Energy 288 (2024) 129795

Contents lists available at ScienceDirect

Energy
journal homepage: www.elsevier.com/locate/energy

An evolutionary deep learning model based on EWKM, random forest

algorithm, SSA and BiLSTM for building energy consumption prediction
Lei Lei a, c, Suola Shao a, c, *, Lixia Liang b, c
a
School of Civil Engineering and Architecture, Zhejiang Sci-Tech University, Hangzhou, 310018, China
b
Zhejiang Academy of Building Research and Design Company Limited, Hangzhou, 310018, China
c
Zhejiang Engineering Research Center of Green and Low Carbon Technologies in Buildings, Hangzhou, 310018, China

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

Handling Editor: X Zhao Accurate prediction of building energy consumption is crucial to the rational scheme of building energy.
Combining the entropy-weighted K-means (EWKM) with the random forest (RF) method, a feature selection
Keywords: (EWKM-RF) method is proposed in this paper. Based on the proposed EWKM-RF method, the classification and
Building energy consumption prediction feature selection of the energy consumption influencing factors can be achieved exclusively. Meanwhile, based
Deep learning
on the EWKM-RF method and the bi-directional long-short-term memory neural network (BiLSTM) optimized by
Bidirectional long short-term memory
the sparrow search algorithm (SSA), an RF-SSA-BiLSTM prediction model for building energy consumption is
Feature selection
Sparrow search algorithm established in this paper. As the weight, learning rate, and hidden layer node parameters of the BiLSTM neural
network are optimized with the SSA, the constraints for manually adjusting parameters are avoided in the
proposed prediction model. To examine the accuracy of the proposed model, energy consumption data of a civil
public building in Dalian city are collected and tested. Results show the prediction error of RF-SSA-BiLSTM after
feature selection is reduced by 24.55 % in high and low energy consumption months. Compared with RF-BiLSTM,
RF-PSO-BiLSTM, and RF-CNN-BiLSTM, the RF-SSA-BiLSTM has strong robustness. The average MAE, RMSE and
MAPE values of energy consumption prediction in the four seasons are 1.30, 1.63 and 0.02.

1. Introduction data-driven method extracts hidden rules and patterns from a large
number of historical energy consumption data, which has higher pre­
According to the report of the International Energy Agency (IEA), the diction accuracy and wider application scenarios. Therefore,
energy consumption of the construction industry accounts for more than data-driven methods are popular in the field of building energy con­
33 % of the total global energy consumption [1,2]. Among existing sumption prediction.
technologies, building energy consumption prediction is an effective Data-driven methods can be divided into statistical model methods
means to improve energy efficiency, playing an important role in [17–19] and machine learning methods [20–24]. In recent years, ma­
ensuring the smooth operation of building energy systems and formu­ chine learning methods have become popular in building energy con­
lating reasonable operation strategies for building managers [3–5]. sumption predictions for high prediction accuracy and the ability to
The building energy consumption prediction methods are divided solve nonlinear problems. In big sample data sets and nonlinear data
into two categories: physical modeling methods and data-driven sets, the machine learning methods have higher prediction accuracy
methods [6–11]. Based on physical principles and engineering exper­ than the statistical methods. Compared with the shallow machine
tise, the physical modeling method analyzes building energy consump­ learning model, the deep learning model [25–30] has more layers of
tion by calculating the relationship between energy consumption network structure and can extract more information when processing
influencing factors and building energy consumption. It has been widely periodic energy consumption data. Nowadays, building energy con­
used in common building energy simulation software, such as Ener­ sumption prediction models have developed from traditional physical
gyplus, eQuest, DesT, and DOE-2 [12–15]. However, when detailed and models and statistical models to deep learning models. The deep
accurate building and environmental input parameters cannot be ob­ learning model can predict energy consumption quickly and accurately,
tained, the model performance will be poor [16]. In contrast, the but there are a large number of hyperparameters in the model that need

* Corresponding author. School of Civil Engineering and Architecture, Zhejiang Sci-Tech University, Hangzhou, 310018, China.
E-mail address: [email protected] (S. Shao).

https://doi.org/10.1016/j.energy.2023.129795
Received 4 July 2023; Received in revised form 26 September 2023; Accepted 25 November 2023
Available online 29 November 2023
0360-5442/© 2023 Elsevier Ltd. All rights reserved.
L. Lei et al. Energy 288 (2024) 129795

Fig. 1. Research framework.

2
L. Lei et al. Energy 288 (2024) 129795

to be set. The setting of hyperparameters will have a direct impact on the out the important energy consumption influencing factors and reduce
results of the prediction model. The traditional trial and error method is the dimension of the input parameters of the prediction model. Finally,
less efficient in setting parameters, and so is the enumeration method. the SSA is used to optimize the BiLSTM, and the SSA-BiLSTM building
Introducing an intelligent optimization algorithm is an effective solution energy consumption prediction model is established. The prediction
to solve the problem of hyperparameter setting in a prediction model. process is shown in Fig. 1.
Wang et al. [31] used the improved artificial ecosystem-based optimi­
zation (AEO) algorithm to optimize the parameters of the time convo­ 2.1. Cluster analysis of building energy consumption influencing factors
lution network. Compared to traditional methods such as manual tuning
and grid search, the prediction model has higher computational effi­ EWKM is a fast and efficient high-dimensional data clustering
ciency and prediction performance. Qiao et al. [32] used the improved method [37]. Compared with the hard subspace clustering method, the
Aquila Optimizer algorithm to optimize multiple parameters in the time clustering accuracy has been improved, and it has better stability and
convolution network to further improve the performance of the pre­ flexibility. EWKM solves the problem that traditional clustering methods
diction model. Suo et al. [33] used an improved Chimp Optimization cannot reveal the potential relationship between influencing factors
algorithm to optimize the learning rate and hidden layer of the when dealing with energy consumption data by introducing Mahala­
Bi-directional Gated Recurrent Unit. The results show that the model has nobis distance. After collecting the energy consumption influencing
the best prediction performance in comparison with other benchmark factors of n buildings, the data set X = {x1, …,xj, …,xn} is obtained,
models. In this study, a new sparrow search algorithm is used to opti­ where xj=(xj1, …,xji, …,xjm) is the m energy consumption influencing
mize parameters in the energy consumption prediction model. factors set of the jth building. The clustering process of building energy
Compared with other traditional optimization algorithms, this algo­ consumption influencing factors is: first determine the number of clus­
rithm has the characteristics of fast convergence speed and strong global ters K and the positive parameterγ, and set the initial weight of all
search ability. influencing factors to 1/m. Then, the objective function F(W, Z, Λ) is
In addition, building energy consumption data has strong periodicity calculated by updating the partition matrix W, cluster centers Z and
and is affected by many characteristics, such as building envelope dimension weights Λ. Finally, until F(W, Z, Λ) reaches the local mini­
characteristics, meteorological characteristics, and occupant behavior mum, the clustering results and the weight of the influencing factors are
characteristics. The influencing factors increase the computational output. The expressions of the objective function F(W, Z, Λ) and the
burden of the prediction model and affect the accuracy of the prediction influence factor weight λki are as follows:
results. To solve this problem, feature selection is widely used in [ ]
∑K ∑n ∑ m ∑
m
building energy consumption prediction research [34]. Qiao et al. [35]. F(W, Z, Λ) =
( )2
ξkj λki zki − xji + τ λki log10 λki (1)
used EMD and Boruta feature selection (BFS) methods to select the k=1 j=1 i=1 i=1
features of the influencing factors. Results show that the proposed
( )
method can effectively identify the influencing factors with significant Dki

correlation. The accuracy of the prediction model is significantly

− γ
e
improved. Ding et al. [36]. used the recursive feature elimination (RFE) λki = ( ) (2)
method to delete redundant features. Results show that too many
Dki
∑
m − γ
e
building energy consumption influencing factors will not continuously i=1
improve the prediction accuracy. This is because the upper limit of the
performance of the prediction model is limited by the quality of the data ∑
n
( )2
Dki = ωkj zki − xji (3)
set, but the prediction model can achieve the best performance by
j=1
optimizing the algorithm and feature selection. At present, how to use
appropriate feature selection methods to process energy consumption where W is the partition matrix, Z is cluster centers, Λ is dimension
influencing factors and improve the accuracy and calculation speed of weights, k is the number of clusters, 1≤k ≤ K, j is the number of
prediction models is a key problem. buildings, 1≤j ≤ n, i is the number of energy consumption influencing
Based on the entropy weighting k-means (EWKM) algorithm coupled factors, 1≤i ≤ m; ξkj is the membership degree of the jth influence factor
with the random forest (RF) method and the bidirectional long-short- set belonging to the kth cluster, λki is the weight of the first influence
term memory (BiLSTM) neural network optimized by the sparrow factor in the kth cluster, zki is the cluster center of the first influence
search algorithm (SSA), a building energy consumption prediction factor in the kth cluster, xji is the first influence factor in the building, the
model is proposed in this study to solve the problems existing in the positive parameter τ is used for the size of control right redistribution,
existing research and improve the accuracy of prediction models of Dki is the sum of the distances from the ith influence factor to the cluster
building energy consumption. Compared with the traditional prediction center zki in the kth cluster.
method, this study uses the EWKM coupled random forest method to Since the selection of the number of clusters k and the weight dis­
select the impact factors of building energy consumption, which reduces tribution coefficient λ will directly affect the clustering results of the
the data dimension and improves the overall calculation efficiency of the influencing factors, the internal verification method Davies-Bouldin
prediction model. To fully leverage the time characteristics of energy index (DBI) is used to determine the optimal values of k and λ. DBI
consumption data, BiLSTM is employed for energy consumption data evaluates the clustering effect by calculating the ratio of the distance
prediction. BiLSTM utilizes both historical and future energy con­ between different clusters to the internal distance of different clusters.
sumption data for training to enhance the accuracy of energy con­ The smaller the DBI value, the better the clustering effect. The DBI
sumption data prediction. Addressing the influence of subjective factors calculation formula is as follows:
on the selection of prediction model parameters, BiLSTM is optimized
using the SSA algorithm to alleviate the burden of manual parameter 1 ∑K
Si + Sj
DBI = max (4)
adjustment and enable automatic parameter adjustment, thereby K i=1 j∕
=i ‖ zi − zj ‖2

enhancing the model’s generalization ability and practicality.

Among them, K is the number of clusters, Si and Sj are the average
Euclidean distance from the influence factor to the cluster centers zi and
2. Methodology
zj in the ith and jth clusters, and ‖zi-zj‖2 is the Euclidean distance be­
tween the ith cluster center and the jth cluster center.
In this paper, the EWKM is used to classify the influencing factors of
building energy consumption. Then, the RF algorithm is used to screen

3
L. Lei et al. Energy 288 (2024) 129795

influencing factors with importance scores lower than or equal to zero

are eliminated to reduce the data dimension.

2.3. Building energy consumption prediction model

In this study, the sparrow search algorithm is used to optimize the

parameters of the bidirectional long short-term memory network, aim­
ing to improve the prediction accuracy of building energy consumption.
By embedding the sparrow search algorithm into the training process of
BiLSTM, the automatic optimization of model parameters is realized,
and the feasibility and effectiveness of the prediction model are
improved.

2.3.1. Sparrow search algorithm

In 2020, Xue et al. [40] proposed a group intelligence optimization
algorithm, the sparrow search algorithm (SSA), inspired by the foraging
and anti-predatory behavior of sparrows. The algorithm has the ad­
vantages of fast convergence speed, strong global search ability, and
fewer setting parameters. In the search process, the sparrow population
is divided into two categories according to a certain proportion: the
producers and the scroungers. The producers are responsible for
adjusting the search direction and scope of the entire population, and
the scroungers searches by using the search information of the pro­
ducers. The flowchart of the SSA is shown in Fig. 2. For the C population
consisting of b sparrows as follows:
⎡ ⎤
c1,1 c1,2 … c1,d
Fig. 2. The flowchart of the sparrow search algorithm.
⎢ c2,1 c2,2 … c2,d ⎥
C=⎢ ⎣… … … …⎦
⎥ (6)
cb,1 cb,2 … cb,d
2.2. Feature selection of building energy consumption influencing factors
where b is the number of sparrows, d is the number of parameters to be
The redundant features and irrelevant features in the building energy
optimized in the neural network, cb,d is the position information for the
consumption influencing factors will increase the complexity of machine
bth only responsible for the dth parameter search sparrow The fitness
learning. Therefore, the selection of building energy consumption
function Gc of all sparrows is:
influencing factors helps to reduce the computational cost and moder­
⎡ ([ ]⎤
ately improve the prediction performance of the model. Random forest g c1,1 (c[1,2 ⋯ c1,d
is an ensemble learning algorithm, which was first proposed by Brieman ⎢
Gc = ⎢ ] g c2,1 c2,2 ⋯ ⎥ ⎥ (7)
in 2001 [38]. This method can reduce the dimension of the influencing ⎣ c2,d ⋮ ⎦
( [⋮
factors by feature selection. Because random forests repeatedly sample ⋯ ⋮ g cb,1 cb,2
some influencing factors when constructing decision trees, random
forests can effectively avoid overfitting and improve the generalization where g is the fitness value. Producers with better fitness in the search
performance of the model compared with other methods for calculating process will preferentially obtain food and provide foraging directions
the importance of features [39]. In the training process of a single de­ for other scroungers. The iterative update formula of the location of the
cision tree, the unused influence factors are called out of the bag (OOB). producers is:
Using the out-of-bag data, the feature importance score of each influ­ ⎧
⎪
(
i
)
⎨ ca exp −
encing factor under the current decision tree can be calculated. Finally, , R< ST
(8)
i,j
ca+1
i,j =
α itermax
by sorting the feature importance scores of all decision trees in a random ⎪
⎩ cai,j + QL, R≥ ST
forest, the importance ranking of influencing factors is obtained.
In the process of feature importance calculation, the OBB error is
where a is the number of current iterations, itermax is the maximum
calculated for N decision trees in the random forest, and the OBB error of
number of iterations, ca i,j is the ith sparrow to the jth parameter opti­
the ith influence factor is recorded as E1 i. Secondly, under the condition
mization in ath iteration, 1≤i ≤ b, 1≤j ≤ d. The α is a random number of
that other influence factors remain unchanged, the selected influence
(0,1]. The R is the warning value of [0,1]. The ST is the safety value of
factor value is randomly modified, and the OBB error is calculated again,
[0.5,1]. The Q is a random value of the normal distribution. The L is a
which is recorded as E2 i. Finally, by comparing the difference between
matrix with dimension and all elements are 1. When R < ST, there is no
E1 i and E2 i, the importance score of the influencing factors is obtained.
predator around, and the producers continue to search. When R ≥ ST, it
The formula is as follows:
means that there are predators around, and all sparrows search for other
1 ∑
N
( 2 ) safe areas. The scroungers positions were updated as follows:
VI = Ei − Ei1 (5) ( a )
N i ⎧ cworst − cai,j
⎪
⎨ Q⋅exp , if i > b/2
2
According to Eq. (5), if the value of E2 i increases significantly, it ca+1
i,j =
i (9)
⎪ ⃒ ⃒
indicates that this feature has an important influencing factor on energy ⎩ a+1 ⃒ a
cP + ⃒ci,j − ca+1
⃒ +
⃒⋅A ⋅L, otherwise
consumption. If the E2 ivalue does not change significantly, it shows that
P

the feature is not related or negatively correlated with energy con­

where cp is the optimal position of the current producers. The cworst is the
sumption. The importance scores of all energy consumption influencing
current global worst position. The A is an 1 × d dimensional matrix,
factors are sorted in descending order, and the energy consumption
where each element is randomly assigned − 1 or 1, and A+ = AT(AAT)− 1.

4
L. Lei et al. Energy 288 (2024) 129795

⎧
⎪
⎪ ( ) ( )
⎨→ ̅̅̅→ → ← ← ←
h t = LSTM h t− 1 , xt , t ∈ [1, T]h t = L STM h t− 1 , xt , t ∈ [T, 1] (11)
⎪
⎪
⎩

→ ←
ybt =μ h t + ν h t (12)

→ ←
where h t and h t are the energy consumption output at the time t of the
→ ←
forward LSTM and the backward LSTM respectively. h t− 1 and h t− 1 are
Fig. 3. Structure of the BiLSTM network. the energy consumption output of the forward LSTM and the backward
LSTM at time t-1. The yb t is the energy consumption prediction output
of the BiLSTM. The values of forward factorμand backward factorνare
obtained by training the energy consumption data. The T is the length of
the time series. The LSTM structure of each layer in the BiLSTM is shown
in Fig. 4, including the forgetting gate, input gate, output gate, and unit
state.

2.4. Performance evaluation of building energy consumption prediction

model

A variety of performance indicators is used to comprehensively

evaluate the performance of the energy consumption prediction model.
It includes the mean absolute error (MAE), the root mean square error
(RMSE), the mean absolute percentage error (MAPE), and the coefficient
of variation of the root mean squared error (CV-RMSE). The MAE index
is used to measure the error between the predicted value and the actual
value of energy consumption. The RMSE index is used to reflect the
Fig. 4. Structure of the LSTM.
degree of deviation between the predicted energy consumption value
and the actual value of the prediction model under the same data set.
The MAPE index is used to evaluate the percentage of error between the
If the current individual fitness value and the current best fitness
predicted value of energy consumption and the actual value. The CV-
value fall into the local optimum, the sparrow population will update its
RMSE index is a root mean square error index for energy consumption
position according to the following formula:
data under cross-validation, which reflects the deviation degree of en­
⃒ ⃒
⎧ Ca + p⋅⃒⃒Ca − Ca ⃒⃒, g > g ergy consumption data error.
i b
⎪
⎨ best i,j best
⃒ 1∑ n
a+1
Ci,j = (⃒⃒Ca − Ca ) (10) MAE = yi|
|yi − ̂ (18)
⎪
⎩ a i,j worst n i=1
Ci,j + U⋅ , gi = gb
(fi − fw ) + ε
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
1∑ n
where cbest is the current global optimal position. The p is the step control RMSE = (yi − ̂y i )2 (19)
n i=1
parameter, which is a normal distribution random number with a mean
of 0 and a variance of 1. U ∈ [-1,1] is the direction of sparrow movement.
n ⃒ ⃒
The gi is the individual fitness value of the current sparrow. The gw is the 1∑ ⃒̂ ⃒
⃒ y i − yi ⃒
MAPE = (20)
global worst fitness value. The gb is the global optimal fitness value. The n i=1 ⃒ ̂y i ⃒
ԑ is a very small constant. √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ /
1∑ n
1∑ n

2.3.2. Bidirectional long short-term memory neural network CVRMSE = (yi − ̂ y i )2 yi

̂ (21)
n i=1 n i=1
The BiLSTM neural network proposed by Graves and Schmidhuber in
2005 combines forward propagation LSTM network and backward where n is the total energy consumption data. The yi is the predicted
propagation LSTM network [41]. The bidirectional mechanism helps to value of the ith energy consumption. The ̂
y i is the actual value of the ith
comprehensively utilize the characteristics of energy consumption time energy consumption.
series data and can process historical and future energy consumption
data separately, improving the accuracy of energy consumption data
prediction. The structure of the BiLSTM neural network is shown in 2.5. The process of predicting energy consumption in buildings
Fig. 3.
Compared with the traditional LSTM, the BiLSTM can simulta­ First, the EWKM algorithm is applied to cluster the normalized en­
neously process the energy consumption data in two directions before ergy consumption data and assign corresponding category labels. Then,
and after the time series. Forward LSTM uses past energy consumption the random forest algorithm is used to eliminate the redundant energy
to predict future energy consumption, while backward LSTM uses future consumption influencing factors. The factors that have effects on
energy consumption to fit past energy consumption. Through the bidi­ building energy consumption are selected. Finally, the SSA is used to
rectional transmission mode of the BiLSTM, the dependence between optimize the weight attenuation coefficient, learning rate, and the
energy consumption data and influencing factors can be effectively number of hidden layer nodes of the BiLSTM neural network. The search
captured and extracted. The output of the BiLSTM is as follows: range for the weight attenuation coefficient of the BiLSTM neural
network is [10− 10, 10− 2], the search range for the learning rate is
[0.0001, 0.002], and the search range for the number of hidden layer

5
L. Lei et al. Energy 288 (2024) 129795

Fig. 5. Energy consumption prediction process.

Table 1
Description of variables from the original dataset.
Category No. Variable Abbreviation Unit Range
o
Meteorological parameter 1 Outdoor temperature OT C [26.86,38.52]
2 Relative humidity RH % [34.1,97.4]
3 Wind speed WS m/s [0.2,1.2]
4 Solar radiation SR W/m2 [55.6,566.7]
Architectural information 5 Number of floors NF – [4,7]
6 Gross floor area GFA m2 [3051,9856]
7 Building orientation BO – [231,285]
8 Window wall ratio WWR % [0.14,0.26]
9 Heat transfer coefficient of external walls HTCW W/(m2⋅K) [0.6,1.56]
10 Shading coefficient SC – [0.49,0.96]
11 Building length-to-width ratio BLR – [1.5,3.19]
12 Heat transfer coefficient of roof HTCR W/(m2⋅K) [0.5,1]
Occupancy activity 13 Personnel density PD person/m2 [0.25,1.2]
Building services and energy systems 14 Lighting power density LPD W/m2 [4.69,15.9]
15 Fresh air volume per capita FAVPC m3/h. person [26,52]
16 COP COP – [4.89,6.75]
o
17 Air supply temperature AST C [17.2,26.1]
18 Fan efficiency FE % [0.55,0.91]
19 Water pump efficiency WPE % [0.29,0.91]
o
Indoor environment 20 Room temperature RT C [18.16,26.75]

nodes is [10, 50]. The RF-SSA-BiLSTM energy consumption prediction university in Dalian are used to train and test the energy consumption
model is established. The main workflow of building energy consump­ prediction model.
tion prediction is shown in Fig. 5.
3.1. Data collection for the selection of energy consumption influencing
3. Data collection factors characteristics

The measured data is divided into two parts: The relevant data In this study, data collection was conducted on 100 civil public
collected from 100 civil public buildings in Dalian are applied to the buildings in Dalian. Using building electricity consumption as the
cluster analysis and feature selection of energy consumption influencing building energy consumption value, a total of 100 sets of data on
factors. The energy consumption data of an experimental building in a building electricity consumption and 20 building energy consumption

6
L. Lei et al. Energy 288 (2024) 129795

Fig. 6. Dalian experimental building electricity consumption: (a) Hourly energy consumption; (b) Monthly electricity consumption.

influencing factors were collected. The time interval for data collection 3.2. Data collection of building energy consumption prediction
is 1 h. Twenty building energy consumption influencing factors were
collected from five aspects: Architectural information, meteorological The electricity consumption data of an experimental building in a
parameter, occupancy activity, building services and energy systems, university in Dalian is collected as the test data of the energy con­
and indoor environment, as shown in Table 1. The architectural infor­ sumption prediction model. The experimental building is a frame shear
mation is obtained according to the measured data of each civil public structure, five floors above ground and one floor underground. The total
building, including the number of floors, gross floor area, building area of the building is 9533 m2, and the air conditioning area is 8580 m2.
orientation, window wall ratio, heat transfer coefficient of external From 1 January 2018 to 7 December 2018, the hourly energy con­
walls, shading coefficient, building length-to-width ratio, and heat sumption data of the experimental building were collected, and 8176
transfer coefficient of the roof. The meteorological parameters were sets of valid data were obtained, as shown in Fig. 6 (a). Fig. 6 (b) shows
collected during the test period, including outdoor temperature, relative the distribution of electricity consumption in different months of the
humidity, wind speed, and solar radiation. The building services and experimental building. In summer, the power consumption increases,
energy systems are measured and calculated, including lighting power and in spring, the power consumption is less. Especially in February, it is
density, fresh air volume per capita, COP, air supply temperature, fan the Chinese Spring Festival holiday, and the number of people on
efficiency, and water pump efficiency. The personnel density is based on campus decreases, which becomes the lowest month of electricity con­
investigation and calculation, which is used to represent the occupancy sumption. Fig. 7 is the hourly meteorological data of the area where the
activity. The room temperature is measured during the collection experimental building is located. In the short-term prediction, the en­
period, which is used to represent the indoor environment. ergy consumption data of the last week of the first month of each quarter
is used to test the prediction model, and the remaining data is used as the
training set. The ratio of the training set to the test set is about 3:1. In the

7
L. Lei et al. Energy 288 (2024) 129795

Fig. 7. Meteorological parameters of Dalian in 2018: (a) hourly outdoor air temperature, (b) hourly relative humidity, (c) hourly solar horizontal radiation, (d)
hourly wind speed.

value of the meteorological data in a day is used as the meteorological

parameters of the day. There are 341 sets of data, of which 311 sets are
used for training and 30 sets of data are used for testing.

4. Results and discussion

4.1. Cluster analysis of building energy consumption influencing factors

After normalizing the 20 energy consumption influencing factors,

the EWKM clustering algorithm is used to cluster the energy consump­
tion influencing factors. The implementation of the EWKM needs to
define the number of clusters k, weight distribution parameters λ, and
convergence value δ. In general, it is a scientific and effective way to
determine the number of clusters by calculating the clustering perfor­
mance index to avoid subjective division when the data set does not
have a clear number of clusters. The search range of the number of
clusters k is set to the interval [2,9], and DBI is selected as the perfor­
mance index of clustering. According to Reference [48], the search
range of weight distribution parameters λ can be set as the interval, and
the convergence δ value is 0.00001. The number of clusters k and the
weight distribution parameters λ are searched exhaustively to calculate
the DBI values of different parameter combinations. The number of
clusters corresponding to the minimum DBI value is the optimal number
of clusters. The clustering parameter optimization results of the energy
consumption influencing factors are shown in Fig. 8 (a) is the DBI value
under different combinations of cluster number k and distribution pa­
rameters λ. As the number of clusters increases, the DBI value shows an
increasing trend, which indicates that too many clusters will not
improve the clustering effect. It can be seen from Fig. 8 (b) that the
optimal parameter combination is the number of clusters k = 4, the
Fig. 8. EWKM clustering performance: (a) Optimization process of the EWKM weight distribution parameter λ = 2.2, and the DBI index is the global
parameter k; (b) DBI value when the number of clusters k is 4. minimum of 5.96, indicating that the clustering effect is the best.
Therefore, 100 buildings are divided into four categories.
medium-term prediction, the hourly data are processed into daily data. To present the clustering results more intuitively, the power con­
For example, the hourly energy consumption in a day is added to obtain sumption during the test period is selected as the clustering results. As an
the daily building energy consumption. At the same time, the average intuitive representation of the actual energy consumption of buildings,

8
L. Lei et al. Energy 288 (2024) 129795

Fig. 9. Electricity consumption clustering results.

Fig. 11. Ranking the importance of variables.

influence of different decision tree numbers on the feature importance

results, and weigh the calculation time and calculation performance, the
search range of the number of decision trees is set to [1, 500].
Fig. 10. Random forest training process. Comparing the calculation results of each out-of-bag data sample with
the category labels, the relationship curve between the number of de­
electricity consumption can effectively reveal the energy consumption cision trees and the Out-of-Bag Error is obtained. The results are shown
differences between different types of buildings. Fig. 9 is the violin di­ in Fig. 10. When the number of decision trees is in the range of [0,47],
agram of building electricity consumption. The energy consumption with the increased number of decision trees, the OOB error shows a
influencing factors of similar buildings are similar, with similar elec­ downward trend. When the number of decision trees exceeds 47, the
tricity consumption behavior and high similarity of electricity con­ OOB error shows a slow upward trend, and then it is in a stable state. It
sumption. The energy consumption influencing factors of different can be seen that when the number of decision trees mtree = 47, the OOB
categories of buildings are quite different, resulting in significant dif­ error of the decision tree model reaches the global minimum of 0.11.
ferences in building electricity consumption behavior and electricity Therefore, the optimal parameter combination of the number of split
consumption. According to Fig. 9, the energy consumption data of features and the number of decision trees in this study is mtry = 4, mtree
Category A buildings and Category D buildings show a single peak dis­ = 47.
tribution. The electricity consumption of Category A buildings is To avoid the error caused by randomness in the random forest al­
concentrated at about 35 kW h, and the electricity consumption of gorithm, the feature importance of the influencing factors is repeated
Category D buildings is concentrated at about 100 kW h. The energy five times. To ensure the reliability of the results, the average value of
consumption data of Category B buildings and Category C buildings the five calculation results is taken as the basis for the sorting and the
show a bimodal distribution. The power consumption of Category B feature selection of the importance score. The importance scores of en­
buildings is concentrated between 50 kW h and 70 kW h, and the power ergy consumption influencing factors are sorted in descending order,
consumption of Category C buildings is concentrated between 80 kW h and the results are shown in Fig. 11. There are significant differences in
and 100 kW h. The maximum power consumption of 42 Category A the feature importance scores of different influencing factors. Among
buildings during the test period is 48.8 kW h, the minimum power them, the characteristic importance scores of relative humidity (RH),
consumption is 32.4 kW h, and the average power consumption is 37.0 outdoor temperature (OT), shading coefficient (SC), room temperature
kW h. The power consumption distribution range of 17 Category B (RT), air supply temperature (AST), window wall ratio (WWR), building
buildings is [49.6,73.2] kW⋅h, 25 Category C buildings is [75.2,98.4] length-to-width ratio (BLR), fan efficiency (FE), COP, wind speed (WS),
kW⋅h, and 16 Category D buildings is [101.2,148.8] kW⋅h. building orientation (BO), heat transfer coefficient of roof (HTCR) and
water pump efficiency (WPE) were all greater than 0, which were
identified as important energy consumption influencing factors. The
4.2. Feature selection of energy consumption influencing factors
importance scores of lighting power density (LPD), heat transfer coef­
ficient of external walls (HTCW), Fresh air volume per capita (FAVPC),
Based on the clustering results of energy consumption influencing
solar radiation (SR), number of floors (NF), personnel density (PD) and
factors, 20 energy consumption influencing factors and category labels gross floor area (GFA) were all less than 0, which were identified as
are input into the random forest algorithm to calculate the importance
redundant energy consumption influencing factors.
score of each energy consumption influencing factor and perform
The feature information collected in this study comes from building
feature selection. The implementation of the random forest algorithm
energy consumption monitoring platforms and on-site testing. The
needs to set two key parameters: the number of split features mtry and
collected feature information should be as complete as possible to
the number of decision trees mtree. The number of splitting features is
guarantee that the important influencing factors can be selected through
usually determined by an empirical formula, where m is the total
feature selection. If the feature information is incomplete, the redundant
number of influencing factors. There are 20 influencing factors in this
influencing factors in the incomplete feature information will result in
study, so the number of splitting features mtry = 4 is set. The number of
fewer important influencing factors, which reduces the accuracy of the
decision trees can determine the optimal number of decision trees by
prediction model.
exhaustive search. When setting the search range, to fully explore the

9
L. Lei et al. Energy 288 (2024) 129795

Fig. 12. RF-SSA-BiLSTM and SSA-BiLSTM short-term prediction results: (a) Low energy consumption season, (b) High energy consumption season, (c) 80–100 h in
high energy consumption season.

Fig. 13. The prediction errors of RF-SSA-BiLSTM and SSA-BiLSTM: (a) Low energy consumption season, (b) High energy consumption season.

4.3. Building energy consumption prediction and reliability of the building energy consumption prediction model
under different energy consumption levels are evaluated. The short-term
The annual electricity consumption data of an experimental building energy consumption prediction results of the RF-SSA-BiLSTM model and
in a university in Dalian are collected in this study paper to conduct the SSA-BiLSTM model are shown in Fig. 12. Fig. 12 (a) shows the
short-term prediction and medium-term prediction respectively. prediction results of the two prediction models in Spring, in which the
trend of the energy consumption prediction curve is consistent with the
4.3.1. Short-term prediction of building energy consumption actual energy consumption curve. Fig. 12 (b) shows the prediction re­
sults of the two prediction models in Summer. At the peak and trough of
4.3.1.1. Impact of feature selection on building energy consumption pre­ energy consumption, the predicted energy consumption curve is poorly
diction. To evaluate the impact of the feature selection method used in fitted with the actual energy consumption curve. To make the compar­
this study on the performance of the building energy consumption ison between the prediction models clearer and more intuitive, the
prediction model, the SSA-BiLSTM prediction model with feature se­ prediction results from the 80th hour to the 100th hour in Summer are
lection (i.e., RF-SSA-BiLSTM) and the SSA-BiLSTM prediction model enlarged, as shown in Fig. 12 (c). It can be seen from Fig. 12 (c) that
without feature selection are used for short-term prediction of energy there is a significant deviation between the predicted value and the
consumption. By comparing the prediction results of the low energy actual value of the peak energy consumption of the two prediction
consumption season represented by Spring and the high energy con­ models from the 86th hour to the 90th hour. The prediction error be­
sumption season represented by Summer, the prediction performance tween the prediction results of the two models and the actual energy

10
L. Lei et al. Energy 288 (2024) 129795

Table 2
Quantitative evaluation of the RF-SSA-BiLSTM model and SSA-BiLSTM model.
Model Low energy consumption High energy consumption
season season

MAE RMSE MAPE MAE RMSE MAPE

RF-SSA-BiLSTM 1.055 1.270 0.021 1.544 1.846 0.026

SSA-BiLSTM 1.651 1.854 0.035 1.776 2.237 0.028

Table 3
Comparison of calculation time between RF-SSA-BiLSTM and SSA-BiLSTM.
Model The number of input impact factors The time of calculation

RF-SSA-BiLSTM 13 73.78 s
SSA-BiLSTM 20 142.4 s

consumption is shown in Fig. 13. Fig. 13 (a) is the prediction error curve
in Spring, and Fig. 13 (b) is the prediction error curve in Summer. The
prediction error of the RF-SSA-BiLSTM model is smaller than that of the
SSA-BiLSTM model, and it is more stable as a whole, showing better
calculation accuracy. The results show that the prediction model can
capture the key influencing factors more accurately by feature selection
of energy consumption influencing factors, avoid the interference of
redundant influencing factors on the prediction model, achieve more
stable prediction, and improve the reliability and accuracy of the pre­
diction model.
To further evaluate the prediction performance of the RF-SSA-
BiLSTM model and the SSA-BiLSTM model, the MAE, RMSE, and
MAPE of the two models were calculated and compared in Table 2. In
the low energy consumption season, the MAE, RMSE and MAPE indexes
of the RF-SSA-BiLSTM model with the feature selection were
1.055,1.270, and 0.021, respectively. Compared with the SSA-BiLSTM
model without the feature selection, the three indexes decreased by
about 36.1 %, 31.4 %, and 40 %, respectively. In the high energy con­
sumption season, the MAE, RMSE, and MAPE indexes of the RF-SSA-
BiLSTM prediction model were 1.544,1.846, and 0.026, respectively.
The MAE, RMSE, and MAPE indexes of the SSA-BiLSTM prediction
model were 1.776, 2.237, and 0.028, respectively. Three indexes
decreased by about 13.0 %, 17.5 %, and 7.1 %, respectively. Results
show that the feature selection method can effectively simplify the en­
ergy consumption data samples. The prediction performance of the RF-
SSA-BiLSTM prediction model with feature selection is better than that
of the SSA-BiLSTM prediction model, which can predict building energy
consumption more accurately.

4.3.1.2. Performance comparison and analysis of optimization methods in

prediction models. To evaluate the influence of feature selection on the
calculation speed of the building energy consumption prediction model,
the calculation speed of the RF-SSA-BiLSTM prediction model with
feature selection is compared to that of the SSA-BiLSTM prediction Fig. 14. Short-term prediction results of building energy consumption: (a)
model without feature selection. The results are presented in Table 3. Spring, (b) Summer, (c) Fall, (d) Winter.
After the selection of building energy consumption impact factors, the
input impact factors for RF-SSA-BiLSTM are reduced from 20 to 13, BiLSTM, the RF-PSO-BiLSTM, and the RF-CNN-BiLSTM prediction
resulting in a training time of 73.78 s. For the SSA-BiLSTM prediction models, while the building energy consumption is treated as output
model, the building energy consumption impact factor is directly input values. The short-term energy consumption is predicted for the last week
for training without feature selection, and the total time is 142.4 s. RF-
SSA-BiLSTM improves the training speed by 48.2 % compared to SSA-
Table 4
BiLSTM. These results indicate that the prediction model’s training
Energy consumption prediction results of RF-SSA-BiLSTM and RF-BiLSTM.
speed is effectively improved after feature reduction.
Model Index Spring Summer Autumn Winter

4.3.1.3. Performance comparison and analysis of optimization methods in RF-SSA-BiLSTM MAE 1.06 1.54 1.12 1.47
prediction models. In the short-term prediction of building energy con­ RMSE 1.27 1.85 1.50 1.91
MAPE 0.02 0.03 0.02 0.02
sumption, there are 8176 valid data sets, which are divided into test data
RF-BiLSTM MAE 2.40 3.23 2.43 3.20
and training data by quarter. The 13 influencing factors after feature RMSE 2.82 3.45 3.06 3.57
selection are used as the input values of the RF-SSA-BiLSTM, the RF- MAPE 0.07 0.06 0.08 0.06

11
L. Lei et al. Energy 288 (2024) 129795

Fig. 15. Energy consumption prediction results of RF-SSA-BiLSTM and RF-BiLSTM: (a) MAE, (b) RMSE, (c) MAPE.

When dealing with the energy consumption data sets of different sea­
Table 5
sons, the burden of manual parameter adjustment is effectively reduced
Prediction results of RF-SSA-BiLSTM, RF-PSO-BiLSTM and RF-CNN-BiLSTM.
by optimizing the parameters, so that the model can achieve a higher
Model Index Spring Summer Autumn Winter accuracy level.
RF-SSA-BiLSTM MAE 1.06 1.54 1.12 1.47 The calculation results of MAE, RMSE, and MAPE indexes of the RF-
RMSE 1.27 1.85 1.50 1.91 SSA-BiLSTM, the RF-PSO-BiLSTM, and the RF-CNN-BiLSTM prediction
MAPE 0.02 0.03 0.02 0.02
models are shown in Table 5 and Fig. 16. By comparing the MAE indexes
RF-PSO-BiLSTM MAE 1.92 1.56 2.22 1.63
RMSE 2.41 2.04 2.69 2.30
of the RF-SSA-BiLSTM model and the RF-PSO-BiLSTM model in Tables5
MAPE 0.06 0.04 0.06 0.03 and it can be seen that in spring and autumn prediction, the MAE values
RF-CNN-BiLSTM MAE 1.45 1.53 1.79 1.81 of the RF-SSA-BiLSTM model are 1.06 and 1.12, which are 44.8 % and
RMSE 1.67 2.02 2.16 2.30 49.5 % lower than that of the RF-PSO-BiLSTM model. In summer and
MAPE 0.03 0.03 0.04 0.04
winter prediction, the RMSE values of the RF-SSA-BiLSTM model are
1.85 and 1.91, which are 9.31 % and 16.96 % lower than that of the RF-
of the first month of each quarter. The prediction results are shown in PSO-BiLSTM model. The area enclosed by the three indicators of the RF-
Fig. 14. SSA-BiLSTM model in each season is smaller than that of the RF-PSO-
The MAE, RMSE, and MAPE are used to evaluate the performance BiLSTM model in Fig. 16, indicating that the RF-SSA-BiLSTM model
improvement of the BiLSTM neural network optimized by the SSA. The has better prediction performance than the RF-PSO-BiLSTM model.
results of MAE, RMSE, and MAPE of the RF-SSA-BiLSTM and the RF- Compared with the classical PSO algorithm, the SSA has a better effect
BiLSTM prediction models are compared in Table 4. It can be seen on search accuracy, convergence speed, and stability. The strong global
that the MAE index of RF-SSA-BiLSTM in the four seasons decreased by search ability of the SSA avoids the problem that the algorithm falls into
54 % on average, the RMSE index decreased by between 46.3 % and 55 the local optimum in the process of searching for BiLSTM parameters
%, and the MAPE index decreased to varying degrees, especially in the Therefore, the RF-SSA-BiLSTM model can more accurately capture the
autumn forecast. The MAPE index achieved a maximum 75 % reduction. characteristics and long-term trend of building energy consumption data
The prediction performance of the model optimized by the sparrow than the RF-PSO-BiLSTM model, and achieve more accurate prediction.
search algorithm is improved, and the prediction result is closer to the By comparing the prediction performance of the RF-SSA-BiLSTM
real energy consumption. Fig. 15 intuitively shows the index values of model and the RF-CNN-BiLSTM model in Tables5 and it can be seen
the RF-SSA-BiLSTM model and the RF-BiLSTM models. The area size of that the prediction accuracy of the RF-SSA-BiLSTM model and the RF-
the approximate diamond surrounded by each index value can be used CNN-BiLSTM model is similar in summer. However, in the other three
to compare the prediction performance of the model in four quarters. As seasons, the RF-SSA-BiLSTM model has better prediction performance.
shown in Fig. 15 (a), the area enclosed by the MAE index of the RF-SSA- Taking the autumn energy consumption prediction as an example, it can
BiLSTM prediction model in four seasons is much smaller than that of be seen from Table 5 that the RMSE index of the RF-SSA-BiLSTM model
the RF-BiLSTM model. The results show that the SSA improves the optimized by parameters is 1.5, while the RMSE index of the RF-CNN-
processing ability of the RF-SSA-BiLSTM model for complex building BiLSTM model is 2.16. In addition, the area enclosed by the MAPE
energy consumption data by adjusting the parameters of the RF-BiLSTM index of the RF-SSA-BiLSTM model in Fig. 16 (c) is smaller than that of
neural network, and the prediction accuracy is significantly improved. the RF-CNN-BiLSTM model, indicating that the average absolute per­
centage error between the predicted value and the true value of the RF-

12
L. Lei et al. Energy 288 (2024) 129795

Fig. 16. Energy consumption prediction results of RF-SSA-BiLSTM, RF-PSO-BiLSTM and RF-CNN-BiLSTM: (a) MAE, (b) RMSE, (c) MAPE.

Fig. 17. Prediction performance under different data availability for

each method.

SSA-BiLSTM model in the four seasons is lower than that of the RF-CNN-
BiLSTM model. The RF-SSA-BiLSTM model and the RF-CNN-BiLSTM
model are two prediction models with different structures. Although
the RF-CNN-BiLSTM model introduces a convolutional neural network
as a feature extractor for building energy consumption influencing fac­
tors, in the process of feature extraction, it may lose the global infor­
mation of energy consumption data, resulting in a decline in prediction
performance. In contrast, the RF-SSA-BiLSTM model focuses on
parameter adjustment and model fitting capabilities, focusing on the
relationship between energy consumption data before and after the time Fig. 18. Mid-term prediction results of building energy consumption: (a) Pre­
diction results of the four models, (b) Absolute error of medium-
series. Therefore, the performance of the RF-SSA-BiLSTM model is better
term prediction.
than that of the RF-CNN-BiLSTM model, and it has better calculation
accuracy.
prediction models under different data availability. When the avail­
ability of energy consumption data changes, the prediction results are
4.3.1.4. Robustness analysis of prediction model. By comparing the ac­
curacy of the prediction model under different data availability, the always relatively stable, indicating that the prediction model is more
robust. With the increased data availability, the CVRMSE of the RF-SSA-
robustness of the RF-SSA-BiLSTM prediction model is evaluated.
BiLSTM prediction model is stable at about 3 %. In contrast, the RF-
Robustness means that the prediction model can maintain stable and
BiLSTM model has the highest sensitivity to the amount of training
reliable prediction performance in the face of outliers or incomplete
data, and its CVRMSE maximum difference is about 2 %. In addition, the
data. Fig. 17 shows the trend of prediction accuracy of the four
RF-CNN-BiLSTM and RF-PSO-BiLSTM models have similar robustness,

13
L. Lei et al. Energy 288 (2024) 129795

but the CVRMSE values of the two models are higher than those of the (1) EWKM clustering method coupled with a random forest algo­
RF-SSA-BiLSTM model. In summary, when the availability of energy rithm can effectively select the important factors in building en­
consumption data changes, the prediction model that optimizes the ergy consumption. In comparison with the SSA-BiLSTM model,
BiLSTM network parameters through the sparrow search algorithm can the prediction accuracy of the RF-SSA-BiLSTM model after
still obtain accurate and stable prediction results. Among the four feature selection is significantly improved. It shows that the
models, RF-SSA-BiLSTM has the best robustness. feature selection can improve the accuracy and reliability of the
prediction model and play an important role in building energy
4.3.2. Mid-term prediction of building energy consumption consumption prediction.
In the mid-term prediction of building energy consumption, there are (2) After optimizing the BiLSTM neural network with the SSA algo­
341 sets of valid data, of which 311 sets are used for training and 30 sets rithm, the prediction performances of the model are effectively
are used for testing. To further analyze the performance of the proposed improved. The RF-SSA-BiLSTM model shows excellent prediction
RF-SSA-BiLSTM model in the medium-term prediction, the hourly en­ ability when processing data in different seasons. The proposed
ergy consumption data of an experimental building are processed into model avoids the burden of manual parameter adjustment, re­
daily energy consumption data, and the June data of high energy con­ alizes automatic parameter adjustment, and improves the
sumption are selected for 30 days of medium-term prediction. The mid- generalization ability and practicability of the model.
term prediction results of the RF-SSA-BiLSTM model and RF-BiLSTM, (3) The RF-SSA-BiLSTM model can predict building energy con­
RF-PSO-BiLSTM, and RF-CNN-BiLSTM models are shown in Fig. 18 sumption at different time scales. The proposed model not only
(a). The prediction results of the RF-SSA-BiLSTM model have the best provides strong support for decision-making and improves the
fitting degree with the actual energy consumption. The energy con­ efficiency of building energy consumption, but also reveals the
sumption prediction of the RF-BiLSTM model on the 4th, 12th, and 25th impact of building energy consumption influencing factors on
days has an obvious deviation from the actual situation. The prediction electricity consumption through cluster analysis and feature se­
results of the RF-PSO-BiLSTM model on the 9th day have a large error lection. It helps managers identify and evaluate different influ­
with the actual energy consumption, and the prediction results of other encing factors and take corresponding management measures to
days are acceptable. In the building energy consumption prediction from achieve sustainable building operation and management.
the 25th day to the 30th day, the four models all showed different de­
grees of error fluctuation, but the error fluctuation of the RF-SSA- In future research, the decomposition techniques will be applied to
BiLSTM model was the smallest, and the predicted value was closer to process the data to optimize the energy consumption model. The
the actual energy consumption. To more clearly compare the prediction building energy consumption prediction performance for different
accuracy of the four models in the medium term of building energy geographical locations and different types of buildings under various
consumption, the needle diagram is used to analyze the prediction error feature selection methods also can be further explored.
of each model, as shown in Fig. 18 (b). The prediction error of the RF-
BiLSTM model on individual days is far greater than that of other CRediT authorship contribution statement
models, and the fluctuation is large. In contrast, the prediction error of
the RF-SSA-BiLSTM model changes little and the overall performance is Lei Lei: Funding acquisition, Project administration, Supervision,
relatively stable. The average relative errors of the RF-SSA-BiLSTM Conceptualization, Investigation, Data curation, Software, Validation,
model, the RF-BiLSTM model, the RF-PSO-BiLSTM model, and the RF- Writing - original draft, Resources, Formal analysis, Methodology. Suola
CNN-BiLSTM model are 1.82 %, 3.23 %, 2.52 %, and 2.3 %, respec­ Shao: Investigation, Supervision, Data curation, Validation, Writing -
tively. Results show that the RF-SSA-BiLSTM model has the highest original draft. Lixia Liang: Investigation, Supervision.
prediction accuracy and is not only suitable for building energy
consumption. Declaration of competing Interest
In this study, EWKM and RF are coupled to realize the feature se­
lection of building energy consumption influencing factors, and a pre­ The authors declare that they have no known competing financial
diction model using the sparrow search algorithm to optimize BiLSTM is interests or personal relationships that could have appeared to influence
established, which provides a more novel and advanced approach for the work reported in this paper.
building energy consumption prediction. However, the universality and
robustness of the proposed method need to be further verified for Data availability
different geographical locations and different types of buildings.
Data will be made available on request.
5. Conclusions
Acknowledgments
The RF-SSA-BiLSTM hybrid model based on deep learning is pro­
posed for building energy consumption predictions in this study. Firstly,
This research was supported by Zhejiang Province “spearhead”
the EWKM clustering algorithm is used to cluster the influencing factors
“bellwether” research and development project “The key technology
in building energy consumption, while the DBI index is used to deter­
and equipment development for low-carbon buildings” (Grant No.:
mine the optimal number of clusters. Secondly, the feature importance
2023C03153).
ranking method in the random forest algorithm is used to select the
energy consumption influencing factors. Finally, the SSA is used to
References
optimize the BiLSTM neural network to predict building energy con­
sumption in the short and medium term. Taking the annual electricity [1] Ye Y, Zuo W, Wang G. A comprehensive review of energy-related data for US
consumption of an experimental building in Dalian as the research ob­ commercial buildings. Energy Build 2019;186:126–37.
ject, the prediction model established in this paper is compared with the [2] Cheng YL, Lim MH, Hui KH. Impact of internet of things paradigm towards energy
consumption prediction: a systematic literature review. Sustain Cities Soc 2022;78:
prediction performance of the RF-BiLSTM model, the RF-PSO-BiLSTM 103624.
model, and the RF-CNN-BiLSTM model. The following conclusions are [3] Yoshino H, Hong T, Nord N. IEA EBC annex 53: total energy use in buildings-
drawn: Analysis and evaluation methods. Energy Build 2017;152:124–36.
[4] Wang Z, Srinivasan RS. A review of artificial intelligence based building energy use
prediction: contrasting the capabilities of single and ensemble prediction models.
Renew Sustain Energy Rev 2017;75:796–808.

14
L. Lei et al. Energy 288 (2024) 129795

[5] Karijadi I, Chou SY. A hybrid RF-LSTM based on CEEMDAN for improving the [24] Faiq M, Tan KG, Liew CP, et al. Prediction of energy consumption in campus
accuracy of building energy consumption prediction. Energy Build 2022;259: buildings using long short-term memory. Alex Eng J 2023;67:65–76.
111908. [25] Rahman A, Srikumar V, Smith AD. Predicting electricity consumption for
[6] Amasyali K, El-Gohary NM. A review of data-driven building energy consumption commercial and residential buildings using deep recurrent neural networks. Appl
prediction studies. Renew Sustain Energy Rev 2018;81:1192–205. Energy 2018;212:372–85.
[7] Tamer T, Dino IG, Akgül CM. Data-driven, long-term prediction of building [26] Kim TY, Cho SB. Predicting residential energy consumption using CNN-LSTM
performance under climate change: building energy demand and BIPV energy neural networks. Energy 2019;182:72–81.
generation analysis across Turkey. Renew Sustain Energy Rev 2022;162:112396. [27] Lei L, Chen W, Wu B, et al. A building energy consumption prediction model based
[8] Li X, Chen S, Li H, et al. A behavior-orientated prediction method for short-term on rough set theory and deep learning algorithms. Energy Build 2021;240:110886.
energy consumption of air-conditioning systems in buildings blocks. Energy 2023; [28] Ding Z, Chen W, Hu T, et al. Evolutionary double attention-based long short-term
263:125940. memory model for building energy prediction: case study of a green building. Appl
[9] Zhao D, Zhong M, Zhang X, et al. Energy consumption predicting model of VRV Energy 2021;288:116660.
(Variable refrigerant volume) system in office buildings based on data mining. [29] Shi X, Huang G, Hao X, et al. Sliding window and dual-channel CNN (SWDC-CNN):
Energy 2016;102:660–8. a novel method for synchronous prediction of coal and electricity consumption in
[10] Wang L, Kubichek R, Zhou X. Adaptive learning based data-driven models for cement calcination process. Appl Soft Comput 2022;129:109520.
predicting hourly building energy use. Energy Build 2018;159:454–61. [30] Jang J, Han J, Leigh SB. Prediction of heating energy consumption with operation
[11] Qiao Q, Yunusa-Kaltungo A, Edwards RE. Towards developing a systematic pattern variables for non-residential buildings using LSTM networks. Energy Build
knowledge trend for building energy consumption prediction. J Build Eng 2021;35: 2022;255:111647.
101967. [31] Wang Y, Zhang C, Fu Y, et al. Hybrid solar radiation forecasting model with
[12] Crawley DB, Lawrie LK, Winkelmann FC, et al. EnergyPlus: creating a new- temporal convolutional network using data decomposition and improved artificial
generation building energy simulation program. Energy Build 2001;33(4):319–31. ecosystem-based optimization algorithm. Energy 2023:128171.
[13] Zhu Y. Applying computer-based simulation to energy auditing: a case study. [32] Qiao X, Peng T, Sun N, et al. Metaheuristic evolutionary deep learning model based
Energy Build 2006;38(5):421–8. on temporal convolutional network, improved aquila optimizer and random forest
[14] Yan D, Xia J, Tang W, et al. DeST-An integrated building simulation toolkit Part I: for rainfall-runoff simulation and multi-step runoff prediction. Expert Syst Appl
fundamentals//Building Simulation, vol. 1. Tsinghua Press; 2008. p. 95–110. 2023:120616.
[15] Raoufi K, Wisthoff AK, DuPont BL, et al. A questionnaire-based methodology to [33] Suo L, Peng T, Song S, et al. Wind speed prediction by a swarm intelligence based
assist non-experts in selecting sustainable engineering analysis methods and deep learning model via signal decomposition and parameter optimization using
software tools. J Clean Prod 2019;229:528–41. improved chimp optimization algorithm. Energy 2023;276:127526.
[16] Cuerda E, Guerra-Santin O, Sendra JJ, et al. Understanding the performance gap in [34] Wang Z, Xia L, Yuan H, et al. Principles, research status, and prospects of feature
energy retrofitting: measured input data for adjusting building simulation models. engineering for data-driven building energy prediction: a comprehensive review.
Energy Build 2020;209:109688. J Build Eng 2022:105028.
[17] Fang T, Lahdelma R. Evaluation of a multiple linear regression model and SARIMA [35] Qiao Q, Yunusa-Kaltungo A, Edwards RE. Developing a machine learning based
model in forecasting heat demand for district heating system. Appl Energy 2016; building energy consumption prediction approach using limited data: Boruta
179:544–52. feature selection and empirical mode decomposition. Energy Rep 2023;9:3643–60.
[18] Zheng Z, Chen H, Luo X. A Kalman filter-based bottom-up approach for household [36] Ding Y, Fan L, Liu X. Analysis of feature matrix in machine learning algorithms to
short-term load forecast. Appl Energy 2019;250:882–94. predict energy consumption of public buildings. Energy Build 2021;249:111208.
[19] Bourdeau M, qiang Zhai X, Nefzaoui E, et al. Modeling and forecasting building [37] Jing L, Ng MK, Huang JZ. An entropy weighting k-means algorithm for subspace
energy consumption: a review of data-driven techniques. Sustain Cities Soc 2019; clustering of high-dimensional sparse data. IEEE Trans Knowl Data Eng 2007;19
48:101533. (8):1026–41.
[20] Wei Y, Zhang X, Shi Y, et al. A review of data-driven approaches for prediction and [38] Breiman L. Random forests. Mach Learn 2001;45:5–32.
classification of building energy consumption. Renew Sustain Energy Rev 2018;82: [39] Genuer R, Poggi JM, Tuleau-Malot C. Variable selection using random forests.
1027–47. Pattern Recogn Lett 2010;31(14):2225–36.
[21] Deb C, Lee SE, Santamouris M. Using artificial neural networks to assess HVAC [40] Xue J, Shen B. A novel swarm intelligence optimization approach: sparrow search
related energy saving in retrofitted office buildings. Sol Energy 2018;163:32–44. algorithm. Systems Science & Control Engineering 2020;8(1):22–34.
[22] Zhong H, Wang J, Jia H, et al. Vector field-based support vector regression for [41] Graves A, Schmidhuber J. Framewise phoneme classification with bidirectional
building energy consumption prediction. Appl Energy 2019;242:403–14. LSTM and other neural network architectures. Neural Network 2005;18(5–6):
[23] Shao M, Wang X, Bu Z, et al. Prediction of energy consumption in hotel buildings 602–10.
via support vector machines. Sustain Cities Soc 2020;57:102128.

15