Energy Economics 111 (2022) 106023

Contents lists available at ScienceDirect

Energy Economics
journal homepage: www.elsevier.com/locate/eneeco

What is the effect of weather on household electricity consumption?

Empirical evidence from Ireland
Jieyi Kang a,b ,∗, David M. Reiner a,c
a Energy Policy Research Group, University of Cambridge, United Kingdom
b
Department of Land Economy, University of Cambridge, United Kingdom
c
Judge Business School, University of Cambridge, United Kingdom

ARTICLE INFO ABSTRACT

JEL classification: We explore the links between weather variables and residential electricity consumption using high-resolution
C55 smart metering data. While weather factors have been used for grid-level electricity demand estimations, the
D12 impact of different weather conditions on individual households has not been fully addressed. The deployment
R22
of smart meters enables us to analyse weather effects in different periods of the day using hourly panel
Q41
datasets, which would previously have been impossible. To conduct the analysis, fixed-effects models are
Keywords: employed on half-hourly electricity consumption data from 3827 Irish household meters. We demonstrate
Weather effects that temperature has robust and relatively flat effects on electricity demand across all periods, whereas rain
Residential electricity consumption
and sunshine duration show greater potential to affect individual behaviour and daily routines. The models
Fixed-effects models
show that the most sensitive periods differ for each weather variable. We also test the responses to weather
Smart metering data
factors for weekends and workdays. Weather sensitivities vary with the day of the week, which might be
caused by different household patterns over the course of the week. The methodology employed in this study
could be instructive for improving understanding behavioural response in household energy consumption. By
using only weather indicators, this approach can be quicker and simpler than traditional methods – such as
surveys or questionnaires – in identifying the periods when households are more responsive.

1. Introduction on longer-time frames, such as monthly household usage, or relatively

shorter periods (daily consumption) but at the regional level (Pardo
In recent years, there has been an increase in residential smart meter et al., 2002; Davies, 1958; Atalla and Hunt, 2016; Trotter et al., 2016).
installations in many jurisdictions as they move to modernise their High-frequency individual usage data makes it possible to examine
electricity networks (Eid et al., 2017). The old mechanical metering the price effects during a specific short period during a day rather
systems usually record monthly energy consumptions of households, than using daily or monthly time steps. Although the results of the
which limit the possibility of understanding residential electricity con- efficiency of different price schemes can be contradictory, increasingly
sumptions in depth. Besides, dynamic pricing of electricity is impossible studies have been done in the field to examine the effects from different
using current metering infrastructures, due to the technical constraints perspectives.
of having no real-time usage data. In light of these concerns, the de- However, the influence of weather in residential electricity con-
ployment of Advanced Metering Systems can potentially be part of the
sumption is one area that has not been extensively studied, although
solution to achieve greater energy efficiency. There is one significant
it has been widely accepted as an important factor affecting energy
advantage of smart metering that is widely accepted — The new tech-
demand. The exploration of the relationship between energy consump-
nologies record high-resolution data of household electricity usage and
tion and weather is often seen in two sets of studies: (1) weather as
increase the visibility of energy consumption. As a consequence, the
control variables in models focusing on price or on socio-economic
availability of high volumes of data enables more fine-grained studies
of residential behaviour and consumption patterns (Razavi et al., 2019). effects (Wangpattarapong et al., 2008; Newsham and Bowker, 2010;
Thus, one area that particularly benefits from the installation of Cosmo and O’hora, 2017); (2) alternatively, weather has been used as
smart meters is the study of the effects of pricing structures on elec- the main independent variables but only when investigating the rela-
tricity consumption. Previous studies in this area have focused either tionship between daily or even monthly regional demand and weather

∗ Corresponding author at: Energy Policy Research Group, University of Cambridge, United Kingdom.
E-mail address: jieyi.kang@hotmail.com (J. Kang).

https://doi.org/10.1016/j.eneco.2022.106023
Received 19 July 2021; Received in revised form 29 March 2022; Accepted 8 April 2022
Available online 29 April 2022
0140-9883/© 2022 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
J. Kang and D.M. Reiner Energy Economics 111 (2022) 106023

variables (Moral-Carcedo and Vicéns-Otero, 2005; Costa and Kahn, CDD, HDD, and the lag effects of CDD and HDD. One compelling
2010; Blázquez et al., 2013). Weather variables such as temperature, argument they make is that the effect of temperature on the weekend
precipitation, relative humidity, wind speed, cloud cover, and sun is slightly different than on working days. Furthermore, a model based
duration are the most common variables used in both types of research. on aggregated monthly or annual data might not be able to reveal
In spite of interest in the relationship between electricity consumption differences between the two cases. In addition to temperature, rainfall
and weather, few studies have studied the possible association during is another common variable examined. Hor et al. (2005) investigated
different periods of the day due to limitations on the frequency of monthly electricity demand from 1983 to 1995 in the UK and found a
energy use data (Davies, 1958). Would specific findings hold in every very weak negative relationship between rainfall and monthly demand.
period? For example, will residential customers reduce their consump- However, they argued that the correlation between demand and rainfall
tion in every period of a sunny day? Are the weather responses, in fact, should be stronger but that the weak unexpected negative coefficient
period-dependent? A better understanding of the weather impact on is mainly because they used only national-level data, while rainfall
electricity can assist researchers, policymakers and energy companies. is very location-specific. Davies (1958)’s work considered aggregated
A study of how residential customers respond to weather in different country-level electricity demand in England and Wales arguing that
periods can provide insights into daily patterns of household behaviour, five meteorological elements affect demand: temperature, wind speed,
e.g during which periods a family is more likely to be active or often cloudiness, visibility, and precipitation. Temperature allied with wind
go out. speed determines the need for heat, while the remaining variables
We examine here the weather response at different times of day determine the level of daylight illumination, affecting lighting de-
using fixed-effects models on high-frequency usage data from Ireland’s mand. The study divides daily demand into eight three-hour periods
Smart Metering Electricity Behavioural Trial (Commission for Energy of demand to verify whether the effect of weather is the same across
Regulation (CER), 2012a) combined with weighted weather data from different periods of a day. The results show that temperature has a peak
five weather stations in Ireland. Due to the half-hourly data available influence on demand around 9:00 and a lower coefficient during the
from smart meters, we are able to investigate the household response 17:00 period. However, the direct effects of rainfall are only evident
to weather during different periods. We aim to provide evidence that at 17:00. The findings indicate that the effect of a weather variable
the weather sensitivities are indeed period-dependent and that weather is not constant through a day, and it could be interesting to examine
factors may be good proxies for household behaviour patterns in dif- the differences in the residential sector specifically. Many researchers
ferent periods of a day. In addition, this paper explores the impact of (Pardo et al., 2002; Räsänen et al., 2010; Albert and Rajagopal, 2013)
weather on the differences in electricity demand between weekends
agree with Davies (1958) that weather indices such as humidity, wind
and workdays, thereby demonstrating that the relationships between
speed, cloudiness, and barometric pressure are suitable explanatory
weather and energy demand are not universal.
factors for weather sensitivity, although those variables may have less
The paper continues by reviewing the related literature of weather
significant influence on electricity demand than temperature, rainfall
effects in Section 2 and the details of the dataset used are specified in
and sun duration.
Section 3. The two main models used and the explanation of variable
As discussed above, studies involving weather effects have paid
selection are described in Section 4. Results of the models are presented
more attention to total electricity consumption in a region. There
in two parts in Section 5, and Section 6 provides a discussion of
has been a lack of panel data to support deeper studies of the resi-
potential implications and offers some conclusion.
dential electricity sector — current panel studies concerning weather
and residential electricity are primarily based on aggregated regional
2. Literature review
panel data. Atalla and Hunt (2016) looked at the residential electric-
ity demand in six Gulf Cooperation Council countries using a panel
2.1. Weather effects on demand in general
dataset of annual demand in slightly different periods from country-
The discussion of weather variables often appears in two sets of to-country. CDD and HDD are the only weather indicators used but
studies in this field: one is model establishments for electricity con- do not necessarily have significant impacts on demand. It depends on
sumption forecasting and usually at an aggregated regional/national geographic locations and whether there is variation in the weather
level. For example, Mirasgedis et al. (2006) summarise the studies variable. Blázquez et al. (2013) used aggregate monthly panel data at
paying particular attention to short-term forecasting and the role of the province level for 47 Spanish provinces from 2000 to 2008. The
weather variability. They claim that based on the experience of utilities, authors acknowledge that in panel data analysis, fixed-effects models
the main weather factors affecting electricity consumption are temper- (FE) or random-effects models could be helpful to control unobserved
ature, humidity, and precipitation in decreasing order of importance, heterogeneity, however, neither of these was appropriate for their study
while wind speed and solar radiation is not significant for the Greek since they include a lagged dependent variable in their model. Again,
mainland. Therefore, they only include the two weather variables CDD and HDD are also the only weather conditions considered, which
(temperatures and relative humidity) in the models predicting the is common in panel studies of regional residential electricity consump-
mid-term electricity consumption in Greece. Instead of using outdoor tion. Due to the lack of data at household level, very little research has
temperatures directly, heating degree days (HDD) and cooling degree been done based on non-aggregate residential consumption. Henley and
days (CDD) are used to reflect the non-linear relationship between Peirson (1998) studied residential energy demand and the interaction
temperature and demand, which is particularly common in electricity of price and temperature based on a Time-of-Use (TOU) trial with 150
demand studies (Bessec and Fouquau, 2008; Alberini and Filippini, households between April 1989 and March 1990. Through a fixed-
2011; Boogen et al., 2017). However, in these studies the effects of effects model, they found that the effect of temperature is negative and
weather are based on total consumption including all sectors, not just non-linear, and the magnitudes vary for different periods. Alberini and
on residential consumption specifically. Thus, the importance of these Towe (2015) attempted to estimate residential electricity usage savings
factors still needs to be examined with a particular focus on residential from energy efficiency programmes. They assembled a panel dataset
electricity demand. The second type of research where weather vari- of monthly electricity usage and bills for a sample of about 17,000
ables are often included is in studies of the determinants of regional households in Maryland from 2008 to 2012. They used Difference-in-
electricity consumption. Such studies rarely focus solely on residential Difference’’ and fixed-effects models to capture annual and seasonal
electricity demand but rather on total regional consumption. Trotter household effects, and season-by-year effects. Weather effects are not
et al. (2016) examined the relationship between climate and daily the focus of the study, but CDD and HDD were included for monthly
electricity demand in Brazil where the only weather factors used are consumption control.

2
J. Kang and D.M. Reiner Energy Economics 111 (2022) 106023

2.2. Weather effects in studies using smart metering data Table 1

Residential Time-of-Use tariffs 1st January to 31st December 2010.
Cents per Night Day Peak
In light of the trend of smart meter installation around the world,
kWh 23.00–8.00 8.00–17.00 17.00–19.00
availability of household-level consumption data has begun to change. 19.00–23.00 weekdays (Mon to Fri), excluding
One of the main innovations brought by smart meters is that electric 17.00–19.00 weekends bank holidays
utilities can obtain huge volumes of high-resolution household usage and bank holidays
data. A daily load profile of a household that depicts daily consump- Tariff A 12.00 14.00 20.00
tion trends from midnight to 11:59 p.m can now be easily drawn. Tariff B 11.00 13.50 26.00
Tariff C 10.00 13.00 32.00
High sampling frequencies provide operators with the opportunity to
Tariff D 9.00 12.50 38.00
better understand consumption patterns of their residential customers.
The availability of household consumption data enables researchers to
identify the determinants of residential demand and the difference of
effects on the demand of different periods of a day in more depth. One explored extensively. Furthermore, weather impacts are commonly dis-
main strand of the literature using smart meter data investigates the cussed at an aggregate level, e.g., daily or monthly level. However, as
effects of socio-economic and house-specific variables on load profiles. proposed in some studies (Davies, 1958; Henley and Peirson, 1998), the
Anderson et al. (2017) summarised the existing evidence of household impact of weather indices might differ depending on the time of a day.
characteristics linked to load profiles and categorised those variables The lack of research could be due to the limited availability of high-
into three subgroups: (1) household features, such as number of per- frequency household-level data. Even with greater access to detailed
sons, number of children, and age distribution (Yohanis et al., 2008; household usage data, the focus of studies using smart meter data has
Beckel et al., 2015); (2) dwelling status: e.g. dwelling type, household been on time-of-use tariffs, rather than weather effects. Therefore, a
tenure, number of rooms (Firth et al., 2008; McLoughlin et al., 2012); comprehensive study of the weather effects on residential electricity
and (3) householder characteristics: employment status, social status, demand and household behaviour patterns during different periods of
age and gender. Other Householder variables, such as education level, the day would be helpful to filling the gap.
ethnic group, marital status and household income are also found
to have significant impact on demand and load profiles (McLoughlin 3. Data
et al., 2012; Carroll et al., 2014). Nevertheless, research into electricity
demand and household features have rarely paid attention to weather 3.1. Residential electricity consumption data
variables. There is little evidence of weather effects on residential
demand from household-level data. Kavousian et al. (2013) examine The smart meter dataset used in this paper was collected as part
structural and behavioural determinants of residential consumption of Ireland’s Electricity Smart Metering Consumer Behavioural Trial,
using a dataset of 10-minute interval smart meter readings from 1628 which includes 4000 residential customers (Commission for Energy
households in California. They prove that weather and location are Regulation (CER), 2012a).
among the most important determinants of residential electricity use. Half-hourly readings of usage were recorded by meters installed in
However, the only weather variables, they include in their models are the trial from 15 July 2009 to 31 December 2010. During the bench-
outdoor temperature and climate zone. mark period (July 2009 to Dec 2009) all households were charged
Another set of studies use smart metering data and consider weather a fixed tariff. From 1 January 2010, those who were selected into
variables to identify the effectiveness of time-of-use tariffs. Torriti treatment groups were charged time-of-use (TOU) tariffs. There were
(2012) took advantage of data from a TOU and smart metering trial in 4 TOU tariff periods: peak (17:00-18:59 Monday–Friday, excluding
Northern Italy involving quarter-hourly readings from 1446 households public holidays), day (08:00–16:59; 19:00–22:59 Monday–Friday, plus
from 1 July 2009 to 30 June 2011. The findings show that peak 17:00–18:59 public holidays, Saturday and Sunday) and night (23:00–
load shifting took place for morning peaks and created a split into 07:59) periods. The tariff structure is shown in Table 1. In order to
two peaks for evening periods, while total consumption increased by control the effects caused by the price incentives, our research only
13.69%. The only weather variable, temperature, is used to control includes data recorded from 1 January 2010 onward, where tariffs are
for the effect of weather variation, but the effect is not discussed in constant across all households during that period. In addition, homes
details. Other studies have used data from the large-scale trial smart with average daily consumption of more than 54 kWh are also removed
metering experiment or Consumer Behavioural Trial (CBT) carried out because these outliers may not be residential consumers, but are more
by the Irish Commission for Energy Regulation (CER). Cosmo et al. likely to be small enterprises or home-based enterprises1
(2014) utilise the CBT panel data of over 4000 households to explore
whether the designed TOU is efficient in reducing peak demand. Two 3.2. Weather data
weather variables – sunshine duration and heating degree days – are
included. Their results show that HDD are positively associated with To generate daily weather observations at specific times of day,
consumption, while the opposite relationship is found for sunshine hourly weather data provided by the Irish Meteorology Office (Met
duration for the three periods considered (day, peak, and night). They Éireann) are matched with household electricity consumption data.
only used the weather data from Dublin Airport weather station, as Recorded hourly observations are dry-bulb (air) temperature (◦ C), rela-
detailed information on household location is not available. However, tive humidity (%), wind speed (kph), and fraction of sunshine per hour
considering that the selected households were drawn from across the (%). Since location information is not provided by CER due to privacy
country, a population-weighted weather dataset from different weather concerns, a population-weighted weather dataset should be considered
stations would be more accurate for a study of weather effects. In to reflect consumption response to weather from families living across
addition, the time periods may be too long since weather effects could Ireland (Valor et al., 2001; Auffhammer and Aroonruengsawat, 2012).
change dramatically over the course of a period lasting as long as 10 h. As a result, four weather stations, Dublin Airport, Valentia Observatory,
From the review above It can be easily seen that little research
has focused on weather effects and weather influences are usually
introduced as control variables for other research objectives. Generally, 1
Furthermore, the impacts of daylight saving time (31st October 2010 and
temperature is the main weather factor considered and other variables, 25th March 2010) are taken into account. The data from the second 2 am (end
such as precipitation, wind speed, and sun duration, have not been of DST) is deleted from the dataset to avoid double counting.

3
J. Kang and D.M. Reiner Energy Economics 111 (2022) 106023

Fig. 1. Comparison of county level distribution between acceptances and total population.

Table 2 for a weekend day. It provides a comprehensive view of daily life in

Average correlations between weather variables and household demands.
Ireland and possible behaviour during every 15-minute slot of a day.
Observatory stations Rain Temperature R. Humidity Wind Speed Sun Duration As a result, the findings of the survey can be particularly helpful in
Belmullet 0.051 −0.274 0.088 −0.031 −0.088 two ways: (1) to divide hourly data more accurately and avoid splitting
Cork Airport 0.016 −0.274 0.091 0.049 −0.090
one major daily activity into two periods, which may distort the actual
Dublin Airport 0.047 −0.267 0.194 0.041 −0.132
Valentia 0.024 −0.265 −0.026 0.000 −0.076
response by either exaggerating or underestimating the effects. For
Weighted 0.051 −0.271 0.165 0.020 −0.129 instance, separating 12:00–14:00 into two different periods may cancel
out part of the impacts of lunchtime; (2) to better understand how
people respond to weather changes. For example, if people are less
sensitive to rainfall during 12.00–14.00, it could be a lunchtime effect.
Belmullet and Cork Airport, were chosen. The first three synoptic sta- Therefore, this survey data provides a supplementary tool to explain
tions are the choices of Met Éireann for regular Irish weather statements and confirm the results obtained from the proposed models.
(Met Éireann, 2018) and Cork was selected to ensure enough sufficient
regional representation because a significant number of participating 4. Methodology
households live in Cork (See Fig. 1). Since the distribution of the
final acceptances onto the trial was similar to the total population at As seen in the literature review in Section 2, studies of the effects of
county-level (Fig. 1), the population ratios around the four stations are weather variables have mainly focused on relationships between daily
aggregated and calculated as weights to create a new dataset to match electricity consumption and daily weather change. It is not clear that
with the consumption data. how households respond to weather change at different times of day.
The weights of the population ratios used are 0.535 for Dublin, To investigate the weather sensitivities in different periods of a day, it
0.175 for Cork, 0.16 for Belmullet and 0.11 for Valentia. The method is reasonable to assume that households will not change their behaviour
of how to draw the boundaries of each station is not ideal due to immediately when the weather changes. In order to capture the lagged
the absence of household location information. For example, County effects, the hourly data is aggregated and divided into periods based on
Clare (CE in Fig. 1) can be associated with Cork station or Belmullet patterns of daily activities, rather than using raw hourly data directly.
station. However, as the weather in Ireland is relatively similar, the Although autoregressive models can be used on hourly data to control
boundaries/weights hardly change the final results as we tried different lagged effects, it might complicate the situation and the lag lengths
weights for the analysis. suitable for weather effects are not clear and there is no agreed lag
We investigated which datasets could better reflect the consumption time in the literature.
responses to weather changes through correlation analysis. First, we Two rules are employed in separating the time periods: (1) To
calculated the correlation coefficients of each household to the weather control for possible price effects caused by the TOU tariffs, the time
variables separately. Secondly, the average correlation coefficients for periods chosen should not cross over two different tariffs (i.e., the tariff
the weather factors were obtained for each weather dataset and shown structure shown in Table 1); and (2) A period does not split major
in Table 2. The weighted method seems the most balanced dataset that activities. On the basis of these rules, the tariffs provides natural breaks
is with higher correlations across different weather variables. at the early morning, peak and night periods. However, the day price
The descriptive statistics for the weather variables are described in period (see Table 1) is much longer than the other periods, which may
Table 3.2 obscure the real response, and so needs to be sub-divided. In the end,
9 periods are set as follow with the help of the time use study: early
3.3. Time use study data morning (6:00–8:00), day_1 (8:00–10:00), day_2 (10:00–12:00), day_3
(12:00–15:00), afternoon/day_4 (15:00–17:00), peak (17:00–19:00),
In order to help divide hourly data into several discrete periods, the early evening/evening_1 (19:00–21:00), evening_2 (21:00–23:00) and
Irish National Time-Use Survey 2005 (McGinnity et al., 2005) is used. night (23:00–3:00). The summary of weather variables in these periods
It collected detailed national time-use statistics on over 1000 adult can be seen in Table 3. The period of night_2 from 3:00 to 6:00 is not
participants’ daily activities, which includes two complete diaries of discussed in the main analysis, since no major activities occur during
their activities over a 24-hour period — one for a weekday and another that period and that the estimates for weather variables are generally
less than 0.5% (seen in Table 5). The descriptions of the four main
activities with the highest proportions of people doing on workdays
2
The rules of how these periods are divided will be introduced in Section 4. and weekends are shown in Table 4. The numbers in each cell represent

4
J. Kang and D.M. Reiner Energy Economics 111 (2022) 106023

Table 3
Descriptive Statistics for the weather variables.
Morning Day_1 Day_2 Day_3 Day_4 Peak Evening_1 Evening_2 Night
Mean 7.23 8.64 10.18 11.03 10.75 9.83 8.75 7.91 7.20
◦ Std. Dev. 5.42 5.56 5.37 5.33 5.71 5.88 5.63 5.32 5.25
Temperature ( C)
Min −7.49 −7.49 −5.16 −2.84 −3.60 −4.98 −5.20 −5.85 −6.33
Max 17.52 18.09 19.94 20.57 20.78 19.82 18.69 17.52 17.09
Mean 89.26 84.43 77.98 74.21 74.97 78.61 83.04 86.48 88.75
Std. Dev. 5.08 7.54 9.74 10.48 11.10 10.39 8.42 6.21 4.75
Rel. Humidity (%)
Min 64.67 62.83 51.77 46.53 47.57 51.18 57.25 64.74 71.44
Max 97.52 97.14 96.15 96.32 96.53 96.91 97.27 97.22 97.28
Mean 8.24 9.04 10.06 10.67 10.46 9.60 8.61 8.05 8.02
Std. Dev. 3.56 3.55 3.72 3.77 3.84 3.87 3.82 3.89 3.78
Wind speed (knots)
Min 2.03 2.54 2.91 3.46 2.66 2.86 1.75 1.47 2.06
Max 27.16 27.87 27.07 29.03 28.93 29.79 28.83 27.93 24.34
Mean 0.14 0.36 0.48 0.46 0.33 0.20 0.07 0.00 0.00
Median 0.04 0.31 0.50 0.45 0.25 0.03 0.00 0.00 0.00
Sun duration (% per hour) Std. Dev. 0.23 0.30 0.31 0.30 0.31 0.29 0.15 0.00 0.00
Min 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
Max 0.97 0.98 0.98 0.98 0.98 0.98 0.91 0.03 0.00
Mean 0.08 0.07 0.08 0.08 0.10 0.13 0.12 0.10 0.10
Median 0.01 0.01 0.01 0.01 0.00 0.01 0.00 0.00 0.01
Rainfall (mm) Std. Dev. 0.19 0.18 0.18 0.18 0.23 0.31 0.29 0.23 0.22
Min 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
Max 1.40 1.34 1.32 1.71 1.96 2.27 1.93 1.84 1.82

Table 4
Percentage of four main activities for workdays and weekends.
1st 2nd 3rd 4th
Workdays Weekends Workdays Weekends Workdays Weekends Workdays Weekends
Morning Sleep Sleep Personal Care Personal care Eating Paid Employ Travel Travel
6.00–8.00 46.3–91.9 72.9–92.4 1.7–12.4 1.2–6.2 0.3–9.7 0.6–4 0.6–9 0.4–3.1
Day_1 Paid Employ Sleep Sleep Eating Travel Perscare Personal care Paid Employ
08.00–10.00 11.6–35.8 26.4–61.5 8.8–32 5.5–14.2 9–16 8.1–11.5 4.6–15.9 5.6–10.9
Day_2 Paid Employ Sleep Cleaning Eating Eating Paid Employ Travel Cleaning
10.00–12.00 33–39 6.7–18 7.4–9.1 4.5–12.4 3.1–9 9.1–10.8 6.3–8.3 6.2–10.3
Day_3 Paid Employ Eating Eating Paid Employ Breaks Cooking Shopping Travel
12.00–15.00 17.9–38.7 5.6–23.7 3–26.2 6.8–12.4 1.8–14.5 7.2–11.9 3.4–5.7 5.3–8.9
Day_4 Paid Employ Eating Eating Paid Employ Travel Travel Cleaning Shopping
15.00–17.00 34.9–38.6 3.8–16.3 2.2–9.6 9.2–11 5.3–7.1 7–9.1 4.2–6.6 7.3–8.5
Peak Paid Employ Paid Employ Travel Chatting Cooking Travel Eating TV/Vdieos
17.00–19.00 18.2–35.9 9.1–10.5 8–13.9 6.6–9.6 4.3–11.9 6.1–8.8 2.6–9.7 7.6–8.8
Evening_1 Eating TV/Videos TV/Videos Eating Paid Employ Travel Travel Chatting
19.00–21.00 6.5–20.3 10–18.3 7.8–19.2 6.6–17.8 8.3–12.7 5.2–9.8 4.9–11.9 7.2–9
Evening_2 TV/Videos TV/Videos Resting Eating out Chatting Resting Eating Chatting
21.00–23.00 18.4–33.2 16–26.8 8.6–10.8 5.2–12.4 6.8–8.5 8.7–11.3 2.4–6.9 8–9.7
Night Sleep Sleep TV/Videos TV/Videos Resting Eating out Chating Resting
23.00–03.00 37.2–90.1 26.1–81.3 0.4–18.3 10–17.7 0.8–6.6 6.5–16.3 0.3–5.3 0.9–6.5

the minimum and maximum percentages of people doing the activities The first model (Model 1) explores the effects of selected weather
during each period of a day. variables on electricity consumption and is as follows:
As panel data allows for the exploitation of both time and cross-
∑
5 ∑
12 ∑
7
ℎ
section dimensions, it has the potential to eliminate unobserved het- log (𝑞𝑖,𝑡 ) = 𝛼𝑖ℎ + 𝛿𝑝ℎ 𝑊𝑝,𝑡 + 𝜆ℎ𝑚 𝑀𝑚,𝑡 + 𝜃𝑗ℎ 𝐷𝑗,𝑡 + 𝜁𝑦ℎ 𝐻𝑦 + 𝜀ℎ𝑖,𝑡 (1)
erogeneity in the data (Asteriou and Hall, 2011). As a result, given 𝑝=1 𝑚=2 𝑗=2
the nature of the panel dataset, two fixed-effect models are employed. ℎ ) denotes the loga-
where ℎ = 1 − −9 for each of the 9 periods. log (𝑞𝑖,𝑡
Although random-effects (RE) models are also used in the related rithm of household 𝑖’s daily electricity consumption in kilowatt-hours
literature, fixed effects (FE) models better suit the purposes of this during one period of day 𝑡. As discussed above, there are 9 periods in
study. a day. The model therefore is run for each period separately; 𝑊𝑝,𝑡 are
With FE models, the focus is given to weather variables, while the the five weather variables; 𝑀𝑚,𝑡 are dummies of month indicators, and
effects of variables whose values are consistent across time January is selected as the baseline (when 𝑚=1) ; 𝐷𝑗,𝑡 indicate day of
(Wooldridge, 2013), such as demographics, housing conditions, and week and the reference category is Monday (where 𝑗=1); the coefficient
electric appliance ownership, are captured in a single fixed-effects esti- 𝛿 represents the expected weather effect on consumption, while the
mator since the focus of the study is not on household characteristics. In coefficients 𝜆 and 𝜃 quantify the consumption differences between the
addition, the results of the Hausman test imply that FE models are more expected effect (the month 𝑖 and the day 𝑗) and the baseline (January
suitable, since the null hypotheses of RE models is rejected (p-values and Monday); 𝐻𝑦 is the public holiday dummy; 𝛼𝑖ℎ are household fixed
of 0.0000). effects and 𝜀ℎ𝑖,𝑡 is a stochastic disturbance term. There may also be

5
J. Kang and D.M. Reiner Energy Economics 111 (2022) 106023

Table 5
Estimated coefficients from Model 1 of weather effects.
Variables (1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
Morning Day_1 Day_2 Day_3 Day_4 Peak Evening_1 Evening_2 Night Night_2
6.00–8.00 8.00–10.00 10.00–12.00 12.00–15.00 15.00–17.00 17.00–19.00 19.00–21.00 21.00–23.00 23.00–03.00 3.00–6.00
Weather
Temperature −0.001792*** −0.005751*** −0.01291*** −0.01402*** −0.01326*** −0.01307*** −0.01216*** −0.01015*** −0.009062*** −0.0007606***
Rainfall −0.0002214 0.04434*** 0.04876*** 0.01696*** 0.04374*** 0.01897*** 0.009795*** 0.009304*** 0.004056 0.001362***
Sun duration 0.01921*** 0.01038*** 0.002493 0.01257*** −0.03527*** −0.04913*** −0.09356*** / / /
Relative humidity −0.001364*** −0.0007425*** 0.001294*** 0.002896*** 0.001511*** 0.001598*** 0.001377*** 0.0006098*** 0.001433*** 0.001433***
Wind speed −0.0006325** −0.00001235 0.002379*** 0.003416*** 0.003632*** 0.002928*** 0.003103*** 0.002508*** 0.002866*** 0.0002630*
Weather coefficients above in a readable form
Temperature −0.2% −0.6% −1.3% −1.4% −1.3% −1.3% −1.2% −1.0% −0.9% −0.1%
Rainfall 4.4% 4.9% 1.7% 4.4% 1.9% 1.3% 1.0% 0.4% 0.1%
Sun duration 1.9% 1.0% 0.2% 1.3% −3.5% −4.9% −9.4% \ / /
Relative humidity −0.1% −0.1% 0.1% 0.3% 0.2% 0.2% 0.1% 0.1% 0.1% 0.1%
Wind speed −0.1% 0.2% 0.3% 0.4% 0.3% 0.3% 0.3% 0.3% 0.0%
Time
Sunday Ref Ref Ref Ref Ref Ref Ref Ref Ref
Monday 0.2519*** 0.06085*** −0.3173*** −0.2831*** −0.06142*** 0.1199*** 0.07258*** 0.07611*** 0.009453*** −0.03566***
Tuesday 0.2891*** 0.08829*** −0.3536*** −0.3214*** −0.08644*** 0.1056*** 0.06893*** 0.08320*** 0.06792*** 0.06792***
Wednesday 0.2855*** 0.09368*** −0.3522*** −0.3289*** −0.1049*** 0.07815*** 0.04770*** 0.06515*** 0.01134*** −0.04373***
Thursday 0.2928*** 0.09713*** −0.3717*** −0.3506*** −0.1207*** 0.06379*** 0.03537*** 0.06527*** 0.02789*** −0.04102***
Friday 0.2727*** 0.09912*** −0.3376*** −0.3145*** −0.1003*** 0.03893*** 0.008178*** 0.02278*** 0.1039*** −0.02955***
Saturday 0.05649*** 0.08121*** −0.07385*** −0.09508*** −0.003968 0.05539*** 0.00417 −0.02833*** 0.09864*** −0.01729***
Non-holiday Ref Ref Ref Ref Ref Ref Ref Ref Ref Ref
Holiday −0.08155*** −0.04648*** 0.1033*** 0.09791*** 0.01649*** −0.06722*** −0.03440*** −0.02201*** 0.04221*** 0.03038***
Month
January Ref Ref Ref Ref Ref Ref Ref Ref Ref Ref
February 0.01436*** −0.003099 −0.07754*** −0.09040*** −0.1293*** −0.1381*** −0.04967*** −0.05712*** −0.07450*** −0.05259***
March −0.008113** −0.03512*** −0.08563*** −0.09903*** −0.1810*** −0.2989*** −0.1011*** −0.08255*** −0.1096*** −0.06409***
April −0.08523*** −0.08503*** −0.07490*** −0.1029*** −0.2316*** −0.4113*** −0.3512*** −0.1131*** −0.08343*** −0.07262***
May −0.08081*** −0.09148*** −0.09447*** −0.1194*** −0.2253*** −0.4209*** −0.4227*** −0.2038*** −0.07287*** −0.06053***
June −0.07489*** −0.08472*** −0.04169*** −0.06096*** −0.1848*** −0.4117*** −0.4391*** −0.3027*** −0.03034*** −0.01912***
July −0.1195*** −0.1150*** −0.01284** −0.04466*** −0.2042*** −0.4413*** −0.4714*** −0.2942*** −0.04733*** −0.01373***
August −0.1254*** −0.1164*** −0.003654 −0.03843*** −0.2074*** −0.4392*** −0.4451*** −0.2142*** −0.05090*** −0.01801***
September 0.006025 −0.05519*** −0.06729*** −0.07816*** −0.1684*** −0.3890*** −0.2414*** −0.1378*** −0.09302*** −0.02973***
October −0.02453*** −0.04794*** −0.07490*** −0.09983*** −0.1810*** −0.3057*** −0.1271*** −0.1383*** −0.1027*** −0.06576***
November 0.01848*** −0.01011*** −0.09065*** −0.1065*** −0.06440*** −0.05878*** −0.08408*** −0.09882*** −0.1026*** −0.06209***
December 0.05976*** 0.1034*** 0.1129*** 0.05702*** 0.09065*** 0.01957*** 0.007024* 0.02495*** 0.03491*** 0.0847207***
Constant −0.2103*** 0.2800*** 0.5056*** 0.9000*** 0.5277*** 0.8759*** 0.8964*** 0.7818*** 0.4845*** −0.103125***
Obeservations 1495843 1496318 1495849 1496196 1496368 1496948 1497084 1497236 1497783 1497123
R2 0.5637 0.4846 0.4322 0.4735 0.4702 0.5116 0.523 0.5553 0.6319 0.6722

heating/cooling degree days, hours of sunshine, rainfall, wind speed

and relative humidity are five leading variables that have been used
in the relevant research. Model 1 employs all these variables, apart
from heating degree days (HDD) and cooling degree days (CDD), which
are replaced by air temperature in the equation. The reason for this
substitution is that HDD and CDD are used to reflect the non-linear
relationship between daily electricity demand and daily temperature.
However, although a non-linear response is found in other studies
(Woods and Fuller, 2014; Auffhammer and Mansur, 2014), there is no
clear non-linear relationship, but rather a linear correlation in Irish
houses (Fig. 2). One reason may be that the temperature range in
Ireland is relatively flat and air conditioning uncommon in Ireland. In
addition, Model 1 examines the weather during different periods of the
day, rather than daily changes, so the using CDD and HDD would not
suit the case.
The second model (Model 2) is based on Model 1 but streamlined
to focus on estimating how differently households respond to weather
Fig. 2. Average daily electricity consumption per household. changes on weekends and weekdays. The different consumption pat-
terns can be seen in Fig. 3.
To estimate the differences, the following model is tested:
unobserved household-specific differences in consumer demand, for ex- ∑
3
ℎ
ample, presence of electric dryers or other appliances. The fixed-effects log (𝑞𝑖,𝑡 ) = 𝛼𝑖ℎ + 𝜗ℎ 𝐷𝑥 + 𝜁𝑦ℎ 𝐻𝑦 + 𝛿𝑝ℎ 𝑊𝑝,𝑡
estimator used can handle it well as this household-level heterogeneity 𝑝=1

is constant over time. ∑

3 ∑
12

Although weather has been identified in many studies as an essen- + 𝛽𝑖ℎ 𝑊𝑝,𝑡 × 𝐷𝑥 + 𝜆ℎ𝑚 𝑀𝑚,𝑡 + 𝜀ℎ𝑖,𝑡 (2)
𝑝=1 𝑚=2
tial factor, no agreement has been reached on which weather variables
and in what form they should be added into the modelling. However, The model is similar to Model 1, apart from the following changes:

6
J. Kang and D.M. Reiner Energy Economics 111 (2022) 106023

Fig. 3. Average daily electricity consumption per household.

1. Day of week dummies are replaced by a workday dummy 𝐷𝑥 5. Results

to estimate the difference between workdays and weekends.
It should be highlighted that the definition of working days 5.1. General relationships between weather factors and demands
varies depending on the period of the day. Before 19:00 (peak
period), the definition remains the same as the typical sense that Analysing the weather sensitivities on periods of day basis will
Monday to Friday are working days. However, the definitions allow us to answer different questions. First, do consumers change
of working days from 19:00–03:00 are slightly different. For their behaviour alongside changes in weather and the seasons? And
the evening_1 and evening_2, periods workdays are defined as if so, which weather variables affect the consumption behaviour most
Monday to Thursday, which means 3 days for each weekend significantly? Is there any particular period in which the effect of one
because it is sensible to treat Friday evening as the start of
specific weather factor dominates? The weather effects on electricity
a weekend. Additionally, before 23:00 on a Sunday can also
consumption in Ireland do not reflect behaviour change related to heat-
be regarded as part of a weekend. However, it may be logical
ing demand since natural gas is the main heating source in Ireland and
to assume that the behaviour/life pattern for the night period
electric heating appliance ownership is low, around 10% (Commission
(23:00–3:00) on a Sunday is more similar to a workday. During
for Energy Regulation (CER), 2012a). Instead, the changes in demand
late evenings/evening_2 on weekends eating out is still the sec-
ond most common activity (Table 4) and so Sunday evenings reflect how weather factors will affect households’ daily behaviour
should not be treated as workdays. Therefore, the definition for electricity-intensive activities such as lighting, cooking, and other
of workdays for the night period is Sunday to Thursday. The household appliances (washing machines, dishwashers, dryers, televi-
analysis for holidays/public holidays applies the same rule. sions, etc.) and so will reflect both variation in household chores and
2. Only three weather variables are included in this model. Wind activities as well as whether people are at home or whether have gone
speed and relative humidity are excluded as they have less outside or away from home.
impact on demand. In addition, the objectives of this model As the dependent variable is log transformed, the coefficients of
are to examine the differences in response in the main weather independent variables present the percentage change in demand (the
factors between weekdays and weekends. Adding variables that dependent variable) for a one unit increase in that variable. The table
have limited effect can overfit the model and may lead to biased presents the estimated coefficients from the models for the ten periods
results. As a result, Model 2 only keeps three weather variables. (Table 5). We keep the coefficients in this form to show the significance
It is because: (a) the results shown in Model 1 prove that level and original results. To make the results more readable, we
humidity and wind speed have the least and almost negligible converted the coefficients into percentages and showed them below the
effects on demand. A model including the two variables would original results. For example, a coefficient of −0.01291 indicates the
weaken the model; (b) To ensure no significant variables are demand decreases by almost 1.3%.
missed, we tested the model with all five weather variables and
their interactions, which shows that with or without relative 5.1.1. Temperature
humidity and wind speed included, the results for the other three Temperature always has a negative effect on household electricity
variables remain almost the same. Therefore, the more concise consumption. This is in line with many previous research findings
model with better explanatory power is employed. (Blázquez et al., 2013; Cosmo et al., 2014) that daily electricity demand
In Model 2, the main coefficients of concern are 𝛽𝑖ℎ and 𝛿𝑝ℎ . 𝛽𝑖ℎ decreases when the daily temperature rises. This result holds across
represents the difference in demand between workdays and week- all periods, not just for average daily consumption. The reduction in
ends/holidays caused by weather variable 𝑊𝑝,𝑡 . 𝛿𝑝ℎ indicates the pos- demand with rising temperature could be caused by various drivers
sible effects of weather variable 𝑊𝑝,𝑡 on weekends/holidays. Note that including engaging in more outdoor activities and lower heating de-
together holidays plus weekends act as the reference category and that mand. Considering that the Irish heating system largely depends on
holidays are not separated from weekends because less than 10 days in natural gas (Commission for Energy Regulation (CER), 2012b), with a
a year are treated as holidays. Results from the interaction between the higher possibility that the reduction from temperature is from spending
holiday dummy and weather factors may be biased due to the limited more time outside. By contrast, the reason for the negative effect on
sample size. mornings (6:00-8:00) may be different, since most households should

7
J. Kang and D.M. Reiner Energy Economics 111 (2022) 106023

be still asleep or in the bed. The response to temperature can be 5.2. Behaviour difference between weekends and workdays
seen as the sensitivity of the activities in this period to temperature
change (warm/cold weather. Hence, a higher sensitivity represents Based on the overview of the effects of the weather factors, this sec-
the activities/behaviour in that period are more likely to be outdoor tion attempts to identify and answer the following questions: are there
activities. From Table 5, it can be seen that night (23:00-03:00), and differences in demand between weekends and workdays in weather
especially early morning (6:00-8:00) are far less sensitive than other sensitivities? What differences in daily routine between weekends and
periods with less than a 2% reduction. The highest coefficient is in the workdays can cause any discrepancies? Estimated results are shown
early afternoon (12:00-15:00), which indicates that the activities in that in Table 6. Here, the interactions of weather variables and workday
period can be most sensitive to warmer weather. dummies indicate the differences between workdays and weekends.
Likewise, to make the results easier to understand, we calculated and
converted them into percentage (%) form for estimates. The workday
5.1.2. Rain results are calculated by adding the differences between weekends and
In terms of rainfall, our prior expectations were that higher rainfall workdays, (i.e coefficients for workdays) to the baseline/reference (i.e,
could be associated with an increase in electricity demand for all coefficients for weekends). For example, the response to temperature
periods. It seems reasonable that the heavier it rains the less likely on workday mornings (−0.002) equals to the coefficient for week-
that people would go outside. As expected, all periods show a positive ends (0.002524) plus the coefficient for workdays (−0.004587). The
relationship between rainfall and consumption, except for mornings transformed results are shown below the original results.
(6:00-10:00) and late nights. The reversed sign in the early mornings
(6:00-8:00) may indicate that the households wake up later when it 5.2.1. Temperature
is raining outside. Moreover, the electricity usage in mornings (8:00- As expected, the results suggest that temperature have a negative
10:00) and late nights are rarely affected by rain. By the time many peo- effect on the demand among both weekends and workdays of all
ple have left home to work, while those who stay at home may not be periods. However, the exception is weekend early mornings. The reason
ready to go out immediately for shopping or exercises after breakfasts. for this unexpected impact of temperature is not particularly clear. Of
all periods, early weekend mornings (6:00-10:00) have the least impact,
Relatively few households are awake after 23:00, most households
which suggests that the early morning is the most insensitive period.
tend to go to bed earlier on workdays and even on weekends many
The behaviour during that period is robust and less likely to be changed
households will not stay up beyond midnight. This assumption can be
by temperature.
verified by the Time-Use Survey in Ireland (McGinnity et al., 2005),
Furthermore, weekends are in general more sensitive to temperature
which showed that more than 50% percent of people are sleeping at
change than weekdays. This difference can be explained by more ac-
23:00–23:59. The figure soars to 85% for 0:00–1:00.
tivities occurring indoors. However, the difference in the early evening
(19:00–21:00) seems negligible. It could be explained by that limited
5.1.3. Sun activities would occur during the post-dinner time on both weekends
The results of sun duration suggest two clear patterns over the and workdays, since many would enjoy an indoor relaxing time after
course of a day. The turning point seems to be 15:00, which is con- dinner. The largest difference appears at night (23:00–3:00), which is
sistent with the findings of Harold et al. (2015) and Cosmo and O’hora in line with expectations. People would be more likely to go out later
(2017). The more sunshine observed In that period, the less electricity and stay out later on weekends/holidays, especially on warmer days,
whereas people tend to go to sleep earlier on workdays even on warmer
consumed by households. Before 15:00 it has an opposite effect with the
days.
model producing a positive coefficient. The positive effect may be due
to the different nature of the activities during the two periods. Since,
5.2.2. Rainfall
as discussed, electricity consumption does not reflect heating demand,
The effects of rain represent to what percentage the electricity
the results would indicate that more indoor activities (e.g chores, DIY,
demand would change by the rainfall. From the results it is possible
gardening) tend to occur over 6:00-15:00 whereas there was greater
to infer how flexible plans or activities are in a given period. A period
chance of outdoor activities (e.g., shopping, sports) occurring in the with higher sensitivity to rain that there may be more outdoor activities
late afternoon and early evening. Furthermore, the larger coefficients or households prefer to go out during that period.
in the early evening (17:00-21:00) reveal that for the half year that The midday (10:00-15:00) and early evening (19:00-21:00) periods
has sunshine in the early evening (mainly late Spring and summer), on workdays are the only time slots with greater sensitivities than their
willingness to go out is particularly sensitive to sunshine during that counterparts on weekends. It is noteworthy that the sensitivities during
period. the midday period (10:00-15:00) on weekdays are exceptionally higher
than any other period in either weekends or workdays. It indicates
5.1.4. Humidity and wind speed that stay-at-home family members tend to go out during that period
on weekdays. While on weekends, the households may not be able
Relative humidity and wind speed show similar patterns in affecting
to go out early due to more chores and family care. This period
residential electricity demand. They increase demand for electricity
actually shows the largest gap between weekends and workdays, which
for all periods after 10:00. In terms of wind speed, it has limited
indicates the large underlying difference in daily routines between
impact on electricity demand in the early mornings (6:00-10:00) with
weekends and workdays are in the midday period. For example, people
less than a 0.005% reduction in demand during 6:00-8:00 and with might regularly go out on workdays while stay at home on weekends
an insignificant coefficient at 8:00-10:00. On the other hand, relative during this period. Likewise, the significant high coefficients of over 0.1
humidity has a negative relationship with consumption during the same are also found on weekend mornings (8:00-10:00) and nights (23:00-
period. Humidity have a compounding effect with temperature, where 3:00). The unusually high sensitivities on weekend mornings may be
air temperature with higher humidity may give a colder apparent tem- due to more chores done or sports activities.
perature. However, all the impacts from humidity and wind speed are Interestingly, in spite of a smaller difference compared to the 10:00-
of negligible magnitude with under 0.5% change in demand. Therefore, 15:00 period, workdays in the early evening (19:00-21:00) are surpris-
these two variables will be removed in the following model where the ingly more sensitive than on weekends, whereas a plausible hypothesis
focus is to identify the differences between weekdays and weekends for would be that evenings should be more sensitive on weekends. It could
each of the main weather factors. be a result of the timing of outdoor activities on weekdays since people

8
J. Kang and D.M. Reiner Energy Economics 111 (2022) 106023

Table 6
Estimated coefficients from Model 2 Workday and Weekend differences.
Variables (1) (2) (3) (4) (5) (6) (7) (8) (9)
Morning Day_1 Day_2 Day_3 Day_4 Peak Evening_1 Evening_2 Night
6.00–8.00 8.00–10.00 10.00–12.00 12.00–15.00 15.00–17.00 17.00–19.00 19.00–21.00 21.00–23.00 23.00–03.00
Weather
Temperature (weekend as ref) 0.002524*** −0.003289*** −0.01518*** −0.01767*** −0.01343*** −0.01192*** −0.01074*** −0.009082*** −0.01038***
Temperature * Workday −0.004587*** −0.003138*** 0.002516*** 0.005478*** 0.001465*** 0.0008027** 0.0003072 0.001683*** 0.005085***
Rainfall (weekend as ref) 0.04134*** 0.1401*** 0.02043** 0.03680*** 0.08389*** 0.04746*** 0.03440*** 0.04573*** 0.1062***
Rainfall * Workday −0.07466*** −0.1499*** 0.1075*** 0.09673*** −0.02118*** −0.01767*** 0.01490*** −0.02919*** −0.09189***
Sun duration (Weekend as ref) 0.04544*** 0.04637*** 0.03262*** −0.004638 −0.07735*** −0.08104*** −0.09148*** \ \
Sun duration * Workday −0.02227*** −0.04787*** −0.05150*** −0.01570** 0.02682*** 0.009996 −0.04168*** \ \

Weather coefficients above in a readable form

Temperature (weekend as ref) 0.25% −0.33% −1.52% −1.77% −1.34% −1.19% −1.07% −0.91% −1.04%
Temperature * Workday −0.46% −0.31% 0.25% 0.55% 0.15% 0.08% 0.03% 0.17% 0.51%
Rainfall (weekend as ref) 4.13% 14.01% 2.04% 3.68% 8.39% 4.75% 3.44% 4.57% 10.62%
Rainfall * Workday −7.47% −14.99% 10.75% 9.67% −2.12% −1.77% 1.49% −2.92% −9.19%
Sun duration (Weekend as ref) 4.54% 4.64% 3.26% −0.46% −7.74% −8.10% −9.15% s \
Sun duration * Workday −2.23% −4.79% −5.15% −1.57% 2.68% 1.00% −4.17% \ \

Time
Weekend Ref Ref Ref Ref Ref Ref Ref Ref Ref
Workday 0.3123*** 0.1146*** −0.3311*** −0.3451*** −0.1178*** 0.04710*** 0.05132*** 0.06453*** −0.1198***
Holiday 0.04419*** −0.01954*** −0.03182*** −0.02297*** −0.01975*** −0.02584*** −0.01376*** 0.000299 −0.0006169

Month
January Ref Ref Ref Ref Ref Ref Ref Ref Ref
February 0.004319 −0.008879** −0.07065*** −0.09739*** −0.1371*** −0.1472*** −0.05878*** −0.06146*** −0.08921***
March −0.004226 −0.02927*** −0.1076*** −0.1486*** −0.2024*** −0.3194*** −0.1171*** −0.08663*** −0.1314***
April −0.1016*** −0.08898*** −0.07714*** −0.1317*** −0.2542*** −0.4451*** −0.3794*** −0.1315*** −0.1149***
May −0.1026*** −0.09503*** −0.09476*** −0.1480*** −0.2462*** −0.4544*** −0.4558*** −0.2277*** −0.1127***
June −0.1000*** −0.09125*** −0.03643*** −0.08592*** −0.2083*** −0.4538*** −0.4787*** −0.3347*** −0.07309***
July −0.1447*** −0.1189*** 0.002915 −0.05617*** −0.2208*** −0.4777*** −0.5048*** −0.3240*** −0.07985***
August −0.1503*** −0.1238*** 0.0003798 −0.05829*** −0.2266*** −0.4726*** −0.4770*** −0.2396*** −0.09119***
September −0.01313*** −0.05696*** −0.05770*** −0.09163*** −0.1853*** −0.4187*** −0.2734*** −0.1618*** −0.1238***
October −0.04151*** −0.05426*** −0.06378*** −0.1088*** −0.1861*** −0.3289*** −0.1444*** −0.1533*** −0.1291***
November 0.007114** −0.01446*** −0.08299*** −0.1136*** −0.06349*** −0.06272*** −0.09159*** −0.1012*** −0.1115***
December 0.04578*** 0.09784*** 0.1216*** 0.06538*** 0.08648*** 0.01402*** 0.001041 0.02436*** 0.02734***
Constant −0.3460*** 0.2122*** 0.6075*** 1.1735*** 0.7033*** 1.0694*** 1.0534*** 0.8668*** 0.8850***
Observations 1495843 1496318 1495849 1496196 1496368 1496948 1497084 1497236 1493682
R2 0.5647 0.4844 0.4323 0.4732 0.4699 0.5106 0.5227 0.5549 0.6315

would only be able to go out during that period on weekdays while they continue on weekends during this period. The increased demand re-
could choose other time periods on weekends. In addition, households verse the common idea that families or individuals are more willing
may have dinner at a slightly later period on weekends. For many, this to spend time outside, especially on a sunny weekend. However, this
period may be post-dinner on workdays (19:00-21:00) but may actually may be capturing an effect of preferences of specific activities/routines
be dinner time for weekends. Therefore, whether there is rain or not on weekend mornings. The positive results could be due to the fact
may have a greater effect on workdays, due to the lower probability of that households have propensities to carry out housework on weekend
going out in the evenings on workdays. mornings, before heading out in the afternoons. Additionally, some
The only negative effects are for early mornings (6:00-10:00) on types of chores are more likely to give a rise to electricity consumption
workdays. It is possible that the heavier it rains, the earlier people on a sunny day. For example, roughly 30% of households do not own
may feel to leave houses to avoid the traffic jam, although the effect a dryer (Leahy et al., 2012), so they would choose to do laundry
significantly drops from −3.3% to −0.8% at 8:00-10:00. It is a solid on a sunny day and even households with dryers might choose to
proof that the negative effect mainly comes from the behaviour of reduce their bills and dry their clothes outside. Thus, the positive
workers in the house since the effect falls to nearly zero when it reaches effects may reflect the behavioural habits on weekend mornings. It
the start of work hours. should be highlighted that the positive impacts are decreasing from the
8:00-10:00 morning period and becomes insignificant by mid-afternoon
5.2.3. Sun (12:00-15:00), which is the only insignificant period. The reason may
The prior expectation was that longer sun duration should be associ- be that on weekends, family meals are common at lunchtime and sun
ated with decreased electricity demand, as people are more likely to go duration does not affect these behaviour patterns. After that point, sun
out on a sunny day. However, contrary to this expectation, the opposite sensitivities on weekends gradually increase from −7.6% to −9% during
findings are found in the weekend mornings (6:00–12:00) and the the 15:00–21:00 period. This may indicates more sun-related outdoor
early mornings on workdays (6:00–8:00). The increased consumption activities later on weekend days, compared to ‘‘housework mornings’’.
in sunny early mornings for both weekends and workdays could be On the other hand, negative effects are seen during almost every
partly explained by a relatively early wake-up times. This effect is weekday period. However, it is still important to note that compared
especially clear on weekends since on workdays people are more likely to a relatively constant sensitivity of −1.8% in the period 8:00-15:00
to maintain their routines in the early mornings and may not change on workdays, a clear increasing pattern is shown for the period after
their behaviour easily in response to greater sunlight. 15:00. The sensitivities are much higher than during the first half of
In the mornings, from 8:00 until 12:00, interesting and unusual day, with values over 5%. This provides strong evidence that sim-
differences between weekends and workdays appear. While sunshine ilar to behavioural patterns on weekends, afternoons are generally
hours now have a negative effect on workdays, the positive effects more flexible on workdays. Higher negative coefficients imply that the

9
J. Kang and D.M. Reiner Energy Economics 111 (2022) 106023

households have more free or flexible time and are more likely to go sensitivity gradually increases from late afternoon (15:00) and peaks
out. Nevertheless, it should be pointed out that a weaker sensitivity in early evening (19:00–21:00), compared to the small and positive
to sunshine duration does not necessarily mean that people are less sensitivities shown in the mornings.
likely to go out during that period. Unlike for rain, whether or not The responses to weather factors for weekends and workdays are
there is sunshine would generally not affect people’s movements or tested in the second model. The differences are caused by differ-
activities. For instance, if one is used to shopping for groceries for ent household patterns between weekends and workdays. In terms of
the family in the morning, he/she would not cancel or delay the temperature sensitivities, weekends are more sensitive than workdays
shopping just because of cloudy weather. Therefore, a relatively smaller because households have more available time to spend, while the
sensitivity should be interpreted as a higher possibility that one’s time sensitivity difference is minimal. The biggest difference is seen on the
is occupied by regularly scheduled plans, which could be either indoors night period (23:00–3:00), where people would care less about the cold
or outdoors. weather and be more likely to go out on weekends. Moreover, the
Similar to the results shown in the rain effects above, early evening rainfall results suggest two clear patterns: before 15:00 workdays are
(19:00-21:00) is the only period when workdays are more sensitive more sensitive than weekends, although the effects of workday morn-
than holidays. It should be kept in mind that only half of a year (mainly ings are insignificant; after 15:00, weekends show higher sensitivities
late spring and summer) has sunshine during the period. The findings in than workdays, apart from 19:00–21:00. The findings imply that more
this period therefore largely limit and reflect the behaviour in summer. rain-sensitive activities occur before mid-afternoon during weekdays,
As suggested in the rain section, only in this evening period are people while these activities (e.g. outdoor activities) occur at 15:00 afterward
still able and more willing to go out on workdays, compared to late in general. The difference in life patterns between workdays and week-
evenings and nights. Another interesting finding is that sun duration ends are also revealed by sun duration. In the mornings (6:00–12:00),
in this period (19:00-21:00) of workdays has the largest effect among while sunlight has positive effects on weekends: the longer the duration
all other sunshine effects on encouraging people to go out. Note that of sunshine, the greater the consumption during those periods on the
as no sunshine exists after 21:00, no sun effect can be tested for those weekend. It could be associated with more sun-sensitive chores on
periods. the weekend mornings. The pattern changes after 15:00 — households
seem more flexible at this time on weekends. And both weekends and
6. Discussion and conclusion workdays reveal increasing sensitivities during that period, especially
workdays, which soars after 15:00 from −1.6% to a maximum of 13%.
This study set out to examine the behaviour of residential customers One especially interesting finding is that early evening (19:00–21:00) is
exposed to different weather conditions in different periods of a day the one period when weekends are less sensitive to all weather factors
than workdays. This may be unexpected but could be explained by the
using unbalanced panel data from the Irish Smart Metering Electricity
fact: Compared to weekends, early evenings on weekdays might be the
Consumer Behavioural Trial (Commission for Energy Regulation (CER),
most flexible time where outdoor activities are possible, especially for
2012a). To conduct the analysis, half-hourly electricity consumption
those employed households, so the period is more sensitive to weather.
data from 3827 household meters over one year were aggregated into
The study could be instructive for understanding household energy
daily usage for every period of a day. Together with the weather
consumption behaviour. First, the weather sensitivity analysis provides
variables, fixed-effects models with robust standard errors clustered at
an overview of households’ behaviour/life pattern without the assis-
the household level were used to control for unobserved household-
tance of a survey. Especially, sunshine and rain sensitivities may be
specific factors, which gives a better understanding of households’
considered as proxies of whether a period is with more flexibility and
response to weather factors at different times of the day.
whether people tend to leave home (or use less) at certain periods re-
Overall, this paper has demonstrated from the first model that in
spectively. Furthermore, analysing the differences in patterns between
general although temperature has robust and relatively flat effects on
weekdays and weekends can help identify which periods on weekends
electricity demand across all periods, rain and sunshine duration show
or workdays are more sensitive and flexible. With more knowledge
greater potential to affect individual behaviour and daily routines. The
of people’s life pattern among different periods the tariff structure
demand response to temperature could be interpreted as warm/cold
design could be more efficient in shifting energy demand. Secondly,
sensitivities of the activities in that period. As expected, the periods
with deeper analysis on individual level, for example, combing the
from 10:00–21:00 present higher sensitivities than early mornings and
attitude and behaviour data in the survey with the weather sensitivity
nights, since more activities occur in those periods. Although night patterns, it could create an initial profile of a family’s daily activities.
time periods (21:00–3:00) have smaller sensitivities than daytime, they For instance, if a family displays relatively higher weather sensitivities,
are still much more sensitive than early mornings. Not many activities this may reflect greater flexibility in their living patterns. A target tariff
occur over 6:00–10:00 when most people are getting up and going off aimed at those families may help shift peak electricity demand. These
to work. The rainfall sensitivity may act as an indicator of whether potential implications lead to possible future research in improving
outdoor activities occur more often in that period. It should be noted residential customers’ consumption profiles. It would be interesting to
that the results mainly reflect the behaviour of the households who are categorise the households by their weather sensitivities and to examine
in the house during day-time, and the proportion of these households if the weather sensitivities are associated with certain demographic fac-
account for over 68% of the sample. One of the lowest rainfall sensi- tors, which may provide a cheaper and faster means of understanding
tivities appears at 12:00–15:00 which is cooking and lunchtime that a household’s social-economic profile. Data mining tools are helpful
the possibility of going outdoor would be relatively small. This finding in this case to cluster and classify the residential customers, which
is consistent with the Irish Time Use Survey (Table 3) that for those offers a new angle to summarise and depict customers’ activity patterns
who are not working (61%–82%), eating is one of the main activities. by using only weather and consumption data. By using only weather
Apart from the similar pattern of lower sensitivities at the start and indicators this approach can be faster and simpler than traditional
end of a day, the relatively high coefficients at two periods 10:00– methods – such as surveys or questionnaires – in identifying which
12:00 and 15:00–17:00 reveal that individuals could be more used to period are more flexible at the household level.
or prefer going out during these periods. On the other hand, the impact
of sunshine on households’ behaviour differs from rainfall, although CRediT authorship contribution statement
both affect the chances of going out. Negative sunshine sensitivities
represent the time availability and willingness to go out of households, Jieyi Kang: Conceptualisation, Methodology, Software, Validation,
in other words, how flexible the period so that one can response to Formal analysis, Data curation, Writing – original draft, Visualisation.
good weather. The results strongly support the interpretation that the David M. Reiner: Writing – review & editing, Supervision.

10
J. Kang and D.M. Reiner Energy Economics 111 (2022) 106023

Appendix A. Supplementary data Eid, C., Hakvoort, R.A., Jong, M.de., 2017. Global trends in the political economy of
smart grids. Political Econ. Clean Energy Trans. 1 (2019), 250–270.
Firth, S., et al., 2008. Identifying trends in the use of domestic appliances from
Supplementary material related to this article can be found online
household electricity consumption measurements. Energy Build. 40 (5), 926–936.
at https://doi.org/10.1016/j.eneco.2022.106023. http://dx.doi.org/10.1016/J.ENBUILD.2007.07.005.
Harold, J., Lyons, S., Cullinan, J., 2015. The determinants of residential gas demand
References in Ireland. Energy Econ. 51, 475–483. http://dx.doi.org/10.1016/J.ENECO.2015.
08.015.
Alberini, A., Filippini, M., 2011. Response of residential electricity demand to price: Henley, A., Peirson, J., 1998. Residential energy demand and the interaction of price
The effect of measurement error. Energy Econ. 33 (5), 889–895. http://dx.doi.org/ and temperature: British experimental evidence. Energy Econ. 20 (2–6), 157–171.
10.1016/j.eneco.2011.03.009. http://dx.doi.org/10.1016/s0140-9883(97)00025-x.
Alberini, A., Towe, C., 2015. Information v. energy efficiency incentives: Evidence from Hor, C.L., Watson, S.J., Majithia, S., 2005. Analyzing the impact of weather variables
residential electricity consumption in Maryland. In: Energy Economics. vol. 52, on monthly electricity demand. IEEE Trans. Power Syst. 20 (4), 2078–2085. http:
Elsevier B.V., pp. S30–S40. http://dx.doi.org/10.1016/j.eneco.2015.08.013. //dx.doi.org/10.1109/TPWRS.2005.857397.
Albert, A., Rajagopal, R., 2013. Building dynamic thermal profiles of energy consump- Kavousian, A., Rajagopal, R., Fischer, M., 2013. Determinants of residential electricity
tion for individuals and neighborhoods. In: Proceedings - 2013 IEEE International consumption: Using smart meter data to examine the effect of climate, building
Conference on Big Data, Big Data 2013, http://dx.doi.org/10.1109/BigData.2013. characteristics, appliance stock, and occupants’ behavior. Energy 55, 184–194.
6691644. http://dx.doi.org/10.1016/j.energy.2013.03.086.
Anderson, B., et al., 2017. Electricity consumption and household characteristics: Leahy, E., Lyons, S., Walsh, S., 2012. Electrical appliance ownership and usage in
Implications for census-taking in a smart metered future. Comput. Environ. Urban Ireland. Papers WP421, Economic and Social Research Institute (ESRI), Dublin.
Syst. 63, 58–67. http://dx.doi.org/10.1016/j.compenvurbsys.2016.06.003. McGinnity, F., Russell, H., Williams, J., Blackwell, S., 2005. Time Use in Ireland, 2005:
Asteriou, D., Hall, S.G., 2011. Applied Econometrics, second ed. Palgrave Macmillan, Survey Report. ESRI and Department of Justice, Equality and Law Reform, Dublin.
Hampshire. McLoughlin, F., Duffy, A., Conlon, M., 2012. Characterising domestic electricity con-
Atalla, T.N., Hunt, L.C., 2016. Modelling residential electricity demand in the GCC sumption patterns by dwelling and occupant socio-economic variables: An Irish case
countries. Energy Econ. 59, 149–158. http://dx.doi.org/10.1016/j.eneco.2016.07. study. Energy Build. 48, 240–248. http://dx.doi.org/10.1016/J.ENBUILD.2012.01.
027. 037.
Auffhammer, M., Aroonruengsawat, A., 2012. Simulating the impacts of climate change, Met Éireann, 2018. Weather statements. Accessed from: https://www.met.ie/climate/
prices and population on California’s residential electricity consumption. Cli- past-weather-statements.
mate Change 109 (S1), Available at: https://ei.haas.berkeley.edu/research/papers/. Mirasgedis, S., et al., 2006. Models for mid-term electricity demand forecasting
(Accessed 5 January 2018). incorporating weather influences. Energy 31 (2–3), 208–227. http://dx.doi.org/10.
Auffhammer, M., Mansur, E.T., 2014. Measuring climatic impacts on energy con- 1016/j.energy.2005.02.016.
sumption: A review of the empirical literature. Energy Econ. 46, 522–530. http: Moral-Carcedo, J., Vicéns-Otero, J., 2005. Modelling the non-linear response of Spanish
//dx.doi.org/10.1016/j.eneco.2014.04.017. electricity demand to temperature variations. Energy Econ. 27 (3), 477–494. http:
Beckel, C., et al., 2015. Automated customer segmentation based on smart meter data //dx.doi.org/10.1016/J.ENECO.2005.01.003.
with temperature and daylight sensitivity. In: 2015 IEEE International Conference Newsham, G.R., Bowker, B.G., 2010. The effect of utility time-varying pricing and load
on Smart Grid Communications, SmartGridComm, pp. 653–658. http://dx.doi.org/ control strategies on residential summer peak electricity use: A review. Energy
10.1109/SmartGridComm.2015.7436375. Policy 38, 3289–3296. http://dx.doi.org/10.1016/j.enpol.2010.01.027.
Bessec, M., Fouquau, J., 2008. The non-linear link between electricity consumption Pardo, A., Meneu, V., Valor, E., 2002. Temperature and seasonality influences on
and temperature in Europe: A threshold panel approach. Energy Econ. 30 (5), Spanish electricity load. Energy Econ. 24 (1), 55–70. http://dx.doi.org/10.1016/
2705–2721. http://dx.doi.org/10.1016/j.eneco.2008.02.003. S0140-9883(01)00082-2.
Blázquez, L., Boogen, N., Filippini, M., 2013. Residential electricity demand in Spain: Räsänen, T., et al., 2010. Data-based method for creating electricity use load profiles
New empirical evidence using aggregate data. Energy Econ. 36, 648–657. http: using large amount of customer-specific hourly measured electricity use data. Appl.
//dx.doi.org/10.1016/J.ENECO.2012.11.010. Energy 87 (11), 3538–3545. http://dx.doi.org/10.1016/J.APENERGY.2010.05.015.
Boogen, N., Datta, S., Filippini, M., 2017. Demand-side management by electric utilities Razavi, R., et al., 2019. Occupancy detection of residential buildings using smart meter
in Switzerland: Analyzing its impact on residential electricity demand. Energy Econ. data: A large-scale study. Energy Build. 183, 195–208. http://dx.doi.org/10.1016/
64, 402–414. http://dx.doi.org/10.1016/j.eneco.2017.04.006. J.ENBUILD.2018.11.025.
Carroll, J., Lyons, S., Denny, E., 2014. Reducing household electricity demand through Torriti, J., 2012. Price-based demand side management: Assessing the impacts of time-
smart metering: The role of improved information about energy saving. Energy of-use tariffs on residential electricity demand and peak shifting in northern Italy.
Econ. 45, 234–243. http://dx.doi.org/10.1016/j.eneco.2014.07.007. Energy 44 (1), 576–583. http://dx.doi.org/10.1016/J.ENERGY.2012.05.043.
Commission for Energy Regulation (CER), 2012a. CER Smart Metering Project - Trotter, I.M., et al., 2016. Climate change and electricity demand in Brazil: A stochastic
Electricity Customer Behaviour Trial, 2009-2010 [Dataset], first ed. Irish Social approach. Energy 102, 596–604. http://dx.doi.org/10.1016/j.energy.2016.02.120.
Science Data Archive, SN: 0012-00. www.ucd.ie/issda/CER-electricity. Valor, E., Meneu, V., Caselles, V., 2001. Daily air temperature and electricity load
Commission for Energy Regulation (CER), 2012b. CER Smart Metering Project - Gas in Spain. J. Appl. Meteorol. 40 (8), 1413–1421. http://dx.doi.org/10.1175/1520-
Customer Behaviour Trial, 2009-2010. [Dataset], first ed. Irish Social Science Data 0450(2001)040<1413:DATAEL>2.0.CO;2.
Archive, SN: 0013-00. www.ucd.ie/issda/CER-gas. Wangpattarapong, K., et al., 2008. The impacts of climatic and economic factors on
Cosmo, V. Di, Lyons, S., Nolan, A., 2014. Estimating the impact of time-of-use pricing residential electricity consumption of Bangkok Metropolis. Energy Build. 40 (8),
on Irish electricity demand. Energy J. 35 (2), 117–136. http://dx.doi.org/10.5547/ 1419–1425. http://dx.doi.org/10.1016/j.enbuild.2008.01.006.
01956574.35.2.6. Woods, J., Fuller, C., 2014. Estimating base temperatures in econometric models that
Cosmo, V., O’hora, D., 2017. Nudging electricity consumption using TOU pricing include degree days. Energy Econ. 45, 166–171. http://dx.doi.org/10.1016/j.eneco.
and feedback: evidence from Irish households. J. Econ. Psychol. 61, 1–14. http: 2014.06.006.
//dx.doi.org/10.1016/j.joep.2017.03.005. Wooldridge, J.M., 2013. Introductory Econometrics. Cengage Learning, http://dx.doi.
Costa, D.L., Kahn, M.E., 2010. Why has California’s Residential Electricity Consumption org/10.1016/j.jconhyd.2010.08.009.
Been so Flat Since the 1980s?: A Microeconometric Approach. NBER Working Paper Yohanis, Y.G., et al., 2008. Real-life energy use in the UK: How occupancy and dwelling
No. 15978. 15978, Cambridge, MA. characteristics affect domestic electricity use. Energy Build. 40 (6), 1053–1059.
Davies, M., 1958. The relationship between weather and electricity demand. Proc. IEE http://dx.doi.org/10.1016/j.enbuild.2007.09.001.
C 106 (9), 27. http://dx.doi.org/10.1049/pi-c.1959.0007.

11