with a schema that purely relies on artificial lighting to
maintain the setpoint defined by the user at the desk level. As regards to HVAC systems, it is
possible to make a comparison with a policy using the setpoints for times and temperatures
recommended by EnergyStar [U.S. Environmental Protection Agency (EPA) 2013], a joint
D

program of the U.S. Environmental Protection Agency and the U.S. Department of Energy.
This is the solution chosen by the works proposed in [Lu et al. 2010b; Gao and Whitehouse
2009]. Another good comparison term for environmental conditions is the standard proposed
by ASHRAE both with respect to the temperature and ventilation system. This is the
solution adopted in [Erickson et al. 2011], where the authors compare their control strategy
for HVAC systems – based on the predicted patterns of user occupancy – to the baseline
provided by the ASHRAE set points. The ASHRAE baseline operates as a purely reactive
system which activates the HVAC system only after detecting user presence.

ACM Computing Surveys, Vol. V, No. N, Article A, Publication date: January YYYY.
Intelligent Management Systems for Energy Efficiency in Buildings: A Survey – Appendix A:5

A mere comparison among the performance of the di↵erent strategies might however not
be sufficient. The authors of [Taherian et al. 2010] state that the analysis of building energy
consumptions requires a more in-depth inspection, by isolating the non-reducible fraction
due to appliances that can never be deactivated. For instance, such “baseline energy use”
includes the energy due to: home security systems, alarm systems, the refrigeration unit,
and the IT infrastructure. Hence, the authors claim that the comparison should be made
over the ine↵ectiveness factor, which is obtained by excluding the “baseline energy use”
from the observed energy consumption. This factor represents the amount of energy due
to the habits of the human occupants, i.e., the quantity presenting the largest room for
improvement; this quantity is also defined as the Human-driven Energy Use.

T
1.3. Inference Accuracy
If the system includes “intelligent” modules for inferring some environmental conditions
starting from the available sensory data, it is advisable to individually estimate the accuracy
of those components, in order to assess their potential influence on the energy consumption
of the BMS and the residual margin for improvement. The only way to assess the accuracy
of the intelligent modules is to compare the results of the inference with the ground truth; in
case the system infers discrete concepts, its correctness may be measured by the percentage
of correct inferences, and in terms of false positives and false negatives. This approach is
AF
adopted for instance in [Lu et al. 2010b] to evaluate the accuracy of the HMM used to deduce
the presence of users. If the inferred concept is measurable by a continuous indicator, it is
advisable to provide at least the average error and the statistical variance. If the intelligent
modules use a probabilistic approach to represent the belief over a specific characteristic of
the world, then it is possible to use metrics which allow to express the distance between
the belief probability distribution and the ground truth. To this aim, the Kullback-Leibler
divergence or the Jenesen-Shannon divergence may be used, as proposed, for instance,
in [Erickson et al. 2011] to estimate the accuracy of the system for user presence prediction
by comparing the statistical model of their system with the distribution associated with the
ground truth.
When the inference system has to deal with user presence detection or with on-going
activity detection, a preliminary evaluation of its accuracy may be obtained by relying on
one of the several publicly available databases, provided by research groups active in the
field. The use of such datasets allows to address the issue of ground truth extraction from
sensory data, which is typically a very long and tiring task. One of such datasets is the
Tulum dataset [Cook and Schmitter-Edgecombe 2009], created within the WSU CASAS
R
project, by monitoring the occupants of a house for a period of approximately one month,
recording the sensory readings obtained through a set of sensors pervasively installed in
the monitored areas, and annotating the DB with the activities performed by the users.
Similarly structured, and also quite popular, is the Kasteren dataset [van Kasteren et al.
2011]. Works exploiting those public datasets include [Lu et al. 2010b; Gao and Whitehouse
2009]. A survey on more elaborate criteria for the assessment of user activity recognition
systems may be found in [Tapia et al. 2004].

2. RESEARCH CHALLENGES AND OPEN ISSUES

D

2.1. BMS Architectures

Some of the issues characterizing the design of BMS architectures have been identified by
the authors of [Cook and Das 2004], although many of them still remain open or have not
been conclusively resolved. Among these, the support for heterogeneity has a preminent role.
As previously described, there are di↵erent approaches in the literature for the definition
of middlewares allowing for an abstraction from the physical layer, but a standard solution
is still missing. This limits the interoperability and the ability to compose truly modular

ACM Computing Surveys, Vol. V, No. N, Article A, Publication date: January YYYY.
A:6 A. De Paola et al.

systems. The definition of a standard would be a big plus for several stakeholders of BMS,
notably for producers of sensors, actuators and appliances for BMSs. Such a connection
architecture should allow the automatic discovery of resources available, allowing for device
plug-and-play, and minimizing the burden well. Each device would be commercialized to-
gether with its software driver, and the final user would have the possibility of selecting the
best solution for his requirements.
2.2. Measuring Energy Consumption
Many solutions have been presented in the literature, concerning the design of BMSs capable
of monitoring energy consumption. Nevertheless, more research e↵orts need to be devoted
to the definition of configuration procedures which minimize human intervention. The ideal

T
goal is a system whose installation is performed by simply deploying energy sensors near to
the significant appliances, and by indicating through a friendly interface in which room, and
near which appliances each device is installed. The energy consumption models should be
automatically learned simply by exploiting collected measurements. Moreover, a mechanism
capable of supporting the final user in the design phase of the energy monitoring system is
missing in the current literature.
2.3. Intelligent Support Techniques
AF
The main challenge related to the adoption of intelligent support techniques is the reduction
of the burden required for the learning phase. Many techniques require a labeled training set
for tuning the system behavior, but gathering these data may cause a relevant discomfort
to the user. We believe that much e↵ort have to be devoted to the definition of semi-
automated labeling procedure which would allow to reduce such discomfort. Obviously, in
order to learn preferences or habits of the users, it is mandatory to interact with them,
but it is also necessary to make this interaction as natural as possible in order to develop
intelligent BMS truly able of “disappearing into the background” [Weiser 1991].
ACRONYMS

6LoWPAN IPv6 over Low power Wireless Personal Area Networks

AI Artificial Intelligence
AICO Artificial Intelligence Compliant Object
AmI Ambient Intelligence
API Application Programming Interface
API Aggregate Power Index
R
BMS Building Management System
BN Bayesian Network
CBC Case-Based Reasoning
CoC Code of Conduct
EMD Energy Management Device
FSMFT Finite State Machines based on Fuzzy Transitions
GPS Global Positioning System
HMM Hidden Markov Model
HRL Hierarchical Reinforcement Learning
HTTP HyperText Transfer Protocol
D

HVAC Heating, ventilation and Conditioning

IoT Internet of Things
IP Internet Protocol
IR Infrared
LAN Local Area Network
NALM Non-intrusive Application Load Monitoring
NILM Non-Intrusive Load Monitoring
OSGi Open Service Gateway initiative

ACM Computing Surveys, Vol. V, No. N, Article A, Publication date: January YYYY.
Intelligent Management Systems for Energy Efficiency in Buildings: A Survey – Appendix A:7

PDA Personal Digital Assistant

PIR Passive Infrared
PLC Power Line Communication
RECAP RECognition of electrical Appliances and Profiling in real-time
REST Representational state transfer
RFID Radio-Frequency IDentification
RL Reinforcement Learning
SaaS Software as a service
sMAP Simple Measurement and Actuation Profile
SMPS Switch Mode Power Supply
SOAP Simple Object Access Protocol
SWRL Semantic Web Rule Languages

T
TCP Transmission Control Protocol
WoT Web of Things
WSN Wireless Sensor Network

ADDITIONAL REFERENCES
Ashrae Standards. 2013. (2013). http://www.ashrae.org/standards-research--technology/
standards--guidelines
AF
S. Attia, L. Beltrán, A. De Herde, and J. Hensen. 2009. Architect Friendly: a comparison of ten di↵erent
building performance simulation tools. In Proc. of the 11th Int. IBPSA Conf. 204–211.
D.J. Cook and S.K. Das. 2004. Smart environments: Technology, protocols and applications.
D.J. Cook and M. Schmitter-Edgecombe. 2009. Assessing the Quality of Activities in a Smart Environment.
Methods of informaion in medicine 48, 5 (2009), 480–485.
D.B. Crawley, L.K. Lawrie, F.C. Winkelmann, W.F. Buhl, Y.J. Huang, C.O. Pedersen, R.K. Strand, R.J.
Liesen, D.E. Fisher, M.J. Witte, and others. 2001. EnergyPlus: creating a new-generation building
energy simulation program. Energy and Buildings 33, 4 (2001), 319–331.
F. Doctor, H. Hagras, and V. Callaghan. 2005. A type-2 fuzzy embedded agent to realise ambient intelligence
in ubiquitous computing environments. Information Sciences 171, 4 (2005), 309–334.
P.G. Ellis, B. Griffith, N. Long, P. Torcellini, and D. Crawley. 2006. Automated multivariate optimization
tool for energy analysis. In Proc. of the 2006 IBPSA-USA Conf. (SimBuild 2006). 42–48.
P.G. Ellis and P.A. Torcellini. 2005. Simulating tall buildings using EnergyPlus. In Proc. of the 9th Int.
IBPSA Conf. 279–286.
V.L. Erickson, M.A. Carreira-Perpinan, and A.E. Cerpa. 2011. OBSERVE: Occupancy-based system for effi-
cient reduction of HVAC energy. In Proc. of the 10th ACM/IEEE Int. Conf. on Information Processing
in Sensor Networks (IPSN 2011). 258–269.
A. Fernández-Montes, L. Gonzalez-Abril, J.A. Ortega, and F.V. Morente. 2009. A study on saving energy
R
in artificial lighting by making smart use of wireless sensor networks and actuators. IEEE Network 23,
6 (2009), 16–20.
G. Gao and K. Whitehouse. 2009. The self-programming thermostat: optimizing setback schedules based
on home occupancy patterns. In Proc. of the 1st ACM Work. on Embedded Sensing Syst. for Energy-
Efficiency in Buildings. 67–72.
F. Hammad and B. Abu-Hijleh. 2010. The energy savings potential of using dynamic external louvers in an
office building. Energy and Buildings 42, 10 (2010), 1888–1895.
M.K. Hasan, K.A.P. Ngoc, Y.K. Lee, and S. Lee. 2009. Preference learning on an OSGi based home gateway.
IEEE Trans. on Consumer Electronics 55, 3 (2009), 1322–1329.
Integrated environmental Solutions - Virtual Environments (IES-VE). 2013. (2013). http://www.iesve.com/
D

W. Kastner, M.J. Kofler, and C. Reinisch. 2010. Using AI to realize energy efficient yet comfortable smart
homes. In Proc. of the 2010 8th IEEE International Work. on Factory Communication Syst. (WFCS).
169–172.
A.H. Khalili, C. Wu, and H. Aghajan. 2010. Hierarchical preference learning for light control from user feed-
back. In Proc. of the 2010 IEEE Computer Society Conf. on Computer Vision and Pattern Recognition
Work. (CVPRW). 56–62.
D. Kolokotsa, K. Niachou, V. Geros, K. Kalaitzakis, G.S. Stavrakakis, and M. Santamouris. 2005. Imple-
mentation of an integrated indoor environment and energy management system. Energy and Buildings
37, 1 (2005), 93–99.

ACM Computing Surveys, Vol. V, No. N, Article A, Publication date: January YYYY.
 A:8 A. De Paola et al.

K. Lee and J.E. Braun. 2006. Evaluation of methods for determining demand-limiting setpoint trajectories in
commercial buildings using short-term data analysis. In Proc. of the 2006 IBPSA-USA Conf. (SimBuild
2006). 107–114.
J. Lu, D. Birru, and K. Whitehouse. 2010a. Using Simple Light Sensors to Achieve Smart Daylight Har-
vesting. In Proc. of the 2nd ACM Work. on Embedded Sensing Syst. for Energy-Efficiency in Building.
73–78.
J. Lu, T. Sookoor, V. Srinivasan, G. Gao, B. Holben, J. Stankovic, E. Field, and K. Whitehouse. 2010b. The
smart thermostat: using occupancy sensors to save energy in homes. In Proc. of the 8th ACM Conf. on
Embedded Networked Sensor Syst. (SenSys ’10). 211–224.
B.W. Olesen and K.C. Parsons. 2002. Introduction to thermal comfort standards and to the proposed new
version of EN ISO 7730. Energy and buildings 34, 6 (2002), 537–548.
M.S. Pan, L.W. Yeh, Y.A. Chen, Y.H. Lin, and Y.C. Tseng. 2008. A WSN-based intelligent light control
system considering user activities and profiles. IEEE Sensors Journal 8, 10 (2008), 1710–1721.

T
J.Ü. Pfa↵erott, S. Herkel, D.E. Kalz, and A. Zeuschner. 2007. Comparison of low-energy office buildings in
summer using di↵erent thermal comfort criteria. Energy and Buildings 39, 7 (2007), 750–757.
A. Schumann, M. Burillo, and N. Wilson. 2010. Predicting the desired thermal comfort conditions for shared
offices. In Proc. of the Int. Conf. on Computing in Civil and Building Engineering (ICCCBE 2010).
95–96.
S. Taherian, M. Pias, G. Coulouris, and J. Crowcroft. 2010. Profiling energy use in households and office
spaces. In Proc. of the 1st Int. Conf. on Energy-Efficient Computing and Networking. 21–30.
E. Tapia, S. Intille, and K. Larson. 2004. Activity recognition in the home using simple and ubiquitous
AF sensors. In Pervasive Computing. Lecture Notes in Computer Science, Vol. 3001. 158–175.
U.S. Environmental Protection Agency (EPA). 2013. EnergyStar Program. (2013). http://www.energystar.
gov
T. van Kasteren, G. Englebienne, and B. Kröse. 2011. Hierarchical Activity Recognition Using Automatically
Clustered Actions. In Ambient Intelligence. Lecture Notes in Computer Science, Vol. 7040. 82–91.
M. Weiser. 1991. The Computer for the 21st Century. Scientific American 265, 3 (1991), 66–75.
Y.J. Wen and A.M. Agogino. 2008. Wireless networked lighting systems for optimizing energy savings and
user satisfaction. In Proc. of the 2008. IEEE Wireless Hive Networks Conf. (WHNC 2008). 1–7.
L.W. Yeh, C.Y. Lu, C.W. Kou, Y.C. Tseng, and C.W. Yi. 2010. Autonomous light control by wireless sensor
and actuator networks. IEEE Sensors Journal 10, 6 (2010), 1029–1041.
R
D

ACM Computing Surveys, Vol. V, No. N, Article A, Publication date: January YYYY.

View publication stats