Daily Life and Demand: An Analysis of Intra-Day Variations in Residential Electricity Consumption With Time-Use Data

Energy Efficiency (2020) 13:433–458
https://doi.org/10.1007/s12053-019-09791-1
ORIGINAL ARTICLE
Daily life and demand: an analysis of intra-day variations

in residential electricity consumption with time-use data
Aven Satre-Meloy & Marina Diakonova &
Philipp Grünewald
Published online: 24 April 2019

# The Author(s) 2019
Abstract Demand-side flexibility has been suggested temporal characteristics of electricity demand and in-
as a tool for peak demand reduction and large-scale form approaches to shift or reduce it. We stress the
integration of low-carbon electricity sources. Deeper importance of considering consumption as a function
insight into the activities and energy services performed of time of day, and we use our findings to argue that a
in households could help to understand the scope and more nuanced understanding of this relationship can
limitations of demand-side flexibility. Measuring and yield useful insights for residential demand flexibility.
Evaluating Time- and Energy-use Relationships
(METER) is a 5-year, UK-based research project and Keywords Household electricity consumption . Peak
the first study to collect activity data and electricity use demand . Demand flexibility . Time-use data . Regression
in parallel at this scale. We present statistical analyses of analysis
these new data, including more than 6250 activities
reported by 450 individuals in 173 households, and their
relationship to electricity use patterns. We use a regular-
Introduction
ization technique to select influential variables in regres-
sion models of average electricity use over a day and of
The residential sector is the largest end user of electricity
discretionary use across 4-h time periods to compare
in the UK, accounting for 45% of total consumption in
intra-day variations. We find that dwelling and appli-
2017 (BEIS 2018). It is also responsible for up to 50%
ance variables show the strongest associations to aver-
of national peak demand, during which time electricity
age electricity consumption and can explain 49% of the
provision is especially costly and carbon-intensive
variance in mean daily usage. For models of 4-h average
(Ofgem 2010). As evidenced by the share of generation
Bde-minned^ consumption, we find that activity vari-
from renewables increasing to over 29% in 2017, the
ables are consistently influential, both in terms of coef-
UK’s electricity system is increasingly low-carbon, yet
ficient magnitudes and contributions to increased model
much more ambition is needed to achieve national cli-
explanatory power. Activities relating to food prepara-
mate goals in the next decades (BEIS 2018).
tion and eating, household chores, and recreation show
The UK’s 2017 Clean Growth Strategy sets forth a
the strongest associations. We conclude that occupant
set of policies and proposals to further accelerate the
activity data can advance our understanding of the
deployment of low-carbon energy while maintaining
increased economic growth (BEIS 2017). One of the
proposals, the BSmart systems plan,^ aims to help con-
A. Satre-Meloy (*) : M. Diakonova : P. Grünewald
Environmental Change Institute, University of Oxford, South sumers use energy more flexibly. Energy storage will
Parks Rd, Oxford OX1 3QY, UK support integration efforts as the UK pursues rapid
e-mail: [email protected] deployment of renewable energy, but demand flexibility
434 Energy Efficiency (2020) 13:433–458
can reduce the cost of integration and storage require- presents results of full-day and 4-h models of average
ments for low-carbon energy systems (National electricity consumption. In the BDiscussion^ section, we
Infrastructure Commission 2016). discuss the implications of our findings, and in the
Understanding how to make demand more flexible in BConclusions^ section, we conclude.
the residential building sector requires a more detailed
investigation of what electricity is used for in house-
holds (Grunewald and Diakonova 2018a; Powells et al. Literature review
2014). Time-use data is increasingly used for modeling
electricity consumption, but these analyses often use Demand-side flexibility
simulated activity data and proxies for electricity con-
sumption, such as household expenditure data, which Flexibility has been defined in the energy demand liter-
requires assumptions about the link between time-use ature as the ability for consumers to change how, when,
data and electricity consumption patterns. In some and where energy is used—a shift in focus from the
cases, these models are validated against real consump- magnitude of overall consumption to the timing of de-
tion data, though the validation sample sizes are in the mand (Powells et al. 2014; Torriti et al. 2015). We
10–20 household range (Richardson et al. 2010; Widén conceive of it here simply as Ba potential to change
et al. 2009; Widén and Wäckelgård 2010). Even as time- power at a certain rate (Watt/hour)^ (Grunewald and
use data has been more recently incorporated into these Diakonova 2018a, p. 59). While much of the focus in
models, however, there remains a notable gap in the demand-side energy research has been on reducing de-
literature of empirical studies on how occupant activities mand through energy efficiency, as the energy system
influence the temporal aspects of demand (Anderson becomes increasingly supplied by variable, low-carbon
2016; Anderson and Torriti 2018; Grunewald and generation, flexibility becomes increasingly important
Diakonova 2018b). for balancing supply and demand. The potential respon-
In this paper, we aim to contribute to filling this gap siveness of demand, or its ability to shift in time to
by constructing regression models of original survey, match high generation from renewables or to flatten
time-use, and electricity consumption data for 173 UK demand during peak periods, is essential for minimizing
households collected as part of an ongoing study. We the costs of transitioning to a low-carbon power system
test the extent to which socio-demographic, dwelling, (Strbac et al. 2012).
appliance, and activity data can explain variations in Demand-side response (DSR) refers to measures that
intra-day household electricity use, from 5 p.m. until 9 provide flexibility to the energy system by shifting or
p.m. on the following day. We then compare the effects reshaping energy loads. McKenna et al. (2017) conceive
of reported time-use activities in households throughout of DSR as a three-dimensional space including
the day by dividing the day into six equal 4-h periods. Btechnology change,^ Bservice expectation change,^
We aim to understand whether and how the inclusion and Bactivity change.^ The technology change dimen-
of categorized activity data improves models of house- sion refers to an automated or remotely controlled shift
hold electricity consumption. We investigate how dif- in energy use, for example, a smart fridge that delivers a
ferent categories of activities are associated with elec- response without affecting the energy service delivered.
tricity consumption during different times of day, and The authors note this has been the main form of DSR
we consider how this knowledge can inform efforts to that has been tested in energy models given the relative
enhance demand-side flexibility and enable more and ease of simulating such technology changes. BService
less costly integration of renewable energy sources. expectation change^ refers to shifting occupants’ ex-
The paper is organized as follows: In the BLiterature pected level of energy service from an appliance or
review^ section, we present a brief review of the flexi- technology (i.e., the thermostat set-point), while an
bility and household electricity modeling literature, in- Bactivity change^ requires a shift in either the timing
cluding both determinants of household electricity con- or type of activity and relies on more active behavioral
sumption and the use of time-use data in energy model- changes among energy users.
ing. In the BData and methodology^ section, we present Grünewald and Diakonova (2018a) more broadly
our methodology for collecting household electricity differentiate these dimensions of DSR as Bappliance
and time-use data in parallel. The BResults^ section led^ and Bactivity led.^ Among Bactivity led^ DSR are
Energy Efficiency (2020) 13:433–458 435
several different types of shift mechanisms: activity activity patterns bear a strong relationship to both the
shifts, or changing the timing of an activity; substituting types of DSR that can take place in homes as well as the
practice, which does not reorder activities in time but capacity of households to participate in DSR programs.
instead substitutes an energy-intensive activity for one Several different approaches have been developed to
that is not (e.g., having a cold rather than hot meal); better understand these relationships in order to inform
substituting metabolic energy, such as mixing dough by more effective DSR. The following sections review
hand rather than with an electric mixer; or changing the these approaches in the context of research on electricity
practitioner, which might entail going out for dinner demand modeling and incorporating time-use data in
rather than cooking at home. these models.
Recent research has begun to investigate the DSR
potential of these and similar types of measures, but still
little is known about both the mechanisms by which Household electricity demand: models and determinants
electricity consumption patterns can be reshaped and the of use
different constraints or motivating factors that can re-
shape them. A review of major DSR trials in the UK The literature on quantitative approaches to model and
found that residential customers are responsive to eco- understand key socio-technical determinants of electric-
nomic incentives to shift demand but that the size of the ity use in households is expansive. Previous studies
shift can vary significantly (DECC 2012). The report employ both top-down and bottom-up approaches.
notes several areas where findings were inconclusive. Top-down approaches consider how national-level data,
These include the responsiveness of vulnerable and low- such as characteristics of the housing stock and macro-
income customers, the impact of non-economic signals, economic factors, influence electricity consumption,
and detailed information on the way consumers shift while bottom-up approaches use detailed household-
their usage in response to incentives. level data on dwelling characteristics and occupant
Further empirical work on the flexibility of house- socio-demographics to identify the primary drivers of
hold activities includes research that qualitatively electricity consumption for a sample of households and
assessed the likelihood of certain practices being per- then extrapolate these findings to the wider population
formed during peak hours (4–8 p.m.) (Powells et al. (Swan and Ugursal 2009). Studies are also varied in
2014). This study found a very high or high likelihood how they measure household electricity use, in some
of TV watching, cooking, computer/Internet use, and cases considering average annual consumption values
dish washing during these hours, with lesser frequencies while in others studying daily or hourly consumption.
of laundry, ironing, vacuuming, and bathing. More in- A recent comprehensive review of the socio-demo-
sight into the types of activities happening during peak graphic, dwelling, and appliance-related determinants of
periods and their contributions to electricity demand can household electricity consumption undertaken by Jones
help identify those that are best suited for DSR. et al. (2015) shows that no fewer than 62 factors have
But how well Bsuited^ an activity is for shifting is not been found to affect household electricity consumption.
the same as how Bflexible^ it might be. Torriti et al. Twenty of these are consistently found to have a signif-
(2015) stress this point by developing a "Flexibility icant, positive correlation with electricity use
Index" composed of five separate indices (synchroniza- (Bconsistently^ is defined by the authors as showing a
tion, variation, non-shared activities, active home occu- statistically significant relationship in more than three
pancy, and spatial mobility). These component indices studies). The socio-demographic factors include number
are calculated for five time periods throughout the day of occupants, presence of teenagers, household income,
for a sample of 153 respondents, and the findings show and disposable income. The impact of occupant age,
that morning peak times feature high levels of synchro- education level, and tenure type is inconclusive in the
nization and shared activities but low variation in activ- studies reviewed. The dwelling factors that are influen-
ities. Evening peak times show less synchronization but tial include dwelling age, dwelling type, number of
higher spatial mobility and variation in activities. rooms, total floor area, and ownership of electric space
Demand-side flexibility is complex and is difficult to heating and cooling systems. Appliance factors include
measure empirically. It is likely that socio-demographic, ownership of a desktop computer, television, electric
dwelling, appliance ownership characteristics, and oven, refrigerator, dishwasher, washing machine, and
tumble dryer, as well as the overall number of appli- mostly driven by variables measuring reported appli-
ances owned. ance ownership and use, especially dishwashers, electric
Huebner et al. (2016) confirm many of these findings power showers, televisions, and desktop computers.
in one of the few representative studies of electricity Notably, the models are less able to explain load factor
consumption in the UK residential sector. Using data and timing of peak demand than total consumption and
from a sample of 845 English households from 2011 to mean daily maximum demand. Key explanatory factors
2012, they find that models containing only appliance for understanding variations in electricity consumption
ownership and use factors explain 34% of the variance at different times of day may therefore be missing in the
in non-heating annual electricity consumption, whereas models.
models containing socio-demographic variables explain Exploiting a similar smart meter data set from the
only 21%, and models containing dwelling and other Irish Commission for Energy Regulation (CER), Ander-
occupant variables are poor in explaining electricity use. son et al. (2016) examine the links between half-hourly
Their model combining these factors shows that dwell- electricity consumption data and household characteris-
ing floor area, number of occupants, wet appliance tics for the purpose of assessing the feasibility of using
ownership, and hours of TV watched per day are statis- high-resolution electricity data to infer household char-
tically significant predictors of annual consumption. acteristics to supplement census-taking efforts. They
As our paper focuses specifically on factors affecting construct multi-level regression models to analyze nu-
intra-day variations in electricity use, a brief review of merous Bprofile indicators^ that describe the temporal
findings from studies with a similar focus is included characteristics of load profiles, such as base load, mean
here. The number of these studies is fewer than those load, 97.5th percentile load (as a proxy for peak load),
included in the review discussed above because in much and several other temporal parameters. Their results
of the previous research, advanced metering data was show that several household characteristics, such as
not as available or accessible. With the growth of smart number of occupants, household income, and presence
meter adoption, however, these data and studies inves- of children, are useful predictors of several of these
tigating them are becoming increasingly common. Bprofile indicators,^ especially those measuring overall
In general, this body of work suggests that appliance magnitude of consumption.
ownership and usage factors have stronger relationships Kavousian et al. (2013) examine structural and be-
with variations in consumption than socio-demographic havioral determinants of daily maximum and minimum
and dwelling factors. In a study monitoring 27 UK demand for a data set of 10-min interval electricity
residential buildings’ electricity consumption at 5-min readings and an extensive household survey of 1628
intervals for a period of 2 years, Firth et al. (2008) find US dwellings in California. The study was non-
that an observed increase in usage between years is random and had a sample biased toward high-income
primarily attributable to increases in the consumption and well-educated participants. They find that while
of standby appliances, such as televisions and other daily minimum consumption is influenced most by
consumer electronics, and Bactive^ appliances, such as weather, location, and physical characteristics of the
kettles, washing machines, electric showers, and building, such as size and type of home, daily maximum
lighting. demand is influenced by high-consumption, intermittent
McLoughlin et al. (2012) study the effect of dwelling appliances, such as electric water heaters or tumble
and occupant characteristics for a representative sample dryers. The authors note the clear difference in their
of 4200 Irish dwellings on several dependent variables, results between drivers of daily minimum and maxi-
including total electricity consumption over a 6-month mum demand, explaining that peak demand is more
period, mean daily maximum demand over that period, dependent on the activity patterns that lead to use of
electrical load factor (ratio of daily mean to daily max- high-consumption appliances. Minimum demand is
imum electrical demand), and timing of peak demand. driven primarily by locational and physical characteris-
They find that several dwelling and socio-demographic tics of the dwelling.
factors, such as number of bedrooms, type of dwelling, Several recent papers have aimed to investigate
and head of household age, are significant in explaining socio-demographic and physical dwelling determinants
load factor variations (i.e., those that are Bpeakier^ or of whole load profile shapes rather than average values
Bflatter^). They also find that timing of peak demand is of electricity consumption across varying time periods
(McLoughlin et al. 2015; Rhodes et al. 2014; Viegas electricity use (De Lauretis et al. 2017; Jalas and
et al. 2016). These studies take an alternative approach Juntunen 2015; Sekar et al. 2018). In many of these
by first clustering load profiles into representative pro- studies, models and results are validated with actual
file classes and then using logistic regression to deter- consumption measurements from a small sample of
mine factors influencing membership in different load households (McKenna and Thomson 2016; Widén
profile classes. While this is a different approach than et al. 2009).
the one used in this paper and in the studies reviewed While this literature shows varying results for the
above, it achieves a similar aim. The findings from these links between time use and electricity consumption,
studies indicate that variables such as working from some common themes emerge. Mealtimes and related
home, time spent watching television, age, and owner- activities are consistently found to be high-consumption
ship of dishwashers and washing machines are most activities (De Lauretis et al. 2017; Druckman et al. 2012;
important for explaining the shape of a household’s Jalas and Juntunen 2015). Studies also find evidence
daily electrical load profile. that housework and personal time are energy-intensive
The literature investigating highly resolved variations activities (Palmer et al. 2013; Widén et al. 2009). Un-
in residential electricity consumption is growing along- surprisingly, sleeping and resting are low-consumption
side the availability of data, but many of the results activities (Druckman et al. 2012; Jalas and Juntunen
discussed here are dependent on the sample nature and 2015), whereas recreational activities can be either high
size and not necessarily generalizable to wider popula- or low energy depending on whether they involve
tions. It is also important to consider the climatic and televisions and computers or reading and socializing.
social contexts in which these studies are undertaken, as In terms of the temporal shifts in the timing of UK peak
these likely influence their findings. demand over the past 40 years, Anderson and Torriti
(2018) find that this is mostly due to changes in food-
Incorporating time-use data in models of electricity related activities, in part due to changes in personal care
consumption and housework shifts, and little to do with changes in
media use.
Social scientists have long researched the dynamics of Using the specific case of laundry practices in the
daily life through studies of how people spend their UK, Anderson (2016) highlights the role of societal
time. Longitudinal studies of thousands of households change in time shifts for energy-using activities. In-
have investigated questions of historical changes in time creases in labor market participation by more females
use and how these relate to trends in technological and in society is likely partially responsible for shifting
social change (eurostat 2009; Gershuny 2008; U.S. laundry energy demand into weekday mornings and
Bureau of Labor Statistics 2018). Research in energy evening peak times but also into Sunday mornings,
demand modeling has acknowledged the need to take which are less problematic for providing energy. The
account of the timing of household activities in order to author concludes that more analyses of the changes to
more accurately represent electricity demand (Torriti routine energy-using practices are necessary for
et al. 2015). In the last decade, occupant activity data assessing the value of various demand response
from time-use studies has been increasingly incorporat- strategies.
ed into high-resolution models of residential electricity A review of time-use modeling of electricity demand
consumption (Torriti 2014). These studies involve vary- undertaken by Torriti (2014) identifies five limitations in
ing approaches for incorporating time-use data. In some the literature. First, time-use data must be aggregated
studies, probabilistic simulations are used to develop across large numbers of households to be statistically
models of active occupancy, occupant activity se- significant. Second, time-use data are often sampled
quences, and appliance usage that are predictive of throughout the year, whereas peak electricity events
electricity load curves (Ellegård and Palm 2011; often happen on specific days, such as during tempera-
Richardson et al. 2010; Widén et al. 2009). In others, ture extremes or national events. Third, the low frequen-
changes in electricity consumption over decades are cy of large time-use surveys means the data quickly
linked to changes in time-use and household expendi- become outdated. Fourth, the timing of occupancy and
ture data, often using decomposition analysis, which typical activities in developed countries is less variable
attributes changes in these factors to overall changes in than other factors that influence consumption, such as
weather or appliance design. Fifth, modeling multiple- incentive for participation is the chance to win a year’s
person households using time-use data, especially for worth of free electricity.
modeling occupancy patterns, is much more challenging When registering for the study, participants complete
than for single-person households, which was con- a survey wherein they provide detailed socio-
firmed empirically by Grünewald and Diakonova demographic information along with data on dwelling
(2018b). The fourth point, however, is contested. characteristics and appliance ownership. Participating
Tabbone et al. (2016) found that even within countries, households are sent a parcel prior to their chosen date,
cultural differences in time-use can affect load profiles. which includes an electricity recorder, activity re-
To these limitations, McKenna et al. (2017) add corder(s), and an instruction booklet. The study encour-
several additional failings of time-use demand models. ages all household members over the age of eight to
First, they note that time-use studies were not originally participate.
designed for the purpose of modeling energy use, and Participants are instructed to attach the electricity
they thus do not differentiate between Benergy- recorder below the household’s electricity meter. The
intensive^ or Blow-energy^ alternatives of the same electricity recorder collects readings every second for
activity. Meal time is a good example, as this can vary 28 h, from 5 p.m. on their chosen day until 9 p.m. the
from relatively low energy (cold meal preparation) to following day. This study length is chosen in order to
very energy intensive (oven and electric hob use). Sec- capture two typical peak demand periods and because
ond, time-use data do not account for overlapping or the electricity recorders have a battery life of around
Bbundled^ activities, such as those that might occur 28 h per charge. Fuel type used for heating and cooking
during multi-tasking. Third, the typical 10-min reporting appliances is collected in the household survey.
window for activities may not capture energy-relevant Activities are recorded with a dedicated app pre-
yet shorter duration activities, such as boiling the kettle. installed on individual devices. The app presents six
This study addresses the limitations highlighted in options per screen that guide participants through
these two reviews through a new approach to collecting recording their activities, starting with the location
time-use data. To the authors’ knowledge, this is the first and concluding with the number of people partaking
study to collect time-use data and electricity readings in the activity and one’s enjoyment of it. Asking
directly from households in parallel at this scale. This about enjoyment was found to increase participant
research contributes to an emerging focus within energy retention in previous time-use research (Gershuny
demand studies on not only the magnitude but also the and Sullivan 2017). Figure 1 shows an example ac-
timing of electricity consumption in households and tivity entry sequence. In contrast to paper-based time-
how it is influenced both by numerous socio- use diaries (eurostat 2009), the app can guide partic-
demographic and dwelling factors that have been stud- ipants to select energy-relevant details, such as appli-
ied in previous research, as well as by the activities of ance use, allowing for detailed descriptions of activ-
occupants, which may contribute additional insights for ities. Instead of capturing the duration of activities, as
determining the timing of electricity use in homes and was done in previous time-use studies, the app re-
the potential for shifting these activities to provide more cords activities reported instantaneously. Users can
flexibility to the energy system. record multiple activities in sequence and are also
given an option to record the Bend^ of activities.
While users are encouraged to report activities when
they actually occur, entries can be made retrospec-
Data and methodology tively and into the future. One test of validity for
reporting accuracy suggests around 80% of activity
Data collection records for boiling the kettle are reported within
10 min of the activity itself (see Grünewald and
Electricity readings and activity records are collected Diakonova (2018b) for more details). More details
from UK households as part of an ongoing study about the functionality of the app is discussed in
(Grünewald and Layberry 2015). Participation is volun- Grünewald et al. (2017), and a description of the data
tary, and participants are recruited online, via radio, and storage and handling procedures can be found in
through campaigns at selected community events. An Grünewald and Diakonova (2019).
Study sample The study is voluntary, and the sample is not fully
representative of the general UK population. Selection
The study sample consists of 173 households and 447 biases include an underrepresentation of renters, low-
individuals who together reported 6265 activities. The income groups, and the elderly. Gas boiler ownership is
process for determining inclusion in the subsequent slightly underrepresented, suggesting a higher preva-
analyses is as follows: lence of all-electric heating in the sample. These biases
are likely to influence both overall consumption as well
& Household that own solar PV are excluded due to as the timing of demand, which is why we caution
concerns about the validity of electricity readings. against overgeneralizing from these results.
& Households that did not complete the full survey are Table 1 also lists the distributions of season and day
excluded. The small occurrence of missing data of week that households participate. A majority of
(11% or 23 cases) did not warrant more complex households (57%) participate during the winter season
handling, such as multiple imputation, so these cases (November–February). Data collection is focused on
are deleted listwise. this season because it coincides with UK annual peak
& Electricity readings below 20 W are excluded from demand (DECC 2014). Households are recruited to start
daily and 4-h averages, as this signals a failure to the study on a weekday, though given the study period
attach the electricity recorder properly. spans 2 days, 22% of the sample starts on a Friday and
& Activities that were reported either before the elec- completes the study on a Saturday. We do not create
tricity meter starts recording or after it stops are separate models for season or day of week in order to
excluded. Similarly, activities that are reported in preserve the sample size across models, but we do
an hour where a valid electricity reading is missing include these as candidate variables during model selec-
are excluded. tion, which is further explained below.
& Only activities reported in the home are included,
since we are interested in those activities that have a
direct influence on household electricity consumption. Activity variables
The app pre-codes activities, following the Harmonized

European Time-Use Studies (HETUS) standards. For
Socio-demographic, dwelling, and appliance variables this analysis, activities are grouped into seven catego-
ries, which have become common in time-use research
The household survey includes questions on occupant (Ellegård et al. 2010; Lader et al. 2006; Stankovic et al.
socio-demographics, physical dwelling characteristics, 2016). Examples of activities that fall into each category
and appliance ownership. In previous studies, these are shown in Table 2.
have shown to have the strongest associations with In the following analyses, activities are aggregated
residential electricity consumption in the UK (Huebner across household members so that each household has a
et al. 2016; Jones and Lomas 2016; Wyatt 2013). Table 1 total count, or frequency, for each activity category.
presents frequencies and descriptive statistics for these Table 2 presents descriptive statistics for each activity
variables for both the sample and the wider UK popu- category for the sample.
lation (where available). Categorical variables are dum- Number and type of activities reported vary through-
my coded prior to analysis, and reference categories are out the day, as shown in Fig. 2a. Most activities are
bolded in Table 1. Estimated monthly electric bill mea- reported in the early evening (5 p.m.–9 p.m.), followed
sures participants’ estimate of their electricity expendi- by the morning period (5 a.m.–9 a.m.). For electricity
ture. While we would expect this to have a positive research, this is helpful since early evenings also have
association to average daily electricity consumption, the highest electricity use. Direct comparison of the
we suspect that monthly estimates on expenditure might early evening period, which is recorded on the first
vary widely from day-to-day variations in consumption. and second day, shows that the number of activities
We include it in our analyses to understand how well reported reduces by around 23%. Respondent fatigue
occupant estimates of electricity consumption do in fact is a possible explanation for some participants. The
predict actual average daily consumption values. relative share of activities reported in each category
Fig. 1 Example of an activity

entrance sequence on the activity
recorder (Grünewald et al. 2017)
remains similar, however, so we do not expect this to variable prior to de-minning and transformation. Boxes
bias the second 5–9 p.m. model. indicate 25th, 50th, and 75th percentiles, while the
Following Mckenna et al.’s (2017) point that the whiskers indicate 1.5 times the interquartile range below
activity taxonomy developed for HETUS does not dif- or above the 25th or 75th percentiles, respectively. Dots
ferentiate between Benergy-intensive^ and Blow- represent outliers beyond these. Mean consumption dur-
energy^ activities, we address this limitation by repeat- ing peak times is characteristically high compared to
ing the analysis described below including only activi- other day times. Some of the outliers in Fig. 2b are
ties we designate as Benergy-intensive.^ Table 3 pre- households with unusually high consumption through-
sents a list of these 26 activities and their descriptive out the day. As Fig. 2 suggests, number of activities
statistics. These activities were reported a total of 2218 reported and average electricity consumption are posi-
times by study participants. In Fig. 3, we show how the tively correlated during peak times, implying that activ-
frequency of these activities varies throughout the day. ity reporting itself may be a useful proxy for understand-
While the total number of activities reported during each ing high usage.
4-h period remains proportional to the relative totals in The sample’s mean daily electricity average of
Fig. 2a, the share of Bfood^ and Brecreation^ activities is 565 W (SD = 340 W) is slightly higher than the
larger, and no Bwork^ or Bother care^ activities are Bmedium^ estimates given by Ofgem’s Typical Domes-
reported. tic Consumption Values (TDCV) for different classes of
residential customers, which range from 350 to 490 W
Dependent variable: average electricity usage (Ofgem 2017). This may be because most households
and average Bde-minned^ usage take part in winter months when electricity use is typi-
cally higher and because of the sample biases previously
We use several dependent variables in the following mentioned.
analyses. First, we model average daily electricity con-
sumption in Watts (W) across the full 28-h study period. Statistical analysis: lasso and OLS regression
Next, we model average daily Bde-minned^ electricity
consumption for both the full study period and for 4-h We use a variant on multiple linear regression to select
periods throughout the day. Electricity readings are av- final models of electricity consumption. This variant is
eraged for the time period to which each model corre- the Bleast absolute shrinkage and selection operator^
sponds (i.e., the full-day models use the average and de- (lasso), developed by Tibshirani (1996). Lasso is one
minned average from 5 p.m. on day 1 until 9 p.m. on day of the several prominent regularization techniques that
2 while the other models use de-minned average elec- have gained popularity for use in regression analyses
tricity usage over each specified 4-h period). De- where one or more of the following situations is present:
minning subtracts each household’s minimum demand the set of possible predictors is large and the analysis
from its average usage in order to remove Bbaseload^ aims to identify those that most contribute to variations
electricity consumption (Jin et al. 2017). We use this in the response variable; the data is Bhigh-dimensional,^
technique because our aim is to characterize intra-day meaning the number of predictors is larger than the
variations in electricity consumption and especially number of observations; or the data suffers from a high
those variations that can be explained by activity data, degree of multicollinearity between predictors, which
which we expect to show stronger associations to de- occurs when one predictor variable can be linearly pre-
minned electricity usage than normal averages. All de- dicted from the others in the model.
pendent variables are log-transformed prior to regres- These situations are often present in electricity de-
sion analysis in order to improve linearity of regression mand modeling. Failing to address them can lead to
residuals. Figure 2b shows boxplots for each dependent poor performance of the model and biased results due
Table 1 Selected socio-demographic, dwelling, appliance, and Table 1 (continued)

study participation date characteristics. National figures are based
on estimates from various sources (CCC 2016; DECC 2013; ONS Description Response Sample Population
2016, 2017, 2018a, b) frequency frequency
(%) (%)
Description Response Sample Population
frequency frequency Electricity display 9
(%) (%) Solar thermal 8
Other TV/computer screens M = 2.39,
Home type Flat/apartment 20 21
appliances SD =
Detached 15 17 (count) 1.33
Semi-detached 23 25 Dehumidifiers M = 0.17,
Terraced 42 37 SD =
0.42
Tenure Rent 20 32 Air conditioners M = 0.06,
Own 80 64 SD =
No. of rooms, 2 or fewer 6 4 0.28
excluding Portable heaters M = 0.33,
3–5 59 54
kitchen and SD =
6 or more 35 42 0.60
bathrooms
Household < £25,000 18 41 Night storage heaters M = 0.12,
income SD =
< £35,000 8 16
0.67
< £50,000 18 17 Power showers M = 0.32,
> £50,000 56 27 SD =
No. of 1–2 55 64 0.65
occupants 3–4 41 30
Season Winter (Nov–Feb) 46
>4 4 7 participated Shoulder (Mar–Apr; 19
Age Under 18 30 23 Sept–Oct)
19–50 52 44 Summer 35
Over 50 18 35 (May–August)
Day of week Sunday–Monday 17
Pet ownership Cat 24 participated Monday–Tuesday 16
Dog 9
Tuesday–Wednesday 8
Fish 5
Wednesday–Thursday 36
Estimated < £400 40
Thursday–Friday 8
monthly < £700 41
electric bill Friday–Saturday 16
> £700 19
Bill Not affordable 15
affordability Somewhat affordable 52 to unreliable model coefficients. Furthermore, conven-
Affordable 33 tional methods for variable selection, such as the step-
Electricity tariff Renewable electricity 18 wise methods, have been shown to cause various ana-
Economy 7 or 10 8 lytical issues, including overestimated R-squared values
Other tariff 14 and regression coefficients and predicted values that are
Do not know 14 too narrow (Harrell 2001). Regularization techniques
Large Washing machine 83 83 address these issues by constraining the magnitudes of
appliances Tumble dryer 31 50 coefficients and by trading a small increase in model
Washer dryer 16 14 bias for a larger reduction in variance (see Satre-Meloy
Underfloor heating 19 (2019) for more details). While regularization tech-
Gas boiler 83 85 niques are the subject of a rich literature in statistics
Heat pump 2 and machine learning and are well-suited to challenges
Electric stovetop 34 38 in statistical modeling of electricity demand, their appli-
Electric vehicle 0.3 0.4 cation in the energy literature thus far has been limited
(Hsu 2015; Huebner et al. 2016).
Table 2 Activity categories, examples of their most frequently reported activities, and descriptive statistics
Time-use category Example activities Mean (SD)
Care for home Arrange things, clear up, washing machine, vacuum cleaning, using tumble dryer 3 (2.7)
Care for others Put child to bed, play with child, being with others, bathing children, with pet 3.2 (3.4)
Care for self Sleep, shower, getting dressed, in bed, getting ready 7.7 (5.8)
Food Hot drink, eat hot meal, cook on hob, wash dishes, eating 12.5 (8.2)
Recreation Watching TV, reading, socializing, Internet, E-mail 7.3 (6.2)
Work Work, main job, work second job, work with machinery, study 0.6 (1.1)
In this study, the data are neither Bhigh-dimensional^ estimates are progressively shrunk toward zero. In this
(n ≫ p) nor do the predictors suffer from a high degree of way, lasso enables the selection of a model that does not
multicollinearity, which is determined by inspecting the overfit the data but still has low error. Both the predic-
variance inflation factors (VIFs) for the data.1 For these tors and the response are standardized to have mean zero
reasons, the primary interest in applying regularization and a standard deviation of one prior to running lasso.
is to perform variable selection for a large predictor set Estimating the optimal value for λ is done using k-
to achieve a sparser model. While one of the other fold cross-validation, where models are constructed
fundamental regularization techniques, ridge regression, using a range of values for λ and each model’s mean-
performs well when dealing with Bhigh-dimensional^ squared error (MSE) after cross-validation is plotted.
data, it does not perform variable selection. The elastic Comparing the MSE for each model at varying λ values
net, a combination of ridge and lasso approaches, can then enables the selection of a parsimonious model with
address issues of multicollinearity but is somewhat more low MSE, a process known as hyperparameter tuning in
complex to implement (Zou and Hastie 2005). Our machine learning disciplines. In cases where model
motivation in applying lasso regression is thus to attain sparsity is a primary aim, it is common to follow the
sparsity and deliver statistical and computational gains Bone standard error^ rule, which says to select the
(Tibshirani 2011). simplest model that has an MSE within one standard
Lasso applies a penalized linear regression model error of the model with the minimum MSE (Friedman
that shrinks the coefficients of some regression covari- et al. 2010, p. 17). We follow this rule in the following
ates while setting others exactly to zero, thus performing analyses to improve model interpretability.
variable selection. While ordinary least squares (OLS) Lasso regression is applied to the full predictor set in
regression estimates coefficients by minimizing the reorder to construct a model for full-day average electric-
sidual sum of squares (RSS), lasso minimizes the RSS ity consumption. The predictor variables selected at this

with an added penalty parameter based on ∑pj¼1 β j for stage are then included in an OLS regression, where
some multiplier λ. Minimizing the RSS is thus given by both unstandardized and standardized coefficients are
the following equation: given. Unstandardized coefficients are measured in the
original units of each independent variable. When the
!2 dependent variable has been log-transformed, it changes
n p
∑ yi −α− ∑ β j xij þ λ ∑ β j ð1Þ by 100× (unstandardized coefficient) percent on average
i¼1 j¼1 j¼1 for each one unit increase in the predictor variable,
holding all other variables in the model constant. Stan-
The amount of penalty or shrinkage that is applied to dardized coefficients are instead measured in terms of
the regression coefficients is controlled by the parameter standard deviation and thus can be compared in magni-
λ. As λ is increased, a larger penalty is applied, and the tude to determine which predictors have stronger or
weaker associations with consumption given that pre-
1
VIFs signal whether regression coefficients are inflated due to corre- dictors are measured in different units and on varying
lation between predictor variables. If they are uncorrelated, VIF = 1. scales. Because our data contain numerous binary or
Traditionally, VIFs greater than 10 indicate high multicollinearity
(Roberts and Thatcher 2009). In the models presented here, no VIFs dummy-coded predictors, we follow Gelman’s (2008)
are found to be greater than 10. suggestion of standardizing all non-binary predictors by
Fig. 2 a Frequency of activities reported by time period and time-use category (N = 6265). b Boxplots of the sample’s average electricity
consumption by time period (N = 173)
Table 3 Time-use codes, categories, and activity descriptions for Benergy-intensive^ activities
Time-use category Time-use code Activity description Category mean (SD)
Care for home 3212 Vacuum cleaning 0.67 (1.02)

3311 Washing machine
3320 Ironing
Care for self 310 Wash and dress 1.73 (1.57)
311 Shower
315 Hygiene/beauty
316 Wash
321 Bathing
8574 Shave
Food 212 Eating hot meal 4.79 (3.70)
214 Hot drink
3110 Prepare meal
3112 Prepare hot meal
3120 Baking
3124 Toaster
3133 Dish washer
3190 Prepare meal
Recreation 3720 Internet 5.16 (4.63)
3725 Online shopping
7170 E-mail
8001 Social media
8215 Watching TV
8221 Screen time
8552 Computer games
8553 TV
Fig. 3 Frequency of Benergy-

intensive^ activities reported by
time period and time-use category
(N = 2218)
dividing by two standard deviations rather than one. electricity consumption (for both Ball^ as well as
This allows for numeric predictors’ coefficients to be Benergy-intensive^ activities), we compare adjusted R-
interpreted on the same scale as the binary inputs, which squared values for models with and without the activity
we leave unstandardized. variables included. R Statistics (R Development Core
Next, we model daily average de-minned electricity Team 2008) and the associated packages dplyr
consumption, again using lasso to select influential pre- (Wickham et al. 2019) and Glmnet (Friedman et al.
dictors. Finally, we use lasso to select separate models 2010) are used for data cleaning and regularized regres-
for each 4-h time period, starting with 5–9 p.m. on day 1 sion analyses. Arm (Gelman et al. 2018) is used to
and ending with 5–9 p.m. on day 2. We choose a 4-h computed standardized coefficients, and ggplot2
duration for each model to balance the aim of modeling (Wickham 2016) is used for plots.
typical patterns of daily life (i.e., peak demand hours,
late evening, overnight, early morning, afternoon, etc.)
while keeping the analysis from becoming overly com- Results
plex, but our approach could be applied to consumption
averages taken over shorter or longer time durations. We The following section presents regression results of full-
investigate the predictors selected in each model and day and 4-h models of electricity consumption using
compare their coefficients in order to identify which household survey and activity data.
predictors are influential in explaining consumption pat- Lasso regression is performed on all survey and
terns during different times of day. Next, using only activity data to select influential variables in models of
activities we designate as Benergy-intensive,^ we repeat daily average electricity consumption and de-minned
this analysis to examine whether using a more energy- average electricity consumption. The penalty parameter
relevant categorization of activities improves model λ is tuned using 10-fold cross-validation. Figure 4
coefficients and/or explanatory power. shows the cross-validation procedure. The model with
Finally, to investigate how much activity data con- the minimum MSE (left dotted line) is found with a λ
tributes to explaining variations in de-minned average value of 0.03, includes 23 predictors, and has an MSE of
0.25. The model with an MSE within one standard error This model drops EV ownership and number of rooms
of the minimum (right dotted line) is found with a λ as predictors.
value of 0.08 and includes 13 predictors. In general, the differences between models 1 and 2
First, we include these selected 13 predictors in an are slight, with model 2 showing slightly lower explan-
OLS regression. Second, we re-run lasso on the full atory power than model 1. Notably, we do not find
data, this time using de-minned average daily elec- additional activity variables are selected in model 2,
tricity consumption as the dependent variable. For the even though we might suspect activities to show a
second model, lasso selects two new variables and stronger association to de-minned rather than average
sets two from the first model to zero, meaning the daily consumption. We return to this finding in the
total number of selected predictors remains at 13. We BDiscussion^ section.
again include these variables in a simple OLS regres- To further investigate how categories of time-use
sion. Table 4 presents regression results, including activities might explain patterns in household electricity
unstandardized coefficients (B) and 95% confidence consumption, we model de-minned average electricity
intervals along with standardized coefficients (β) for use over 4-h time periods. Lasso regression is used to
both full-day models, with coefficients listed in order select models (which can have varying numbers of
of standardized coefficient magnitude. For all regres- predictors), and we again include lasso-selected vari-
sion models, we examine diagnostic plots of fitted ables in an OLS regression for each 4-h period to inves-
values versus residuals and Q-Q plots. With few tigate model coefficients. Table 5 shows variables and
exceptions, these plots confirm normality and linear- unstandardized coefficients with 95% confidence inter-
ity of the residuals. vals. We exclude standardized coefficients from this
Model 1 in Table 4 explains R2 = 49% (adjusted R2 = table for simplicity but include these in a table in the
0.44) of the variance in average electricity consumption, Appendix.
F(13, 159) = 10, p < 0.001. The strongest predictors of Table 5 shows that for all times of day except 1 a.m.–
increased daily consumption are mostly dwelling and 5 a.m., activity category variables are frequently select-
appliance-related variables, especially EV ownership, ed in these models and show strong associations to de-
number of power showers, living in a detached home, minned electricity consumption. Activities in the Bcare
number of rooms, and number of TVs/computers. for home,^ Bfood,^ and Brecreation^ categories, in par-
Socio-demographic variables that correlate to increased ticular, are consistently selected as influential predictors
usage include cat ownership and number of occupants. across 4-h models.
The only activity category variable selected is recrea- Coefficients for activity variables show how different
tion; number of recreation activities reported is associ- types of activities vary in their associations to electricity
ated with increased daily average consumption. consumption at different times of day. Care for the
Ownership of a gas boiler and being on a renewable home, which includes some energy-intensive household
or Bgreen^ electricity tariff are the only variables in the chores, has a stronger association with de-minned con-
model that predict decreases in average daily consump- sumption in the early evening, mid-morning, and after-
tion. The former effect is likely observed because not noon models and a weaker association in the late eve-
owning a gas boiler indicates higher use of electricity for ning and early morning models. It is absent from the
space and water heating. The latter effect may reflect a overnight model (1–5 a.m.) as well as from the day 2 5–
predisposition to conserve electricity among those who 9 p.m. model. Food-related activities follow patterns of
opt in to a renewable tariff. typical mealtimes, with stronger coefficients in the eve-
The second model in Table 4, for which the depen- nings, early morning, and afternoon. The models do not
dent variable is de-minned average daily electricity con- show a consistent result in terms of the influence of
sumption, explains R2 = 43% (adjusted R2 = 0.38) of the different mealtimes on consumption, as the food coeffi-
variance of the dependent variable, F(13, 159) = 9.2, cient for the morning mealtime is larger than in the day 1
p < 0.001. In addition to the variables in model 1, lasso 5–9 p.m. model but smaller than in the day 2 5–9 p.m.
selects being on an Economy 7 or 10 tariff, which model. This finding highlights that electricity consump-
typically charge lower prices for nighttime or Boff-peak^ tion of food preparation and consumption can vary
electricity use, and number of night storage heaters. between days during the same mealtimes. Recreation
Both have positive coefficients but high uncertainty. activities show strong associations to electricity
Fig. 4 Plot of cross-validation MSE for lasso regression of daily error bars show the mean and standard error of the cross-validation
average electricity consumption. The horizontal bottom axis MSE, respectively. The left vertical dotted line indicates the λ
shows the logarithm of the tuning parameter λ, which increases value at which the minimum MSE is found, while the right vertical
in magnitude from left to right. The top horizontal axis shows the dotted line indicates the model with the fewest nonzero coeffi-
number of nonzero coefficients for each model run, and points and cients within one standard deviation of the minimum
consumption except overnight and during the early coefficients can be interpreted in terms of their compar-
morning, but this variable is also absent from the day 2 ison to the reference category, which is winter partici-
5–9 p.m. model. pation. Households that participate in summer show
The only model in which no activity variables are lower 5–9 p.m. de-minned electricity usage than house-
selected is the overnight model, which also happens to holds that participate in winter; similarly, households
be the model with the fewest selected predictors. Only that participate in spring/autumn show slightly increased
EV ownership and being on an Economy 7 or 10 tariff 9 p.m.–1 a.m. consumption compared with households
are selected. Both of these can reasonably be expected to participating in winter.
explain increases in nighttime electricity usage. Regarding model explanatory power, we find that the
Regarding non-activity predictors that are selected in model with the largest proportion of explained variance
the 4-h models, many of these variables are the same as is the 5–9 p.m. day 1 model (R2 = 43%, adjusted R2 =
those selected in the models of 28-h electricity con- 0.40) and that the model with the lowest is the 5–9 p.m.
sumption. Some of these variables, such as living in a day 2 model (R2 = 25 %, adjusted R2 = 0.23). The results
detached home, number of occupants, and estimated for the two comparable 5–9 p.m. models also show
monthly electric bill appear to influence de-minned inconsistencies in terms of variables selected, especially
electricity consumption consistently throughout the in the case of the activity variables. This finding chal-
day, as indicated by their selection in the majority of lenges the notion of Baverage load profiles^ (as com-
the models. Other variables are much more time-specific monly used for settlement, see Elexon (2013)) and
in their associations with de-minned electricity usage. suggests that activity-related factors that influence elec-
For instance, gas boiler ownership shows a large, tricity consumption vary from day to day. This is a
negative association with electricity consumption in preliminary finding, however, and draws from a limited
the late evening and early morning models. This finding sample of days.
may be intuitive, since space heating is typical in the late To address the limitation that these categories of
evening and early morning, and homes that heat with activities do not differentiate between Benergy-
gas rather than electricity will likely have lower electric- intensive^ and Blow-energy^ activities, we repeat this
ity consumption during these hours. analysis including only those activities that we designate
Because households select their own date for partic- as Benergy-intensive.^ Table 6 presents results for these
ipation, we include season and day of week as candidate models. With the exception of several coefficients, we
predictors for each model. Table 5 shows that there does observe in most cases a marked increase in the care for
not appear to be a notable day-of-week effect for our home, food, and recreation activity variables’ coeffi-
sample, but we do find a slight seasonal effect, as the 5– cients. This is especially true for Bcare for home^ activ-
9 p.m. day 1 model includes the Bparticipated in ities reported in the morning, mid-morning, and after-
summer^ predictor with a negative coefficient, and the noon periods, suggesting energy-intensive housework is
9 p.m.–1 a.m. model includes the Bparticipated in spring a strong predictor of de-minned usage during these
or autumn^ predictor with a positive coefficient. These times.
Table 4 Regression results for average electricity usage and average de-minned usage during full 28-h study period
1) Daily average (W) 2) De-minned daily average (W)
B (5–95% CI) β B (5–95% CI) β
EV 0.61** (0.18, 1.04) 0.61

Tariff: Renewable − 0.30** (− 0.50, − 0.11) − 0.30 − 0.30* (− 0.54, − 0.06) − 0.26
Gas boiler − 0.27* (− 0.48, − 0.06) − 0.27 − 0.19 (− 0.46, 0.09) − 0.16
Owns cat 0.25** (0.08, 0.43) 0.25 0.18 (− 0.04, 0.39) 0.24
Tariff: Economy 7 or 10 0.43 (− 0.03, 0.88) 0.24
Recreation activities 0.02** (0.01, 0.03) 0.24 0.03*** (0.01, 0.04) 0.23
No. of power showers 0.15* (0.03, 0.27) 0.19 0.13 (− 0.02, 0.28) 0.18
Detached home 0.19 (− 0.04, 0.43) 0.19 0.45** (0.18, 0.73) 0.34
No. of rooms 0.07* (0.005, 0.14) 0.18
No. of TV/computers 0.06* (0.004, 0.12) 0.17 0.06 (− 0.02, 0.13) 0.18
Under floor heating 0.16 (− 0.03, 0.35) 0.16 0.16 (− 0.07, 0.40) 0.20
Tumble dryer 0.10 (− 0.07, 0.27) 0.10 0.10 (− 0.11, 0.30) 0.08
No. of night storage heaters 0.07 (− 0.11, 0.25) 0.01
Estimated monthly electric bill 0.03 (− 0.02, 0.09) 0.10 0.02 (− 0.05, 0.09) 0.13
No. of occupants 0.03 (− 0.04, 0.10) 0.08 0.08 (− 0.01, 0.16) 0.11
Constant 5.42*** (5.10, 5.74) 6.27 4.97*** (4.61, 5.33) 6.15
Observations 173 173
R2 0.49 0.43
Adjusted R2 0.44 0.38
Residual std. error (df = 159) 0.47 0.57
*p < 0.05; **p < 0.01; ***p < 0.001
In some instances, the energy-intensive activity different categorizations, though the strength of associ-
models either drop or add activity-related variables for ations clearly increases when only energy-intensive ac-
different times of day. While the Bcare for home^ cate- tivities are included.
gory when not filtering for energy-intensive activities is To test the extent to which categorized activity data
selected in the day 1 5–9 p.m. and the 9 p.m.–1 a.m. improves model explanatory power, we compute the
models, it is not selected when only energy-intensive adjusted R2 value for each model excluding the lasso-
activities are included. This may suggest that more selected activity predictors for that model, and we do
energy-intensive activities related to household chores this for both categorizations tested.2 We present these
are not as common in the evenings. It also indicates that results in Table 7. We see that excluding activity data
Bcare for home^ as a category includes non-energy- reduces model explanatory power by an average of nine
intensive activities that are important for explaining percentage points across models (excluding the 1–5 a.m.
electricity consumption during evenings. model). We do not observe, however, a noticeable dif-
Similar distinctions are found for food-related activ- ference between the models including all activity data
ities. The model including only energy-intensive food and those including only Benergy-intensive^ activities.
activities more clearly differentiates between evening These findings provide evidence of the potential im-
and morning meals, as morning food activities show provements to household electricity modeling by in-
less associations with electricity consumption when excluding activity data, even when it is relatively coarsely
cluding non-energy-intensive activities, likely because categorized.
morning meals are often cold and do not involve appli-
ance use. For recreation activities, the timing of when 2
Adjusted R2 is used for the comparison in order to account for the
these are influential does not change between the two varying number of predictors in each model.
Table 5 Regression results for de-minned average electricity use: 4-h models (5 p.m. day 1–9 p.m. day 2)
Predictor 5–9 p.m. (1) 5–9 p.m. (2) 9 p.m.–1 a.m. 1 a.m.–5 5 a.m.–9 a.m. 9 a.m.–1 1 p.m.–5
a.m. p.m. p.m.
Activity categories
Care for home 0.17*** 0.16* 0.16 0.15 0.15
(0.09, 0.24) (0.002, 0.32) (− 0.07, 0.39) (− 0.02, (− 0.03,
0.33) 0.32)
Care for others 0.15 0.35**
(− 0.04, 0.33) (0.10, 0.60)
Food 0.03 0.10*** 0.08* 0.06 0.08
(− 0.02, 0.07) (0.05, 0.15) (0.01, 0.16) (− 0.01, 0.13) (− 0.02,
0.19)
Other activity 0.22*
(0.02, 0.43)
Recreation 0.06* 0.08* 0.21* 0.25*
(0.01, 0.12) (0.01, 0.15) (0.02, 0.39) (0.04, 0.46)
Appliances
EV 2.23***
(1.06, 3.40)
Gas boiler − 0.37* − 0.59**
(− 0.65, (− 0.99,
− 0.09) − 0.20)
Heat pump 0.76
(− 0.20, 1.72)
No. of night storage heaters 0.39**
(0.15, 0.63)
No. of power showers 0.23** 0.27*
(0.06, 0.39) (0.01, 0.53)
No. of TVs/computers 0.07 0.17** 0.09
(− 0.04, (0.04, 0.30) (− 0.04,
0.18) 0.22)
Tumble dryer 0.28 0.37*
(− 0.04, 0.60) (0.01, 0.74)
Under floor heating 0.11 0.34
(− 0.15, 0.37) (− 0.04, 0.72)
Dwelling
Detached home 0.62*** 0.56** 0.20 0.17 0.66**
(0.29, 0.95) (0.15, 0.97) (− 0.10, 0.50) (− 0.25, 0.59) (0.19, 1.14)
No. of rooms in home 0.08 0.10 0.08
(− 0.05, (− 0.04, (− 0.06,
0.20) 0.23) 0.22)
Terraced home − 0.16
(− 0.41, 0.08)
Socio-demographics
Bill affordability − 0.18*
(− 0.35,
− 0.01)
Estimated monthly electric 0.09 0.06 0.09 0.08
bill (− 0.01, (− 0.02, 0.13) (− 0.02, 0.19) (− 0.04,
0.19) 0.21)
No. of occupants 0.13* 0.06 0.09 0.13
(0.03, 0.22) (− 0.03, 0.15) (− 0.06, 0.23)
Table 5 (continued)
Predictor 5–9 p.m. (1) 5–9 p.m. (2) 9 p.m.–1 a.m. 1 a.m.–5 5 a.m.–9 a.m. 9 a.m.–1 1 p.m.–5
a.m. p.m. p.m.
(− 0.02,
0.27)
Owns cat 0.31 0.36** 0.35*
(− 0.01, (0.12, 0.59) (0.02, 0.69)
0.64)
Participated in summer − 0.43***
(− 0.67,
− 0.18)
Participated in 0.34**
spring/autumn (0.09, 0.59)
Tariff: Economy 7 or 10 1.73***
(1.07, 2.40)
Tariff: Renewable − 0.42** − 0.43**
(− 0.72, (− 0.69,
− 0.11) − 0.17)
Tenure (rent vs. own) − 0.38*
(− 0.73,
− 0.02)
Constant 5.82*** 4.55*** 5.04*** 3.92*** 4.53*** 3.53*** 3.42***
(5.38, 6.26) (4.00, 5.10) (4.68, 5.40) (3.75, 4.09) (3.98, 5.07) (2.91, 4.15) (2.79, 4.04)
Observations 171 170 172 172 172 170 170
2
R 0.43 0.25 0.43 0.28 0.33 0.29 0.39
Adjusted R2 0.40 0.23 0.39 0.27 0.28 0.26 0.35
*p < 0.05; **p < 0.01; ***p < 0.001
Discussion These results also, however, identify new variables that

might help predict daily usage, such as electricity tariff
Summary and discussion of results choice or pet ownership. The results provide some evi-
dence that these may be important for understanding
The results show that 13 lasso-selected predictors can electricity consumption patterns.
explain 49% of full-day average household electricity This study goes beyond previous research by investi-
usage. When the dependent variable is Bde-minned^ to gating how the associations between appliance, dwelling,
remove baseload consumption, lasso selects a model that and socio-demographic variables and de-minned electric-
explains 43% of the variance in average de-minned elec- ity consumption vary as a function of time of day. The EV
tricity usage. These results are consistent with or in some ownership variable, for instance, is only selected in the
cases improve on model explanatory power found in nighttime model, which suggests that ownership of an
previous studies of household electricity usage (Huebner EV increases a household’s nighttime consumption more
et al. 2016; Kavousian et al. 2013; McLoughlin et al. so than it increases consumption during other times.
2012). For the models of 4-h average de-minned electricity Similarly, the number of power showers and presence
consumption, explanatory power is somewhat lower, rang- of underfloor heating are associated with consumption
ing from 25 to 43% of variance explained. These models during late evening and morning hours, when showers
include varying numbers of lasso-selected predictors, but are likely more frequent. Number of TVs/computers
we find that removing activity variables from these models owned is a stronger predictor in the morning and evening.
reduces their explanatory power considerably. The study also tests the extent to which different
The results for both full-day models confirm previous types of activities are associated with electricity con-
findings that appliance ownership and occupant socio- sumption during different periods of the day. We find
demographics show the strongest associations to elec- evidence that certain types of activities are more relevant
tricity consumption patterns (Huebner et al. 2016). for understanding electricity consumption patterns and
450
Table 6 Regression results for de-minned average electricity use: 4-h models (5 p.m. day 1–9 p.m. day 2) including only Benergy-intensive^ activities across categories
Predictor 5–9 p.m. (1) 5–9 p.m. (2) 9 p.m.–1 a.m. 1 a.m.–5 a.m. 5 a.m.–9 a.m. 9 a.m.–1 p.m. 1 p.m.–5 p.m.
Activity categories
Care for home 1.05** 0.52* 0.70**
(0.39, 1.71) (0.13, 0.91) (0.24, 1.15)
Care for self 0.11
(− 0.05, 0.27)
Food 0.02 0.18** 0.23** 0.14 0.20*
(− 0.06, 0.11) (0.06, 0.30) (0.06, 0.40) (− 0.07, 0.35) (0.01, 0.39)
Recreation 0.10** 0.20*** 0.35** 0.30*
(0.04, 0.16) (0.10, 0.29) (0.12, 0.58) (0.05, 0.55)
Appliances
EV 2.23***
(1.06, 3.40)
Gas boiler − 0.32* − 0.65***
(− 0.59, − 0.05) (− 1.03, − 0.27)
No. of night storage heaters 0.42***
(0.18, 0.66)
No. of power showers 0.26** 0.22
(0.10, 0.42) (− 0.05, 0.49)
No. of TVs/computers 0.14* 0.08
(0.01, 0.27) (− 0.05, 0.20)
(− 0.03, 0.71) (0.02, 0.75)
Under floor heating 0.12 0.40* 0.38
(− 0.13, 0.38) (0.03, 0.77) (− 0.05, 0.81)
Dwelling
Detached home 0.68*** 0.66** 0.20 0.28 0.18 0.76**
(0.34, 1.03) (0.24, 1.07) (− 0.09, 0.50) (− 0.14, 0.69) (− 0.32, 0.69) (0.28, 1.24)
No. of rooms in home 0.08 − 0.01 0.07
(− 0.04, 0.21) (− 0.16, 0.14) (− 0.07, 0.21)
(− 0.39, 0.13)
Socio-demographics
Estimated monthly electric bill 0.11* 0.04 0.10 0.04
(0.01, 0.21) (− 0.03, 0.11) (− 0.01, 0.20) (− 0.08, 0.17)
No. of occupants 0.16** 0.08 0.15* 0.12 0.15*
(0.06, 0.26) (− 0.01, 0.17) (0.01, 0.29) (− 0.04, 0.28) (0.001, 0.30)
Owns cat 0.33** 0.33
(0.10, 0.56) (− 0.01, 0.67)
− 0.43**
Energy Efficiency (2020) 13:433–458
1 p.m.–5 p.m.
also that certain activities have stronger or weaker asso-
(2.86, 4.11)
ciations at different times of day.
3.48***
In particular, our results show that housework, eating and
0.37
0.34
170
meal preparation, and recreation or media use are strongly
associated with de-minned electricity consumption,
confirming similar findings from previous research
9 a.m.–1 p.m.
(Anderson and Torriti 2018; De Lauretis et al. 2017; Jalas
(2.97, 4.28)
(0.11, 1.01)
and Juntunen 2015; Palmer et al. 2013). The times at which
3.62***
0.56*
0.33
0.28
these activities are stronger predictors of electricity con-
170
sumption also provides insight into when activities and
electricity use are more tightly coupled, such as during
evening peak periods and later evenings for food and recre-
5 a.m.–9 a.m.
ational activities, early mornings for personal care activities,

(4.03, 5.00)
and early and mid-mornings for household chores.

4.52***
0.30
0.26 Our results also highlight the extent to which
172
denoting activities as Benergy-intensive^ can strengthen

these links. We find in general a sizeable increase in the
strength of relationships between number of activities
1 a.m.–5 a.m.
(1.07, 2.40)
(3.75, 4.09)
reported and average de-minned electricity consumption

throughout the day when including only Benergy-
1.73***
3.92***
0.28
0.27
172
intensive^ activities. We do note, however, that these

results are somewhat mixed. For instance, the Bcare for
home^ variable is significant in the day 1 5–9 p.m.
model when including all activity data but not when
(− 0.65, − 0.14)
9 p.m.–1 a.m.
(0.12, 0.61)
(4.63, 5.34)
including only energy-intensive activities. This finding

− 0.40**
suggests other Bcare for home^ activities are important

4.99***
0.36**
0.45
0.41
172
for explaining electricity consumption during this time.

While acknowledging that our designation of activities
as Benergy-intensive^ is subjective, we take this finding
as evidence that the links between activities and elec-
5–9 p.m. (2)
(4.24, 5.31)
tricity consumption may be more complex than simply

4.77***
differentiating activities on the basis of their expected

0.20
0.18
170
use of energy. Further discussion of this point can be

found in Grünewald and Diakonova (2018b).
Two additional results merit further discussion. The
(− 0.69, − 0.17)
(− 0.81, − 0.20)
first is our finding that we do not observe an increase in

5–9 p.m. (1)
(5.24, 5.87)
the number of activity variables selected between the

− 0.51**
5.56***
daily average and daily de-minned average models. We

0.35
0.32
171
do, however, observe frequent selection of activity var-

iables in the 4-h de-minned usage models. We expect
*p < 0.05; **p < 0.01; ***p < 0.001
that this observation results from the duration of time

Participated in spring/autumn
over which we are de-minning electricity consumption.

Over a 28-h period, de-minning does not appear to yield
Participated on Fri–Sat
stronger associations between activities reported and

Table 6 (continued)
Tariff: Renewable
average consumption. Over shorter 4-h periods, howev-

Tariff: Economy
Participated in
er, de-minning does appear to strengthen relationships

Observations
Adjusted R2
summer
7 or 10
Constant
between reported activities and consumption. This find-

Predictor
ing suggests that activity data can improve models of

R2
shorter duration more so than it can models of longer

Table 7 Model explanatory power with and without lasso-selected activity variables for both Ball^ activity data and Benergy-intensive^
activity data models
5–9 p.m. 5–9 p.m. 9 p.m.–1 1 a.m.– 5 a.m.– 9 a.m.–1 1 p.m.–5

(1) (2) a.m. 5 a.m. 9 a.m. p.m. p.m.
BAll^ activity models

Adjusted R2 0.40 0.23 0.39 0.27 0.28 0.26 0.35
Adjusted R2 without 0.28 0.17 0.30 – 0.25 0.13 0.23
activity variables
Difference in adj R2 0.12 0.06 0.09 – 0.03 0.13 0.12
BEnergy-intensive^ activity models

Adjusted R2 0.31 0.18 0.41 0.27 0.26 0.28 0.34
Adjusted R2 without 0.27 0.15 0.30 – 0.22 0.18 0.23
activity variables
Difference in adj R2 0.04 0.03 0.11 – 0.04 0.10 0.11
duration. We expect activities to show increasingly Our novel approach to collect time-use data in paral-
strong associations to more finely resolved electricity lel with electricity consumption can overcome many of
readings, and this is something we are exploring in the limitations and challenges previously discussed in
current research. the literature on time-use or activity-based models of
The second result deserving of further discussion is our energy demand (McKenna et al. 2017; Torriti 2014).
finding that the models for the day 1 5–9 p.m. period and Specifically, this method enables the collection of more
the day 2 5–9 p.m. period show considerable differences energy-relevant activities, which we have shown to
in the variables selected and in model explanatory power. increase the strength and statistical validity of the rela-
We note that fewer activities are reported during the day 2 tionships between household activities and electricity
5–9 p.m. period, but given that lasso standardizes predic- consumption. Given the lack of empirical evidence on
tors prior to selection, this should not influence the algo- these relationships, these results can help improve the
rithm’s selection procedure. Furthermore, we note that the specification of activity-appliance signals in models
1–5 p.m. model includes several activity variables and a incorporating time-use data. Our approach also facili-
greater explanatory power than the day 2 5–9 p.m. model, tates the collection of activity data from multi-occupant
even with far fewer activities reported. We take this result households, which have previously been more difficult
to suggest that the relationship between activities and to account for in time-use models of electricity demand.
electricity consumption may vary considerably in the Second, while some literature has concluded that it is
same households between days and also that peak hours sufficient to model active occupancy states for the purpose
may be especially variable. This finding, too, warrants of constructing more accurate energy demand models, we
further exploration in future research. believe that such an approach fails to more directly link the
types of activities that are being performed during Bactive
occupancy,^ which have important consequences and pol-
Policy implications icy implications for delivering more flexible demand.
Our results show that certain types of activities have
These results have implications both for energy demand stronger relationships to electricity consumption at dif-
models and for policy considerations surrounding de- ferent times of day. Targeting these activities in DSR
mand flexibility and DSR. First, we have shown that interventions could yield larger shifts in demand. Fur-
activity data, whether categorized as Benergy-intensive^ thermore, tailoring DSR interventions based on better
or not, can improve models of household electricity use. evidence about what is actually occurring in households
For at least one of our Bpeak period^ models, we find during peak times and about how these activities vary
that activity data is especially useful for understanding among different segments of the population can im-
variations in de-minned electricity use during these prove their effectiveness while also mitigating their
hours. impact on vulnerable populations.
In terms of potential for shifting demand alone, these We believe there is potential to scale this research to much
results suggest food preparation and meal times, house- larger (> 1000 households), nationally representative
hold chores, and recreational activities should be prior- samples using app-based activity diaries coupled with
itized for activity-led DSR given their strong associa- smart meter data, and preparations are underway for
tions with electricity use. But this is not the only con- scaling this study. We also encourage similar studies
sideration upon which effective DSR strategies should across varying cultural and societal contexts.
be based. Other considerations, such as those compris-
ing Torriti et al.’s (2015) BFlexibility Index^, are also
valuable for determining which activities at which times Conclusions
can be shifted without adverse social impacts.
This paper presents analysis of intra-day electricity con-
Study limitations sumption for a sample of 173 UK households. Electric-
ity provision during peak times is costly and challenging
Some limitations are present in this analysis. The sample in energy systems with high penetrations of renewable
is non-representative, and several sample biases are pres- electricity. Understanding the drivers behind electricity
ent. The sample size is also small relative to the number consumption during peak times can assist in the design
of predictors. Standard errors for most variables in the of more effective intervention strategies to shift electric-
regression models are therefore large in comparison to ity use patterns in the residential sector.
coefficient size. Models are not constructed for different We present regression models of average full-day and
days of week or seasons in order to preserve the sample 4-h de-minned electricity consumption using a wide
size, but these variables are instead included at the model range of predictive factors, including socio-demograph-
selection stage. Furthermore, self-reported activities can ic, physical dwelling, appliance ownership, and catego-
lead to biases and inaccuracies. Measuring activities as rized activity variables. Our analysis tests the ability of
simple frequency counts is a coarse way to include these these variables to explain consumption at different times
data and may obfuscate more complex relationships of day and the strength of these relationships.
between activity type, timing, and electricity Our results show that adding activity data to regres-
consumption. A similar approach as the one taken by sion models can improve their explanatory power of de-
Rhodes et al. (2014) and McLoughlin et al. (2015), who minned electricity consumption during most times of
cluster load profiles and use household characteristics to day. This result holds regardless of whether only activ-
build predictive models of membership in distinctive ities that we reasonably assume to be Benergy-intensive^
load profiles, may be instructive in this context. Time- are included. Given how challenging it is to achieve
use data may be even better suited to this approach than greater explanatory power of highly diverse electricity
socio-demographic, dwelling, or appliance ownership uses in households, a nine-percentage point increase in
data, due to its temporal resolution. adjusted R2 on average is encouraging. The results also
Our initial categorization of activity data relies on a show strong associations between activities and electric-
taxonomy that was not developed for energy modeling ity consumption at different times of day. The impor-
purposes, and although we make an attempt at catego- tance of household chores, food, and recreation activi-
rizing activities as Benergy-intensive,^ this is a subjec- ties for explaining electricity consumption patterns is
tive exercise. We believe much more work could be clearly demonstrated.
done to identify more energy-relevant activity catego- We expect that more nuanced categorizations of ac-
ries with useful implications for energy demand model- tivities and investigations of their relationship to elec-
ing and for estimating the potential of demand response tricity consumption at more finely resolved timescales
(Anderson 2016). will yield improved insights into the flexibility of de-
Finally, more data from a more representative sample mand. This evidence could reduce the cost of integrating
of households could reduce uncertainty surrounding the renewable energy sources into existing energy systems
strength of relationships between predictive variables and by providing a better estimate of the potential of
electricity consumption and could make these results demand-side response in the residential sector, thus
more generalizable. The data collection is ongoing, such contributing to accelerating the low-carbon energy
that more detailed analyses can be conducted in future. transition.
Funding information This research received funding from the Compliance with ethical standards The University of Ox-
Engineering and Physical Sciences Research Council (EPSRC) ford’s Central University Research Ethics Committee (CUREC)
under grant EP/M024652/1.Data availabilityThe data that support has approved this study.
the findings of this study are available from the corresponding
author upon request. The tools used for time-use and electricity Conflict of interest The authors declare that they have no con-
data collection are publicly available at https://goo.gl/81mvCB. flicts of interest.
Appendix
Table 8 Standardized regression coefficients for Table 5
5–9 p.m. (1) 5–9 p.m. (2) 9 p.m.–1 a.m. 1 a.m.–5 a.m. 5 a.m.–9 a.m. 9 a.m.–1 p.m. 1 p.m.–5 p.m.
Predictor β β β β β β β
Activity categories
Care for home 0.46*** 0.21* 0.22 0.32 0.28
Care for others 0.23 0.49**
Food 0.16 0.52*** 0.25* 0.25 0.31
Other activity 0.37*
Recreation 0.33* 0.25* 0.42* 0.44*
Appliances
EV 2.23***
Gas boiler − 0.37* − 0.59**
Heat pump 0.76
No. of night storage heaters 0.53**
No. of power showers 0.29** 0.35*
No. of TVs/computers 0.19*** 0.46** 0.24
Under floor heating 0.11 0.34
Dwelling
Detached home 0.62*** 0.56** 0.20 0.17 0.66**
No. of rooms in home 0.20** 0.25 0.21
Socio-demographics
Bill affordability − 0.24*
Estimated monthly electric bill 0.27*** 0.18 0.26 0.25
No. of occupants 0.30* 0.14 0.20 0.29
Owns cat 0.31 0.36** 0.35*
Participated in summer − 0.43***
Participated in spring/autumn 0.34**
Tenure (rent vs. own) − 0.38*
Constant 6.19 5.66 5.66 4.10 5.28 4.99 4.75
Observations 171 170 172 172 172 170 170
R2 0.43 0.25 0.43 0.28 0.33 0.29 0.39
Adjusted R2 0.40 0.23 0.39 0.27 0.28 0.26 0.35
*p < 0.05; **p < 0.01; ***p < 0.001

Table 9 Standardized regression coefficients for Table 6
5–9 p.m. (1) 5–9 p.m. (2) 9 p.m.–1 a.m. 1 a.m.–5 a.m. 5 a.m.–9 a.m. 9 a.m.–1 p.m. 1 p.m.–5 p.m.
Predictor β β β β β β β
Activity categories
Care for home 1.05** 0.44* 0.50**
Care for self 0.21
Food 0.04 0.43** 0.28** 0.23 0.38*
Recreation 0.40** 0.43*** 0.53** 0.44*
Appliances
EV 2.23***
Gas boiler − 0.32* − 0.65***
No. of night storage heaters 0.56***
No. of power showers 0.34** 0.29
No. of TVs/computers 0.38* 0.20
Under floor heating 0.12 0.40* 0.38
Dwelling
Detached home 0.69*** 0.66** 0.20 0.28 0.18 0.76**
No. of rooms in home 0.22 − 0.02 0.19
Socio-demographics
Estimated monthly electric bill 0.34* 0.12 0.29 0.13
No. of occupants 0.39** 0.19 0.35* 0.28 0.35*
Owns cat 0.33** 0.33
Participated in summer − 0.43**
Participated in spring/autumn 0.36**
Participated on Fri–Sat 0.56*
Constant 6.19 5.72 5.62 4.10 5.34 4.69 4.74
Observations 171 170 172 172 172 170 170
R2 0.43 0.25 0.43 0.28 0.33 0.29 0.39
Adjusted R2 0.40 0.23 0.39 0.27 0.28 0.26 0.35
*p < 0.05; **p < 0.01; ***p < 0.001
Open Access This article is distributed under the terms of the Anderson, B., & Torriti, J. (2018). Explaining shifts in UK elec-
Creative Commons Attribution 4.0 International License (http:// tricity demand using time use data from 1974 to 2014.
creativecommons.org/licenses/by/4.0/), which permits unrestrict- Energy Policy, 123, 544–557. https://doi.org/10.1016/j.
ed use, distribution, and reproduction in any medium, provided enpol.2018.09.025.
you give appropriate credit to the original author(s) and the source, Anderson, B., Lin, S., Newing, A., Bahaj, A., & James, P. (2016).
provide a link to the Creative Commons license, and indicate if Electricity consumption and household characteristics: impli-
changes were made. cations for census-taking in a smart metered future.
Computers, Environment and Urban Systems, 63, 58–67.
https://doi.org/10.1016/j.compenvurbsys.2016.06.003.
References BEIS. (2017). The clean growth strategy: leading the way to a low
carbon future. Department for Business, Energy & Industrial
Strategy.
Anderson, B. (2016). Laundry, energy and time: Insights from 20
years of time-use diary data in the United Kingdom. Energy BEIS. (2018). DUKES chapter 5: statistics on electricity from
Research & Social Science, 22, 125–136. https://doi. generation through to sales. In Digest of UK Energy Statistics
org/10.1016/j.erss.2016.09.004. (DUKES): electricity. London: Department for Business,
Energy & Industrial Strategy https://www.gov. Gelman, A. (2008). Scaling regression inputs by dividing by two
uk/government/statistics/electricity-chapter-5-digest-of-united- standard deviations. Statistics in Medicine, 27(15), 2865–
kingdom-energy-statistics-dukes. Accessed 11 April 2019 2873. https://doi.org/10.1002/sim.3107.
CCC. (2016). Annex 2: heat in UK buildings today. In Next steps Gelman, A., Su, Y.-S., Yajima, M., Hill, J., Pittau, M. G., Kerman,
for UK heat policy. London: Committee on Climate Change J., et al. (2018). arm: data analysis using regression and
https://www.theccc.org.uk/wp-content/uploads/2017/01 multilevel/hierarchical models. https://CRAN.R-project.
/Annex-2-Heat-in-UK-Buildings-Today-Committee-on- org/package=arm.
Climate-Change-October-2016.pdf . Accessed 12 February Gershuny, J. (2008). Time-use studies: daily life and social
2019. change: Full Research Report (ESRC End of Award
De Lauretis, S., Ghersi, F., & Cayla, J.-M. (2017). Energy consump- Report, RES-000-23-0704-A). Swindon: ESRC. https://s3-
tion and activity patterns: an analysis extended to total time and eu-west-1.amazonaws.com/esrc-files/outputs/DXDg-
energy use for French households. Applied Energy, 206, 634– pyFREKpOk_6bEon8Q/gWE89bsrZEu6BF9Azmg_Cw.pdf
648. https://doi.org/10.1016/j.apenergy.2017.08.180. Gershuny, J., & Sullivan, O. (2017). United Kingdom time use
DECC. (2012). Demand side response in the domestic sector - a survey, 2014–2015. [data collection]. UK Data Service. SN:
literature review of major trials. London: Department of 8128. https://doi.org/10.5255/UKDA-SN-8128-1
E n e r g y a n d C l i m a t e C h a n g e h t t p s : / / w w w. g o v. Grunewald, P., & Diakonova, M. (2018a). Flexibility, dynamism
uk/government/uploads/system/uploads/attachment_ and diversity in energy supply and demand: a critical review.
data/file/48552/5756-demand-side-response-in-the- Energy Research & Social Science, 38, 58–66. https://doi.
domestic-sector-a-lit.pdf. Accessed 11 April 2019. org/10.1016/j.erss.2018.01.014.
DECC. (2013). Report 9: Domestic appliances, cooking & cooling Grunewald, P., & Diakonova, M. (2018b). The electricity footprint
equipment (No. BRE report number 288143). London: of household activities - implications for demand models.
Department of Energy and Climate Change https://www. Energy and Buildings, 174, 635–641. https://doi.org/10.1016
gov.uk/government/uploads/system/uploads/attachment_ /j.enbuild.2018.06.034.
data/file/274778/9_Domestic_appliances__cooking_and_ Grünewald, P., & Diakonova, M. (2019). METER data. Energy-
cooling_equipment.pdf. Accessed 16 March 2018. use.org. http://www.energy-use.org/docs/data/. Accessed 11
DECC. (2014). Seasonal variations in electricity demand. April 2019.
London: Department of Energy and Climate Change Grünewald, P., & Layberry, R. (2015). Measuring the relationship
h t t p s : / / a s s e t s . p u b l i s h i n g . s e r v i c e . g o v. between time-use and electricity consumption. In
uk/government/uploads/system/uploads/attachment_ Proceedings of the 2015 ECEEE Summer Study on Energy
data/file/295225/Seasonal_variations_in_electricity_ Efficiency in Buildings (Vol. 9, pp. 2087–96). Presqu’île de
demand.pdf . Accessed 12 February 2019. Giens, France: ECEEE. www.eceee.org/library/conference_
proceedings/eceee_Summer_Studies/2015/9-dynamics-of-
Druckman, A., Buck, I., Hayward, B., & Jackson, T. (2012). Time,
consumption/measuring-the-relationship-between-time-use-
gender and carbon: a study of the carbon implications of
and-electricity-consumption/.
British adults’ use of time. Ecological Economics, 84, 153–
Grünewald, P., Diakonova, M., Zilli, D., Bernard, J., & Matousek,
163. https://doi.org/10.1016/j.ecolecon.2012.09.008.
A. (2017). What we do matters – a time-use app to capture
Elexon. (2013). Load profiles and their use in electricity settlement energy relevant activities. In Proceedings of the 2017 ECEEE
(p. 31). Elexon www.elexon.co.uk/wp-content/uploads/2013 Summer Study on Energy Efficiency in Buildings (Vol. 9, pp.
/11/load_profiles_v2.0_cgi.pdf. 2085–93). Presqu’île de Giens, France: ECEEE. www.eceee.
Ellegård, K., & Palm, J. (2011). Visualizing energy consumption org/library/conference_proceedings/eceee_Summer_
activities as a tool for making everyday life more sustainable. Studies/2017/9-consumption-and-behaviour/what-we-do-
Applied Energy, 88(5), 1920–1926. https://doi.org/10.1016/j. matters-8211-a-time-use-app-to-capture-energy-relevant-
apenergy.2010.11.019. activities/.
Ellegård, K., Vrotsou, K., & Widén, J. (2010). VISUAL- Harrell, F. (2001). Regression modeling strategies: with applica-
TimePAcTS/energy use – a software application for visual- tions to linear models, logistic regression, and survival anal-
izing energy use from activities performed. In Conference ysis. New York: Springer-Verlag www.springer.
proceedings 3rd International Scientific Conference on com/gb/book/9781441929181.
BEnergy Systems with It^ (pp. 1–17). http://citeseerx.ist. Hsu, D. (2015). Identifying key variables and interactions in
psu.edu/viewdoc/download?doi=10.1.1.468.5526&rep=rep1 statistical models of building energy consumption using reg-
&type=pdf. ularization. Energy, 83, 144–155. https://doi.org/10.1016/j.
eurostat. (2009). Harmonised European time use surveys: 2008 energy.2015.02.008.
guidelines. Luxembourg: Office for Official Publications of Huebner, G., Shipworth, D., Hamilton, I., Chalabi, Z., &
the European Communities. Oreszczyn, T. (2016). Understanding electricity consump-
Firth, S., Lomas, K., Wright, A., & Wall, R. (2008). Identifying tion: a comparative contribution of building factors, socio-
trends in the use of domestic appliances from household demographics, appliances, behaviours and attitudes. Applied
electricity consumption measurements. Energy and Energy, 177, 692–702. https://doi.org/10.1016/j.
Buildings, 40(5), 926–936. https://doi.org/10.1016/j. apenergy.2016.04.075.
enbuild.2007.07.005. Jalas, M., & Juntunen, J. K. (2015). Energy intensive lifestyles:
Friedman, J., Hastie, T., & Tibshirani, R. (2010). Regularization time use, the activity patterns of consumers, and related
paths for generalized linear models via coordinate descent. energy demands in Finland. Ecological Economics, 113,
Journal of Statistical Software, 33(1), 1–22. 51–59. https://doi.org/10.1016/j.ecolecon.2015.02.016.
Jin, L., Lee, D., Sim, A., Borgeson, S., Wu, K., Spurlock, C. A., & ONS. (2017). Families and households: 2017. https://www.ons.
Todd, A. (2017). Comparison of clustering techniques for gov.uk/peoplepopulationandcommunity/birthsdeathsand
residential energy behavior using smart meter data (p. 7). ?marriages/families/bulletins/familiesandhouseholds/2017.
Presented at the thirty-first AAAI conference on Artifical Accessed 26 February 2019.
intelligence. ONS. (2018a). English housing survey: headline report,
Jones, R. V., & Lomas, K. J. (2016). Determinants of high electri- 2016–17. Ministry of Housing, Communities &
cal energy demand in UK homes: appliance ownership and L o c a l G o v e r n m e n t . h t t p s : / / w w w. g o v.
use. Energy and Buildings, 117, 71–82. https://doi. uk/government/uploads/system/uploads/attachment_
org/10.1016/j.enbuild.2016.02.020. data/file/675942/2016-17_EHS_Headline_Report.
Jones, R. V., Fuertes, A., & Lomas, K. J. (2015). The socio- pdf. Accessed 26 February 2019.
economic, dwelling and appliance related factors affecting ONS. (2018b). Percentage of households with durable goods, UK:
electricity consumption in domestic buildings. Renewable Table A45. https://www.ons.gov.uk/peoplepopulation
and Sustainable Energy Reviews, 43, 901–917. https://doi. ?andcommunity/personalandhouseholdfinances
org/10.1016/j.rser.2014.11.084. ?/expenditure/datasets/percentageofhouseholds
Kavousian, A., Rajagopal, R., & Fischer, M. (2013). Determinants ?withdurablegoodsuktablea45. Accessed 26 February 2019.
of residential electricity consumption: using smart meter data Palmer, J., Terry, N., Kane, T., Firth, S., Hughes, M., Pope, P., et al.
to examine the effect of climate, building characteristics, (2013). Further analysis of the household electricity use
appliance stock, and occupants’ behavior. Energy, 55, 184– survey - electrical appliances at home: tuning in to energy
194. https://doi.org/10.1016/j.energy.2013.03.086. saving (No. 475/09/2012). Cambridge Architectural
Lader, D., Short, S., & Gershuny, J. (2006). The time use survey, Research Limited, element energy, and Loughborough
2005 (p. 94). London: Office for National Statistics www. University. https://assets.publishing.service.gov.
timeuse.org/sites/ctur/files/public/ctur_report/1905/lader_ uk/government/uploads/system/uploads/attachment_
short_and_gershuny_2005_kight_diary.pdf. data/file/275484/electricity_survey_2_tuning_in_to_energy_
McKenna, E., & Thomson, M. (2016). High-resolution stochastic saving.pdf.
integrated thermal–electrical domestic demand model. Powells, G., Bulkeley, H., Bell, S., & Judson, E. (2014). Peak
Applied Energy, 165, 445–461. https://doi.org/10.1016/j. electricity demand and the flexibility of everyday life.
apenergy.2015.12.089. Geoforum, 55, 43–52. https://doi.org/10.1016/j.
geoforum.2014.04.014.
McKenna, E., Higginson, S., Grunewald, P., & Darby, S. J. (2017).
R Development Core Team. (2008). R: a language and environ-
Simulating residential demand response: improving socio-
ment for statistical computing. Vienna, Austria: R
technical assumptions in activity-based models of energy
Foundation for Statistical Computing www.R-project.org.
demand. Energy Efficiency, 11, 1583–1597. https://doi.
Rhodes, J. D., Cole, W. J., Upshaw, C. R., Edgar, T. F., & Webber,
org/10.1007/s12053-017-9525-4.
M. E. (2014). Clustering analysis of residential electricity
McLoughlin, F., Duffy, A., & Conlon, M. (2012). Characterising demand profiles. Applied Energy, 135, 461–471. https://doi.
domestic electricity consumption patterns by dwelling and org/10.1016/j.apenergy.2014.08.111.
occupant socio-economic variables: an Irish case study. Richardson, I., Thomson, M., Infield, D., & Clifford, C. (2010).
Energy and Buildings, 48, 240–248. https://doi.org/10.1016 Domestic electricity use: a high-resolution energy demand
/j.enbuild.2012.01.037. model. Energy and Buildings, 42(10), 1878–1887.
McLoughlin, F., Duffy, A., & Conlon, M. (2015). A clustering https://doi.org/10.1016/j.enbuild.2010.05.023.
approach to domestic electricity load profile characterisation Roberts, N., & Thatcher, J. (2009). Conceptualizing and testing
using smart metering data. Applied Energy, 141, 190–199. formative constructs: tutorial and annotated example. ACM
https://doi.org/10.1016/j.apenergy.2014.12.039. SIGMIS Database, 40(3), 9–39.
National Infrastructure Commission. (2016). Smart Power Satre-Meloy, A. (2019). Investigating structural and occupant
(National Infrastructure Commission Report). https://www. drivers of annual residential electricity consumption using
gov.uk/government/uploads/system/uploads/attachment_ regularization in regression models. Energy, 174, 148–168.
data/file/505218/IC_Energy_Report_web.pdf. Accessed 18 https://doi.org/10.1016/j.energy.2019.01.157.
March 2018. Sekar, A., Williams, E., & Chen, R. (2018). Changes in time use and
Ofgem. (2010). Demand side response. A discussion paper (p. 87). their effect on energy consumption in the United States. Joule.,
London: Office of Gas and Electricity Markets https://www. 2, 521–536. https://doi.org/10.1016/j.joule.2018.01.003.
ofgem.gov.uk/ofgem-publications/57026/dsr-150710.pdf. Stankovic, L., Stankovic, V., Liao, J., & Wilson, C. (2016).
Accessed 11 April 2019. Measuring the energy intensity of domestic activities from
Ofgem. (2017). Typical domestic consumption values for gas and smart meter data. Applied Energy, 183, 1565–1580.
electricity. London: Office of Gas and Electricity Markets https://doi.org/10.1016/j.apenergy.2016.09.087.
https://www.ofgem.gov.uk/system/files/docs/2017/08/tdcvs_ Strbac, G., Aunedi, M., Pudjianto, D., Djapic, P., Teng, F., Sturt,
?2017_open_letter.pdf . Accessed 20 February 2019. A., et al. (2012). Strategic assessment of the role and value of
ONS. (2016). Population estimates for UK, England and Wales, energy storage systems in the UK low carbon energy future.
Scotland and Northern Ireland: mid-2015. https://www.ons. London: Imperial College http://www.efn-uk.org/l-
gov.uk/peoplepopulationandcommunity/populationand street/politics-lib/gov-agents/index_files/ICLenergy-
m i g r a t i o n / p o p u l a t i o n e s t i m a t e s / b u ll e t i n s / a n n u a l storage2012.pdf . Accessed 24 February 2019.
midyearpopulationestimates/mid2015. Accessed 29 Swan, L. G., & Ugursal, V. I. (2009). Modeling of end-use energy
February 2019. consumption in the residential sector: a review of modeling
techniques. Renewable and Sustainable Energy Reviews, based on surveys and smart metering data. Energy, 107, 804–
13(8), 1819–1835. https://doi.org/10.1016/j.rser.2008.09.033. 817. https://doi.org/10.1016/j.energy.2016.04.065.
Tabbone, L., Ravalet, E., Durand-Daubin, M., Kaufmann, V., & Wickham, H. (2016). ggplot2: elegant graphics for data analysis.
Cayla, J.-M. (2016). Spatial location of activities and energy New York: Springer-Verlag http://ggplot2.org.
consumption of households in France. Presented at the Wickham, H., François, R., Henry, L., & Müller, K. (2019). dplyr:
DEMAND Centre Conference, Lancaster. http://www. a grammar of data manipulation. https://CRAN.R-project.
demand.ac.uk/wp-content/uploads/2016/05/DEMAND2016_ org/package=dplyr.
Full_paper_131-Tabbone-v2.pdf. Accessed 11 April 2019. Widén, J., & Wäckelgård, E. (2010). A high-resolution stochastic
Tibshirani, R. (1996). Regression shrinkage and selection via the model of domestic activity patterns and electricity demand.
lasso. Journal of the Royal Statistical Society: Series B: Applied Energy, 87(6), 1880–1892. https://doi.org/10.1016/j.
Methodological, 58, 267–288. apenergy.2009.11.006.
Tibshirani, R. (2011). Regression shrinkage and selection via the Widén, J., Lundh, M., Vassileva, I., Dahlquist, E., Ellegård, K., &
lasso: a retrospective. Journal of the Royal Statistical Society, Wäckelgård, E. (2009). Constructing load profiles for house-
Series B: Statistical Methodology, 73(3), 273–282. hold electricity and hot water from time-use data—modelling
Torriti, J. (2014). A review of time use models of residential approach and validation. Energy and Buildings, 41(7), 753–
electricity demand. Renewable and Sustainable Energy 768. https://doi.org/10.1016/j.enbuild.2009.02.013.
Reviews, 37, 265–272. https://doi.org/10.1016/j.
Wyatt, P. (2013). A dwelling-level investigation into the physical
rser.2014.05.034.
and socio-economic drivers of domestic energy consumption
Torriti, J., Hanna, R., Anderson, B., Yeboah, G., & Druckman, A.
in England. Energy Policy, 60, 540–549. https://doi.
(2015). Peak residential electricity demand and social prac-
org/10.1016/j.enpol.2013.05.037.
tices: deriving flexibility and greenhouse gas intensities from
time use and locational data. Indoor and Built Environment, Zou, H., & Hastie, T. (2005). Regularization and variable selection
24(7), 891–912. https://doi.org/10.1177/1420326 via the elastic net. Journal of the Royal Statistical Society,
X15600776. Series B: Statistical Methodology, 67(2), 301–320.
U.S. Bureau of Labor Statistics. (2018). American time use survey:
handbook of methods. Washington, D.C.: U.S. Department of Publisher’s note Springer Nature remains neutral with regard
Labor https://www.bls.gov/opub/hom/atus/pdf/atus.pdf . to jurisdictional claims in published maps and institutional
Accessed 26 February 2019. affiliations.
Viegas, J. L., Vieira, S. M., Melício, R., Mendes, V. M. F., & Sousa,
J. M. C. (2016). Classification of new electricity customers

Daily Life and Demand: An Analysis of Intra-Day Variations in Residential Electricity Consumption With Time-Use Data

Uploaded by

Copyright:

Available Formats

Daily Life and Demand: An Analysis of Intra-Day Variations in Residential Electricity Consumption With Time-Use Data

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Daily Life and Demand: An Analysis of Intra-Day Variations in Residential Electricity Consumption With Time-Use Data

Uploaded by

Copyright:

Available Formats

Energy Efficiency (2020) 13:433–458

Daily life and demand: an analysis of intra-day variations

Published online: 24 April 2019

The app pre-codes activities, following the Harmonized

Fig. 1 Example of an activity

Table 1 Selected socio-demographic, dwelling, appliance, and Table 1 (continued)

Time-use category Example activities Mean (SD)

Time-use category Time-use code Activity description Category mean (SD)

Care for home 3212 Vacuum cleaning 0.67 (1.02)

Fig. 3 Frequency of Benergy-

1) Daily average (W) 2) De-minned daily average (W)

B (5–95% CI) β B (5–95% CI) β

EV 0.61** (0.18, 1.04) 0.61

*p < 0.05; **p < 0.01; ***p < 0.001

*p < 0.05; **p < 0.01; ***p < 0.001

Discussion These results also, however, identify new variables that

(Anderson and Torriti 2018; De Lauretis et al. 2017; Jalas

ational activities, early mornings for personal care activities,

and early and mid-mornings for household chores.

denoting activities as Benergy-intensive^ can strengthen

reported and average de-minned electricity consumption

intensive^ activities. We do note, however, that these

including only energy-intensive activities. This finding

suggests other Bcare for home^ activities are important

for explaining electricity consumption during this time.

tricity consumption may be more complex than simply

differentiating activities on the basis of their expected

use of energy. Further discussion of this point can be

first is our finding that we do not observe an increase in

the number of activity variables selected between the

daily average and daily de-minned average models. We

do, however, observe frequent selection of activity var-

that this observation results from the duration of time

over which we are de-minning electricity consumption.

stronger associations between activities reported and

average consumption. Over shorter 4-h periods, howev-

er, de-minning does appear to strengthen relationships

between reported activities and consumption. This find-

ing suggests that activity data can improve models of

shorter duration more so than it can models of longer

5–9 p.m. 5–9 p.m. 9 p.m.–1 1 a.m.– 5 a.m.– 9 a.m.–1 1 p.m.–5

BAll^ activity models

BEnergy-intensive^ activity models

Table 8 Standardized regression coefficients for Table 5

*p < 0.05; **p < 0.01; ***p < 0.001

Table 9 Standardized regression coefficients for Table 6

*p < 0.05; **p < 0.01; ***p < 0.001

You might also like

p < 0.05; p < 0.01; p < 0.001

p < 0.05; p < 0.01; p < 0.001

p < 0.05; p < 0.01; p < 0.001

p < 0.05; p < 0.01; p < 0.001