Renewable and Sustainable Energy Reviews 137 (2021) 110591

Contents lists available at ScienceDirect

Renewable and Sustainable Energy Reviews

journal homepage: http://www.elsevier.com/locate/rser

A deep learning framework for building energy consumption forecast

Nivethitha Somu a, *, Gauthama Raman M R b, Krithi Ramamritham a
a
Smart Energy Informatics Lab (SEIL), Department of Computer Science and Engineering, Indian Institute of Technology-Bombay, Mumbai, 400076, Maharashtra, India
b
iTrust-Centre for Research in Cyber Security, Singapore University of Technology and Design, Singapore

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

Keywords: Increasing global building energy demand, with the related economic and environmental impact, upsurges the
Energy consumption need for the design of reliable energy demand forecast models. This work presents kCNN-LSTM, a deep learning
Deep learning framework that operates on the energy consumption data recorded at predefined intervals to provide accurate
Buildings
building energy consumption forecasts. kCNN-LSTM employs (i) k − means clustering – to perform cluster
Clustering
Convolutional neural network
analysis to understand the energy consumption pattern/trend; (ii) Convolutional Neural Networks (CNN) – to
Long short term memory extract complex features with non-linear interactions that affect energy consumption; and (iii) Long Short Term
Memory (LSTM) neural networks – to handle long-term dependencies through modeling temporal information in
the time series data. The efficiency and applicability of kCNN-LSTM were demonstrated using a real time
building energy consumption data acquired from a four-storeyed building in IIT-Bombay, India. The performance
of kCNN-LSTM was compared with the k-means variant of the state-of-the-art energy demand forecast models in
terms of well-known quality metrics. It is also observed that the accurate energy demand forecast provided by
kCNN-LSTM due to its ability to learn the spatio-temporal dependencies in the energy consumption data makes it
a suitable deep learning model for energy consumption forecast problems.

consumption, energy costs, and environmental impact (carbon emis­

sions). According to the World Energy Council, “While overall per capita
1. Introduction energy demand would begin to fall, demand for electricity would double
by 2060”, which necessitates the need for larger investments in smart
1.1. Energy, buildings, and environment infrastructures that promote energy efficiency [11]. From the smart grid
perspective, buildings have become more intelligent with the integra­
A nation’s total energy demand can be estimated by aggregating tion of advanced information and communication technologies, electric
three main economic sectors: buildings, industry, and transport [1,2]. vehicles, decentralized generation and storage systems, and energy
According to World Watch Institute data, buildings as the largest energy management systems. Therefore, the design and implementation of
consumer, accounts for 40% of the global annual energy consumption smart technologies for power grids and buildings to meet the global
and 36% of the total carbon emissions, especially in urban areas [3,4]. energy demand in an effective and economically sustainable way with
The growth rate of building energy consumption in Organisation for reduced carbon emissions have become extremely important.
Economic Co-operation and Development (OECD) and non-OECD na­
tions for 2012 and 2040 is 1.5% and 2.1% per year, respectively [5].
Specifically, the total energy consumption of higher education institu­ 1.2. Energy consumption forecast
tional and commercial buildings are 45% and 30% higher than the
residential buildings [6]. The proper use of energy through the imple­ Recently, research on energy consumption forecast in buildings has
mentation of appropriate energy management systems and end-user become increasingly significant as buildings are equipped with smart
energy efficiency control strategies [7] provides lower operational meters to monitor energy consumption of buildings at fine-grained in­
costs by reducing energy use and avoiding penalties imposed by the tervals. Massive and high dimensional energy consumption data from
utilities [8–10]. In this regard, several nations have augmented the the smart meters collected at different granularities helps in under­
implementation of energy regulations and codes for buildings, such that standing the energy consumption patterns for its application to energy
new energy-efficient building designs ensure reduced energy demand forecasting, demand response, heating ventilation and air

* Corresponding author.
E-mail addresses: [email protected] (N. Somu), [email protected] (G. Raman M R), [email protected] (K. Ramamritham).

https://doi.org/10.1016/j.rser.2020.110591
Received 4 April 2020; Received in revised form 30 October 2020; Accepted 17 November 2020
Available online 14 December 2020
1364-0321/© 2020 Elsevier Ltd. All rights reserved.
N. Somu et al. Renewable and Sustainable Energy Reviews 137 (2021) 110591

List of abbreviations σ Logistic sigmoid function

xt Input
Terms Explanation ct Cell state
k CNN-LSTM k-means clustering based convolutional neural c˜t New candidate cell state
networks and long short term memory Cnij Output of the convolution operations in the convolutional
CNN Convolutional neural networks layer
LSTM Long short term memory ckn Convolution kernel matrix
KReSIT Kanwal Rekhi School of Information Technology fmij Convolution filter matrix
IIT Indian Institute of technology n nth feature map
MAE Mean absolute error i and j Number of steps of convolution filter in horizontal
MAPE Mean absolute percentage error and vertical directions
MSE Mean squared error bn and bfc Bias
RMSE Root mean squared error δ Activation function
OECD Organization for economic co-operation and development Pnij Output of the pooling operation in the pooling layer
HVAC Heating ventilation and air conditioning
pfmij Pooling filter matrix
BEMS Building energy management system
Wfc Weight matrix of the fully connected layer
ISCOA-LSTM Improved version of sine cosine optimization
algorithm based LSTM FC Output matrix of the fully connected layer
x Input to the fully connected layer
Et = {e1t , e2t , …, eit , …, ent } Energy consumed at different timestamps
E = {T1 , T2 , …, TM } Energy consumption dataset
MIMO Multi-input multi-output
C = {C1 , C2 , …, Ck } Clusters
InputL Input window
T1 and T2 Time series data
OutputL Output window
n Length of T1
if = (Sn − InputL − a)/OutputL Total number of input and forecast
m Length of T2
windows
ETrain Train dataset
Sn Total number of samples
EVal Validation dataset
a Forecast interval
ETest Test dataset
SInput and SOutput Input and the output window
MQTT Message queuing telemetry transport
InputL and OutputL Size of the input and output window
x Energy consumption value
SSE Sum of squared error
avg(x) Average value of x
LSTM Long short term memory
std(x) Standard deviation of x
RNN Recurrent neural networks
min(x) and max(x) Minimum and maximum values of x
it Input gate
n Total number of data points in the dataset
ot Output gate
yi and ̂y i Actual and the forecasted energy consumption value
ft Forget gate
at timestamp i
Wf , Wc , Wi , andWo Weight matrices
bf , bi , bc , andbo Bias vector

conditioning (HVAC) optimization, and fault diagnosis and detection. 1.3. Machine learning based energy forecasting
Accurate and reliable energy demand forecasts enable the utilities to
plan resources and balance supply-demand, thereby ensuring stability Energy consumption forecast models in the literature fall into three
and security of the power grid & reliability of service provisions [12]. categories: (i) Engineering methods: Uses physical and thermody­
Overestimation and underestimation of the energy demand lead to a namic laws & require complex building and environmental parameters;
severe impact on the economic and industrial developments [13]. Ac­ Difficult and time-consuming; Examples-EnergyPlus, Ecotect, etc [2].
curate modeling and prediction of energy demand help in efficient en­ (ii) Statistical methods: Correlates energy consumption with relevant
ergy management in smart buildings, accurate demand response factors like climate data, occupancy, etc.; Lacks accuracy and flexibility;
strategies, electricity supply management, and context aware control Examples-Time series (autoregressive models) [14] and regression
strategies [14]. Therefore, energy consumption forecast models have models (linear regression) [16,17], and (iii) Artificial intelligence
become an integral part of Building Energy Management System (BEMS) methods: Learns the consumption patterns from the historical energy
to improve the buildings’ energy efficiency for a sustainable economy consumption data, i.e., discover the non-linear relationship between the
through a conservation-minded society, reasonable use of available input (historical data) and output (target consumption) [18,19];
energy resources, and efficient national energy strategy [15]. However, Examples-artificial neural network [20], support vector regression [4],
the non-linear, dynamic, and complex nature of energy consumption etc. Among these, artificial intelligence approaches have become ‘active
data, along with the presence of trend, seasonal & irregular patterns, and research hotspot’ due to their efficiency and flexibility over engineering
dependence on various exogenous factors like climatic conditions, na­ and statistical methods (Table 1) [21,22].
ture of the day, socio-economic factors, etc. presents accurate and reli­ From Table 1, it is evident that ANN and its variants have been
able energy consumption forecast as an interesting research problem. widely explored and applied for energy consumption forecast since they
are non-linear, self-adaptable, and can approximate any function, given

2
N. Somu et al. Renewable and Sustainable Energy Reviews 137 (2021) 110591

Table 1 Table 1 (continued )

Related works. Authors Techniques Dataset Evaluation
Authors Techniques Dataset Evaluation metrics
metrics
Jatin Bedi • LSTM Electricity load RMSE,
Muhammad • Deep ELM 4 residential MAE, RMSE, and and Durga • MIMO data- correlation
Fayaz and • Adaptive neuro-fuzzy buildings, MAPE Toshniwal, Chandigarh, coefficient, and
DoHyeun inference system Seoul, South 2019 [35] India MAPE
Kim, 2018 • ANN Korea Yunsun Kim • ARIMA Institutional RMSE and MAPE
[23] et al., 2019 • ARIMA-GARCH building,
Sehrish Malik • ANN 4 residential Number of [36] • Holt–Winters’ double Chungang
and • PSO buildings, particles used, multiple seasonal University,
DoHyeun Seoul, South prediction exponential Seoul, Korea.
Kim, 2018 Korea accuracy, and smoothing
[24] number of epochs • ANN
Jatin Bedi • K-means Electricity load RMSE and MAPE Hanane • ANN Case study: 5 RMSE and MAPE
and Durga • Empirical mode data- Dagdougui • Learning algorithm: buildings
Toshniwal, decomposition Chandigarh, et al., 2019 Bayesian (residential,
2018 [25] • LSTM India [37] regularization and commercial,
Chengdong Li • DBN Retail store and MAE, RMSE, and Levenberg-Marquardt and office),
et al., 2018 • Contrastive office building, mean relative Urban
[26] divergence Fremont, CA, error educational
USA. district,
Jihoon Moon • PCA Private MAE, MAPE, and Montreal
et al., 2018 • ANN university in RMSE Dac-Khuong • Electromagnetism Single family Linear
[27] • SVR Seoul, South Bui et al., based firefly algorithm house-Istanbul, correlation
Korea 2020 [2] • ANN Turkey and coefficient, R2 ,
Samir Gradient boosting Northern and MSE and RMSE twelve RMSE, MAE, and
Touzani machine Central buildings MAPE
et al., 2018 California, simulated by
[28] Washington, D. Ecotect
C. and Seattle Lulu Wen Deep RNN with gated Residential RMSE, MAE,
K.P. Amber • Multiple regression South Bank RMSE, MAE, et al., 2020 recurrent unit building, Pearson
et al., 2018 • Genetic programming Technopark, MRE, MAPE, and [38] Austin, Texas, correlation
[6] • ANN London, U⋅K normalized U.S.A coefficient, and
• DNN RMSE MAPE
• SVM Duc-Hoc Tran Evolutionary neural Residential Coefficient of
Wei Wang • LSTM Southeast MAPE and RMSE et al., 2020 machine inference model building, Ho Chi determination,
et al., 2019 • ANN University, [19] - Least squares SVR + Minh City, Viet MAE, RMSE, and
[3] • SVR Nanjing, China RBF neural network Nam MAPE
Aowabin LSTM-based deep RNN Public Safety RMSE and
Rahman Building, Salt Pearson
et al., 2018 Lake coefficient the sufficient number of hidden layers and nodes in the hidden layer
[29] City, U.S. state [39]. However, the inability to handle historical data dependencies and
of Utah and
several factors like parameter initialization, trap at local minima, slow
Residential
buildings, convergence, and scalability of network architecture provoked the re­
Austin, Texas, searchers to actively work on these issues or explore deep learning ap­
U.S.A proaches for building energy consumption forecast. Further, a
Zhaoyang Ye Levenberg–Marquardt Shopping mall, RMSE considerable amount of information collected over various sensors
and Moon BPNN Dalian, China
Keun Kim,
deployed in buildings has transformed energy forecasting research into a
2018 [30] “big data” research problem [40]. Interestingly, recent advancements in
Tianhao • Sample data selection Case study: Coefficient of deep learning theory have resulted in efficient tools to handle massive
Yuanet al., method- grey Commercial determination, and high dimensional energy consumption data, which can outperform
2018 [31] correlation method + and residential absolute
traditional machine learning tools [41]. In this way, this work presents
entropy weight building, percentage error,
• BPNN Tianjin, China MAPE, and RMSE kCNN-LSTM, a deep learning framework which employs k − means
• Multiple linear clustering, CNN, and LSTM for trend analysis, energy-related feature
regression identification, and modeling long term dependencies in the energy
Federico Ensemble: Base learning Energy MRE, MAE, consumption data, respectively.
Divina model-Regression trees consumption symmetric
et al., 2018 based on evolutionary data, Spain MAPE, RMSE,
[32] algorithms, ANN, and and R2 1.4. Research gaps, novelty, and contributions
Random forest; Top
learning model: Gradient
boosting machine From the brief literature survey on the building energy consumption
Tanveer • Compact decision tree Office building, MAE, RMSE, and forecast models designed for residential, academic, and commercial
Ahmad and • Fit k-nearest classifier Beijing, China MAPE buildings, the following research gaps were identified.
Huanxin • Simple and stepwise
Chen, 2018 linear regression
[33] model 1. Most of the energy consumption data used in the literature are multi-
Shidrokh • ARIMA Library MAE, RMSE, and featured (occupancy, temperature, humidity, building schedule, etc.)
Goudarzi • SVR building, average relative and data-rich (massive and high dimensional data). The research
et al., 2019 • PSO Universiti error challenge in building energy demand forecasting is to achieve high
[34] • False nearest Teknologi
neighbours Malaysia
forecast accuracy using the energy consumption data recorded at
• K-means each timestamp.
2. The energy consumption in an academic building does not reveal the
trend and seasonality during the time series analysis. It might be due

3
N. Somu et al. Renewable and Sustainable Energy Reviews 137 (2021) 110591

to several factors like high dynamics in schedules, occupancy related 7. The effectiveness of kCNN-LSTM for reliable building energy con­
status, operations, etc. sumption forecast is validated through a case study using the real
3. Moreover, the existing energy forecasting models follow static time building operational data acquired from the BEMS deployed at
learning, where its performance is entirely dependent only on the KReSIT, IIT-Bombay. Further, the performance of KCNN-LSTM over
historical energy consumption data. However, the inclusion of recent the k − means variant of existing energy demand forecast models was
observation with the historical data with the help of a sliding win­ assessed in terms of MAE, MSE, MAPE, and RMSE for the considered
dow approach would result in better forecast accuracy. year, weekdays, and weekend.
4. Further, the existing research works on LSTM based building energy
consumption forecast model operates over static data (benchmark 1.5. Organization
datasets) instead of real time building operation data obtained from
the BEMS. The paper follows the following structure. Section 2 introduces the
basic principles of k − means, energy consumption forecasts, LSTM
What makes kCNN-LSTM different from the existing building energy neural networks, and convolutional neural networks. Section 3 presents
consumption models? The distinctive features of kCNN-LSTM are. a brief description of the formulation of the energy consumption forecast
problem and the proposed deep learning framework to forecast the en­
1. Feature generation from timestamp: The considered building en­ ergy consumption of the buildings. Section 4 provides a detailed analysis
ergy consumption data is a nxm dimensional data, where n denotes of the performance of kCNN-LSTM and state-of-the-art energy demand
the rows (energy consumption records) and m represents the col­ forecast models in terms of MAE, MSE, MAPE, and RMSE. Section 5
umns (m = 2; timestamp in dd-mm-yyyy HH: MM: SS format and concludes the paper.
energy consumption). As a data preprocessing step, seven features
(day of the year, season, month, day of the week, hour of the day, 2. Preliminaries
minute of the hour, type of the day) were generated from the time­
stamp, which enables the learning model to gain better insight on the This section presents a clear view of the formulation of energy con­
trend and seasonality of the energy consumption data. sumption forecast problem, k − means clustering, long short term
2. Clustering algorithms: An academic building’s energy consump­ memory, and convolutional neural networks.
tion is quite complex and dynamic, which does not exhibit an
apparent trend and seasonality in the initial analysis. In such cases, 2.1. Energy consumption forecast problem
the application of clustering algorithms to the energy consumption
data before data modeling provides better insights into the trend and A multi-featured building energy demand forecasting problem can be
seasonal characterization of the data through the generation of defined as the energy consumed by several components (e.g., plugs,
clusters. lights, fans, air conditioners, servers, computers, etc.) and environ­
3. Multi-input and multi-output sliding window: The application of mental factors (humidity, temperature, etc.) which are monitored and
multi-input and multi-output sliding window to kCNN-LSTM pro­ recorded by the sensors (e.g., smart meters, temperature sensors, etc.)
vides robust and reliable forecasting by moving through the window installed at various levels of the considered building. The energy
of historical and recent energy consumption observations. consumed at timestamp t can be represented as in Eqn. (1).
4. Static or live data: kCNN-LSTM has been implemented as an energy { }
Et = e1t , e2t , …, eit , …, ent (1)
consumption forecast model in the BEMS designed for Kanwal Rekhi
School of Information and Technology (KReSIT), IIT-Bombay, India.
where, eit is the energy consumption logged by the ith sensor at time­
The key contributions of this paper are. stamp t. This work uses Multi-Input and Multi-Output (MIMO) sliding
window approach for energy consumption forecast to achieve better
1. kCNN-LSTM, a deep learning framework, is presented to provide forecast accuracy. Therefore, let {InputL , OutputL } ∈ N represents the
reliable and accurate building energy consumption forecasts. size of the input and output window. Further, the total number of input
2. As a data preprocessing step, seven timestamp based features were and output window can be defined as if = (Sn − InputL − a)/OutputL ,
generated to enrich the energy consumption data recorded at regular where Sn and a denotes the number of samples and output intervals,
time intervals as parameter rich data. respectively.
3. The complex trend and seasonality in the energy consumption data Input window (SInput ) of size (InputL ) and output window (SOutput ) of
are analyzed using the k − means clustering algorithm that uses the size (OutputL ) are represented as in Eqn. (2) and Eqn. (3) (Fig. 1).
LB-Keough distance metric to identify the similarity between the {
SInput = Et , Et+1 , …, Et+if
}
(2)
time series (energy consumption data) of different months in the
considered annual energy consumption data. { }
4. The spatio-temporal dependencies in the energy consumption data SOutput = Ẽt , Ẽt+1 , …, Ẽt+if (3)
are learned and modeled by the convolutional neural networks and
long short term memory neural networks, respectively. In this work, CNN-LSTM is modeled as an energy forecaster and
5. Multi-input and multi-input sliding window mechanism is deployed therein an approximation function (f) which relates SInput (Eqn. (2)) and
to provide accurate and reliable energy consumption forecast. SOutput is defined in Eqn. (3).
6. Further, effective modeling of higher order and non-linear de­ ( )
SOutput = f SInput , f : PInputL xn →POutputL xn (4)
pendencies in the energy consumption data enables kCNN-LSTM to
provide accurate forecasts in real time for a long period without Eqn. (4) states that given an input window (InputL ), the model (f)
retraining. learns to forecast the energy consumption values of the output window
(OutputL ) with minimal forecast error (Error) as defined in Eqn. (5).

4
N. Somu et al. Renewable and Sustainable Energy Reviews 137 (2021) 110591

Fig. 1. Energy consumption forecast – Input and output window size.

1∑ n ⃒
⃒ i t⃒
⃒ i. Forget gate (f t ) :Decides the information that should be discarded
Error = ⃒et − ẽt ⃒ (5) from the cell state represented by ft , based on the last hidden state
n i=1
(ht− 1 )and the new input (xt ), i.e., the output of the previous cell state
where, eit and ẽtt are the real and forecasted energy consumption value of and input of the current cell state, respectively (Eqn. (7)).
( )
the ith sensor at t th timestamp. ft = σ Wf .[ht− 1 , xt ] + bf (7)

2.2. k-means clustering where, Wf and bf are the weight matrices and bias vector of the forget
gate, respectively; σ is the logistic sigmoid function. The degree of in­
k − means clustering algorithm is a widely used partitioning cluster formation retention relies on the value of the forget gate, which lies in
analysis method. It is an iterative hill-climbing algorithm that groups m the range of [0,1] (‘0’-forget all; ‘1’-remember all).
data points into k (user-defined parameter) clusters to optimize the Sum
of Squared Error (SSE) that measures the intra-cluster similarity or inter- ii. Input gate (it ) :Decides the information of input (xt ) that should be
cluster dissimilarity. Each data point is assigned to one of the cluster stored in the cell state represented by the input gate (it ), where the
centroids (initialized randomly) based on its minimum distance from the information in the input gate (it ) and the new candidate cell state (c˜t )
centroid. Next, each cluster centroid is updated by obtaining the mean of are updated (Eqn. (8) and Eqn. (9)). The new cell state (ct ) is updated
the data points presented in the individual clusters. This procedure is by combining the previous cell state (ct− 1 ) and c˜t with the impact of
repeated until SSE (Eqn. (6)) between cluster centroids and the data forget gate (ft ) and input gate (it ) (Eqn. (10)).
points is minimized.
it = σ(Wi .[ht− 1 , xt ] + bi ) (8)
∑
K ∑
SSE = D P i − Cm (6) c̃t = tanh(Wc .[ht− 1 , xt ] + bc ) (9)
m=1 D P i ∈Cm

The k − means algorithm can be applied to massive, high- ct = ft * ct− 1 + it *c̃t (10)
dimensional, and non-linear time-series data (sequential values
measured at equal time intervals) to obtain consistent individual groups where, Wi and bi are the weight matrices and bias vector of the input
of time-series for accurate predictions. In this case, clustering parame­ gate, respectively; Wc and bc are the weight matrices and bias vector of
ters like distance measure, cluster evaluation measure, etc. should be the cell state, respectively; * is the point-wise multiplication; tanh is a
decided before clustering. hyperbolic tangent function with the range of [-1,1].

2.3. Long short term memory iii. Output gate (ot ) :Decides the information in the cell state (ct )
that should flow as the output of the output gate (ot ). Eqn. (11)
Long Short Term Memory Neural Networks (LSTM), the special and evaluates which part of the cell state is to be exported, and Eqn.
improved architecture of Recurrent Neural Networks (RNN) employs (12) computes the final output.
gate units and ‘self-connected memory cells’ to extract the underlying ot = σ (Wo .[ht− 1 , xt ] + bo ) (11)
complex temporal dependencies in long and short time-series data,
thereby addressing the ‘vanishing gradient problem’ of RNN [42]. LSTM ht = ot .tanh(ct ) (12)
consists of a memory block which is responsible for determining the
addition and deletion of information through three gates, namely input where, Wo and bo are the weight matrices and bias vector of the output
gate (it ), forget gate (ft ), and output gate (ot ) (Fig. 2) [43,44]. The gate, respectively; The activation functions (σ (.) and tanh(.)) used to
memory cell in the memory block remembers temporal state informa­ express the non-linearity of the LSTM network can be defined as in Eqn.
tion about current and previous timesteps. The workflow of LSTM at (13) and Eqn. (14).
timestep t is detailed as [45,46]:

5
N. Somu et al. Renewable and Sustainable Energy Reviews 137 (2021) 110591

Fig. 2. Long short term memory.

1 ( )
σ (x) = x
(13) Cijn = sum ckn ⊗ fmij + bn (15)
1 + e−

ex − e− x Y n = δ(Cn ) (16)
tanh(x) = (14)
ex + e− x
where, Cnij is the output of the convolution operations in the convolu­
2.4. Convolutional neural networks tional layer; ckn is the convolution kernel matrix; fmij is the filter matrix;
sum( *) adds all elements in *; n is the nth feature map; i and j are the
Yann LeCun, et al. designed CNN or ConvNets, a class of deep feed- number of steps of convolution filter in horizontal and vertical di­
forward neural network which has proven its significance in extract­ rections; bn is the bias; δ() is the activation function.
ing the spatial features in time series data, image recognition, and With the completion of convolution, there are n feature maps that
classification tasks [47]. serve as input to the pooling layer. Pooling reduces the dimension of
The overall architecture of CNN comprises of convolutional layer, feature maps by reducing the redundant features without losing
pooling layer, and fully connected layer to model complex data (Fig. 3) important information (Fig. 4(b)). The pooling operations reduce the
[48,49]. In general, CNN has several hierarchies of convolutional and computational burden of the model by compressing the input feature
pooling layers, wherein several convolution runs are done to extract the map. The average, max, and sum pooling can be computed using Eqn.
important features from the input data. In the convolutional layer, (17) – Eqn. (19), respectively.
neurons of different layers of the network are locally connected through ( )
a weight sharing technique. Further, the convolutional layer forms the Pnij = AverageP pfmij (17)
core of CNN, which performs convolution and activation operations on ( )
the input data to create a feature map. A sliding window of convolu­ Pnij = MaximumP pfmij (18)
tional filter (matrix) of size equal to the size of the convolutional kernel ( )
moves across the horizontal and vertical directions of the 2 − Pnij = SumP pfmij (19)
dimensional input data (Fig. 4(a)). The size of the convolutional filter
should be equal to the size of the convolutional kernel. The convolution where, Pnij is the output of the pooling operation in the pooling layer; i
operation is completed by the kernel through the construction of a and j are the number of steps of pooling filter in horizontal and vertical
feature map, a 2D representation of the kernel generated by calculating directions; pfmij is the pooling filter matrix; AverageP ( *) provides the
the dot product of the convoluted kernel and the convolution filter. The average of all the elements in *; MaximumP ( *) provides the maximum
convolution operation and activation operation in the convolutional element in *; SumP ( *) provides the sum of all the elements in *. Finally,
layer can be defined as in Eqn. (15) and Eqn. (16). the role of the flatten layer is to flatten the 2D data into a 1D vector
representation, which is fed as input to the fully connected layer (Fig. 4

Fig. 3. Convolutional neural network architecture.

6
N. Somu et al. Renewable and Sustainable Energy Reviews 137 (2021) 110591

Fig. 4. (a) 2D convolution operation; (b) Pooling operation with pooling size = [2,2]; (c) Flatten operation [48].

(c)). The main idea of the fully connected layer is to connect the adjacent where, Wfc is the weight matrix; FC is the output matrix of the fully
layers to integrate the features in providing the linear output for connected layer; bfc is the bias; x is the input to the fully connected layer.
regression problems [50]. The calculation in the fully connected layer is
defined in Eqn. (20).
( )
FC = δ Wfc x + bfc (20)

7
N. Somu et al. Renewable and Sustainable Energy Reviews 137 (2021) 110591

Fig. 5. kCNN-LSTM building energy consumption forecast model-Workflow.

Table 2
Preprocessed energy consumption dataset.
Year Day of the year Season Month Day of the week Hour of the day Minute of the hour Type of the day Energy Consumed

2017 1 3 1 1 0 0 1 256
2017 1 3 1 1 0 15 1 258
… … … … … … … … …
2019 334 3 11 49 23 45 0 426

3. CNN-LSTM: proposed deep learning framework for building 3.2. Data clustering and analysis layer
energy consumption forecast
Given an energy consumption dataset E with M objects, E = {T1 , T2 ,
Fig. 5 presents the architecture of kCNN-LSTM to forecast the energy …, TM }, where Ti is a time series. Time series clustering of E into clusters
consumption of buildings. The overall workflow of kCNN-LSTM consists C = {C1 , C2 , …, Ck } is done by grouping the data points based on the
of (i) Data source and preprocessing layer: Transform the raw energy similarity in the consumption trends (E = Uki=1 Ci , where i ∕ = j and
consumption data into a compatible format, (ii) Data clustering and Ci ∩Cj = ∅) [51]. The procedure for the selection of the cluster param­
analytics layer: cluster the data into different groups for trend analysis, eters like distance measure, evaluation measures, etc. is detailed below.
(iii) Dataset generation layer: generate datasets for train, valid, and
test, and (iv) Model building and evaluation layer: create models to 3.2.1. Distance measure
forecast energy consumption and evaluate their performance using Euclidean distance and Dynamic Time Warping (DTW) form the most
various error metrics. kCNN-LSTM energy consumption forecast model widely used distance metric for time series clustering due to the effi­
is then integrated with the BEMS to forecast building energy consump­ ciency of Euclidean distance (simple, fast, and parameter-free) and the
tion for the user-specified intervals (time slot, day, week, month, season, effectiveness of DTW.
etc.).
(i) Euclidean distance measure: It provides one-to-one matching
3.1. Data source and preprocessing layer between the timestamps of the time series data T1 = {t11 , t12 , …
, t1n } and T2 = {t21 , t22 , …, t2n } as defined in Eqn. (21).
The energy consumption data of the KReSIT building at IIT, Bombay, √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
√∑
contains more than 30 electrical related features. Out of these, this study √ n ( )2
Euclidean(T1,T2) = √ t1i − t2j (21)
considers timestamp (dd-mm-yyyy HH: MM: SS) and energy consump­ i,j=1
tion columns for processing, as it is available across different hierarchies
of buildings. Since the yearly energy consumption data of 2018 is However, Euclidean distance is not the best choice for time series
considered for analysis, the available per second granularity data is data since it depends on the domain & time series characteristics of the
aggregated to 15 min granularity. Totally, there are 96 energy con­ data and disability to capture the distortions in the time domain, i.e.,
sumption values for each day in the selected year (Block0 - 00:00 to sensitive to shifts in the time axis.
00:15, Block1 - 00:15 to 00:30, …, Block 95–23:45 to 00:00). For
techniques employed to handle missing values, outliers, and high (ii) DTW distance measure [52]: DTW finds the optimal non-linear
magnitude values, refer Section 4.2. Further, a simple Python code is alignment between two time series by calculating their matching
designed to generate seven features (day of the year (0–365), season (0- error as follows:
Summer; 1-Monsoon; 2-Winter), month, day of the week (0–52), hour of a. Consider two time series of the same length n, T1 = {t11 , t12 ,
the day (0–24), minute of the hour (0–60), type of the day (0-Holiday; 1- …, t1n } and T2 = {t21 , t22 , …, t2n }, where t1i and t2i are the
Working day)) from the available timestamp. Table 2 shows the struc­
ture of the energy consumption data used for experimentation.

8
N. Somu et al. Renewable and Sustainable Energy Reviews 137 (2021) 110591

energy consumption value of time series T1 and T2 at ith 3.4. Model building and evaluation layer
timestamp.
b. Construct a cost matrix of n x n dimension, whose ith and jth kCNN-LSTM, a building energy consumption forecast model is
element represents the Euclidean distance between t1i and t2j . designed to model the spatial correlation between the time stamp based
c. Find the path (P), i.e., optimal alignment between T1 and T2 generated variables and temporal information in the irregular con­
through the constructed cost matrix that minimizes the cu­ sumption patterns for better forecast accuracy. The model architecture
mulative distance (Eqn. (22)). of kCNN-LSTM consists of convolutional layer, pooling layer, LSTM
⎛√ ̅̅̅̅̅̅̅̅̅̅̅̅⎞ ⎛√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅⎞ layer, and fully connected layer. The convolutional and pooling layers of
√ K
√∑ √∑
√ n ( )2 CNN extracts the spatial characteristics of the multivariate data and pass
*
P = argminP ⎝√ pk = argminP ⎝√
⎠ t1i − t2j ⎠ (22) on the identified features as input to the LSTM. The LSTM models the
irregular trends in the energy consumption data based on the spatial
k=1 i,j=1

features provided by the CNN. Later then, a fully connected layer re­
where, P = (p1 ,p2 ,…,pK ), each element P denotes the distance between
ceives and decodes the output of LSTM to provide the forecasted energy
ith and jth data point in T1 and T2 . consumption data. The multi-input and multi-output sliding window
mechanism is employed in such a way that kCNN-LSTM learns the input
d. Find the optimal path using a recursive dynamic programming data for every block (15 min).
function. The input of size 60 × 7 is fed to the convolutional layers with filters,
kernels, padding, and activation of 64, 2 × 1, ‘same’, and ReLU,
DTW has a complexity of O(nm), where n and m are the length of first respectively. The output of the convolutional layer is passed to an LSTM
and second time series, respectively. layer with 64 units and ‘tanh’ activation. With the series of fully con­
nected layer, the number of units in the dense layers are 32 and 60 to
(iii) Lower bound Keough distance measure [53,54]: Due to high produce the forecast of the next 60 min energy demand. The hyper­
complexity, the recursive application of DTW on a long time se­ parameters of each layer in kCNN-LSTM were fine-tuned using ISCOA,
ries data is expensive. Imposing locality constraint using a an enhanced variant of SCOA to improve the learning speed and per­
threshold determined by the window size and Lower Bound formance of the learning model. ISCOA uses Haar wavelet based mu­
Keough (LB Keough) distance metric provides a way to speed up tation operator to enhance the divergence nature of the algorithm
DTW. LB Keough is based on the fact that DTW uses global path towards the global optimum. For more details related to hyperparameter
constraints when comparing two time series. LB Keough for two tuning, refer [55].
time series T1 = {t11 , t12 , …, t1n } and T2 = {t21 , t22 , …, t2n } is
defined as in Eqn. (23). 4. Case study
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
√ ⎧
√ ⎪
√∑ ⎪ T2i − U 2i if T2i > Ui This section presents a detailed description of the considered energy
√ N ⎨
LB − Keough(T1 , T2 ) = √ T − L2i if T2i < Li (23) consumption data, data preprocessing techniques, and evaluation met­
√ ⎪ ⎪ 2i
i=1 ⎩
0 Otherwise rics. Further, a detailed analysis of the performance of kCNN-LSTM over
the k − means variant of the state-of-the-art building energy consump­
where, Ui and Li are the lower and upper bound of time series T1 , which tion forecast model.
can be defined as Ui = max(T1i− r : T1i+r ) and Li = min(T1i− r : T1i+r ); r de­
pends on the type of path constraint used (e.g., Itakura parallelogram 4.1. Dataset
and Sakoe-Chiba band). The complexity of LB Keough is O(m).
This work uses k − means clustering algorithm and LB-Keough dis­ This work uses the energy consumption data of KReSIT, IIT-Bombay,
tance metric to cluster the energy consumption patterns for trend India. The flow of electricity consumption data from sensing to sharing
analysis. comprises several layers, physical and software components (Fig. 6). The
functionalities of each layer are detailed below.
3.2.2. Optimal value of k
The most primary and essential step in any unsupervised algorithm is (i) Physical layer: It corresponds to the actual building environ­
to find the optimal number of clusters to which the data points are ment for which electricity consumption is measured. KReSIT, IIT-
added. Several methods like Elbow, average Silhouette, gap statistics, B is a four-storeyed academic building with three wings (A, B, and
etc. have been proposed in the literature to identify the optimal number C) for each floor, which consists of smart classrooms, office
of clusters (k). This work employs the Elbow method to identify the rooms, auditoriums, lecture halls, research laboratories, and
optimal value of k for precise clustering results. The elbow method plots server rooms. The location of the considered academic building is
various SSE values for different k values through iterative runs of k − in Mumbai. The average temperature of Mumbai falls within 17 −
means clustering algorithm. The basic idea is that as the value of k in­ 34o C for three seasons, namely summer (April–June), monsoon
creases, the average distortion decreases with fewer elements in the (July–September), and winter (October–March). The energy and
cluster. Therefore, k value at which the distortion decreases the most is temperature profiles of KReSIT are monitored and controlled by
the elbow point, which is chosen as the optimal k for the considered its own BEMS built for “Minimal power consumption in an
dataset. occupied room” and ‘Zero consumption during zero occupancy”.

3.3. Dataset generation layer Since the major objective of this research is to forecast the overall
energy consumption of the considered buildings, the consumption data
The energy consumption data of each cluster is partitioned into provided by the smart meter installed at the MAINS is used for
training (ETrain ), validation (EVal ), and testing (ETest ) in the ratio of experimentations.
60:20:20 for overall, weekdays, weekend, and day analysis. For experimental purposes, the energy consumption data from
January 1, 2018 to December 31, 2018 at 15 min granularity was
considered. For better visualization, each of the 15 min interval is
mapped to a block number, i.e., Block 0: 00:00 to 00:15, Block 2: 00:15
to 00:30, …, Block 95: 23:45 to 0:00. Fig. 8 provides the statistical detail

9
N. Somu et al. Renewable and Sustainable Energy Reviews 137 (2021) 110591

Fig. 6. BEMS, KReSIT, IIT-B: Architecture.

(i) Sensing layer: A three-phase smart meters deployed at various facilities of KReSIT records the various electrical parameters through a MODBUS protocol over
RS485 cable. This study uses the energy consumption variable, which provides energy consumed by air conditioners, plugs, lights, and fans at per-second
granularity in the considered physical space. The energy consumption data is queried from the smart meter by a Raspberry Pi at predefined intervals in a
round robin fashion. A python script is executed in Raspberry Pi to collect the smart meter data and transmit to the central server using Message Queuing
Telemetry Transport (MQTT) protocol. The sensor nodes communicate via. serial or low-range RF. Further, a gateway (NodeMCU and RPi) is used to send the
sensor data to the central server for storage and processing.
(ii) Communication layer: MQTT, a fast and minimal overhead transmission protocol, is used to publish the data from the smart meters to the remote ingestion
servers. It uses ‘publish-subscribe’ concept, which enables multiple subscribers to receive the selected data stream by registering with the central broker.
(iii) Ingestion and storage layer: The data stream of the smart meters are published on the topic of structure <channel>/<data type>/<sensor identifier>, where
different channels used to publish data are specified by the topic, data type is used by the ingestion engine to find the schema of the published data stream, and
sensor identifier corresponds to the global unique ID assigned to each smart meter. The smart meter data is stored in MySQL database and CSV format for
visualization and archival.
(iv) Visualization layer: Grafana, an open source data visualization engine, is used to display the real-time electricity consumption of various appliances (ACs,
plugs, lights, and fans), which enables the consumers to visualize their energy consumption. The academic building dataset can be accessed from Ref. [56] for
research purposes. The live power consumption data can be visualized, as shown in Fig. 7.

Fig. 7. Data visualization of KReSIT overall power consumption.

of the energy consumption data for the months of 2018. ⎧x + x

i− 1 i+1
⎪
⎪ xi ∈ NaN; xi− 1 , xi+1 ∕
∈ NaN
⎪
⎪ 2
⎨
4.2. Data preprocessing f (xi ) = xi− 1 + xi− 2 (24)
⎪
⎪ xi ∈ NaN; xi− 1 , xi− 2 ∈ NaN; xi+1 ∈ NaN
∕
⎪
⎪ 2
⎩
In a real world scenario, energy consumption data is susceptible to xi xi ∕
∈ NaN
several discrepancies like missing data, incomplete data, noise, etc.
which might lead to poor data analysis. The techniques used to pre­ where, xi represents the energy consumption value at ith timestamp.
process the energy consumption data are detailed below.
(ii) Outlier detection: “Three sigma rule of thumb” is used to
(i) Missing value: Data interpolation method is employed to impute recover the erroneous value in the energy consumption dataset
the missing values in the energy consumption dataset (Eqn. (24)). (Eqn. (25)).

10
N. Somu et al. Renewable and Sustainable Energy Reviews 137 (2021) 110591

where, n is the number of data points in the dataset; yi and ̂

y i are the
actual and the forecasted energy consumption value at timestamp i,
respectively.

(ii) Root mean squared error: It is defined as the standard deviation

of the differences between actual and the forecasted value (Eqn.
(28)).
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
√ ( )2
√
√ n y
̂ − y
√∑ i i
RMSE = √ (28)
i=1
n

Fig. 8. Statistics of KReSIT energy consumption data for 2018. (iii) Mean absolute error: It measures the absolute difference be­
tween the actual and the forecasted value (Eqn. (29)).
{
avg(x) + 2.std(x) if xi > avg(x) + 2.std(x) n ⃒ ⃒
f (xi ) = (25) 1∑ ⃒ ⃒
xi otherwise MAE = ⃒̂y i − yi ⃒
⃒ ⃒ (29)
n i=1
where, x is a vector that consists of xi ; avg(x) is the average value of x;
std(x) is the standard deviation of x. (iv) Mean absolute percentage error: It is the measure of the
amount of deviation of the forecasted value from the actual value
(iii) Normalization: Min-max normalization technique is used to (Eqn. (30)).
regularize the energy consumption data to avoid inaccurate ⃒ ⃒
forecasts due to the high magnitude in the energy consumption 1∑ n ⃒ ⃒
⃒̂y i − yi ⃒
⃒ ⃒*100 (30)
data (Eqn. (26)). MAPE =
n i=1 ⃒⃒ yi ⃒⃒
xi − min(x)
f (xi ) = (26)
max(x) − min(x) 5. Results and discussions

where, min(x) and max(x) are the minimum and maximum values of x,
The implementation of kCNN-LSTM and the experimental analysis
respectively.

4.3. Performance metrics

There exist several statistical measures to evaluate the performance

of the learning model based on the difference between the actual and
predicted value. The performance evaluation metrics used in this paper
are defined below.

(i) Mean squared error: It is defined as the average squared dif­

ference between the actual and the forecasted value (Eqn. (27)).
( )2
∑n ̂y i − yi
MSE = (27)
i=1
n
Fig. 9. Optimal number of cluster-Elbow method.

Table 3
LB-Keough distance matrix.
Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec

Jan 0 0.2573 0.4000 0.5253 0.7467 0.1313 0.3930 0.5967 0.2966 0.7075 0.3134 0.4040
Feb 0.9143 0 1.1917 0.2607 1.2914 0.2951 1.1596 0.9738 0.1183 0.5721 0.2396 1.0829
Mar 0.6175 0.2488 0 0.3578 0.2298 0.1580 0.2730 0.4172 0.3766 0.3179 0.5215 1.0556
Apr 1.2903 1.0702 1.3483 0 1.5387 1.0032 1.3159 1.1997 1.1093 0.8165 1.1244 1.2989
May 1.0323 0.4832 0.1398 0.4926 0 0.2995 0.5324 0.5502 0.6560 0.4026 0.7946 1.4301
Jun 0.7177 0.3172 0.7714 0.3668 0.8568 0 0.7803 0.6995 0.3958 0.5197 0.4220 0.9394
Jul 0.3790 0.0414 0.1022 0.3927 0.4171 0.1514 0 0.2753 0.1328 0.3566 0.2659 0.8025
Aug 0.4490 0.0510 0.1279 0.0746 0.2597 0.0475 0.1072 0 0.1659 0.0817 0.2027 0.6976
Sep 0.5486 0.1005 0.7276 0.3096 0.8489 0.1709 0.6779 0.5866 0 0.4288 0.1016 0.7053
Oct 0.7388 0.2619 0.7102 0.0632 0.9346 0.3679 0.6912 0.4949 0.4222 0 0.3892 0.8452
Nov 0.6150 0.0536 0.9079 0.2452 1.0596 0.2228 0.8744 0.7127 0.0116 0.4468 0 0.6693
Dec 0.0934 0.1955 0.4630 0.3906 0.8102 0.2701 0.4003 0.3970 0.1838 0.4857 0.1660 0

11
N. Somu et al. Renewable and Sustainable Energy Reviews 137 (2021) 110591

Fig. 10. Cluster analysis (a) Cluster 1 (b) Cluster 2 (c) Cluster 3.

(ii) Phase 2: Energy consumption trend analysis.

12
N. Somu et al. Renewable and Sustainable Energy Reviews 137 (2021) 110591

Fig. 11. Energy consumption patterns in Cluster 1 for different days (a) Monday (b) Wednesday (c) Friday (d) Sunday.

Fig. 12. Energy consumption patterns in Cluster 2 for different days (a) Monday (b) Wednesday (c) Friday (d) Sunday.

13
N. Somu et al. Renewable and Sustainable Energy Reviews 137 (2021) 110591

Fig. 13. Energy consumption patterns in Cluster 3 for different days (a) Monday (b) Wednesday (c) Friday (d) Sunday.

In general, the application of clustering to any data results in the

Table 4
identification of groups or clusters with similar characteristics. In this
Parameter setting.
work, k-means clustering, a partitioning based clustering algorithm is
S. Model Description Range applied to group the data points (time series data - energy consumption
No.
value recorded at each timestamp) into clusters based on the similarity
1 ISCOA Number of populations 5 in the energy consumption patterns. Due to the simplicity and minimal
Number of iterations 20
complexity, LB Keough distance metric is used to compute the similarity
2 CNN- Epoch 100
LSTM Batch size 1 between the time series data (months). Table 3 provides the distance
Optimizer Adam matrix computed using LB Keough measure for the time series data of all
Convolution and pooling [1,3] months in 2018.
layers Further, the Elbow method is used to identify the optimal value of k.
LSTM layers [1,3]
Filters [64,128]
For the same, SSE is used as an evaluation metric, which when plotted
Kernel (2,1) against the number of clusters, helps in the identification of the optimal
Units in fully connected layer [5100] k (Fig. 9). For the considered energy consumption data of 2018, the
Hidden units in LSTM [5100] optimal k as 3.
Activation function ReLu (CNN) and tanh
From the clustering results, it is obvious that the energy consumption
(LSTM)
patterns of the year 2018 can be divided into three clusters, namely
Cluster 1, Cluster 2, and Cluster 3. An in-depth analysis of the obtained
was carried out in Operating system - Windows 10; System configuration clusters (Cluster 1, Cluster 2, and Cluster 3), provides detailed insight
- i7 processor and 64 GB RAM; and Framework - Python 3.6. The ex­ into the energy consumption patterns in KReSIT building across
periments carried to analyze the efficiency of kCNN-LSTM for the energy different months in 2018 (Fig. 10(a)-(c)).
consumption forecast problem can be divided into three phases, namely In this study, a set of contextual features were generated based on the
(i) Data clustering, (ii) Energy consumption trend analysis, and (iii) timestamp in the considered building energy consumption data <
Energy demand forecast analysis. timestamp, energy>. However, to gain deeper insight on the variations
in the energy consumption trend, it is vital to analyze the consumption
(i) Phase1: Clustering patterns in each cluster. For the same, daily analysis was performed on

14
N. Somu et al. Renewable and Sustainable Energy Reviews 137 (2021) 110591

Table 5
Cluster 1 – Performance analysis of the state-of-the-art machine learning and deep learning energy demand forecast models.
Type Metrics Models

ARIMA DBN MLP LSTM CNN CNN-LSTM

Overall MSE 0.0102 0.0183 0.0189 0.6226 0.3313 0.0095

RMSE 0.0982 0.1354 0.1376 0.7890 0.5756 0.0974
MAE 0.1098 0.1063 0.0998 0.7310 0.5429 0.0711
MAPE 0.8128 0.3550 0.3161 2.0425 1.7206 0.2697
Weekdays MSE 0.0356 0.0966 0.0805 0.0267 0.0778 0.0168
RMSE 0.1887 0.3107 0.2837 0.1633 0.2790 0.1297
MAE 0.1527 0.2842 0.2547 0.1240 0.2594 0.1113
MAPE 0.2701 0.8009 1.9979 0.4433 0.7682 0.3872
Weekend MSE 0.0293 0.0040 0.0045 0.0048 0.0048 0.0034
RMSE 0.1712 0.0635 0.0667 0.0693 0.0693 0.0580
MAE 0.1471 0.0547 0.0537 0.0570 0.0602 0.0481
MAPE 0.1447 0.1737 0.1580 0.1725 0.1826 0.1425
Monday MSE 0.0151 0.0286 0.0073 0.0091 0.0030 0.0013
RMSE 0.1232 0.1690 0.0853 0.0953 0.0551 0.0364
MAE 0.1031 0.1585 0.0580 0.0805 0.0404 0.026
MAPE 0.3119 0.5046 0.1320 0.2486 0.1166 0.0728
Wednesday MSE 0.0116 0.0039 0.0023 0.0066 0.0021 0.00100
RMSE 0.1077 0.0624 0.0477 0.0812 0.0455 0.0324
MAE 0.0925 0.0509 0.0398 0.0635 0.0387 0.0267
MAPE 0.2506 0.1635 0.1102 0.1851 0.1128 0.0765
Friday MSE 0.3271 0.0180 0.0095 0.0038 0.0042 0.0024
RMSE 0.5720 0.1343 0.0975 0.0618 0.0650 0.0491
MAE 0.5024 0.1229 0.0825 0.0511 0.0551 0.0368
MAPE 0.2444 0.3860 0.1860 0.1455 0.1601 0.1033
Sunday MSE 0.0064 0.0022 0.0011 0.0010 0.0009 0.0008
RMSE 0.0803 0.0474 0.0336 0.0324 0.0306 0.0285
MAE 0.0679 0.0419 0.0273 0.0273 0.0252 0.0232
MAPE 0.0966 0.1502 0.0867 0.0924 0.0856 0.0783

Table 6
Cluster 2 – Performance analysis of the state-of-the-art machine learning and deep learning energy demand forecast models.
Type Metrics Models

ARIMA DBN MLP LSTM CNN CNN-LSTM

Overall MSE 0.1006 0.1303 0.1116 0.0618 0.0639 0.0212

RMSE 0.1625 0.3609 0.3341 0.2485 0.2528 0.1456
MAE 0.1138 0.3215 0.2931 0.2133 0.2278 0.0997
MAPE 0.2062 0.2549 0.2375 0.5785 0.2358 0.2054
Weekdays MSE 0.1007 0.1007 0.0384 0.0504 0.0687 0.0251
RMSE 0.2066 0.3174 0.1958 0.2246 0.2621 0.1586
MAE 0.1456 0.2718 0.1438 0.1911 0.2206 0.1397
MAPE 0.1918 0.3015 0.4496 0.1855 0.1926 0.1653
Weekend MSE 0.0101 0.0416 0.0156 0.0057 0.0334 0.0038
RMSE 0.2134 0.2039 0.1251 0.0757 0.1828 0.062
MAE 0.1329 0.1892 0.1052 0.0579 0.1683 0.0433
MAPE 0.8885 0.5277 0.2774 0.1704 0.4541 0.1425
Monday MSE 0.0113 0.0272 0.0230 0.0122 0.0147 0.0110
RMSE 0.1190 0.1648 0.1516 0.1106 0.1211 0.1050
MAE 0.0829 0.1391 0.1230 0.0771 0.0943 0.0721
MAPE 0.2252 0.4564 0.2893 0.2287 0.2838 0.2169
Wednesday MSE 0.0094 0.0229 0.0095 0.0099 0.0126 0.0002
RMSE 0.0971 0.1512 0.0975 0.0995 0.1124 0.0164
MAE 0.1040 0.1125 0.0714 0.0708 0.0865 0.0707
MAPE 0.2576 0.4807 0.1733 0.1817 0.2370 0.1551
Friday MSE 0.0218 0.1466 0.1293 0.1218 0.1222 0.0013
RMSE 0.1372 0.1478 0.0969 0.1118 0.0967 0.0871
MAE 0.1162 1.0291 0.2317 0.8481 0.9251 0.1137
MAPE 0.2259 0.3521 0.2725 0.2838 0.2915 0.2346
Sunday MSE 0.9105 0.0653 0.0167 0.0148 0.0149 0.0068
RMSE 0.0203 0.0048 0.0041 0.0039 0.0046 0.0036
MAE 0.1182 0.0695 0.0640 0.0626 0.0679 0.0600
MAPE 0.3103 0.0542 0.0470 0.0448 0.0512 0.0418

15
N. Somu et al. Renewable and Sustainable Energy Reviews 137 (2021) 110591

Table 7
Cluster 3 – Performance analysis of the state-of-the-art machine learning and deep learning energy demand forecast models.
Type Metrics Models

ARIMA DBN MLP LSTM CNN CNN-LSTM

Overall MSE 0.0019 0.0708 0.0380 0.0113 0.0124 0.0010

RMSE 0.1047 0.2660 0.1950 0.1049 0.1114 0.0303
MAE 0.0812 0.2079 0.1450 0.0880 0.0893 0.0165
MAPE 0.2086 0.3882 0.3107 0.1998 0.2101 0.1670
Weekdays MSE 0.1006 0.0552 0.4857 0.2612 0.1323 0.0142
RMSE 0.3264 0.2350 0.2204 0.1616 0.1193 0.1150
MAE 0.2180 0.1892 0.1678 0.1348 0.9167 0.0922
MAPE 3.3808 0.3675 0.3686 0.3208 0.2199 0.2134
Weekend MSE 0.0024 0.0019 0.0015 0.0016 0.0013 0.0013
RMSE 0.1207 0.0433 0.0389 0.0397 0.0366 0.0362
MAE 0.1138 0.0357 0.0308 0.0304 0.0283 0.0275
MAPE 0.8961 0.0847 0.0702 0.0700 0.0677 0.0646
Monday MSE 0.1006 0.0482 0.0417 0.0290 0.0292 0.0256
RMSE 0.3250 0.2196 0.2042 0.1702 0.1708 0.1601
MAE 0.2177 0.1641 0.1566 0.1347 0.1379 0.1333
MAPE 0.2332 0.2704 0.4117 0.2303 0.2337 0.2257
Wednesday MSE 0.1006 0.0662 0.0463 0.0314 0.0218 0.0153
RMSE 0.1253 0.2572 0.2151 0.1771 0.1478 0.1235
MAE 0.2169 0.2010 0.1623 0.1443 0.1186 0.1045
MAPE 0.5244 0.3756 0.4868 0.2569 0.2460 0.1947
Friday MSE 0.1317 0.0507 0.0578 0.0280 0.0192 0.0101
RMSE 0.3630 0.2251 0.2403 0.1673 0.1386 0.1005
MAE 0.3126 0.1769 0.1885 0.1339 0.1180 0.0778
MAPE 0.4956 0.3336 0.6999 0.2457 0.2384 0.1473
Sunday MSE 0.0251 0.0351 0.0201 0.0109 0.0104 0.0075
RMSE 0.1586 0.1873 0.1419 0.1045 0.1018 0.0867
MAE 0.1397 0.1604 0.1144 0.0835 0.0796 0.0611
MAPE 0.3280 0.3643 0.3987 0.1942 0.1929 0.1404

all days of weekdays and weekend. The experimental design and vali­ forecast problem was validated through a comparative analysis with
dation are presented with the daily analysis on the selected days of k-means clustering variant of ARIMA, Deep Belief Network (DBN),
weekdays (Monday, Wednesday, and Friday) and weekends (Sunday) of Multi-Layer Perceptron (MLP), CNN and LSTM. The hyperparameters of
the identified clusters. The analysis and results of the other days of the kCNN-LSTM (momentum, dropout, weight decay, learning rate, strides,
weekdays and weekends are provided in the Appendix (Fig. A1, A2 and filters, etc.) were fine-tuned through the enhanced variant of sine cosine
A3; Table A1, A2 and A3). To understand the fluctuations in the energy optimization algorithm to achieve better forecast accuracy (Table 4).
consumption demand of KReSIT and to provide an accurate demand The hyperparameters of the contrast models were fine-tuned using a
forecast, the average energy consumption over the selected days across sequential grid search algorithm.
the months in each cluster was analyzed (Figs. 11, 12 and 13(a)-(d)). The above stated architecture of kCNN-LSTM was implemented
The daily analysis enables a better understanding of the energy using Keras framework with Tensorflow backend. The average error
consumption patterns and aids the deep learning model to provide better values obtained from 30 independent runs were taken for the analysis of
forecast of the energy demands for the user-specified day, month, and kCNN-LSTM over k-means variant of the state-of-the-art energy demand
interval. forecast models. Tables 5–7 provides average error values of the
considered machine learning and deep learning models for cluster wise,
(iii) Phase 3: Future energy demand forecast analysis weekdays, weekends, and selected days of the week in the identified
clusters.
Finally, the performance of kCNN-LSTM for energy consumption From Tables 5–7, it is evident that kCNN-LSTM outperforms k-means
variants of the contrast models. The main reason behind this is that,
apart from learning the load trend characterization, kCNN-LSTM com­
Table 8 bines the benefit of the contextual features generated from the time­
Computation time analysis.
stamp and the temporal information in the historical energy
Computation time (Secs) consumption data to achieve better forecast accuracy. The computation
Models Cluster 1 Cluster 2 Cluster 3 time of kCNN-LSTM and the considered models reveal that the proposed
energy demand forecast model provides the best computational effi­
ARIMA 79 85 34
DBN 97 102 62 ciency (Table 8).
MLP 84 95 51
LSTM 88 110 45 6. Conclusions
CNN 81 105 43
CNN-LSTM 75 90 40
This work presented kCNN-LSTM, a deep learning framework for

16
N. Somu et al. Renewable and Sustainable Energy Reviews 137 (2021) 110591

Fig. A1. Energy consumption patterns in Cluster 1 for different days (a) Tuesday (b) Thursday (c) Saturday

17
N. Somu et al. Renewable and Sustainable Energy Reviews 137 (2021) 110591

Fig. A2. Energy consumption patterns in Cluster 2 for different days (a) Tuesday (b) Thursday (c) Saturday

18
N. Somu et al. Renewable and Sustainable Energy Reviews 137 (2021) 110591

Fig. A3. Energy consumption patterns in Cluster 3 for different days (a) Tuesday (b) Thursday (c) Saturday

robust and reliable building energy consumption forecast. kCNN-LSTM Table A1

employed k − means clustering, CNN, and LSTM for trend character­ Cluster 1 – Performance analysis of the state-of-the-art machine learning and
ization, energy-related feature identification, and to model temporal deep learning energy demand forecast models (Tuesday, Thursday, and
information in the energy consumption data. The hyperparameters of Saturday)
kCNN-LSTM model were optimized using ISCOA, which uses Haar
wavelet based mutation operator to update position, thereby avoiding Type Metrics Models

premature convergence. A case study using the real time building energy ARIMA DBN MLP LSTM CNN CNN-LSTM
consumption data acquired from the BEMS deployed at KReSIT building, Tuesday MSE 0.0158 0.0113 0.0051 0.0076 0.0065 0.0045
IIT-Bombay was presented. The performance of kCNN-LSTM was RMSE 0.1257 0.1065 0.0715 0.0870 0.0804 0.0668
compared with the k means variant of the state-of-the-art energy de­ MAE 0.1066 0.0778 0.0573 0.0663 0.0563 0.0442
mand forecast models in terms of MSE, RMSE, MAE, and MAPE. The MAPE 0.2674 0.2401 0.1429 0.1886 0.1620 0.1291
Thursday MSE 0.0144 0.0178 0.0051 0.0008 0.0016 0.0005
experimental results demonstrate the efficiency of kCNN-LSTM model RMSE 0.1202 0.1335 0.0716 0.0277 0.0397 0.0218
over the existing demand forecast models in providing accurate energy MAE 0.0960 0.1239 0.0540 0.0212 0.0333 0.0174
consumption demand forecasting. Therefore, the implementation of MAPE 0.4005 0.3963 0.1297 0.0619 0.0967 0.0514
kCNN-LSTM at the electricity network and user level can aid in decision Saturday MSE 0.0251 0.0122 0.0016 0.0008 0.0005 0.0005
RMSE 0.1586 0.1104 0.0400 0.0276 0.0220 0.0213
making, demand management programs, and energy efficiency aspects.
MAE 0.1397 0.0872 0.0349 0.0214 0.0170 0.0156
The future directions of this research include (i) Application of kCNN- MAPE 0.3280 0.2460 0.1034 0.0708 0.0556 0.0513
LSTM to residential buildings and (iii) Analyze the performance of
various optimization algorithms for hyperparameter tuning of kCNN-
Table A2
LSTM.
Cluster 2 – Performance analysis of the state-of-the-art machine learning
and deep learning energy demand forecast models (Tuesday, Thursday,
Credit author statement
and Saturday)
Nivethitha Somu: Conceptualization, Methodology, Investigation,
Type Metrics Models
Data curation, Writing (Original, Review and Editing), and Visualiza­
tion, Gauthama Raman: Conceptualization, Software, Writing (Review ARIMA DBN MLP LSTM CNN CNN-LSTM
and Editing), Krithi Ramamritham: Conceptualization, Validation, Re­ Tuesday MSE 0.0087 0.0166 0.0175 0.0089 0.0095 0.0003
sources, Writing (Writing – review & editing), Supervision, and Funding RMSE 0.0933 0.1287 0.1323 0.0944 0.0972 0.0192
acquisition. MAE 0.0566 0.1000 0.0904 0.0585 0.0600 0.0119
MAPE 0.5184 0.4124 0.2506 0.2466 0.2558 0.1877
Thursday MSE 0.1003 0.0237 0.0209 0.021 0.0197 0.0189
RMSE 0.1376 0.1538 0.1447 0.1451 0.1405 0.0192
(continued on next page)

19
N. Somu et al. Renewable and Sustainable Energy Reviews 137 (2021) 110591

Table A2 (continued ) [7] Chou JS, Tran DS. Forecasting energy consumption time series using machine
learning techniques based on usage patterns of residential householders. Energy
Type Metrics Models
2018;165:709–26. https://doi.org/10.1016/j.energy.2018.09.144.
ARIMA DBN MLP LSTM CNN CNN-LSTM [8] Ocampo Batlle EA, Escobar Palacio JC, Silva Lora EE, Martínez Reyes AM, Melian
Moreno M, Morejón MB. A methodology to estimate baseline energy use and
MAE 0.1119 0.1176 0.1100 0.1105 0.0968 0.0046 quantify savings in electrical energy consumption in higher education institution
MAPE 0.4184 0.4390 0.3958 0.3715 0.4048 0.2877 buildings: case study, Federal University of Itajubá (UNIFEI). J Clean Prod 2019.
Saturday MSE 0.0048 0.0067 0.0045 0.0064 0.0045 0.0035 https://doi.org/10.1016/j.jclepro.2019.118551.
RMSE 0.0695 0.0816 0.0672 0.0802 0.0673 0.0141 [9] Neto AH, Fiorelli FAS. Comparison between detailed model simulation and
MAE 0.0489 0.0611 0.0471 0.0602 0.048 0.0093 artificial neural network for forecasting building energy consumption. Energy
MAPE 0.1766 0.1787 0.1299 0.1738 0.1484 0.0604 Build 2008;40:2169–76. https://doi.org/10.1016/j.enbuild.2008.06.013.
[10] Afzal Muhammad, Huang Qi, Amin Waqas, Umer Khalid, Raza Asif,
Blockchain MN. Enabled distributed demand side management in community
Table A3 energy system with smart homes. IEEE Access 2020;8:37428–39.
[11] Global demand for energy will peak in 2030, says World Energy Council | Business
Cluster 3 – Performance analysis of the state-of-the-art machine learning
| The Guardian n.d.
and deep learning energy demand forecast models (Tuesday, Thursday, [12] Hassan S, Khosravi A, Jaafar J, Khanesar MA. A systematic design of interval type-2
and Saturday) fuzzy logic system using extreme learning machine for electricity load demand
forecasting. Int J Electr Power Energy Syst 2016;82:1–10. https://doi.org/
10.1016/j.ijepes.2016.03.001.
Type Metrics Models [13] Yaslan Y, Bican B. Empirical mode decomposition based denoising method with
ARIMA DBN MLP LSTM CNN CNN-LSTM support vector regression for time series prediction: a case study for electricity load
forecasting. Meas J Int Meas Confed 2017;103:52–61. https://doi.org/10.1016/j.
Tuesday MSE 0.1040 0.0463 0.0333 0.0298 0.2187 0.0222 measurement.2017.02.007.
RMSE 0.2640 0.2151 0.1824 0.1726 0.1568 0.1489 [14] Deb C, Zhang F, Yang J, Lee SE, Shah KW. A review on time series forecasting
MAE 0.3250 0.1636 0.1300 0.1353 0.2152 0.1218 techniques for building energy consumption. Renew Sustain Energy Rev 2017;74:
MAPE 0.4844 0.3006 0.3458 0.2503 0.2404 0.2314 902–24. https://doi.org/10.1016/j.rser.2017.02.085.
Thursday MSE 0.2029 0.0429 0.0532 0.0234 0.1154 0.0126 [15] Xiao J, Li Y, Xie L, Liu D, Huang J. A hybrid model based on selective ensemble for
energy consumption forecasting in China. Energy 2018;159:534–46. https://doi.
RMSE 0.1546 0.2070 0.2306 0.1531 0.1741 0.1123
org/10.1016/j.energy.2018.06.161.
MAE 0.1212 0.1608 0.1671 0.1160 0.8047 0.0883
[16] Wei N, Li C, Peng X, Zeng F, Lu X. Conventional models and artificial intelligence-
MAPE 0.3397 0.3125 0.5654 0.2409 0.8758 0.2138
based models for energy consumption forecasting: a review. J Petrol Sci Eng 2019;
Saturday MSE 0.0108 0.0209 0.0257 0.0130 0.1077 0.0081 181:106187. https://doi.org/10.1016/j.petrol.2019.106187.
RMSE 0.1285 0.1445 0.1603 0.1139 0.1880 0.0900 [17] Ciulla G, D’Amico A. Building energy performance forecasting: a multiple linear
MAE 0.2159 0.1155 0.1328 0.0906 0.0681 0.0662 regression approach. Appl Energy 2019;253:113500. https://doi.org/10.1016/j.
MAPE 0.2617 0.2544 0.4789 0.1918 0.1574 0.1480 apenergy.2019.113500.
[18] Liu T, Tan Z, Xu C, Chen H, Li Z. Study on deep reinforcement learning techniques
for building energy consumption forecasting. Energy Build 2020;208:109675.
https://doi.org/10.1016/j.enbuild.2019.109675.
[19] Tran D, Luong D, Chou J. Nature-inspired metaheuristic ensemble model for
Declaration of competing interest forecasting energy consumption in residential buildings. Energy 2019:116552.
https://doi.org/10.1016/j.energy.2019.116552.
[20] Raza MQ, Khosravi A. A review on artificial intelligence based load demand
The authors declare that they have no known competing financial forecasting techniques for smart grid and buildings. Renew Sustain Energy Rev
interests or personal relationships that could have appeared to influence 2015;50:1352–72. https://doi.org/10.1016/j.rser.2015.04.065.
the work reported in this paper. [21] Mat Daut MA, Hassan MY, Abdullah H, Rahman HA, Abdullah MP, Hussin F.
Building electrical energy consumption forecasting analysis using conventional and
artificial intelligence methods: a review. Renew Sustain Energy Rev 2017;70:
Acknowledgments 1108–18. https://doi.org/10.1016/j.rser.2016.12.015.
[22] Amasyali K, El-gohary NM. A review of data-driven building energy consumption
prediction studies. Renew Sustain Energy Rev 2018;81:1192–205. https://doi.org/
The authors would like to thank The Indian Institute of Technology-
10.1016/j.rser.2017.04.095.
Bombay (Institute Postdoctoral Fellowship-AO/Admin-1/Rect/33/ [23] Fayaz M, Kim D. A prediction methodology of energy consumption based on deep
2019); The Ministry of Human Resource and Development, India – extreme learning machine and comparative analysis in residential buildings.
Impacting Research Innovation and Technology (IMPRINT - Electron 2018;7. https://doi.org/10.3390/electronics7100222.
[24] Malik S, Kim DH. Prediction-learning algorithm for efficient energy consumption in
16MOPIMP002), New Delhi, India; Prof. Kannan Krithivasan, Dean, smart buildings based on particle regeneration and velocity boost in particle swarm
School of Education, SASTRA Deemed University, Tamil Nadu, India optimization neural networks. Energies 2018;11. https://doi.org/10.3390/
(TATA Realty—SASTRA Srinivasa Ramanujan Research Cell, India) and en11051289.
[25] Bedi J, Toshniwal D. Empirical mode decomposition based deep learning for
SEIL members. electricity demand forecasting. IEEE Access 2018;6:49144–56. https://doi.org/
10.1109/ACCESS.2018.2867681.
References [26] Li C, Ding Z, Yi J, Lv Y, Zhang G. Deep belief network based hybrid model for
building energy consumption prediction. Energies 2018;11:1–26. https://doi.org/
10.3390/en11010242.
[1] Yang Y, Chen Y, Wang Y, Li C, Li L. Modelling a combined method based on ANFIS
[27] Moon J, Park J, Hwang E, Jun S. Forecasting power consumption for higher
and neural network improved by DE algorithm: a case study for short-term
educational institutions based on machine learning. J Supercomput 2018;74:
electricity demand forecasting. Appl Soft Comput J 2016;49:663–75. https://doi.
3778–800. https://doi.org/10.1007/s11227-017-2022-x.
org/10.1016/j.asoc.2016.07.053.
[28] Touzani S, Granderson J, Fernandes S. Gradient boosting machine for modeling the
[2] Bui DK, Nguyen TN, Ngo TD, Nguyen-Xuan H. An artificial neural network (ANN)
energy consumption of commercial buildings. Energy Build 2018;158:1533–43.
expert system enhanced with the electromagnetism-based firefly algorithm (EFA)
https://doi.org/10.1016/j.enbuild.2017.11.039.
for predicting the energy consumption in buildings. Energy 2019:116370. https://
[29] Rahman A, Srikumar V, Smith AD. Predicting electricity consumption for
doi.org/10.1016/j.energy.2019.116370.
commercial and residential buildings using deep recurrent neural networks. Appl
[3] Wang W, Hong T, Xu X, Chen J, Liu Z, Xu N. Forecasting district-scale energy
Energy 2018;212:372–85. https://doi.org/10.1016/j.apenergy.2017.12.051.
dynamics through integrating building network and long short-term memory
[30] Ye Z, Kim MK. Predicting electricity consumption in a building using an optimized
learning algorithm. Appl Energy 2019;248:217–30. https://doi.org/10.1016/j.
back-propagation and Levenberg–Marquardt back-propagation neural network:
apenergy.2019.04.085.
case study of a shopping mall in China. Sustain Cities Soc 2018;42:176–83. https://
[4] Jain RK, Smith KM, Culligan PJ, Taylor JE. Forecasting energy consumption of
doi.org/10.1016/j.scs.2018.05.050.
multi-family residential buildings using support vector regression: investigating the
[31] Yuan T, Zhu N, Shi Y, Chang C, Yang K, Ding Y. Sample data selection method for
impact of temporal and spatial monitoring granularity on performance accuracy.
improving the prediction accuracy of the heating energy consumption. Energy
Appl Energy 2014;123:168–78. https://doi.org/10.1016/j.apenergy.2014.02.057.
Build 2018;158:234–43. https://doi.org/10.1016/j.enbuild.2017.10.006.
[5] Conti J, Holtberg P, Diefenderfer J, LaRose A, Turnure JT, Westfall L. International
[32] Divina F, Gilson A, Goméz-Vela F, Torres MG, Torres JF. Stacking ensemble
energy outlook 2016 with projections to 2040. 2016. https://doi.org/10.2172/
learning for short-term electricity consumption forecasting. Energies 2018;11:
1296780.
1–31. https://doi.org/10.3390/en11040949.
[6] Amber KP, Ahmad R, Aslam MW, Kousar A, Usman M, Khan MS. Intelligent
techniques for forecasting electricity consumption of buildings. Energy 2018;157:
886–93. https://doi.org/10.1016/j.energy.2018.05.155.

20
N. Somu et al. Renewable and Sustainable Energy Reviews 137 (2021) 110591

[33] Ahmad T, Chen H. Utility companies strategy for short-term energy demand term memory neural network for wind speed forecasting. Energy Convers Manag
forecasting using machine learning based models. Sustain Cities Soc 2018;39: 2019;185:783–99. https://doi.org/10.1016/j.enconman.2019.02.018.
401–17. https://doi.org/10.1016/j.scs.2018.03.002. [45] Song X, Liu Y, Xue L, Wang J, Zhang J, Wang J, et al. Time-series well performance
[34] Goudarzi S, Anisi MH, Kama N, Doctor F, Soleymani SA, Sangaiah AK. Predictive prediction based on Long Short-Term Memory (LSTM) neural network model.
modelling of building energy consumption based on a hybrid nature-inspired J Petrol Sci Eng 2019;186:106682. https://doi.org/10.1016/j.petrol.2019.106682.
optimization algorithm. Energy Build 2019;196:83–93. https://doi.org/10.1016/j. [46] Bai Y, Zeng B, Li C, Zhang J. An ensemble long short-term memory neural network
enbuild.2019.05.031. for hourly PM2.5 concentration forecasting. Chemosphere 2019;222:286–94.
[35] Bedi J, Toshniwal D. Deep learning framework to forecast electricity demand. Appl https://doi.org/10.1016/j.chemosphere.2019.01.121.
Energy 2019;238:1312–26. https://doi.org/10.1016/j.apenergy.2019.01.113. [47] Yann LeCun YBGH. Deep learning 2015:436–44. https://doi.org/10.1038/
[36] Kim Y, Son H gu, Kim S. Short term electricity load forecasting for institutional nature14539.
buildings. Energy Rep 2019;5:1270–80. https://doi.org/10.1016/j. [48] Li J, Li X, He D. A directed acyclic graph network combined with CNN and LSTM
egyr.2019.08.086. for remaining useful life prediction. IEEE Access 2019;7:75464–75. https://doi.
[37] Dagdougui H, Bagheri F, Le H, Dessaint L. Neural network model for short-term org/10.1109/ACCESS.2019.2919566.
and very-short-term load forecasting in district buildings. Energy Build 2019;203. [49] Lu X, Lin P, Cheng S, Lin Y, Chen Z, Wu L, et al. Fault diagnosis for photovoltaic
https://doi.org/10.1016/j.enbuild.2019.109408. array based on convolutional neural network and electrical time series graph.
[38] Wen L, Zhou K, Yang S. Load demand forecasting of residential buildings using a Energy Convers Manag 2019;196:950–65. https://doi.org/10.1016/j.
deep learning model. Elec Power Syst Res 2020;179:106073. https://doi.org/ enconman.2019.06.062.
10.1016/j.epsr.2019.106073. [50] Sadaei HJ, de Lima e Silva PC, Guimarães FG, Lee MH. Short-term load forecasting
[39] Wei Y, Zhang X, Shi Y, Xia L, Pan S, Wu J, et al. A review of data-driven approaches by using a combined method of convolutional neural networks and fuzzy time
for prediction and classification of building energy consumption. Renew Sustain series. Energy 2019;175:365–77. https://doi.org/10.1016/j.energy.2019.03.081.
Energy Rev 2018;82:1027–47. https://doi.org/10.1016/j.rser.2017.09.108. [51] Aghabozorgi S, Ying Wah T, Herawan T, Jalab HA, Shaygan MA, Jalali A. A hybrid
[40] Molina-Solana M, Ros M, Ruiz MD, Gómez-Romero J, Martin-Bautista MJ. Data algorithm for clustering of time series data based on affinity search technique. Sci
science for building energy management: a review. Renew Sustain Energy Rev World J 2014;2014. https://doi.org/10.1155/2014/562194.
2017;70:598–609. https://doi.org/10.1016/j.rser.2016.11.132. [52] Minnaar A. Time series classification and clustering with Python | alex minnaar’s
[41] Kim TY, Cho SB. Predicting residential energy consumption using CNN-LSTM blog. 2014.
neural networks. Energy 2019;182:72–81. https://doi.org/10.1016/j. [53] Rath TM, Manmatha R. Lower-bounding of dynamic time warping distances for
energy.2019.05.230. multivariate time series 2 . Multivariate time series extension. 2002.
[42] Hochreiter S, Schmidhuber J. Long short-term memory. Neural Comput 1997;9: [54] Karydis I, Nanopoulos A, Papadopoulos AN, Manolopoulos Y. Evaluation of
1735–80. https://doi.org/10.1162/neco.1997.9.8.1735. similarity searching methods for music data in P2P networks. Int J Bus Intell Data
[43] Ghimire S, Deo RC, Raj N, Mi J. Deep solar radiation forecasting with convolutional Min 2005;1:210–28. https://doi.org/10.1504/IJBIDM.2005.008363.
neural network and long short-term memory network algorithms. Appl Energy [55] Somu N, GR MR, Ramamritham K. A hybrid model for building energy
2019;253:113451–70. https://doi.org/10.1016/j.apenergy.2019.113541. consumption forecasting using long short term memory networks. Appl Energy
[44] Chen Y, Zhang S, Zhang W, Peng J, Cai Y. Multifactor spatio-temporal correlation 2020;261:114131. https://doi.org/10.1016/j.apenergy.2019.114131.
model based on a combination of convolutional neural network and long short- [56] Laboratory SEI. Academic building dataset. n.d. http://seil.cse.iitb.ac.in/datasets/.
[Accessed 4 December 2019].

21