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

net/publication/320148957

Short-term residential load forecasting: Impact of calendar effects and

forecast granularity

Article in Applied Energy · November 2017

DOI: 10.1016/j.apenergy.2017.07.114

CITATIONS READS

121 1,865

4 authors, including:

Peter Lusis Kaveh R. Khalilpour

Monash University (Australia) University of Technology Sydney (Australia)
8 PUBLICATIONS 125 CITATIONS 90 PUBLICATIONS 1,444 CITATIONS

SEE PROFILE SEE PROFILE

Ariel Liebman
Monash University (Australia)
66 PUBLICATIONS 297 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

PhD student opportunity: Renewable Hydrogen Supply Chain: Integrated multi-scale design of materials, process system, and market View project

Climate Change and its Effects on the Electricity Sector View project

All content following this page was uploaded by Ariel Liebman on 09 November 2017.

The user has requested enhancement of the downloaded file.

 Applied Energy 205 (2017) 654–669

Contents lists available at ScienceDirect

Applied Energy
journal homepage: www.elsevier.com/locate/apenergy

Short-term residential load forecasting: Impact of calendar effects and MARK

forecast granularity
⁎
Peter Lusisa,b, Kaveh Rajab Khalilpoura,c, , Lachlan Andrewa, Ariel Liebmana,c
a
Faculty of Information Technology, Monash University, Australia
b
Technical University of Denmark, Anker Engelunds Vej 1 Bygning 101A, 2800 Kgs. Lyngby, Denmark
c
Monash Energy Materials and Systems Institute (MEMSI), New Horizons Research Centre, Monash University, Australia

H I G H L I G H T S

• Investigation of the impact of data resolution and calendar effects.

• Support vector regression yields a higher accuracy for a day-ahead load forecast.
• The forecast error can be reduced by using coarser forecast granularity.
• Calendar effects added to the model as dummy variables have little predictive power.
• One year of historical data is sufficient to develop a load forecast model.

A R T I C L E I N F O A B S T R A C T

Keywords: Literature is rich in methodologies for “aggregated” load forecasting which has helped electricity network op-
Short-term load forecasting erators and retailers in optimal planning and scheduling. The recent increase in the uptake of distributed gen-
Residential load eration and storage systems has generated new demand for “disaggregated” load forecasting for a single-cus-
Calendar effects tomer or even down at an appliance level. Access to high resolution data from smart meters has enabled the
Granularity
research community to assess conventional load forecasting techniques and develop new forecasting strategies
Distributed generation and storage
suitable for demand-side disaggregated loads.
management
Disaggregated load This paper studies how calendar effects, forecasting granularity and the length of the training set affect the
accuracy of a day-ahead load forecast for residential customers. Root mean square error (RMSE) and normalized
RMSE were used as forecast error metrics. Regression trees, neural networks, and support vector regression
yielded similar average RMSE results, but statistical analysis showed that regression trees technique is sig-
nificantly better.
The use of historical load profiles with daily and weekly seasonality, combined with weather data, leaves the
explicit calendar effects a very low predictive power. In the setting studied here, it was shown that forecast errors
can be reduced by using a coarser forecast granularity. It was also found that one year of historical data is
sufficient to develop a load forecast model for residential customers as a further increase in training dataset has a
marginal benefit.

1. Introduction conventional power plants involved little uncertainty. Utilities focused

Electricity demand depends on several factors including weather,
time, and socio-economic constraints [1]. Load forecasting considers
these factors to facilitate the decision-making process of unit commit-
ment, economic dispatch, and power system operation [2]. At low
voltage level, demand forecasting improves optimal load control and
circuit switching [3]. The traditional centralized power generation from
on the statistical accuracy of a cluster of loads rather than a single
household [4]. This is changing due to the transition to distributed
energy generation from intermittent energy sources, the decentraliza-
tion of the electricity market, and the rising number of demand side
control systems. The scale of management has been moved down to
microgrid level and single households [5]. Around 30% of the global
electricity demand is related to the power consumption in the

⁎
Corresponding author at: Faculty of Information Technology, Monash University, Australia.
E-mail addresses: lusis.peteris@gmail.com (P. Lusis), kr.khalilpour@monash.edu (K.R. Khalilpour), lachlan.andrew@monash.edu (L. Andrew),
ariel.liebman@monash.edu (A. Liebman).

http://dx.doi.org/10.1016/j.apenergy.2017.07.114
Received 28 March 2017; Received in revised form 29 June 2017; Accepted 26 July 2017
0306-2619/ © 2017 Elsevier Ltd. All rights reserved.
P. Lusis et al.
residential sector [6]. Photovoltaic (PV) systems are the most wide-
spread distributed generation technology in the residential sector by
far, reducing the demand reliance on conventional power plants and
providing the peak load shaving [7].
Rooftop PV systems in the residential sector require new operational
strategies to maximize households’ self-sufficiency and minimize the
negative effect of the afternoon PV feed-in spikes on the grid [8]. To
mitigate the adverse effects of load fluctuations and voltage instability,
as well as store excess PV output, the PV system can be combined with
battery storage [9]. The battery is charged when the surplus energy
from PV panels is available. When the battery is full and the local de-
mand is met, the surplus power is fed into the grid. Considering that this
process occurs in multiple households, a power feed-in peak occurs
requiring the grid operator to pursue load balancing procedures. As
there is a growing solar power generation, it is necessary to find a trade-
off between the amount of the solar energy fed into the grid and optimal
PV-battery system operation from the household perspective [8].
Households’ demand changes quickly as appliances for domestic
chores are turned on and off. While the system-wide load forecast takes
advantage of the load smoothing effect from multiple households, at the
individual household level rapid fluctuations cannot be avoided,
making the load prediction more challenging [10]. Until the recent
deployment of smart meters, there was a shortage of high resolution
data from individual households [3]. A rapid increase in distributed
generation and research in demand side management created a need for
disaggregated data with a high sample rate enabling research in
household load forecasting.
The scheduling of batteries is strongly influenced by errors in input
variables [11]. Whether the household PV-battery system management
is carried out locally or in the cloud environment by a third party, load
forecasting has an important role [5]. The forecast accuracy is even
more critical for single stand-alone systems or microgrids, where un-
foreseen, high-frequency irradiance fluctuations can result in severe
voltage fluctuations [12]. A better voltage control can be achieved by
including the battery state of charge in the microgrid operation model
as this allows the battery to absorb some of the forecast error [13].
PV-battery system scheduling for residential customers is optimized
with complex mathematical programming algorithms or relying on
more efficient geometrical methodology. Case studies for PV-battery
system optimization are often based on historical data rather than solar
irradiation and load forecasts [14]. Since household load is less pre-
dictable than the overall system forecast, excluding load forecast errors
from PV-battery management gives a perfect, but unrealizable solution.
This paper examines how various calendar effects, forecast granularity,
and forecasting strategies affect the load forecast accuracy with dif-
ferent techniques. The aim is to select a short-term load forecast model
that should be deployed in the optimization process of household dis-
tributed generation and storage systems, including PV-battery.
2. Progress in household load forecasting
The load of any single residential customer is less predictable than a
more aggregated load [3]. New forecasting approaches, which arise
from the smoothing effect of household load aggregation, have been
proposed. Humeau et al. [15] analyzed the load consumption of 782
households and found that the normalized forecast error decreased with
the growing number of households in the cluster. Gajowniczek et al.
[10] proposed a blind source separation approach for households with a
similar load pattern; an improved forecast was achieved by decom-
posing the original forecasts into a set of independent components and
classifying and eliminating some of the noise [10].
Historical load and weather data are at the heart of load forecast
models. Several studies have focused on developing new or advanced
features in order to improve the forecast model. Beccali et al. [16] in-
troduced the “humidex index” that accounts for heating and cooling
demand due to the thermal discomfort felt by household residents.
655
Applied Energy 205 (2017) 654–669
Soliman et al. [17] defined “wind chill index” for winter months based
on wind speed and air temperature. Taieb et al. [18] addressed weather
and electricity demand uncertainty by proposing probabilistic forecasts
based on quantile regression. Rodrigues et al. [19], in contrary, avoided
the uncertainty of weather forecast by using only household physical
and demographic data in combination with calendar effects. Sandels
et al. [20] generated realistic household electricity consumption pat-
terns combining behavioral models of residents with their electricity,
hot water and space heating usage. Javed et al. [21] obtained a higher
forecast accuracy when adding socio-economic factors such as the
number of occupants, the age of occupants and the hours of day spent at
home. Tascikaraoglu et al. [22] proposed a spatio-temporal approach
considering the correlation between the energy usage in the target
house and the houses surrounding it. Jain et al. [23] studied the ac-
curacy of a step-ahead forecast in residential buildings examining
multiple spatial levels within a building and various temporal granu-
larities.
It is a common practice to apply multiple forecasting techniques and
compare them against a perfect-forecast model, which is based on ob-
served load values [12]. This benchmark is used to assess the accuracy
of the proposed model. An alternative benchmark is a persistence model
that takes an advantage of the fact that the load remains relatively
constant for a short period. Therefore, it is often difficult to beat a
persistence model in a short-term.
Traditional load forecasting techniques are based on time series or
regression analysis. Time series models such as exponential smoothing
[21], autoregrssive integrated moving average (ARIMA) models [24],
and seasonal ARIMA [22,25], largely rely on correlation between the
load and its past values. Other traditional load forecasting techniques
include statistical methods such as regression trees [26], and multiple
linear regression [5,22]. To tackle the non-linear and highly dynamic
load fluctuations of residential customers, artificial intelligence tech-
niques have become popular in load forecasting [27]. The main tech-
niques include neural networks (NN) [16,19,21] and support vector
machines (SVM) [15,23,27–29]. While artificial intelligence model
tends to provide slightly better forecast [30], this comes at a cost of
longer computational times. The optimal number of layers and neurons
in neural network model has to be determined empirically [25]. A more
accurate solar generation forecast model was achieved by [31], when
combining several forecasting algorithms into a single forecasting
technique. Stephen et al. [32] demonstrated that a forecast ensemble
consisting of neural networks, Gaussian load profile, ARIMA, persistent
and flat forecast models significantly outperforms the forecast model
built on a single forecasting technique for aggregated load forecasting.
Liu et al. [33] proposed a hybrid short-term load forecasting model with
parameter optimization and focused on the model’s implementation in
the microgrid management.
Calendar effects comprise any changes in load consumption related
to calendar periods. The use of calendar effects in load forecasting
captures weekly and seasonal energy consumption patterns [34] and
facilitates the prediction of peak demand [35]. Fewer studies have fo-
cused on the interaction between the residential load and calendar
variables. Various approaches can be found in the implementation of
calendar features into forecasting models. To capture the similarities in
load variation in different time periods, dummy explanatory variables
are used to represent each day of week [29], split weekdays and
weekends or assign variables for different parts of day [5]. Instead of
using dummy variables, the historical data can be split into subsets
according to the same day of week [36] or by certain hours of day. It is
not clear which of two approaches will yield a higher forecast accuracy.
Similarly, seasonal variability (winter-summer) can be addressed
through any of these approaches.
Past studies have built forecasting models using datasets from
60 days [30] to two years [25]. A common approach is to use all
available data or, if the dataset is incomplete, select the longest period
with complete information to build a forecasting model [36]. As the
P. Lusis et al. Applied Energy 205 (2017) 654–669

datasets grow bigger over time it may be questioned if all data should
be used for the model. This is particularly important when using fore-
cast models based on the relationship between independent and de-
pendent variables. Additional data that does not improve the forecast
should be disregarded to reduce required computational time and
power. More importantly, it should be disregarded because it can
mislead the model and actually degrade performance.
Short-term load forecasting models are typically trained and tested
at 30-min [37] or 60-min forecast intervals [19]. High resolution data
from digital interval meters also provides an opportunity to develop
forecast models at shorter periods [8,38]. To minimize frequency and
voltage fluctuations in individual buildings, minute by minute demand
forecast would be beneficial [37]. To author‘s knowledge, only Veit
et al. [36] have considered varying forecast granularities in order to
compare the performance of different forecasting techniques and stra-
tegies.
The forecast is usually provided for 24-h as a day-ahead forecast or
hour-by-hour. The latter assumes that the last observation is always
available before providing the forecast for the next hour. The choice of
a 24-h horizon could be related to the operation of the day-ahead
wholesale electricity market. For a residential customer with a PV-
battery system, a forecast for the next day is equally important in de-
cision-making process of battery operation. Finally, several authors
express concerns that traditional error evaluation approaches, such as
mean absolute percentage error (MAPE), are not suitable in household
load forecasting due to a high error at values close to zero or scaling
issues with due to different load levels among households. In order to
achieve a better understanding of results multiple error evaluation
metrics should be applied [15] or an accuracy measure can be used
instead of statistical evaluation methods [30]. In acknowledgement of
the growing importance of household load forecast and the progress
made in the existing studies, this paper:

1. presents the impact of intra-day, weekly, and seasonal calendar ef-

fects on household load forecasting accuracy considering efficient
implementation strategies and various combinations of calendar
predictors.
2. studies how the length of the training dataset affects the accuracy of
multiple linear regression, regression trees, neural networks and
support vector regression techniques (explained in Section 3.5).
3. evaluates how the forecasting accuracy depends on the granularity
used to generate the forecast.
3. Methodology
This section explains the components of a forecast model and de-
scribes the feature selection for load forecasting scenarios. The load
predictors are categorized as weather features, historical load features,
and calendar effects. The last part of this section introduces the fore-
casting techniques used in the study. Home electricity consumption
data for 27 households in New South Wales were acquired from an
Australian distribution system operator (Ausgrid). The dataset includes
30-min interval data for a period between July 2010 and June 2013.
The households are located within Greater Sydney. Fig. 1 presents a
methodology used to construct a load forecasting model. Two funda-
mental elements are historical load and weather data. Various calendar
effects are added to capture daily, weekly, and seasonal variations of
load pattern. Other factors, such as socio-demographic data, type of
buildings and environmental conditions can further improve the pre-
dictive power. Due to the lack of relevant data for the target house-
holds, other factors are out of the scope of this study.
Fig. 1. The sequence of considerations in building a household load forecast model.
3.1. Weather
The weather effect on household load can be significant [16]. In this
study, forecast models are trained and tested using measured weather
data instead of weather forecasts. The Australian Bureau of Meteor-
ology provided recorded half-hourly weather data from the Ob-
servatory Hill and Sydney Airport meteorological stations for the cor-
responding period. Five features were added to forecast models
including the dry-bulb temperature (Td ), the dew point (Tp ), and pre-
cipitation rate (mm). To reflect the cooling factor of wind speed during
lower temperatures in winter, a wind chill index W, introduced by [17],
was added:
W = (16.5−Td )·V 1/2 (1)
where V is the wind speed in km/h. For the given households, 16.5°C
was empirically found to represent the comfort temperature above
which the wind chill index has no effect on the load consumption. The
final feature is the humidity index function H reflecting an increased
operation of air conditioning during hot and humid summer days [17]:
H = 0.55·Td + 0.25·Tp + 5.05 (2)
The humidity index is set to zero if the dry-bulb temperature is lower

656
P. Lusis et al.
than 22°C. The introduction of variables W and H in lieu of relative
humidity and wind parameters improves the forecasting ability [16].
Wet-bulb temperature, cloudiness index, air pressure, solar radiation
were also available but had a negligible effect. To keep forecasting
models as parsimonious as possible, these variables were excluded.
3.2. Historical load
The periodic nature of load makes the past values of load the most
important factor in short-term load forecasting [1]. The aim is to pro-
vide a day-ahead forecast with 30-min forecast intervals yt + 1,…,yt + 48 ,
made at time t. The following load features were incorporated in the
training process of the forecast model:
• L1 - load value from the same half-hour of the previous day of the
same day type. Hence, yt − 48 for Tue, Wed, Thu, Fri, and Sun. yt − 144
for Mon and yt − 288 for Sat;
• L2 - the average of the same day and the same half-hour of the
previous three weeks (yt − 336 + yt − 672 + yt − 1008 )/3;
• L3 - the average of the same half-hour of the previous seven days
(yt − 48 + ⋯ + yt − 336 )/7 ;
• L4 - the average load of the previous day (yt − 1 + ⋯ + yt − 48 )/48.
The correlation between the previous few load observations is even
stronger [19]. However, for a day-ahead load forecast the information
is available only until hour yt . Therefore, the last load observation can
be used only for hour yt + 1, but not for the whole 24-h forecast period.
Fig. 15 in the appendix shows the load information for each household
including the mininum and maximum 30-min load.
3.3. Calendar effects
The load consumption is related to the time of the day, the day of
week, and the month of year. The latter represents the seasonal load
variations. The availability of 52,608 meter readings (3 years) offers
two options for how calendar effects can be added to the forecast
models. Dummy variables can be introduced to account for each season,
month, week, or day. An alternative is to split the dataset into subsets
according to similar calendar periods. For example, a forecast for
Sunday would be produced using only data from all previous Sundays
[36].
Fig. 2a illustrates the household load observations for each month
over the length of the study. Three seasonal periods with similar load
pattern were distinguished: a summer period with high power con-
sumption (Nov–Feb); a winter period with a lower median load than
Fig. 2. (a) Half-hourly load observations of target households for each month of the year. (b) Average half-hourly load of target households during weekdays, weekends and holidays.
657
Applied Energy 205 (2017) 654–669
summer, but a higher variance (May–Aug) and the remaining months as
a transition period between seasons. This suggests either adding a
dummy for each group or split the original dataset into seasonal sub-
sets. A distinction was also made between weekdays and weekends by
adding a dummy variable for each of those periods. The alternative
approach considers developing separate forecast models for weekdays
and weekends based on data for the corresponding parts of the week.
Based on (Fig. 2b), all 32 National and State holidays over the three
year period were treated as weekend days. Capturing intraday varia-
tions could further improve the forecast model. Although the intra-day
power consumption variation pattern varies among households, four
periods were defined: 2am–7am, 7am–10am, 10am–4pm, 6pm–10pm.
The remaining hours were combined with another dummy variable.
Combining the weather and load variables with day (5), week (2)
and seasonal (3) dummy variables, the forecast model has a maximum
of 19 load predictors. The dummy variables are added using binary
values. For example, the weekday/weekend dummy variable X can be
defined as follows:
⎧ X = 1, if the observation is on a weekend day
⎩ X = 0, if the observation is not on a weekend day
⎨
To analyze the role of calendar effects in the forecasting model, nine
forecasting scenarios were initially developed (Fig. 3). Scenario S1 in-
cludes all calendar effects and will be referred to as the reference sce-
nario. S2 uses intraday variables only, S3 considers weekdays/week-
ends only, and S4 uses seasons of the year. Scenario S5 excludes
calendar effects. In scenarios S6 to S9, the same calendar effects as in
S1-S4 are applied, but the forecast for every day in the test period is
based on subsets. The forecast scenarios are built using data with the
same calendar periods. For example, to predict a load consumption for
Saturday or Sunday, only data from all previous weekends are used to
build the forecasting model. The load variables L1 to L4 relies on the
order of days. Thus, scenarios with weekday/weekend subsets, such as
S8, included only load variables L2, L3 and L4.
3.4. Length of the dataset and forecast granularity
Granularity refers to the length of each forecast interval. By evalu-
ating the forecast model at different sampling intervals it is possible to
study the effect of granularity on the forecast accuracy. The granularity
cannot be shorter than the rate at which historical load and weather
data are sampled, a 30-min interval in this case. Halving the forecasting
interval doubles the amount of data for training and increases the re-
quired computation power. If a finer granularity does not improve the
P. Lusis et al.
Fig. 3. The list of forecast scenarios considering the time of day, the part of week, and seasons. NCE denotes scenarios with no calendar effects. The scenarios left from NCE illustrate
calendar effects added as dummy variables. The scenarios right from NCE divide the training data into subsets according to the calendar period the forecast is prepared for.
forecasting ability and is not required for load control, longer fore-
casting intervals are preferred. 30-, 60-, 120-min forecast intervals were
considered in this paper creating scenarios 1–27 (Fig. 3).
A longer training dataset provides more information to the fore-
casting algorithm. At the same time, the accuracy can be affected if the
historical data significantly differs from the most recent periods. For
example, if a household has recently switched form a gas to electric
heating system, the load consumption pattern will significantly change
and historical data lose relevance. The length of the dataset was varied
between one, two, and three years shown in scenarios 28–81 (Fig. 3).
3.5. Forecasting techniques
Stepwise multiple linear regression (MLR) is one the simplest algo-
rithms to assess the relationship between the load yt at time t and
predictor variables x i . It is given by:
k also known as SARIMAX, was considered. However, the initial results
yt = ∊t + β0 + ∑ wi βi x i,t
i=1
where wi (i: 1,…,k ) indicate the binary weights, and βi (i: 1,…,k ) denote
non-zero regression parameters. Features are added to the model based
on their statistical significance in a regression. Weights w are estimated
with a null hypothesis, which states that the relationship between the
next predictor and load forecast is no better than the one expected by a
chance. Regression analysis is carried out using F-statistic and p-value.
The first shows if a group of features is jointly significant. P-value tells if
each of the individual features and the load observation is statistically
significant. P and F values below 0.05 allow us to assert that there is at
least 95% chance that the relationship is not random. If there is enough
evidence to reject the null hypothesis, then the input feature is added to
the regression model. The predictors are added one by one based on the
lowest p-value.
Fig. 4. Neural network training and forecast model.
Applied Energy 205 (2017) 654–669
Regression trees are built using the standard Classification and
Regression Trees (CART) algorithm originally introduced by [39]. This
method uses recursive partitioning that examines all possible binary
splits at each stage for a set of predictors. The best splitting point for
each predictor is the point that gives the lowest sum of square errors.
The overfitting of the training data for a single tree was avoided by
creating a regression tree ensemble using bootstrap sampling with re-
placement and using the mean of the ensemble predictions as the load
forecast. An ensemble of 60 individual regression trees was found as an
optimum. A further increase in the number of trees did not reduce the
mean square error. The size of each tree in the ensemble was de-
termined by finding a trade-off between the cross-validation error and
the number of leaf nodes. Cross validation error was minimized by
adjusting the minimum number of observations per leaf between 5 and
30 data points. This process was repeated for each household.
Various time series models are available for load forecasting.
Seasonal ARIMA(p,d,q)- (P,D,Q)48 model with exogenous variables,
showed that SARIMAX model performed significantly worse than other
(3)
forecasting techniques for the household load forecast. As a time-series
model, the strength of SARIMAX is in the ability to learn from the
correlation between the present load and its historical observations.
However, SARIMAX is not able to react fast enough to rapid load
changes that deviate from the historical pattern leading to a larger
forecast error. Analyzing predictor importance, the incorporation of
weather variables did not improve the model. This problem has been
addressed by [40] who proposed HP Filter-SARIMA-Revised ARMAX
with regression analysis. Due to the fact that the errors were 1.5 times
as large as the next worse algorithm, SARIMAX model was excluded
from the further study.
A neural network (NN) is a multistage non-linear filter. Each stage, or
“layer”, after the input layer consists of nodes that compute a linear (or,
strictly, affine) combination of their inputs. The output of the node is a
658
P. Lusis et al.
hyperbolic tangent sigmoid “activation” function of this linear combi-
nation. We consider a network consisting of one “hidden” layer and a
single-node output layer. Training the network consists of learning the
weights vj,i given by each hidden layer node j to each regressor input x i ,
and the weight wj given by the output node to the output of each hidden
layer node. bj and bt are bias parameters for the input and hidden layers,
respectively. Manual design choices include the activation function fsig
and the number J of hidden nodes to use. With this notation, the pre-
dicted load is
J n
⎛ ⎞
yt ̂ = ∑ wj·fsig ⎜∑ vj,i x i + bj ⎟ + bt
j=1 ⎝ i=1 ⎠ (4)
After testing a wide range of values, the lowest mean square error
was obtained with J equal to 24. Initial weights were assigned ran-
domly, and trained using Levenberg-Marquardt backpropagation from
the Matlab Neural Network Toolbox. The training process was carried
out using a purely feed-forward model using the previously described
feature set as well as last load observations (Fig. 4). Instead of using a
random sampling that could destroy the correlation between the cur-
rent load and the lagged load, we use one week before each testing
period as the NN validation set. Since the actual values are not available
for a day-ahead recurrent neural network model was used in fore-
casting. This model feeds back the predicted load before providing the
forecast for the next interval. The weight and bias parameters obtained
in the training process were applied to the recurrent forecast model.
The forecast was made recursively for each 24-h period in the test.
Support vector machines. The main concept behind classical support
vector regression (SVR) is to find the decision function yt ̂ that gives as flat
as possible decision boundary. It is given by:
yt ̂ = f (x ) = 〈w,ϕ (x ) 〉 + b, (5)
where w denotes the weight vector controlling the smoothness of the
model and b is bias illustrating the separation between the hyperplane
and the origin (0,0). ϕ (x ) denotes a features mapping that converts the
features x into a high-dimensional feature space. The mapping is re-
quired to convert the problem from nonlinear regression case to a linear
regression and, thus allow us to solve it using a quadratic programming
formulation [28]. Parameters w and b were found by minimizing the
regularized risk function R (Eq. 6), which consists of the regularization
term and the empirical risk [41]:
1
n
minimizeR 2
||w||2 + C ∑ (ζi + ζi∗)
w,b,ζ ,ζ ∗
i=1
subject to yi −〈w,x i 〉−b ⩽ ∊ + ζi i = 1,…,n
〈w,x i 〉 + b−yi ⩽ ∊ + ζi∗ i = 1,…,n
ζi,ζi∗ ⩾ 0 i = 1,…,n
(6)
where Vapnik’s ∊-insensitive loss function ∊ is estimated as one-tenth of
the standard deviation using the interquartile range IQR of the load
observations. The forecast values within the tube have an error of zero,
whereas the values outside the tube have an error ζ equal to the dis-
tance between the data point and the tube wall [28]. The empirical risk
term uses slack variables ζ and ζ ∗ to capture the residuals outside the
tube and ensure that all constraints of the objective function R are met.
This provides confidence that the data samples are trained correctly
[41]. By minimizing the regularization term, we maximize the geo-
metric margin between support vectors to reduce the complexity of the
function and increase flatness. Constant C helps to prevent overfitting
659
Applied Energy 205 (2017) 654–669
Fig. 5. The training and testing datasets.
by giving the trade-off between the regularization term and empirical
risk; C is estimated as IQR/1.349.
The optimal regression f(x) is shown in Eq. (7), where K (x i,x j ) is a
Gaussian kernel function and l is the number of support vectors. It was
estimated from the dual optimization problem with non-negative La-
grange multipliers αi and αi∗. The optimal solution of α gives data points
x i that are the closest to the epsilon margin, hence, support vectors.
Features mapping uses kernel function to convert the problem from
nonlinear regression case to a linear regression. The full optimization
process is described in [42].
l
yt ̂ = f (x ) = ∑ (αi−αi∗) K (x i,x j ) + b
i,j = 1 (7)
3.6. Forecast evaluation
All four forecasting techniques were applied for the 81 scenarios of
Fig. 3. The forecast models were trained for each household separately
giving them unique neural network models and regression coefficients.
All forecasting techniques and all scenarios were tested on data com-
prising 28 days. In order to assess the model performance over different
seasons and diversify the possible load patterns the test dataset com-
prised a randomly selected week in June 2013 (Week 1), March 2013
(Week 2), December 2012 (Week 3), September 2012 (Week 4) (Fig. 5).
The data not included in the test set comprised the data for the training
process, including the data samples after weeks 1,2, and 3. This means
that to produce forecast for week 1, we have used the data that in
practice would not yet be available.
The test on the mean, maximum load, and the variance for the same
months of each year showed that a stable seasonal pattern is present for
24 of 27 households. Since we estimate the average error for all
households, the process can be considered stationary in the sense that
the load remains constant for the same months in different years. For
neural networks, seven days before each test week was selected for
model validation. Forecasts were made at midnight and the forecast
model was applied for each 24-h period in the test set.
Forecast evaluation methods focus on error calculation. In this
study, two error metrics are used. One method is normalized root mean
square error (NRMSE)
2
1 n y −y ̂
NRMSE = · ∑ ⎛⎜ t t ⎞⎟
n t=1
⎝ yt ⎠ (8)
where yt is a load observation at hour t, yt ̂ is the forecast of yt , and n is
the number of observations. Similar to MAPE, NRMSE is sensitive when
the observed load is close to or equal to zero resulting in a very large
error. For instance, if the observed load is 0.05 kWh, but the forecast
was 0.1 kWh, the error is 100%, while the absolute error is relatively
small. This can be used to assess the model performance during the
periods with typically low power consumption, hence, the night or
noon hours. An alternative measure is the root mean square error
(RMSE) given by:
1 n
RMSE = · ∑ (yt −yt )̂ 2 (kWh)
n t=1 (9)
If the interval size is doubled, the RMSE is also expected to double.
Thus, RMSE is better in evaluating the model accuracy during morning
and evening peak periods. In order to compare the forecast accuracy
P. Lusis et al.
Fig. 6. Forecast model development.
considering different granularities, RMSE is expressed in kWh. There
are substantial differences in the mean load and variance among target
households. Consequently, a forecast error for a household with a
higher load consumption will contribute more to the average RMSE,
Fig. 7. The top figure shows the 30-min mean load, maximum load and the RMSE for target households. The average RMSE is given in the legend. The bottom figure illustrates the
NRMSE of target households with the average RMSE given in the legend.
660
Applied Energy 205 (2017) 654–669
making it a potentially misleading indicator of forecast performance. If
of importance, this could be resolved by dividing the final error for each
household by their total load over the training period. In practice, the
best forecast model is expected to yield a low RMSE and NRMSE con-
currently. Fig. 6 illustrates the modeling steps for the forecast strategy
that could be used to reproduce our proposed methodology.
4. Results
This section will consider 30 scenarios which illustrate all of the
important features observed in the full set of results. Fig. 7 illustrates
RMSE and NRMSE for the reference model S1 for each household over
the forecast period. The values in the legends are the average RMSE and
NRMSE scores for each forecasting technique over all target house-
holds. The mean and maximum load of each household are also given. A
variation of the RMSE between 0.2 kWh and 1.1 kWh can be observed
(Fig. 7). A high RMSE for household H26 is largely due to a higher daily
peak load than other households. Fig. 8 illustrates household H26 load
forecast and observed values for a randomly selected four-day period.
All forecasting techniques fail to capture some of the peaks or mis-
predict the occurrence of peak load. H2, in contrast, has a lower peak
load and mean load resulting in the RMSE of 0.4 kWh and the NRMSE
value of 1.8. The average correlation coefficient between RMSE and
peak load, and between RMSE and the mean load for all households are
0.781 and 0.684 respectively.
The NRMSE facilitates understanding of the forecast accuracy as it
gives the magnitude of the difference between observations and fore-
cast values. Households H4 and H14 have large NRMSE values.
Remembering the drawback of NRMSE at low load values, this suggests
that models tend to overpredict the load during night. The particularly
large error of H4 is caused by the occurrence of extremely low loads of
0.01 kWh on June 27 in the test period. Since the load tends to not fall
under 0.2 kWh for this household, the forecast models did not antici-
pate this occurrence, resulting in the large error. The average NRMSE
using MLR and neural network techniques is noticeably higher than
with other forecasting techniques. The left graph for household H26
clearly shows that MLR tends to overpredict the load at night hours.
When looking at NRMSE for neural networks, a weaker performance
P. Lusis et al.
Fig. 8. Left: 4-day forecast (Thursday to Sunday) for a household with the largest RMSE (H26). Right: 4-day forecast (Thursday to Sunday) for a household with low RMSE and NRMSE
(H2).
can be related to households with a low average load, such as H4, H10,
H14, H19 and H21.
4.1. Performance comparison of STLF techniques
The top row in Fig. 7 shows that all forecasting techniques yield a
similar RMSE, except MLR that performed worse. Similarly, MLR
yielded a higher NRMSE, while support vector regression had the
lowest NRMSE suggesting a high forecast accuracy during periods with
low power consumption. Fig. 9 illustrates the error distribution of each
forecasting technique. The bottom gives an insight into the distribution
of large errors. The error statistics are summarized in Table 1. SVR has a
higher frequency of very accurate load predictions than other techni-
ques, especially MLR. A larger standard deviation for MLR also reflects
the weaker performance of MLR. Comparing the load forecasts with
observed values, the average forecast error of SVR is slightly lower than
for other techniques when underpredicting the load. When each
Fig. 9. The error distribution of forecasting techniques (Scenario S1).
661
Applied Energy 205 (2017) 654–669
forecasting technique overpredicts the load, SVR has an average error
of −0.094 followed by regression trees with −0.125 kWh and neural
networks with −0.154 kWh. This is compared to the frequency of load
overpredictions and underpredctions. As it is extremely difficult to
predict the exact load value, a forecast with an error less than
± 0.04 kWh from the observed load is considered acceptable. The fre-
quencies of errors reassure that SVR performs better than other fore-
casting techniques. A higher number of SVR overpredictions can be
excused due to yielding a lower error than other forecasting techniques
for these cases.
A positive skewness reflects a large number of positive errors, but
positive errors reflect the underestimation of the load. This is clearly
seen in the bottom row in Fig. 9 where large positive errors dominate
negative errors. Kurtosis is a measure of how outlier-prone a distribu-
tion is, with a value of 3 being equivalent to a normal distribution. A
high kurtosis value can be interpreted as a failure to predict high load
occurrences. This can be also seen in the distribution of hourly errors
P. Lusis et al. Applied Energy 205 (2017) 654–669

Table 1
The statistics of error distribution (Scenario S1).

Multiple linear regression Regression trees Support vector regression Neural networks

Standard deviation 0.297 0.274 0.279 0.278

Mean underpredicted load (kWh) 0.219 0.207 0.191 0.200
Mean overpredicted load (kWh) −0.174 −0.125 −0.094 −0.154
Underpredictions (> 0.04 kWh) 19,880 17,383 11,078 19,166
Overpredictions (< −0.04 kWh) 9776 7703 10,697 7793
Acceptable ( ± 0.04 kWh) 6632 11,202 14,513 9329
Skewness 2.01 2.44 3.31 2.23
Kurtosis 16.81 21.91 25.93 20.69

Fig. 10. The distribution of absolute errors for each hour.

Fig. 11. RMSE scores for forecast scenarios.

662
P. Lusis et al.
Fig. 12. NRMSE scores for forecast scenarios.
for all techniques in Fig. 10. The forecast distribution interval is wider
during morning and evening hours when most peaks occur. Finally,
variance-stabilizing transformation using the log of observations for
each household showed that the errors for each time interval are dis-
tributed normally in 99% cases for SVR, 98.8% - Regression Trees,
97.7% for NN and 96.7% for MLR. We compared all techniques and did
not observe any systematic behavior.
The average RMSE and NRMSE of target households for different
forecast models over the entire testing period are presented in Figs. 11
and 12. For scenario S1, the regression trees yielded the RMSE of
0.516 kWh followed by SVR and NN with the RMSE of 0.53 kWh and
0.531 kWh respectively. MLR performed slightly worse with the RMSE
of 0.561 kWh. The RMSE diminished for all techniques with an in-
creasingly coarse forecast granularity. Different granularities do not
change the fact which forecast technique performs better. The lowest
NRMSE score was achieved by SVR in all scenarios followed by re-
gression trees and neural networks.
4.2. Forecasting granularity
All forecast models performed better when granularity is changed
from 30-min to 60-min to 120-min. This can be explained as being due
to the smoothing of load fluctuations. For example, a power intensive
electric oven may be used for 30 min in total each morning between 7
am and 9 am in a household with four students. At 30-min forecast
granularity, the time and magnitude of the peak will be different for
each 30-min interval. If the 120-min granularity is used instead, the
peak will be always observed at the same two-hour block. The forecast
model will capture this load and learn to predict it. With 30-min fore-
cast granularity, this peak event may occur in any of adjacent half-hour
intervals from day to day hindering the predictability of the peak. To
minimize the forecast error at 30-min granularity, all forecast techni-
ques will act in a conservative manner. Hence, the forecast models will
issue a forecast value close to the average between a peak occurring and
not occurring for that forecast period. This explains why 60 and 120-
min granularity gives better, but not much better results.
4.3. Length of training period
Little improvement in the forecast accuracy was achieved by
663
Applied Energy 205 (2017) 654–669
increasing the training dataset from one to three years suggesting that
the historical data for one year is sufficient for a household forecast
model. Comparing training sets of one and three years, the RMSE fell by
0.005 kWh for regression trees and 0.015 kWh for NN. This gap in-
creased at 60 and 120-min granularity. Regression trees are less able to
take an advantage of an increase in the training set. This is due to the
fact that regression trees in the ensemble are grown using bootstrap
aggregated sampling.
In contrast, MLR scored the lowest RMSE when using two years of
input data. This is partly due to the changes in the average 30-min
power consumption between the second and the third year. A lower
consumption was observed in households H4 (−0.18 kWh, −19%),
H22 (−0.16 kWh, −27%), H25 (−0.34 kWh, −24%), while perma-
nent changes in the daily load pattern was noticed in H16.
4.4. Calendar effects
The RMSE results show that calendar effects have a very low pre-
dictive power for the residential load forecast. In fact, none of the
forecast models with dummy variables showed a consistent forecast
improvement against the scenario without calendar effects. Assigning
dummy variables for weekdays and weekends as well as for different
seasons did not improve the forecast. The slightly lower error when
achieved including all calendar effects. The addition of all calendar
effects slightly improved NRMSE. Fig. 11, scenarios S9, S18, and S27
show that a forecast based on seasonal subsets has a higher RMSE for all
forecasting techniques and all granularities. Subsetting by the time of
day and the day of week yielded significantly worse results than sub-
setting by season. While joining consecutive input data into subsets
according to a similar type of calendar effects creates groups with more
distinguished load and weather characteristics, the irregularity of the
peak loads led to larger RMSE and NRMSE than when using a single
dataset.
4.5. Statistical analysis
We test the statistical significance of five forecast models on 27
different datasets that vary in length, forecast granularity, and forecast
technique. All but one include various calendar effects as dummy
variables. MLR results are excluded from the statistical analysis.
P. Lusis et al. Applied Energy 205 (2017) 654–669

Table 2
Dummy predictors coefficients for scenario S1 with and without the lagged-load predictors.

Weekday Weekend Night Morning Day Evening Off-hours Winter Summer Off-season L1 L2 L3 L4

w/o Lagged-load
MLR −0.003 0.011 −0.153 −0.049 −0.019 0.127 −0.034 0.002 0.067 −0.009 – – – –
Regression Trees 1.810 1.824 4.226 3.976 2.759 3.242 2.212 1.207 1.736 1.506 – – – –

w/Lagged load
MLR −0.006 0.005 −0.046 −0.018 −0.014 0.036 −0.013 0.001 0.006 0.002 0.096 0.217 0.407 0.208
Regression Trees 1.533 1.603 1.426 1.288 1.787 1.359 1.292 1.017 1.144 1.140 2.963 3.027 3.296 4.050

Following Demsar [43], we use the Holm-Bonferroni multiple hypoth- techniques.

esis test, based on one-sided Wilcoxon signed-rank tests with approxi-
mately normal distribution to compare if any of four forecast models
with calendar effects (e.g. S1, S2, S3, S4 in Fig. 3) is significantly better
than the model without calendar effects (e.g. S5). The null hypotheses
state that there is no difference in the models’ performance. We test
each hypothesis at 0.05 alpha level against the alternative hypothesis
that a forecast model with calendar effects is significantly better. The
corresponding p-values are 0.0189, 0.9800, 0.9978, 0.9978. To control
the familywise error rate (FWER) related to multiple hypothesis tests,
the Holm-Bonferroni tests the p-values in increasing order, with the kth
(of n) compared against the reduced alpha value:
alphak = alpha/(n−rank + 1),
Since p1(0.0189) > alpha1 = 0.05/(4-1+1) = 0.0125, we cannot re-
ject the null hypothesis. The testing then stops and all subsequent hy-
potheses are non-significant. Thus, we cannot reject the null hypothesis,
and infer that in the given setup none of the forecast models including
calendar effects is significantly better than the forecast model without
calendar effects.
Regression trees (RT) also appear better than support vector re-
gression (SVR) and neural networks (NN). To test this, we compared the
RMSE of RT against the minimum of the RMSE of NNs and the RMSE of
SVR. This allows us to use a single (one-sided Wilcoxon) hypothesis test
to say that RT is the best; if we use Holm-Bonferroni, we could only
conclude that either NN or SVR is worse than RT. The null hypothesis is
that the median of the two data sets is the same, and the alternative is
that RT has a lower median RMSE. From nine datasets with all calendar
effects, we find a p-value of 0.002, thus the null hypothesis is rejected
and we concluded that RT is significantly better than the other
Table 2 gives the importance of predictors of calendar effects for
MLR and regression trees techniques considering the reference scenario
S1 and S1 with historical load information L1-L4 removed. The coeffi-
cients for calendar effects predictors are up to a factor of ten higher
when the historical load is not accounted for. This suggests that the
deployment of historical load variables extracts all calendar informa-
tion from data leaving low predictive power for dummy variables.
However, the exclusion of historical load data from the forecast model
makes the household load forecasting less accurate. Comparing load
features, it can be seen that MLR adds the most importance on the
average load of the same half-hour of the previous seven days L3, while
the average load of the previous day L4 has the highest coefficient for
(10)
regression trees.
4.6. Load aggregation
It was also examined if calendar effects have a larger predictive
power for a cluster of households. The historical loads for all target
households were aggregated together and used to train the forecast
models. The aggregated load has less variation than the single house-
hold load with more distinct load cycles. Fig. 13 illustrates the load
observations and forecast at 30-min intervals for a seven-day period in
winter for a randomly selected household and the cluster of all
households. Scenario S1 is presented (including all calendar effects).
The aggregated load figure shows that all forecasting techniques are
better in capturing the peak load and have a closer fit to the actual load.
The advantage of load aggregation can also be seen in error metrics
(Fig. 14). The RMSE is significantly lower than that of the forecasts for
individual households. Comparing scenario S1, aggregated load yielded

Fig. 13. Load forecast for a single household (top) and for a cluster including all target households (bottom). The date number shows the beginning of that day. June 29 and June 30
represents the weekend days.

664
P. Lusis et al.
Fig. 14. RMSE for aggregated load forecast (expressed as per household).
RMSE between 0.129 and 0.147 kWh for different techniques, while the
RMSE of individual households for the same scenarios is between 0.53
and 0.561 kWh. The load aggregation reduced the error per household
fourfold. The NRMSE vales for aggregated load forecast are between six
and seven times lower than for individual households. Again, all fore-
casting techniques yielded similar RMSE and NRMSE values except for
MLR. It had a larger NRMSE error due to overpredicting power con-
sumption at night. However, the aggregated load forecast did not show
a higher importance of calendar effects.
5. Discussion
Calendar effects are implemented in most load forecasting models. At
a utility scale, a lower power consumption during weekends can be
observed in the industrial and comsimercial sectors. Changes in heating
and cooling demand can be associated with seasonal changes. The re-
sidential sector experiences the peak in the evenings when people re-
turn home. Calendar variables are usually implemented in the same
manner regardless of the application. Since the forecast error is higher
for individual households, calendar effects are often seen as an addition
that improves the forecast. In contrast to expectations, the use of
dummy variables for certain calendar periods, outlined in this study,
added little or no predictive power to the forecast model. It may be that
improvement could be gained by substituting intra-day and seasonal
predictors with dummies for each hour of the day and each month of
the year. More variables for shorter calendar periods may provide a
smoother transition between any two hours or months respectively. It
was found that clustering significantly improved the forecast due to a
load smoothing effect rather than calendar effects. A case study could
be developed to observe if a further increase in the number of house-
holds in a cluster would increase the importance of calendar effects,
thus, contributing to a lower error.
Splitting the original dataset into subsets according to calendar
periods yields higher error than any model without calendar effects.
The mean and maximum load consumption for the same type of ca-
lendar periods is similar. However, there is still a large variation in load
pattern observed from day to day due to power consumption irregula-
rities making the model training process more challenging. To minimize
the forecast error all techniques tend to be conservative and under-
predict peak load to avoid a potential misprediction. With subsetting,
665
Applied Energy 205 (2017) 654–669
models become less robust to outliers and thus lead to a higher error. A
reduced amount of data in subsets can also adversely affect perfor-
mance. The results also showed that the current forecast models mis-
predicted the peak load. If accurate peaks in individual houses are of
interest, then new forecasting techniques must be tailored to that task.
5.1. Forecast granularity
As expected, the forecasting error reduced with coarsening forecast
granularity. This is an important finding for the management of a
household PV-battery system. A forecast at coarser granularity can be
combined with the time-of-use (ToU) tariff, which is financially more
attractive for households with PV-battery systems than the traditional
flat-fee electricity tariff. Whether the time-of-use (ToU) tariff is estab-
lished on the demand side [44] or the supply side [45], utilities remark
this as a tool to incentivize households to reduce or shift their peak
demand. Today most residential battery management systems do not
consider load and power generation forecasts. Instead, the decision of
when to charge and discharge battery is made according to the power
level in batteries. The battery control can be changed from the setpoint
state-of-charge (SOC) to a mode that considers the varying price levels
as well as the load and generation forecasts for each tariff block [46]. A
lower error can also minimize the risk to lose the control when the
battery is fully charged or discharged [13]. Thus, an improved forecast
accuracy from a lower granularity is adequate and desirable for max-
imizing the savings on energy bill and increasing the system reliability.
5.2. The amount of available data
The results showed that changing the length of the training dataset
between one and three years does not improve the forecast accuracy.
This suggests that one year of data is enough to develop a load forecast
model for a residential customer, but more data may be beneficial for
some households. For instance, if a household uses an air conditioner a
few times a year, its usage pattern may only become clear from multiple
years’ worth of data. Only MLR was sensitive to the changes in the
training dataset. It was found that in several households the mean and
peak load had changed after the first year. The simpler MLR algorithm
estimates predictor coefficients using all data points. Thus, the best
length of the training set with MLR should be determined manually.
P. Lusis et al. Applied Energy 205 (2017) 654–669

Correlation analysis and visualization of historical load can help to
choose the appropriate length of the dataset. Other techniques were
robust to historical data and had the capability to extract only sig-
nificant information, even if the load pattern in a household has
changed over time. The highest RMSE for MLR of 0.561 for 30-min
granularity versus the RMSE of 0.516 kWh for regression trees is ap-
proximately equal to a 50 watt TV always on. This is a large enough gap
to disregard MLR as an appropriate technique for household load
forecasting.
5.3. Computation time
Computational time of any forecast model is important for utilities
serving large numbers of customers. If the online training is used, where
the load observations from each day are added to the training dataset
and the model is retrained, computational time also plays a crucial role
[47]. When conducting a load forecast for multiple households, the
computational time depends on the number of households, forecasting
horizon, and the selected forecasting technique. In this study, parallel
computations greatly reduced the computational time. The load fore-
casting process in any household is not affected by the forecast for other
households but depends on the predictors only. This makes iterations
independent of each other and enables the distribution of households
among processor cores. The computational time is reduced pro-
portionally to the number of available workers. This was true for all
techniques except neural networks. It was not possible to carry out
parallel computing for neural networks. This is due to a fact that a
unique neural network with different weights and bias factors was
trained and applied for each household and for each scenario. Matlab
Neural Network Toolbox permits to train only one forecasting model at
the time. The computational time of MLR reference model was 14 s
versus 247 s for regression trees, 286 s for SVR, and 2015 s for neural
networks conducted on Intel i7-4712MQ 2.3 GHz processor with 12 GB
of RAM. While computational time is not in the focus of this paper, the
results show that model run-time may be an important factor when
choosing between forecasting techniques with a similar performance.
6. Conclusions
Calendar effects and historical load data are being deployed in most
residential load studies as a self-explanatory factor. This paper
demonstrated how various calendar effects and load forecast granu-
larity affect the accuracy of the day-ahead household load forecast
considering different forecasting techniques. We showed that the sub-
setting approach has a lower forecast accuracy as less training data
reduces the quality of forecast models. One might argue that the access
to a larger dataset might resolve this. The concern with multi-year data
is that households generally have dynamic life styles and over time the
population or consumption pattern may change. Then, adding socio-
demographic factors could help to address these changes. The calendar
effects added as binary variables in this study yielded a lower error than
the subsetting approach for most scenarios, yet statical analysis did not
highlight a significant difference between the scenarios with and
without calendar effects making calendar effects redundant. A much
larger number of households would result in a smoother load profile
and, thus, would make periodic load patterns (intra-day, weekly, sea-
sonal) more distinct and easier to predict.
In any forecasting model, only significant predictors should be used
to minimize the required computational power and time. While ca-
lendar effects do not require external data, this study shows that ca-
lendar effects become more important if the historical load is not
available. Increasing forecast intervals reduced the forecast error
showing that the 120-min granularity should be preferred for a
household battery management system with load forecasts. Through
statistical analysis it was found that regression trees technique performs
significantly better than the other techniques. Support vector regression
had the lowest NRMSE value showing a better overall ability to predict
the household load for the next 24 h. The scenario analysis also showed
that one year of historical load data can be sufficient to develop a load
forecast since a longer dataset did not improve the forecast accuracy. In
future, the knowledge from this paper will be used to study hybrid
forecasting techniques and applied to households with PV-battery
system optimization to evaluate the effect of load forecast on optimal
battery operation.
Acknowledgements
Special thanks go to Pierre Pinson from the Technical University of
Denmark and Souhaib Ben Taieb from Monash University for reviewing
the paper and providing feedback. Dr. Khalilpour and Dr. Liebman
would like to acknowledge the partial support of this project through a
funding from the Australia-Indonesia Centre.

Appendix A

Fig. 15 shows the 30-min mean (rounded) and maximum observed load as well as annual power consumption for each of the target households
over three years. Several households including H8 and H26 consume three times more electricity than others. Peak load varies between 1.493 and
5.55 kWh for a 30-min period. Figures are scaled between zero and the maximum load for the corresponding household. The red period represents
the period with the highest load. The high monthly loads in June and July represent increased consumption due to heating. Households H1, H3, H6
and H10 have relatively flat consumption pattern over the year with peaks in the summer. This suggests that these households have a gas heating
system and peaks occur due to the usage of air conditioners.
The day of week charts show a great variety of daily load consumption. In 14 households the peak consumption occurs on Saturday or Sunday.
The houses with weekday peaks have little in common and peaks can occur at any weekday. Households, such as H2, H14, and H27, have noticeable
higher power consumption during weekdays or weekends, but it is difficult to cluster other households by similar characteristics. The normalized
average daily load curve at 30-min intervals can be seen in the last column. In general, two peaks can be distinguished, a lower peak corresponding
to the morning hours and another, higher peak, corresponding to the evening hours. The duration of peaks and exact time of occurrence are not the
same. For H14, H15, and H16, the peak occurs around midnight, while for most households the evening peak can be observed between 7–9 pm.
There are also exceptions such as H8 with the highest power consumption during the daytime. This suggests that residents are home during the day.
H13, in contrary, has the highest power consumption after midnight and decreases during the daytime. The variety of load characteristics among
households illustrates how challenging is the household load forecasting. The lack of data from individual household appliances and information
about household demographics limits the information that can be extracted from the daily load pattern.
Figs. 16 and 17 illustrate the 30-min load profile for each household for the summer and winter months respectively. The load is expressed in
relative terms based on each household’s highest average 30-min load consumption. There is more distinct morning and evening peaks for more
households during winter. The increased color intensity in afternoon hours during summer corresponds to the usage of air conditioning.

666
P. Lusis et al. Applied Energy 205 (2017) 654–669

Fig. 15. Household load profiles for different calendar periods.

667
P. Lusis et al.
Fig. 16. Household 30-min mean load for summer months.
Fig. 17. Household 30-min load profile for winter months.
References
[1] Fahad MU, Arbab N. Factor affecting short term load forecasting. J Clean Energy
Technol 2014;2(4):305–309.. http://dx.doi.org/10.7763/JOCET.2014.V2.145.
[2] Wan C, Zhao J, Member S, Song Y. Photovoltaic and solar power forecasting for
smart grid energy management. J Power Energy Syst 2015;1(4):38–46. http://dx.
doi.org/10.17775/CSEEJPES.2015.00046.
[3] Hong T, Shahidehpour M. Load forecasting case study. Tech. rep., National
Association of Regulatory Utility Commisioners, Washington, DC, USA; 2015.
[4] McKenna E, Thomson M. High-resolution stochastic integrated thermal-electrical
domestic demand model. Appl Energy 2016;165:445–61. http://dx.doi.org/10.
1016/j.apenergy.2015.12.089.
[5] Iwafune Y, Yagita Y, Ikegami T, Ogimoto K. Short-term forecasting of residential
building load for distributed energy management. In: Energy conference
(ENERGYCON), 2014 IEEE international; 2014. p. 1197–1204. doi: http://dx.doi.
org/10.1109/ENERGYCON.2014.6850575.
[6] Daden. Immersive visual analytics 44, September; 2016.
[7] Bayod-Rújula AA. Future development of the electricity systems with distributed
668
Applied Energy 205 (2017) 654–669
generation. Energy 2009;34(3):377–83. http://dx.doi.org/10.1016/j.energy.2008.
12.008.
[8] Weniger J, Bergner J, Quaschning V. Integration of PV power and load forecasts into
the operation of residential PV battery systems. In: 4th Solar integration workshop;
2014. p. 383–90. doi: http://dx.doi.org/10.13140/2.1.3048.9283.
[9] Song J, Krishnamurthy V, Kwasinski A, Sharma R. Development of a Markov-Chain-
based energy storage model for power supply availability assessment of photo-
voltaic generation plants. IEEE Trans Sustain Energy 2013;4(2):491–500. http://dx.
doi.org/10.1109/TSTE.2012.2207135.
[10] Gajowniczek K, Zabkowski T, Szupiluk R. Blind source seperation for improved load
forecasting on individual household level. Adv Intell Syst Comput
2016;403:609–17. http://dx.doi.org/10.1007/978-3-319-26227-7.
[11] Michiorri A, Bossavy A, Kariniotakis G, Girard R. Impact of PV forecasts uncertainty
on batteries management in microgrids. PowerTech – towards carbon free society
through smarter grids. Grenoble: IEEE; 2013.
[12] Habib AH, Pecenak ZK, Disfani VR, Kleissl J, Callafon RAd. Reliability of dynamic
load scheduling with solar forecast scenarios, no. April. In: IEEE, systems con-
ference (SysCon); 2016. doi: http://dx.doi.org/10.1109/SYSCON.2016.7490593.
[13] Pascual J, Barricarte J, Sanchis P, Marroyo L. Energy management strategy for a
P. Lusis et al.
renewable-based residential microgrid with generation and demand forecasting.
Appl Energy 2015;158:12–25. http://dx.doi.org/10.1016/j.apenergy.2015.08.040.
[14] Zhong Q, Khalilpour R, Vassallo A, Sun Y. A logic-based geometrical model for the
next day operation of PV-battery systems. J Energy Storage 2016;7:181–94. http://
dx.doi.org/10.1016/j.est.2016.06.008.
[15] Humeau S, Wijaya TK, Vasirani M, Aberer K. Electricity load forecasting for re-
sidential customers: Exploiting aggregation and correlation between households.
2013 Sustainable internet and ICT for sustainability, SustainIT; 2013. doi: http://
dx.doi.org/10.1109/SustainIT.2013.6685208.
[16] Beccali M, Cellura M, Lo Brano V, Marvuglia A. Short-term prediction of household
electricity consumption: assessing weather sensitivity in a Mediterranean area.
Renew Sustain Energy Rev 2008;12(8):2040–65. http://dx.doi.org/10.1016/j.rser.
2007.04.010.
[17] Soliman SA-H, Al-Kandari AM. Electrical load forecasting: modeling and model
construction. Elsevier; 2010.
[18] Taieb SB, Huser R, Hyndman RJ, Genton MG. Forecasting uncertainty in electricity
smart meter data by boosting additive quantile regression. IEEE Trans Smart Grid
2016;7(5):2448–55. http://dx.doi.org/10.1109/TSG.2016.2527820.
[19] Rodrigues F, Cardeira C, Calado JMF. The daily and hourly energy consumption and
load forecasting using artificial neural network method: a case study using a set of
93 households in Portugal. Energy Proc 2014;62:220–9. http://dx.doi.org/10.
1016/j.egypro.2014.12.383.
[20] Sandels C, Widén J, Nordström L. Forecasting household consumer electricity load
profiles with a combined physical and behavioral approach. Appl Energy
2014;131:267–78. http://dx.doi.org/10.1016/j.apenergy.2014.06.048.
[21] Javed F, Arshad N, Wallin F, Vassileva I, Dahlquist E. Forecasting for demand re-
sponse in smart grids: an analysis on use of anthropologic and structural data and
short term multiple loads forecasting. Appl Energy 2012;96:150–60. http://dx.doi.
org/10.1016/j.apenergy.2012.02.027.
[22] Tascikaraoglu A, Sanandaji BM. Short-term residential electric load forecasting: a
compressive spatio-temporal approach. Energy Build 2016;111:380–92. http://dx.
doi.org/10.1016/j.enbuild.2015.11.068.
[23] Jain RK, Smith KM, Culligan PJ, Taylor JE. Forecasting energy consumption of
multi-family residential buildings using support vector regression: investigating the
impact of temporal and spatial monitoring granularity on performance accuracy.
Appl Energy 2014;123:168–78. http://dx.doi.org/10.1016/j.apenergy.2014.02.
057.
[24] Nataraja C, Gorawar MB, Shilpa GN, Harsha Jr S. Short term load forecasting using
time series analysis: a case study for Karnataka, India. Int J Eng Sci Innovat Technol
(IJESIT) 2012;1(2):45–53.
[25] Cao X, Dong S, Wu Z, Jing Y. A data-driven hybrid optimization model for short-
term residential load forecasting. In: IEEE int. conf. computer and information
technology/ubiquitous computing and communications/dependable, autonomic
and secure computing/pervasive intelligence and computing; 2015. doi:http://dx.
doi.org/10.1109/CIT/IUCC/DASC/PICOM.2015.41.
[26] Mori H, Kosemura N. Optimal regression tree based rule discovery for short-term
load forecasting. In: 2001 IEEE power engineering society winter meeting.
Conference proceedings (Cat. No.01CH37194) 2 (C); 2001. p. 421–6. doi:http://dx.
doi.org/10.1109/PESW.2001.916878.
[27] Hussain L, M SN, Ali Shah SA. Short term load forecasting system based on support
vector kernel methods. Int J Comput Sci Inform Technol 2014;6(3):93–102. http://
dx.doi.org/10.5121/ijcsit.2014.6308.
[28] Türkay BE, Demren D. Electrical load forecasting using support vector machines. In:
7th International conference on electrical and electronics engineering (ELECO
669
View publication stats
Applied Energy 205 (2017) 654–669
2011); 2011. p. 49–53.
[29] Ceperic E, Ceperic V, Baric A. A strategy for short-term load forecasting by support
vector regression machines. IEEE Trans Power Syst 2013;28(4):4356–64. http://dx.
doi.org/10.1109/TPWRS.2013.2269803.
[30] Gajowniczek K, Zabkowski T. Short term electricity forecasting using individual
smart meter data. Proc Comput Sci 2014;35(C):589–97. http://dx.doi.org/10.1016/
j.procs.2014.08.140.
[31] Samanta M, Srikanth BK, Yerrapragada JB. Short-term power forecasting of solar PV
systems using machine learning techniques; 2014. Accessed Online: [2017-02-12].
[32] Stephen B, Tang X, Harvey PR, Galloway S, Jennett KI. Incorporating practice
theory in sub-profile models for short term aggregated residential load forecasting.
IEEE Trans Smart Grid 2015;PP(99):1–8. http://dx.doi.org/10.1109/TSG.2015.
2493205.
[33] Liu N, Tang Q, Zhang J, Fan W, Liu J. A hybrid forecasting model with parameter
optimization for short-term load forecasting of micro-grids. Appl Energy
2014;129:336–45. http://dx.doi.org/10.1016/j.apenergy.2014.05.023.
[34] Chitsaz H, Shaker H, Zareipour H, Wood D, Amjady N. Short-term electricity load
forecasting of buildings in microgrids. Energy Build 2015;99(587):50–60. http://
dx.doi.org/10.1016/j.enbuild.2015.04.011.
[35] Sigauke C, Chikobvu D. Short-term peak electricity demand in South Africa. African
J Business Manage 2012;6(32):9243–9. http://dx.doi.org/10.5897/AJBM11.2607.
[36] Veit A, Goebel C, Tidke R, Doblander C, Jacobsen H-A. Household electricity de-
mand forecasting. Proceedings of the 5th international conference on future energy
systems – e-Energy ’14 Juni 2014:233–4. http://dx.doi.org/10.1145/2602044.
2602082.
[37] Wright A, Firth S. The nature of domestic electricity-loads and effects of time
averaging on statistics and on-site generation calculations. Appl Energy
2007;84(4):389–403. http://dx.doi.org/10.1016/j.apenergy.2006.09.008.
[38] Abdulla K, De Hoog J, Muenzel V, Suits F, Steer K, Wirth A, et al. Optimal operation
of energy storage systems considering forecasts and battery degradation. IEEE Trans
Smart Grid 2016;PP(99):1–11. http://dx.doi.org/10.1109/TSG.2016.2606490.
[39] Breiman L, Friedman J, Stone CJ, Olshen R. Calsiffication Regress Trees 1984.
[40] Azadeh A, Ghaderi SF, Sohrabkhani S. A simulated-based neural network algorithm
for forecasting electrical energy consumption in Iran. Energy Policy
2008;36(7):2637–44.
[41] Hu Z, Bao Y, Xiong T. Electricity load forecasting using support vector regression
with memetic algorithms. Sci World J 2013. http://dx.doi.org/10.1155/2013/
292575.
[42] Kavaklioglu K. Modeling and prediction of Turkey’s electricity consumption using
support vector regression. Appl Energy 2011;88(1):368–75. http://dx.doi.org/10.
1016/j.apenergy.2010.07.021.
[43] Demšar J. Statistical comparisons of classifiers over multiple data sets. J Machine
Learn Res 2006;7:1–30. http://dx.doi.org/10.1016/j.jecp.2010.03.005.
[44] Li R, Wang Z, Gu C, Li F, Wu H. A novel time-of-use tariff design based on Gaussian
mixture model. Appl Energy 2016;162:1530–6. http://dx.doi.org/10.1016/j.
apenergy.2015.02.063.
[45] Essential Services Commission Victoria. The energy value of distributed generation.
Tech. rep.; August 2016.
[46] Khalilpour KR, Vassallo A. Community energy networks with storage: modeling
frameworks for distributed generation. Springer; 2016.
[47] Cheng Q, Yao J, Wu H, Chen S, Liu C, Yao P. Short-term load forecasting with
weather component based on improved extreme learning machine. In: Chinese
automation congress; 2013. p. 316–21. doi:http://dx.doi.org/10.1109/CAC.2013.
6775750.