CAPSTONE PROJECT
Day-Ahead Load-Priority-based Energy Prediction and
Optimization for Net Zero Energy Buildings
By
Rishibha Jain (102104021)
Shrijal Sharma (102284007)
Ritik Raj (102104010)
Harshpreet Singh (102284002)
Mentors
Dr. Mukesh Singh (Professor,EIED)
Dr. Neeru Jindal (Associate Professor, DECE)
Electrical and Instrumentation Engineering Department, EIED
THAPAR INSTITUTE OF ENGINEERING AND TECHNOLOGY
(DEEMED TO BE A UNIVERSITY), PATIALA, PUNJAB INDIA
1
ACKNOWLEDGEMENT
We wish to express our sincere gratitude to Dr. Mukesh Singh, Professor, Electrical and
Instrumentation Engineering Department, and Dr. Neeru Jindal (Associate Professor, DECE),
Thapar Institute of Engineering and Technology, Patiala, Punjab, for providing us an
opportunity to undertake a Project on, Day-Ahead Load-Priority-based Energy Prediction and
Optimization for Net Zero Energy Buildings.
We thank our mentors for their valuable assistance and guidance throughout the project.
2
CONTENTS
S.No. CONTENTS Page No.
1 Introduction 4
2 Literature Review 5
3 Research Gaps 6
4 Objectives 7
5 Deliverables 8
6 Methodology 9-12
7 Design 13-14
8 Standards 15
9 Gantt Chart 15
10 Components selection criterion 16
11 Results 17-19
12 References 20
3
I. INTRODUCTION
In Net-Zero Energy Buildings (NZEBs), load prioritization offers enormous potential for
minimizing energy use, cutting expenses, and fostering sustainability. Load prioritization
effectively manages energy supply and demand in residential settings while maintaining user
comfort and satisfaction. It combines machine learning algorithms, smart grid technology, and
user-centric techniques. Day-ahead load prioritizing, which incorporates demand response
tactics, occupant behaviour modelling, intelligent building technology, renewable energy
generation, and life cycle assessment methodology, is essential to NZEBs. Through integration,
NZEBs may guarantee occupant comfort and well-being while maintaining net-zero energy
performance.[1]
Further investigation and creativity are required to solve lingering issues and fully realize the
advantages of load prioritization for household energy conservation and the shift to a
sustainable built environment. Household load scheduling aims to increase energy and cost
efficiency while considering the users' comfort limits. Dynamic priority allocation and
scheduling for appliances with consumer comfort limitations are presented, considering day-
ahead weather forecasting and renewable source availability prediction[2]. We look upon
optimization techniques that focus on energy saving and user convenience.
A hybrid approach that uses GWO and PSO is formulated. Through appliance energy
prediction, proactive energy optimization was made possible. An LSTM model was created to
verify the appliances' energy projections. Smart home appliances could be efficiently and
proactively regulated through prediction and optimized control. Initially, we looked at the
predictive model's RMSE score and discovered that the suggested model produces low RMSE
values. Second, after running many simulations, we found that the suggested optimization
results might save energy costs, which could be utilized to manage appliances and maintain the
desired indoor environment in an intelligent house. The optimization tactics were applied to
energy cost reduction goals and assessed for seasonal and monthly data trends to verify the
results. LSTM networks, however, are still susceptible to the exploding gradient issue.
Extended Short-Term Memory Networks are sequential neural networks with deep learning
capabilities that preserve information. It is a unique recurrent neural network that can solve the
RNN's vanishing gradient issue. Traditional RNNs can track arbitrary long-term dependencies
in the input sequences. The problem with classic RNNs is computational (or practical). When
training a classic RNN via back-propagation, the long-term gradients may "vanish" or tend to
zero as a result of minimal numbers inadvertently entering the computations. This effectively
stops the model from learning. Because LSTM units permit gradients to flow with little to no
attenuation, RNNs that employ them can partially overcome the vanishing gradient problem.
4
II. LITERATURE REVIEW
Finding a workable solution is essential when tackling combinatorial problems in the real
world. Notwithstanding the current limits of the employed technique, a workable optimal
solution for a given situation can be produced and solved using optimization techniques.
Additionally, population-based optimization techniques are currently of attention and have led
to the development of numerous new and better methods for solving a wide range of problems.
Various optimization techniques include ACO (Ant Colony Optimization)[10], WAO (Whale
Optimization)[11], GA(Genetic Algorithm)[12], PSO(Particle Swam Optimization)[5],
GWO(Grey Wolf Optimization)[4], etc.
GWO is based on grey wolves' behavior hunting mechanism, and they were placed at the top
of the food chain. The grey wolves' predator instinct makes this technique one of the most
promising optimization techniques. Based on the experiment results, the GWO outperformed
other methods such as PSO, GA, ABC, and SA. However, some problems cannot be solved by
GWO alone. Hence, we will use multiple algorithms together to compare the obtained
results.[4]
The population-based metaheuristic approach known as particle swarm optimization (PSO)
leverages the social behavior of birds to solve optimization issues. A swarm of birds approaches
their meal goal by fusing social and self-experience, rearranging and updating their positions
to make the best possible arrangement. When solving problems, this social-psychological
behavior is used. Researchers have developed several improvements to the original PSO
algorithm to reduce difficulties like premature convergence performance concerns, and
enhance its efficacy and efficiency.[5]
When the PSO algorithm's starting location is distant from the global minimum, its great
exploration ability allows it to converge to a local minimum. During iterations, the program
selects random sites with a low probability of particles to lower this danger. This strategy,
nevertheless, can be dangerous since it might align with a problematic area of the search space.
The PSO process is supported by the Grey Wolf Optimizer (GWO) algorithm, which begins
with high exploration and replaces certain particles with fresh ones found during a few
iterations of the GWO algorithm.[6]
A straightforward evolutionary technique that has demonstrated better performance in global
continuous optimization is differential evolution (DE). It primarily makes use of the differential
information to direct its further exploration. It mainly uses the distance and direction
information from the current population to guide its further search. [7]
Ensemble learning is a machine learning approach that integrates predictions from numerous
separate models to get a superior predictive performance compared to any single model. By
combining the predictions of several models—each of which may have advantages and
disadvantages of its own—ensemble learning aims to capitalize on the wisdom of the crowd.
Better performance and generalization may result from this.
Numerous multi-objective optimization problems (MOPs) arise in real-world applications,
including training, investment, control, scheduling, planning, and weapon selection, that must
be resolved in dynamic or unpredictable situations. These issues are known as dynamic multi-
objective optimization problems (DMOPs). To lower energy loads in residential buildings, a
reliable forecast method must be developed.[8]
5
III. RESEARCH GAPS
Ø Current Optimization models perform with limited accuracy due to various columns in
datasets such as weather conditions, occupancy patterns, and building characteristics.
Ø Research is needed to develop comparison of optimization models (on the same dataset)
that can accurately optimize energy consumption patterns in buildings.
Ø Need for integrating machine learning algorithms with building management systems
to obtain better-optimized energy consumption.
Ø Need to develop more efficient and robust machine learning-based optimization
algorithms that can handle large datasets of building energy systems.
Ø Many of the optimization models in the literature have been used but they overlook
user's convenience, and use limited dataset for optimization.
Novelty: We used Long Short Term Memory (LSTM) and compared the three
optimization techniques.
6
IV. OBJECTIVES
Ø Development of Energy forecasting algorithm based on Machine learning and
optimization-based techniques.
Ø Compare the Simulation results of Prediction and Optimization using LSTM - GWO :
PSO : DE
Ø To develop Prototype hardware based on IoT for energy optimization of appliances
(Plug and Play Module).
7
V. DELIVERABLES
Ø The results of GWO (Grey wolf optimisation) and PSO(Particle Swam Optimisation)
are used on an LSTM based model to compare the energy optimization.
Ø Prototype hardware for plug-and-play appliances, for energy optimization.
8
VI. METHODOLOGY
Figure1: Software-Hardware framework of our project
1. Design of the System
Our process starts with designing a plug-and-play solution for smart home appliances. In order
to minimise switching losses that happen when devices are turned on or off, this system was
created. The system will be simple to install and operate, and it will be made to work with a
variety of household appliances.
The integration of multiple gadgets in the age of smart homes presents issues with user
convenience, energy efficiency, and compatibility. Our project's main goal is to create a plug-
and-play solution that minimises switching losses while integrating devices into smart home
networks. Furthermore, our goal is to maximise appliance scheduling in order to reduce energy
usage and improve user experience.
Users may effortlessly add new gadgets to their smart homes without worrying about
compatibility problems or interfering with current operations thanks to the development of a
plug-and-play system. This system will make use of sophisticated control techniques to reduce
switching losses and guarantee effective energy use.
We use sophisticated optimisation algorithms to try and optimise appliance scheduling in
parallel. Long Short-Term Memory (LSTM), Particle Swarm Optimisation (PSO), and Genetic
Algorithm (GWO) are applied in the optimisation process to generate schedules that take time
limitations, user preferences, and energy usage into account.
2. Information Gathering
The gathering of data is the next stage. Data from several household appliances will be gathered
by us. Included in this data will be details regarding the amount of electricity that each
9
appliance uses, as well as the times that they are most frequently utilised. This information will
be utilised to build an extensive dataset on the use of household appliances.
3. Systems with Plug-and-Play Design
The plug-and-play solution we have created consists of multiple essential parts that work
together to minimise switching losses and enable smooth device integration. The following are
included in the system architecture:
• Device Recognition and Compatibility Verification: Based on functional requirements
and communication protocols, machine learning techniques are used to recognise and
validate compatible devices.
• Seamless Integration: Ensuring device interoperability across manufacturers by putting
in place standardised protocols and communication interfaces.
• Intelligent Control Mechanisms: By coordinating device operations and optimising
power consumption, real-time monitoring and control algorithms are used to reduce
switching losses.
To optimise energy usage and save operating costs, the system's control mechanisms will make
use of methods like demand response strategies, dynamic pricing analysis, and load forecasting.
4. Preparing Data
After the data is gathered, preprocessing will be done to prepare it for usage with our
optimisation methods. This could entail separating the data into training and testing sets,
normalising the data to make sure it is in an appropriate format, and cleaning the data to get rid
of any errors or inconsistencies.
5. Strategies for Optimisation
Next, we will utilise the data to apply two optimisation techniques: Particle Swarm
Optimisation (PSO) and Grey Wolf Optimiser (GWO). By reducing power usage and avoiding
switching losses, the ideal schedule for every device will be determined using these methods.
6. Appliance Scheduling Optimisation Methods
In smart homes, optimising appliance scheduling is essential to achieving optimal energy
efficiency and user comfort. We provide a hybrid strategy that successfully tackles the complex
scheduling problem by combining GWO, PSO, and LSTM.
PSO (particle swarm optimisation) and GWO (genetic algorithm) are population-based
optimisation algorithms that drew inspiration from social insects' and animals' behaviour.
These algorithms will be used to iteratively improve candidate solutions based on fitness, in an
effort to find the best appliance schedules.
Memory for Long Short Term (LSTM): Recurrent neural networks (RNNs) of the long-term
dependency (LSTM) type can recognise long-term dependencies in sequential input. Within
our situation, Recurrent neural networks (RNNs) with long short-term memory (LSTM) may
identify long-term dependencies in sequential input. To enable proactive scheduling decisions,
10
LSTM will be utilised in our scenario to forecast appliance consumption trends based on past
data.
Dynamic and adaptive appliance scheduling that takes into account both short-term user
preferences and long-term trends in energy use is made possible by the integration of GWO
and PSO with LSTM.
7. The LSTM Framework
After optimisation, the data will be fed into an LSTM model (long short-term memory). Based
on past data, this model will be trained on the optimised data and used to forecast the best
timetable for every appliance.
8. Integration of Systems
The plug and play system will incorporate the LSTM model when it has been trained and
evaluated. By doing this, the system will be able to schedule each appliance's operation
automatically according to the LSTM model's predictions, guaranteeing ideal power usage and
avoiding switching losses.
We want to use the suggested plug-and-play system and optimisation strategies in a simulated
smart home scenario to verify the efficacy of our approaches. The following will be part of the
experimental setup:
• Dataset Selection: To train the LSTM model and assess schedule optimisation
strategies, real-world appliance consumption data is used.
• Metrics for Evaluation: evaluating the system's performance in terms of reduced energy
use, user happiness, computational efficiency, and scalability.
• Experimental Results: Comparative studies of energy usage before and after
optimisation are shown, together with user comments regarding scheduling ease.
After interpreting the experimental data, we will talk about the advantages and disadvantages
of our techniques and suggest further research and development directions. Some potential
subjects for further research are:
• more algorithmic improvements and parameter tweaking for scheduling methods.
• including other optimisation goals like using renewable energy sources and reducing
peak demand.
• extension of the plug-and-play system to accommodate new protocols and technologies
for smart homes.
9. Examining and Assessing
Testing and assessment are the last steps in our technique. To assess the system's performance,
we will test it in an actual environment. Additionally, in order to find any areas that can use
improvement, we will gather user input.
This methodology offers a thorough process for creating, putting into practice, and assessing
plug-and-play systems for smart home appliances that minimise power consumption and avoid
switching losses. In addition to being effective and economical, our goal is to develop a system
11
that is simple to use and intuitive for users by utilising machine learning models and
sophisticated optimisation techniques.
Through the creation of a plug-and-play system and cutting-edge optimisation methodologies,
our research seeks to address the difficulties associated with device integration and appliance
scheduling optimisation in smart homes. Energy economy, user comfort, and general smart
home performance should all significantly increase with the smooth integration of devices and
the optimisation of scheduling decisions. This study lays the groundwork for upcoming
advancements in the industry and advances the development of intelligent and sustainable
living environments.
12
VII. DESIGN
Fig.(2) Block diagram of Energy Optimization for Net Zero Energy Buildings
a. Experimental Smart Home Dataset: In this paper, a data set of home appliances
comprises 32 columns. It includes various appliances such as dishwasher, furnance,
wine cellar, microwave, fridge, etc., shown above in the block diagram.
b. Data Preprocessing: This data is preprocessed. Preprocessing includes data reduction
by eliminating unwanted data, cleaning data, and setting data stamp as Indexes. Outliers
are checked. Data trends are researched and studied. Multi Co-linearity Matrix and
Correlation Matrix are made which are then used in the program.
c. Machine Learning: The Matrices calculated and the dataset gives the required
hyperparameters. These hyperparameters are used in programming. Features-
Parameters and Hyperparameters are reshaped, data is splitted. Further training and
testing is done to get the best possible.
d. Optimization Techniques: Various optimization techniques are used to get results which
are then compared graphically and tables are made displaying the outputs obtained from
various algorithms. Some of the performance parameters are used in these papers to
compare the workings of various algorithms. As per the results, the best optimization
algorithm is selected.
• LSTM for Time Series Data
A kind of Recurrent Neural Network (RNN) called Long Short-Term Memory (LSTM) has the
capacity to retain values from previous stages for later usage. The goal of a recurrent neural
network (RNN), a specific type of neural network, is to anticipate the subsequent observational
step based on the steps that have already been seen in the sequence. Time series datasets exhibit
a sequence of dependency among the input variables, in contrast to regression-based modeling.
Recurrent neural networks are incredibly effective at managing input variable dependencies. A
13
particular kind of Recurrent Neural Network (RNN) called LSTM is able to store and process
extended sequences of data. A multi-step univariate forecast technique was created. [2]
• Optimization Techniques
GWO is based on grey wolves' behavior hunting mechanism and they were placed at the top
of the food chain. The predators' instinct of the grey wolves is what makes this technique one
of the most promising optimisation techniques. Based on the results from the experiment, the
GWO was able to outperform other techniques such as PSO, GA, ABC, and SA. But, there are
some problems which cannot be solved by GWO only. Hence, we will use multiple algorithms
together to and compare the obtained results.[3]
The population-based metaheuristic approach known as particle swarm optimization (PSO)
leverages the social behavior of birds to solve optimization issues. A swarm of birds approaches
their meal goal by fusing social and self-experience, rearranging and updating their positions
to make the best possible arrangement. When solving problems, this social-psychological
behavior is used.Researchers have developed several improvements to the original PSO
algorithm in an effort to reduce difficulties like premature convergence, performance concerns,
and enhance its efficacy and efficiency.[4]
When the PSO algorithm's starting location is distant from the global minimum, its great
exploration ability allows it to converge to a local minimum. During iterations, the program
selects random sites with a low probability of particles in order to lower this danger. This
strategy, nevertheless, can be dangerous since it might align with a problematic area of the
search space. The PSO process is supported by the Grey Wolf Optimizer (GWO) algorithm,
which begins with high exploration and replaces certain particles with fresh ones found during
a few iterations of the GWO algorithm.[5]
A straightforward evolutionary technique that has demonstrated better performance in the
global continuous optimization is differential evolution (DE). It primarily makes use of the
differential information to direct its further exploration. It mainly uses the distance and
direction information from the current population to guide its further search. [6]
14
VIII. STANDARDS
• IEC60870-5-104: ESP32 implements communication via IEC60870-5-104 on arduino,
we can use multiple instances of "MASTER" and "SLAVE" objects to connect Arduino
with PLCs or RTU that use this protocol.
• International Electrotechnical Commission (IEC):IEC 61747-22, International
Commission on Illumination (CIE):CIE S 023/E:2013
• IEC 61182-2, Printed board assembly products – Manufacturing description data and
transfer methodology – Part 2: Generic requirements
• IEC 61188-5-1, Printed boards and printed board assemblies – Design and use – Part
5-1: Attachment (land/joint) considerations – Generic requirements
IX. GANTT CHART
15
X. COMPONENTS SELECTION CRITERION
Components selected:
Name of Technical Estimated Status
S. No.
Equipment Specifications Cost (in Rs.) Purchased/Ordered
Quad core
Cortex-A72
1 Raspberry Pi 4 (ARM v8) 64- 4500 -
bit SoC @
1.8GHz
2 Solar Panel 12V, 12W 2000 -
3 PCBs - 1500 -
LEDs, Motors
4 - 1000 -
etc
Miscellaneous Power supply,
5 2000 -
components wires etc.
Criterion:
Raspberry Pi 4 :The Raspberry Pi 4 is selected for applications using machine learning because
of its adaptability, low cost, and strong community. Some more reasons are affordability,
compact size, availability of GPIO pins, and machine learning support.[8]
Solar Panel: The fundamental component of a solar power producing system is solar panels.
These solar panels are made up of several semiconducting material-based solar cells. The
process of converting incoming light into useful power is carried out by these solar cells.
PCBs: PCBs are crucial parts of photovoltaic solar panels. The solar cells are connected by
PCBs, which guarantee effective power collecting and delivery. They also make data
monitoring and control systems for solar panel performance optimisation possible.
LEDs and Motors: We will buy motors of small ratings and other light loads such as LEDs to
show the working of our project hardware on a PCB board.
16
XI. RESULTS
Here, are the results that we have obtained by using various optimization algorithms:
17
In this, various optimization algorithms are studied and compared to minimize energy usage in
Net Zero Energy Buildings and Smart Homes. These algorithms are GWO (Grey Wolf
Optimization), PSO (Particle Swarm Optimization) and differential evolution (DE). First, the
algorithms are run using the dataset of the appliances and simulation results are obtained in the
18
form of graphs and tables which are then compared using the suitable performance parameters.
It is clear from these graphical results that GWO algorithm gives the best results and energy
usage is reduced.
19
REFERENCES
[1] Ahmed, A., Ge, T., Peng, J., Yan, W. C., Tee, B. T., & You, S. (2022). Assessment of
the renewable energy generation towards net-zero energy buildings: A review. Energy
and Buildings, 256, 111755.
[2] Liu, X., Ivanescu, L., Kang, R., & Maier, M. (2012). Real-time household load priority
scheduling algorithm based on prediction of renewable source availability. IEEE
Transactions on Conmer Electronics, 58(2), 318-326.
[3] Siami-Namini, S., Tavakoli, N., & Namin, A. S. (2018, December). A comparison of
ARIMA and LSTM in forecasting time series. In 2018 17th IEEE international
conference on machine learning and applications (ICMLA) (pp. 1394-1401). IEEE.
[4] Hatta, N. M., Zain, A. M., Sallehuddin, R., Shayfull, Z., & Yusoff, Y. (2019). Recent
studies on optimisation method of Grey Wolf Optimiser (GWO): a review (2014–
2017). Artificial intelligence review, 52, 2651-2683.
[5] Jain, M., Saihjpal, V., Singh, N., & Singh, S. B. (2022). An overview of variants and
advancements of PSO algorithm. Applied Sciences, 12(17), 8392.
[6] Şenel, F. A., Gökçe, F., Yüksel, A. S., & Yiğit, T. (2019). A novel hybrid PSO–GWO
algorithm for optimization problems. Engineering with Computers, 35, 1359-1373.
[7] Sun, J., Zhang, Q., & Tsang, E. P. (2005). DE/EDA: A new evolutionary algorithm for
global optimization. Information Sciences, 169(3-4), 249-262.
[8] Al-Rakhami, M., Gumaei, A., Alsanad, A., Alamri, A., & Hassan, M. M. (2019). An
ensemble learning approach for accurate energy load prediction in residential buildings.
IEEE Access, 7, 48328-48338.
[9] Upton, E., & Halfacree, G. (2016). Raspberry Pi user guide. John Wiley & Sons.
[10] Blum, C. (2005). Ant colony optimization: Introduction and recent trends. Physics of
Life reviews, 2(4), 353-373.
[11] Mirjalili, S., & Lewis, A. (2016). The whale optimization algorithm. Advances in
engineering software, 95, 51-67.
[12] Mathew, T. V. (2012). Genetic algorithm. Report submitted at IIT Bombay, 53.
20
