Sustainable Cities and Society 55 (2020) 102000

Contents lists available at ScienceDirect

Sustainable Cities and Society

journal homepage: www.elsevier.com/locate/scs

Using long short-term memory networks to predict energy consumption of T

air-conditioning systems
Chonggang Zhoua, Zhaosong Fanga,*, Xiaoning Xua, Xuelin Zhangb, Yunfei Dinga,
Xiangyang Jiangc, Ying jid
a
School of Civil Engineering, Guangzhou University, Guangzhou, 510006, China
b
Hong Kong University of Science and Technology, Clear Water Bay, Hong Kong
c
Guangzhou Institute of Building Science Co., LTD, Guangzhou, 510440, China
d
College of Architecture and Civil Engineering, Beijing University of Technology, Beijing, 100124, China

A R T I C LE I N FO A B S T R A C T

Keywords: The prediction of energy consumption is important for the efficient operation of building air-conditioning sys-
Air-conditioning system tems. Most predicted models are based on historical energy consumption data and the factors influencing air
Building energy consumption conditioning systems, including weather, time of day, and previous consumption. However, the traditional
Prediction model prediction models, such as the Autoregressive Integrated Moving Average (ARIMA) time series model and back
Long short-term memory
propagation (BP) neural network model, show large errors in their prediction of the energy consumption of air-
conditioning systems. To achieve better prediction, the Long Short-Term Memory (LSTM) model of deep learning
is adopted in this study based on an air-conditioning system of a University Library in Guangzhou. The results
demonstrate that the LSTM model can produce more reliable predictions. The daily energy consumption forecast
reduced by 11.2 % compared to that of the Autoregressive Moving Average model (MAPE). The hourly energy
consumption forecast reduced by 16.31 %. In addition, compared with the BP neural network model, the MAPE’s
daily energy consumption prediction reduced by 49 % and the hourly energy consumption prediction reduced by
36.61 %.

1. Introduction According to an energy report published in 2015, construction op-

erations accounted for 20 % of the total energy consumption in China.
In the current century, more and more people moved to cities for The energy consumption was about 864 million tce (American
work and live (Piselli, Castaldo, Pigliautile, Pisello, & Cotana, 2018), Psychological Association, 2017). According to the International
which leads the population, economy and urban area of many countries Energy Agency, building and electrical energy consumption accounted
to increase sharply (Yao, Yang, Zhuang, Shao, & Yuan, 2018), named for 30 % of the world's total energy consumption in 2017, reaching up
urbanization. However, increasing urbanization would deteriorate the to 120 EJ (International Energy Agency, 2018). Reducing building en-
urban environment (Fang et al., 2018, Fang, Feng, Liu et al., 2019; ergy consumption is a global issue. An effective way to reduce energy
Fang, Feng, Xu et al., 2019; Zhang, Zhou, Zheng, Oladokum, & Fang, consumption is by reasonably adjusting the air-conditioning operating
2020). For addressing environmental issues, the energy consumption mode according to the predicted energy consumption (Zhu, Yu, & Cao,
increased, through related emissions for burning fossils or the extrac- 2019). However, it is difficult to predict the energy consumption of air-
tion of materials (Ren and Cao, 2019a, Ren and Cao 2019b, Ren and conditioning systems due to numerous factors, such as construction
Cao 2020; Wadha, Amin, Matthew, 2018a, Wadha, Amin, Matthew equipment, climatic environment, personnel flow, etc., all of which lead
2018b;). The past two decades have seen an increase in public con- to changes in the energy consumption of air-conditioning systems. In
sciousness about the environmental and operational costs of building addition, the nonlinear nature of air-conditioning operation also in-
energy system (Pinto et al., 2018; C. Ren & Cao, 2019, 2019b). More creases the difficulty in predicting the energy consumption of the air-
than 40 % of global energy consumption and 30 % of global greenhouse conditioning system. There are already diverse methods to predict en-
emission were produced due to both commercial and residential usage ergy consumption by considering multiple factors, such as weather,
in buildings (Ignatius et al., 2015; Fang et al., 2020). time of day, and the previous consumption (Jain, Smith, Culligan, &

⁎
Corresponding author.
E-mail address: [email protected] (Z. Fang).

https://doi.org/10.1016/j.scs.2019.102000
Received 13 May 2019; Received in revised form 4 December 2019; Accepted 4 December 2019
Available online 23 December 2019
2210-6707/ © 2019 Published by Elsevier Ltd.
C. Zhou, et al. Sustainable Cities and Society 55 (2020) 102000

Taylor, 2014; Kavaklioglu, 2011; Kialashaki & Reisel, 2013; Severini, suitable for forecasting energy consumption. However, the BP neural
Squartini, & Piazza, 2013). However, it is hard to apply those methods network takes into account the maximum and minimum outdoor dry
to engineering because the parameters which affect the energy con- bulb temperatures, which are difficult to obtain before the end of the
sumption of air-conditioning systems are difficult to obtain. In actual day.
engineering projects, the energy consumption of the air-conditioning It is clear that the BP neural network has been successfully used for
system is the most accessible parameter that depends on various factors. energy consumption prediction. However, BP neural networks need to
However, it is meaningful and important to predict the energy con- consider the impact of environmental factors. As a result, it is difficult
sumption of the air-conditioning system accurately. to apply them in practice. The LSTM model was first published by
The current popular methods for the prediction of air-conditioning Hochreiter and Schmidhuber (1997) and has been widely used in time
system energy consumption are time series and neural network models. series analysis and prediction such as language translation, image
The most common time series models are the Auto Regressive model analysis, and speech recognition. To date, a few studies have analysed
(AR), the Moving Average model (MA), the Autoregressive Moving and predicted energy consumption using the LSTM model. Kumar,
Average model (ARMA), and the Autoregressive Integrated Moving Hussain, Banarjee and Reza (2018) predicted the electrical load of the
Average model (ARIMA). The common neural network models are the grid using the LSTM model. Marino, Amarasinghe and Manic (2016)
Convolutional Neural Network (CNN), Radial Basis Function (RBF) and proved that LSTM was difficult to predict minute electric energy con-
Back Propagation (BP) Neural Network. In this study, the time series sumption, but it has acceptable performance on hourly electric energy
model is considered. BP neural network usually refers to multi – layer consumption prediction. Their results show that the LSTM model has
forward neural network based on error back propagation algorithm. great potential in energy consumption prediction. Compared with the
(Fang, 2011). The time series model is mainly analogized or extended time series model, the ability of LSTM to solve nonlinear problems and
by the development process, direction, and trend. The model arranges remember historical data is better than that of the BP neural network.
the data according to the development order of time, finds out the Only a limited number of studies have applied the LSTM model to
changing rule of the data, and forecasts the data in the next period. The predict the energy consumption of air-conditioning systems. Kumar
ARIMA model is usually analysed by time series, which has a favourable et al. (2018) and Marino et al. (2016) adopted LSTM model to predict
effect on the prediction of energy consumption (Abdelaal & Algarni, the power consumption of residential users. In addition, Kim and Cho
2014; Ediger & Akar, 2007; Liu, Wu, & Tsao, 2005; Mohamed & Bodger, (2019) used a combination of convolutional neural network (CNN) and
2005). Sen, Roy and Pal (2016) used the ARIMA model to predict the LSTM to predict the power consumption of residential users. However,
energy consumption and greenhouse gas emissions of steel companies. few previous investigations focused on the prediction of the power
The results indicated that the ARIMA model has a good performance in consumption of air conditioning. The primary reason is that air con-
energy consumption prediction. Noh and Rajagopal (2013) applied the ditioners are a part of the electrical equipment and the energy con-
AR model to building energy prediction. They concluded that the AR sumption is not separately statistics. However, the air conditioning
model has a good effect on the hourly energy prediction of high-fre- system of public buildings is relatively complex, including refrigeration
quency data changes. However, the prediction accuracy is lower due to units, cooling water circulation system, chilled water system, etc.
without considering the processing information of white noise sequence Within non-residential buildings, almost half of the electricity con-
in time series in this AR model. The inadequate extraction of relevant sumption is due to Heating, Ventilation and Air-Conditioning (HVAC)
information from the noise sequence makes the AR model difficult to systems (Garnier, Eynard, Caussanel, & Grieu, 2014). Therefore, the air-
deal with mutations caused by human factors. Thus, the prediction conditioning energy consumption of public building system needs to be
effect is poor. De Nadai and Van Someren (2015) combined the ARIMA predicted separately. This study firstly attempts to apply the LSTM
model with artificial neural network to predict natural gas consumption model to predict the daily air-conditioning terminal energy consump-
for achieving a good performance. However, the artificial neural net- tion and hourly air-conditioning system energy consumption of a li-
work increases the complexity of the model. Nichiforov, Stamatescu, brary in Guangzhou. Only energy consumption data is used as an input
Fǎgǎrǎs and Stamatescu (2017) found that the ARIMA model has a parameter. The prediction results are validated with field measurement
better performance in predicting power consumption. Time series data. The BP model and ARIMA model are also applied to predict en-
models require a stable data set of time series. However, the system is ergy consumption in the same scenario and the predicted results are
often affected by factors such as weather and personnel flow. These compared with those obtained by the LSTM model. The results show
factors influence the energy consumption to a large extent. The time that the LSTM model is better than both the ARIMA model and BP
series data set is difficult to stabilize. Thus, the ARIMA model is difficult neural network in predicting air-conditioning energy consumption.
to apply and the predictions are difficult to achieve the desired results.
In addition, the essence of the ARIMA model is the linear regression 2. Model description
model. It is difficult to accurately describe the changes in energy con-
sumption due to the nonlinear variation of the energy consumption. 2.1. LSTM model introduction
Qu, Sun, Fan and Li (2010) and Li and Shang (2008) applied the BP
neural network to predict city power consumption and national energy Air-conditioning energy consumption is always subject to rapid
consumption, which proved the feasibility of the BP neural network in changes due to various factors such as construction equipment, climatic
macroscopic energy prediction. Mena, Rodríguez, Castilla and Arahal environment, and personnel flow. However, the changes often follow
(2014) reported that the prediction needed to consider the impact of periodic laws. If these periodic laws that accurately predict the energy
factors such as temperature, radiation, and working days on building consumption of air-conditioning are correctly mastered, a good energy-
energy consumption. However, most parameters often lacked a record saving operation mode can be obtained. Hochreiter and Schmidhuber
in practice. Therefore, it was difficult to apply this model in reality. (1997) proved that the LSTM model had better robustness to handle
Yuce et al. (2014) applied the BP neural network to predict the energy continuous values and performed well for time series with large lags.
consumption and thermal comfort of the pool, which provided accurate The cell and structure of the LSTM neural network model is shown in
predictions of these parameters by considering various parameters such Figs. 1 and 2.
as water temperature, room temperature, and occupancy. Based on the Graves (2012) indicated that the LSTM model deals with data using
analysis, the BP neural network can predict various building energy sigmoid and tanh function. As shown in Fig. 3, the gate which is used to
consumption patterns well. Alberto et al. (2008) compared the BP screen information should choose sigmoid. The larger the value passed
neural network and a model based on physical principles to predict in the sigmoid function, the more information is retained. In addition,
building energy consumption. Results showed that both models are the model will converge fast due to the tanh function with a high slope

2
C. Zhou, et al. Sustainable Cities and Society 55 (2020) 102000

Fig. 1. LSTM structure.

is processed by the sigmoid function to get ft1. LSTM can remember

large amounts of historical data, by ft1 decided useful data. ft1is calcu-
lated by (2)

ft1 = σ (Wf ⋅X′ t + bf ) (2)

In the input gate, LSTM acquires the new data. X ′t is processed by

the sigmoid function to get it . The it decides the useful data in X ′t . In
addition, X ′t is processed by the tanh function to calculate C′t , which
expresses nonlinear relationships. They are expressed as

it = σ (Wi ⋅X ′t + bt ) (3)

C′t = tanh (Wc⋅X′ t + bc ) (4)

Fig. 2. LSTM structure.
ft 2 = it × C′t (5)
at the original point. In the update gate, Ct is updated. Ct and ft1 are multiplied by the
LSTM is composed of cells with the same structure. In this model, matrix to obtain the historical data. It is added with ft2 to get the output
the data of the next moment is predicted each time by the previous data Ct of this neural unit, which keeps more accurate rules in the cell for
and historical data, which is processed by the cells. Each cell has three accurate prediction. It is expressed as
input parameters, including historically stored information Ct − 1, his-
torical data Xt and ht-1, which represent the prediction results of the last Ct = ft1 × Ct − 1 + ft 2 (6)
cell and input parameter to the cell. Each cell can be divided into four In the output gate, LSTM outputs the result. X ′t is processed by the
parts. sigmoid function to get Ot . Ot decides which Ct needs to be retained as
The data ht-1, which is processed by the previous cell, and the input the result. In addition, Ct is processed by the tanh function to geth′t . h′t
data of the current time Xt are spliced through a matrix to obtain X ′t , andOt are multiplied to obtain the final data ht , which is expressed as
where
Ot = σ (Wo⋅X ′t + bo) (7)
X ′t = [ht − 1, Xt ] (1)
In the forgotten gate, the LSTM acquires the ability to filter data. X ′t ht = Ot *tanh (Ct ) (8)

Fig. 3. Sigmoid and tanh functions.

3
C. Zhou, et al. Sustainable Cities and Society 55 (2020) 102000

Fig. 5. Traditional BP neural network.

2.3. Model comparison

Fig. 4. LSTM model flow chart.
2.3.1. BP neural network
From the above cell processing, it is seen that the LSTM model
2.2. LSTM model flow
learns the law of data changes through historical data, and also retains
historical data. It has the ability to solve the problem of a long-term
The model is built as shown in the Fig. 4. First, the basic LSTM
dependence on the data and can predict the time series better. As shown
model is established, and then the training parameters of the model are
in Fig. 5, when the input parameters enter the hidden layer, each node
set which usually include the number of trainings, the time step, the
of the hidden layer 1 is connected with all the nodes of the hidden layer
length of the time series entering the LSTM, the learning rate, and other
2 so that the model can learn the influence of each input parameter on
such parameters. These parameters affect the final result of the model.
the output. In addition, historical time data often has a great influence
Since there are currently no clear guideline governing the setting of
on the next time data. The BP neural network cannot memorize his-
these parameters, constant experimentation is required to find the best
torical data, and it is more difficult to consider the influence of his-
parameters. When the setup is complete, the raw data is processed. Raw
torical data on the output. Therefore, it is hard to accurately mine the
data usually requires data cleaning and normalization. Data cleaning
rules hidden in time series data sets.
mainly reduces the impact of erroneous data on the model by filling in
In this paper, a four-layer neural network is adopted in BP neural
missing values, smoothing noise data, and deleting outliers. The nor-
network, including an input layer, two hidden layers and an output
malization enables the LSTM model to acquire the ability to handle
layer. When the daily energy consumption is predicted, there are 5
nonlinear problems. As shown in Fig. 3, we can see that the tanh
nodes in the input layer, 3 nodes in the first hidden layer, 1 node in the
function gradually appears as a horizontal straight line after x is greater
second hidden layer, and 1 node in the output layer. For the hourly
than 2. If the original data is directly put into the model, the output is
energy consumption prediction, there are 24 nodes in the input layer,
basically consistent when the data passes the tanh function. The model
10 nodes in the first hidden layer, 1 node in the second hidden layer,
therefore loses the ability to deal with nonlinear problems and becomes
and 1 node in the output layer. The activation function of the hidden
unable to make predictions. After pre-processing the data, the original
layer of the two BP neural networks is tansig. The activation function of
data is divided into training sets and test sets. When the training set is
the output layer is logsig. The objective function is MSE. And the
imported into the model, the model identifies the hidden patterns in the
method of parameter updating is stochastic gradient descent.
data. Finally, the model is verified through the test set. When the ver-
ification result satisfies expectations, the model output is accepted. 2
tansig(n) = −1
When the output result does not meet the expectation, the hyper- 1 + e−2n (12)
parameter is re-set and the model is retrained. To evaluate the obtained
1
results, we use the RMSE (Root Mean Squared Error, MAPE (Mean logsig(n) =
1 + e−n (13)
Absolute Percentage Error) and MAD (Mean Absolute Deviation). They
are expressed as follows.
2.3.2. ARIMA time series model
n
(Yi − Y′ t )2 The basic idea of the ARIMA model is to treat the time series of the
RMSE = ∑ n predicted object as a random sequence and use a mathematical model
i=1 (9)
to approximate the sequence. In this case, the mean variance of the
system does not cause a large change in each data interval and it is
n Yi − Y ′i considered to be a stable time series. However, when the data sets are
∑1 Y ′i
MAPE = × 100% affected by external factors, the change in the data is no longer a
n (10)
random change. The data sets cannot be regarded as a stable time series
when the variance of the mean value has a large change. In this case,
n
∑i = 1 |Yi − Y¯ | the ARIMA model is invalid.
MAD = , The basic form of the ARIMA model is expressed as follows:
n (11)
ft (p , d, q) = (1 + φ1) ft − 1 + (φ2 − φ1) ft − 2 + ⋯+(φp − φ p − 1) ft − p
where Yi is the predicted value, Y ′i is the actual values, and Ȳ is the
mean of predicted value. − φp ft − p − 1 + et − θ1 et − 1 − θ2 et − 2 − ⋯−θq et − q (14)

where t is the number of days, p is the hysteresis of the time series data

4
C. Zhou, et al. Sustainable Cities and Society 55 (2020) 102000

itself used in the prediction model, d is the differential order in which predicted energy consumption in different time steps are consistent
the time series data needs to stabilize the data, q is the hysteresis of the with variations in actual values. When the time step is 24 h, the var-
prediction error used in the prediction model. iations in predicted results are closest to the variations in actual values.
The MAPE decreases to a minimum when the time-step is 24, as shown
3. Data description Fig. 10.
As shown in Fig. 11, when the time step is 24, the variations in
The experimental data comes from the actual energy consumption prediction results of different cycle times are consistent with the var-
of the air-conditioning system of a university library in Guangzhou, iations in actual values. However, at the point of sudden change, the
Guangdong. The library is 31.5 m high, with a building area of 46,774 predicted results and actual value are different. When the epoch is 900,
m2 and an air-conditioning area of approximately 38,500 m2. Most the prediction result is closest to the actual value. From Table 4, it is
building energy collection systems only record the most basic building shown that when the epoch is 900, the RMSE and MAD are at their
energy consumption data. Considering the practical application, this minimum. In addition, the MAPE decreases to a minimum when the
study aims to investigate whether LSTM can predict building energy epoch is 900. When the number of cycles of the model changes slightly,
consumption through the most basic energy consumption data. The li- the accuracy of the model will also change slightly. Therefore, when the
brary air-conditioning system adopts centralized cooling, and the plate- model precision reaches the ideal level, it needs to change the number
changer heat exchange is provided in the building. The device terminal of cycles slightly to find the best model.
of the air-conditioning system is the air-handling unit and fan coil unit.
The inlet and return water temperatures of the air-handling unit and fan 4.3. Model comparison
coil unit are 6.5 °C and 13.5 °C, respectively. The energy consumption of
the central cooling system is mainly composed of the power con- 4.3.1. Model comparison for hourly energy consumption
sumption of the air conditioning terminal and the building cooling When comparing the LSTM model with the traditional BP neural
consumption. In the energy collection system, the cooling consumption network model and the ARIMA model, it is easy find that the predicted
of the building is collected at hourly intervals, and the electricity con- values of the three models are basically consistent with the original
sumption of the air-conditioning terminal is collected at daily intervals. data, as shown in Fig. 12. According to Fig. 13, the BP neural network
There is a positive correlation between terminal energy use and cooling and ARIMA models have poor performance. When the energy con-
consumption. Therefore, the historical energy consumption of the air- sumption changes abruptly, LSTM also has a poor performance. How-
conditioning terminal is forecasted using the daily energy consumption ever, LSTM can reduce the error faster by judging the trend since it
of the air-conditioning terminal, and the historical cooling consumption learns the law from long-term memory data. In addition, LSTM per-
of the building is used to forecast hourly cooling consumption of the forms better than ARIMA and BP Neural Network under the stable
building. These results verify the accuracy of the LSTM model in the energy consumption condition. In Table 5, compared with the ARIMA
prediction of hourly and daily energy consumption of the air-con- and BP neural network model, it is seen that the RMSE, MAD and MAPE
ditioning system (Tables 1 and 2). of the LSTM model are significantly lower. Compared with the BP
neural network, RMSE decreases by 29.02 %, MAD decreases by 32.98
4. Results %, and MAPE decreases by 36.27 %. Compared with the ARIMA model,
RMSE decreases by 16.3 %, MAD decreases by 11.03 %, and MAPE
4.1. Prediction of daily air-conditioning energy consumption decreases by 16.3 %. Thus, LSTM model is better than the ARIMA and
BP neural network models in the prediction of hourly air-conditioning
To analyse the results of the model, different time steps and cycle energy consumption.
times are set when training the LSTM model to predict daily energy
consumption. In Fig. 6, when the number of cycles is 700, the predic- 4.3.2. Model comparison for daily energy consumption
tion results of different time steps are consistent with the actual value From Fig. 14, the predicted values of the three models are basically
trend. However, the differences in predictions among different time consistent with the original data. However, the error of the BP neural
steps are significant when the energy consumption encounters a sudden network is significantly higher than ARIMA and LSTM models, as
change. When the time step is 5, the prediction result is the closest to shown in Fig. 15. The ARIMA model performs better than LSTM in the
the actual value. The MAPE decreases to a minimum when the time step short-term. Nevertheless, with the rise of prediction time, the accuracy
is 5, as shown in Fig. 7. of the ARIMA model begins to decline as it is affected by recent training
In Fig. 8, when the time step is 5, the predicted trend is consistent data. The error of LSTM is low and suitable in the prediction period.
with actual values. The variations are similar with the results shown in When the energy consumption suddenly changes, all of the models have
Fig. 6. From Table 3, the RMSE and MAD at their minimum when the poor performance. Since the energy consumption is affected by varia-
time step is 5. The MAPE is lowest when the epoch is 700. Thus, when tions in weather, all the models are unable to obtain an accurate law
the time step is 5, the epoch should be selected at 700 for the high from historical data. From Table 6, compared with the ARIMA and BP
precision. neural network models, it can be seen that the LSTM model has a sig-
nificant decrease in RMSE, MAD, and MAPE. When LSTM is compared
4.2. Prediction of hourly air-conditioning energy consumption with the BP neural network, RMSE decreases by 38.53 %, MAD de-
creases by 45.39 %, and MAPE decrease by 49 %. When compared with
From Fig. 9, when the number of cycles is 900, the variations in the ARIMA model, RMSE decreases by 1.88 %, MAD decreases by 3.35

Table 1
Data Description.
Type of data period Data set

Daily air-conditioning energy consumption training set 09/01/2016-08/31/2017 361

Daily air-conditioning energy consumption test set 09/01/2017-11/15/2017 74
The hourly air-conditioning cooling load training set 04/10/2014-05/27/2014 1137
The hourly air-conditioning cooling load test set 05/28/2014-06/16/2014 448

5
 Table 2
Detail information of inputs and outputs.
Type of prediction Model Input Energy constitute Output Unit Data types
C. Zhou, et al.

Prediction of Daily Air-conditioning Energy Consumption ARIMA Xn-1, Xn-2, Xn-3 …… Electrical energy consumption of fan coil unit Xn kWh Numeric and continuous
BP neural network Electrical energy consumption of air-handling unit
LSTM Electrical energy consumption of fresh air handling unit
Prediction of Hourly Air-conditioning Energy Consumption ARIMA Yn-1, Yn-2, Yn-3 …… Cooling consumption of fan coil unit Yn kWh Numeric and continuous
BP neural network Cooling consumption of air-handling unit
LSTM Cooling consumption of fresh air handling unit

Xn : The power consumption of air conditioning terminal on day n.

Yn : The cooling consumption of the building at time n.

6
dition of epoch 700.

dition of time-step 5.
Fig. 7. MAPE of different time-step using daily data sets.
Sustainable Cities and Society 55 (2020) 102000

Fig. 8. Daily air-conditioning energy consumption forecast results under con-

Fig. 6. Daily air-conditioning energy consumption forecast results under con-
C. Zhou, et al. Sustainable Cities and Society 55 (2020) 102000

Table 3
Performances of different epoch using daily data sets.
Evaluation index 500 600 700 800 900

RMSE 408.778 404.37 400.92 421.14 411.41

MAPE 9.82 % 9.56 % 8.63 % 10.13 % 9.54 %
MAD 318.50 321.05 297.58 338.74 306.02

Fig. 11. Daily air-conditioning energy consumption forecast results under

condition of time step of 24.

Table 4
Performances of different epoch using hourly data sets.
Evaluation index 700 800 900 1000 1100

RMSE 222.37 217.38 211.15 230.59 221.78

MAPE 8.96 % 9.15 % 8.68 % 9.83 % 9.22 %
Fig. 9. Hourly air-conditioning energy consumption forecast results under MAD 154.26 155.84 150.56 168.93 157.73
condition of epoch 900.

Fig. 10. MAPE of different time steps using hourly data sets. Fig. 12. Different model hour air-conditioning energy consumption prediction
results.

%, and MAPE decreases by 11.2 %. Thus, the LSTM model is better than
the ARIMA and BP neural network model in predicting the daily air- When the number of cycles is less than 700, the error gradually
conditioning energy consumption. decreases. The primary reason is that the model has not fully grasped
the law of energy consumption change. When the number of cycles is
greater than 700, the error increases significantly. The model pays too
5. Discussion much attention to the transformation law of learning the training set
and the generalization ability of the model decreases.
5.1. Daily air-conditioning energy consumption Biswas, Robinson and Fumo (2016) and Neto and Fiorelli (2008)
used the BP neural network to predict the energy consumption of daily
The experimental results indicate that the optimal time step is 5 air-conditioning. They proved that BP neural network has certain ac-
different from the traditional time step of 7. The primary reason is that curacy in daily energy consumption prediction, but outdoor tempera-
the users of the campus library are students and the energy consump- ture should be used as an input parameter. However, the prediction of
tion prediction is for the campus library. The students stay in the library air conditioning energy consumption only based on most basic energy
for a long time on weekdays. Therefore, there is no significant differ- consumption, which causes the low performance of BP neural network.
ence between the energy use of the campus library on weekdays and Amjady (2001) used ARIMA to predict regional power grid load, which
weekends. proves that ARIMA has a good effect on power grid energy consumption

7
C. Zhou, et al. Sustainable Cities and Society 55 (2020) 102000

Fig. 13. Absolute error of partial hour prediction result. Fig. 15. Absolute error of daily prediction result.

Table 5 Table 6
Performances of different models using daily data sets. Performances of different models using daily data sets.

Evaluation index RMSE MAD MAPE Evaluation index RMSE MAD MAPE

LSTM 211.15 150.56 8.68 % LSTM 400.92 297.58 8.63 %

ARIMA 252.37 169.24 10.37 % ARIMA 408.61 307.88 9.72 %
BP 297.57 224.66 13.62 % BP 652.22 544.95 16.92 %

5.2. Hourly air-conditioning energy consumption

The results indicate that when the epoch is 900, the time step should
be selected as 24 for highest precision. In this investigation, the optimal
step size is 24. In campus library, the students also study on weekends.
Therefore the energy use pattern of weekends is similar to weekdays.
Besides, the climate are similar between days, so 24 -h energy con-
sumption prediction works best with a time step of 24.
Ghalehkhondabi, Ardjmand, Weckman and Young (2016) believed
ARIMA is one of the most effective short-term forecasting models. Lai,
Liu, Kuo and Chang (2014) used ARIMA model and BP neural network
to predict the annual energy consumption of buildings. It was found
that BP neural network needed to use appropriate input parameters and
ARIMA was more suitable for the linear relationship. The ARIMA per-
form poorly due to the energy consumption of air conditioning hour
with a nonlinear process. Li, Lin, Wang and Wang (2011) used the BP
neural network model to predict the hourly thermal load. The input
parameters are the historical load and outdoor temperature. But it also
Fig. 14. Daily air-conditioning energy consumption prediction results for dif- has the problem that the prediction result can fluctuate greatly. The BP
ferent models. neural network model performs poorly because the cooling load varies
greatly during the day. In addition, Temperature is as an input para-
prediction. However, compared with power grid energy consumption, meter, which also causes the performace of The BP neural network
air-conditioning energy consumption is easily affected by personnel decline.
flow. Song, Joo and Gunawan (2012) proved that the ARIMA model Based on the analysis, the LSTM, ARIMA and BP neural network
could predict the energy consumption of air conditioners, but the error models have certain ability to predict energy consumption. However,
is large when the energy consumption suddenly changes. The energy the ARIMA and BP neural network models perform poorly in predicting
consumption of air conditioning varies greatly in Spring and Autumn, large fluctuations and trending data (Lai et al., 2014; Li et al., 2011).
so the performance of ARIMA is not as good as LSTM model in this The LSTM model is suitable for predicting the energy consumption of
investigation. air-conditioning systems which have existed for a long period of time
Based on this analysis, LSTM, ARIMA and BP neural network models due to the relatively long historical data and remain useful data by
have certain ability to predict energy consumption. However, the BP forgotten Gate (Liang, Zhen, & Gang, 2017) (Table 7).
neural network model performs poorly for predicting large fluctuations
and trending data. Although the ARIMA model has a good performance 5.3. Differences amony the diffenent models
in initial prediction, the prediction error increases with the increase of
prediction time. The LSTM model is suitable for predicting the energy The ARIMA model was established by Box and Jenkins (1976) in
consumption of air-conditioning systems with a long period of time. 1976. The model only observes the law of the predicted data (Ediger &
Akar, 2007). It only needs to collect a few data to make predictions. It is
always used in the prediction of social problems, morbidity and

8
C. Zhou, et al. Sustainable Cities and Society 55 (2020) 102000

economic growth (Armando, Michel, Angel, & Alberto, 2007; Chen,

The ARIMA model is the linear regression model. It only observes the law of the predicted data without

The LSTM is able to store information for long periods of time. It is well suited for dealing with the
BP neural network has strong nonlinear mapping ability. It is widely applied in nonlinear models.
Yuan, & Shu, 2008; Tseng, Tzeng, Yu, & Yuan, 2001). However, in air-
conditioning system the energy consumption is often affected by en-
vironmental factors with strong fluctuation. Thus, it is difficult to
achieve good predicted performance by using ARIMA model. In
1986,Rumelhart and McCelland (1986) established the BP neural net-
work. The model often has strong nonlinear mapping ability (Dongmei,
Shiqing, Xuhui, & Xue, 2017). The BP neural network is often used to
forecasting financial sector, engineering problem and medical problem
(Chen, Wang, Liu, Luo, & Liu, 2013; Liu, Xie, Xie, Peng, & Xia, 2015;
directly considering the changes of other related random variables.

Sun, Che, & Wang, 2014). It is different from ARIMA in the prediction
with considering the influence of different parameters. The BP neural
network is difficult to widely apply in practical projects. The primary
reason is that only a small amount of parameter data is recorded in BP
neural network. LSTM model was constructed by Hochreiter and
Schmidhuber (1997) in 1997. LSTM is a recurrent neural-network for
specifically designing to overcome the problems of vanishing gradient
problems of highly relevant to time series

(Marino et al., 2016). The model only use observational data to make
predictions and retain the strong nonlinear mapping ability of BP neural
network. The model plays an important role in predictions of en-
gineering, text and image (Srivastava & Lessmann, 2018; Song, Lan,
Xing, Zeng, & Liu, 2018; Yousfi, Berrani, & Garcia, 2016). The pattern
of change in these data could often be found in their own history and
are several different cycles of the changes, which is similar to the
Model Features

characteristics of energy consumption prediction. Thus, LSTM model

can be used to predict the energy consumption with better perfor-
mance.

6. Conclusion
BP neural network is a kind of multilayer feed forward network trained
ARIMA model is generated based on observations to a time series.

LSTM is a recurrent neural-network. It is specifically designed to

Most of the studies used the ARIMA and BP neural network models
to predict the daily and hourly energy consumption of air-conditioning
systems. However, the prediction is not satisfactory, due to the lim-
itations in the prediction model itself. LSTM can efficiently learn and
memorize the previous data in the long-term better than BP neural
overcome the problems of vanishing gradient.

network and has better ability to solve nonlinear problems than ARIMA.
This study applied the LSTM model to predict the hourly energy
by error back propagation algorithm.

consumption and daily energy consumption of air-conditioning systems

in a library. Compared with the BP neural network prediction model
and ARIMA model, the results show that the LSTM model acted more
accurately. The MAPE of the daily energy consumption forecast is re-
duced by 11.2 % and 49 % compared to ARIMA and BP neural network
models respectively. In addition, the MAPE of hourly energy con-
sumption prediction is reduced by 16.31 % and 36.61 %. Hence, the
Core Concepts

prediction accuracy of the LSTM model is higher than the ARIMA model
and BP neural network. In addition, the LSTM model has wide adapt-
ability to predict building energy consumption only by using energy
consumption data. Thus, the LSTM model for predicting the energy
consumption of a building’s central air-conditioning systems has a wide
Hochreiter and Schmidhuber

range of adaptability and accuracy.

Rumelhart and McCelland
Box and Jenkins (1976)

Declaration of Competing Interest

The authors declared that they have no conflicts of interest to this

work. We declare that we do not have any commercial or associative
Authors

(1986)

(1997)
Detail information of the Models.

interest that represents a conflict of interest in connection with the

work submitted.
1976

1986

1997
Time

Acknowledgments
BP neural network

The author(s) disclosed receipt of the following financial support for

the research, authorship, and/or publication of this article. This work
was supported by the National Key R&D Program of China (No.
ARIMA
Table 7

Model

LSTM

2018YFC0705900) And Guangzhou Science and Technology Plan

Project (No. 201804010252).

9
C. Zhou, et al. Sustainable Cities and Society 55 (2020) 102000

References IEEE.
Liu, Y. K., Xie, F., Xie, C. L., Peng, M. J., & Xia, H. (2015). Prediction of time series of npp
operating parameters using dynamic model based on bp neural network. Annals of
Zhang, Y., Zhou, X., Zheng, Z., Oladokum, M. O., & Fang, Z. (2020). Experimental in- Nuclear Energy, 85, 566–575.
vestigation into the effects of different metabolic rates of body movement on thermal Marino, D. L., Amarasinghe, K., & Manic, M. (2016). IEEE IECON 2016 - 42nd Annual
comfort, Vol. 168 106489. Conference of the IEEE Industrial Electronics Society - Florence, Italy (2016.10.23-
Abdelaal, R. E., & Algarni, A. Z. (2014). Forecasting monthly electric energy consumption 2016.10.26)] IECON 2016 - 42nd Annual Conference of the IEEE Industrial Electronics
in eastern Saudi Arabia using univariate time-series analysis. Energy, 22(22), Society - Building Energy Load Forecasting Using Deep Neural Networks, 7046-7051.
1059–1069. Mena, R., Rodríguez, F., Castilla, M., & Arahal, M. R. (2014). A prediction model based on
Amjady, N. (2001). Short-term hourly load forecasting using time-series modeling with neural networks for the energy consumption of a bioclimatic building. Energy and
peak load estimation capability. IEEE Transactions on Power Systems, 16(4), 798–805. Buildings, 82, 142–155.
Armando, A. J., Michel, V. H., Angel, O., & Alberto, G. E. (2007). Forecasting malaria Mohamed, Z., & Bodger, P. (2005). Forecasting electricity consumption in New Zealand
incidence based on monthly case reports and environmental factors in karuzi, bur- using economic and demographic variables. Energy, 30(10), 1833–1843.
undi, 1997–2003. Malaria Journal, 6(1), 1–10. Neto, A. H., & Fiorelli, F. A. S. (2008). Comparison between detailed model simulation
Biswas, M. A. R., Robinson, M. D., & Fumo, N. (2016). Prediction of residential building and artificial neural network for forecasting building energy consumption. Energy and
energy consumption: A neural network approach. Energy, 117, 84–92. Buildings, 40(12), 2169–2176.
Box, G. E., & Jenkins, G. M. (1976). Time series analysis: forecasting and control rev. ed. Nichiforov, C., Stamatescu, I., Fǎgǎrǎs¸, I., & Stamatescu, G. (2017). Energy consumption
Journal of Time, 31(4), 238–242. forecasting using ARIMA and neural network models. International Symposium on
Chen, P., Yuan, H., & Shu, X. (2008). Forecasting crime using the ARIMA model. Electrical & Electronics Engineering IEEE.
International Conference on Fuzzy Systems & Knowledge Discovery. IEEE. Noh, H. Y., & Rajagopal, R. (2013). Data-driven forecasting algorithms for building en-
Chen, L. Y., Wang, C. Q., Liu, H., Luo, Y. Y., & Liu, R. H. (2013). Study on integrity and ergy consumption. Proceedings of SPIE - The International Society for Optical
timing characteristics of shaoyao and gancao decoction on anti-inflammatory effects Engineering, 8692(5), 86920T.
by bp neural network. Acta Pharmacologica Sinica, 34(Suppl 1), S43–S48. Piselli, C., Castaldo, V. L., Pigliautile, I., Pisello, A. L., & Cotana, F. (2018). Outdoor
De Nadai, M., & Van Someren, M. (2015). Short-term anomaly detection in gas con- comfort conditions in urban areas: On citizens’ perspective about microclimate mi-
sumption through ARIMA and Artificial Neural Network forecast. Environmental, tigation of urban transit areas. Sustainable Cities and Society, 39, 16–36.
Energy & Structural Monitoring Systems IEEE. Qu, S., Sun, Z., Fan, H., & Li, K. (2010). BP neural network for the prediction of Urban
Dongmei, H., Shiqing, H., Xuhui, H., & Xue, Z. (2017). Prediction of wind loads on high- building energy consumption based on matlab and its application. Second
rise building using a bp neural network combined with pod. Journal of Wind International Conference on Computer Modeling & Simulation IEEE Computer Society.
Engineering and Industrial Aerodynamics, 170, 1–17. Ren, C., & Cao, S. (2019a). Development and application of linear ventilation and tem-
Ediger, V. S., & Akar, S. (2007). ARIMA forecasting of primary energy demand by fuel in perature models for indoor environment prediction and HVAC systems control.
Turkey. Energy Policy, 35(3), 1701–1708. Sustainable Cities and Society, 51, 101673.
Fang, F. (2011). Research on power load forecasting based on improved BP neural network. Ren, J., & Cao, S. (2019b). (2) incorporating online monitoring data into fast prediction
Doctoral dissertationHarbin Institute of Technology. models towards the development of artificial intelligent ventilation systems.
Fang, Z., Feng, X., Liu, J., Lin, Z., Mak, C. M., Niu, J., et al. (2019). Investigation into the Sustainable Cities and Society, 02(28).
differences among several outdoor thermal comfort indices against field survey in Ren, C., & Cao, S. (2020). Implementation and visualization of artificial intelligent ven-
subtropics. Sustainable Cities and Society, 44, 676–690. tilation control system using fast prediction models and limited monitoring data.
Fang, Z., Feng, X., Xu, X., Zhou, X., Lin, Z., & Ji, Y. (2019). Investigation into outdoor Sustainable Cities and Society, 52 101860.
thermal comfort conditions by different seasonal field surveys in China, Guangzhou. Rumelhart, D., & McCelland, J. (1986). Parallel distributed processing, vol. 1. Cambridge:
International Journal of Biometeorology, 7, 1–12. MIT Press.
Garnier, A., Eynard, J., Caussanel, M., & Grieu, S.éphane (2014). Low computational cost Sen, P., Roy, M., & Pal, P. (2016). Application of ARIMA for forecasting energy con-
technique for predictive management of thermal comfort in non-residential build- sumption and GHG emission: A case study of an Indian pig iron manufacturing or-
ings. Journal of Process Control, 24(6), 750–762. ganization. Energy, 116, 1031–1038.
Ghalehkhondabi, I., Ardjmand, E., Weckman, G. R., & Young, W. A. (2016). An overview Severini, M., Squartini, S., & Piazza, F. (2013). Hybrid soft computing algorithmic fra-
of energy demand forecasting methods published in 2005–2015. Energy Systems. mework for smart home energy management. Soft Computing, 17(11), 1983–2005.
Graves, A. (2012). Long short-term memory. Berlin: Springer - Verlag. Song, L., Joo, I. S., & Gunawan, S. (2012). Next-day daily energy consumption forecast
Hochreiter, S., & Schmidhuber, J. (1997). Long short-term memory. Neural Computation, model development and model implementation. Journal of Solar Energy Engineering,
9(8), 1735–1780. 134(3), 031002.
International Energy Agency (2018). Energy efficiency. France: International Energy Song, S., Lan, C., Xing, J., Zeng, W., & Liu, J. (2018). Spatio-temporal attention based lstm
Agency. networks for 3d action recognition and detection. IEEE Transactions on Image
Jain, R. K., Smith, K. M., Culligan, P. J., & Taylor, J. E. (2014). Forecasting energy Processing, 1.
consumption of multi-family residential buildings using support vector regression: Srivastava, S., & Lessmann, S. (2018). A comparative study of lstm neural networks in
Investigating the impact of temporal and spatial monitoring granularity on perfor- forecasting day-ahead global horizontal irradiance with satellite data. Solar Energy,
mance accuracy. Applied Energy, 123, 168–178. 162, 232–247.
Kavaklioglu, K. (2011). Modeling and prediction of Turkey’s electricity consumption Sun, Q., Che, W. G., & Wang, H. L. (2014). Bayesian regularization BP neural network model
using support vector regression. Applied Energy, 88(1), 368–375. for the stock price prediction. Foundations and applications of intelligent systems. Berlin
Kialashaki, A., & Reisel, J. R. (2013). Modeling of the energy demand of the residential Heidelberg: Springer.
sector in the United States using regression models and artificial neural networks. Tseng, F. M., Tzeng, G. H., Yu, H. C., & Yuan, B. J. C. (2001). Fuzzy arima model for
Applied Energy, 108(8), 271–280. forecasting the foreign exchange market. Fuzzy Sets and Systems, 118(1), 9–19.
Kim, T., & Cho, S. (2019). Predicting residential energy consumption using CNN-LSTM neural. Wadha, A. M., Amin, A. H., & Matthew, W. (2018a). An investigation into domestic en-
Kumar, S., Hussain, L., Banarjee, S., & Reza, M. (2018). Energy load forecasting using ergy consumption behaviour and public awareness of renewable energy in Qatar.
deep learning approach-LSTM and GRU in spark cluster. 2018 Fifth International Sustainable Cities and Society, 41, 639–646.
Conference on Emerging Applications of Information Technology IEEE. Wadha, A. M., Amin, A. H., & Matthew, W. (2018b). An investigation into domestic en-
Lai, S. L., Liu, M., Kuo, K. C., & Chang, R. (2014). Energy consumption forecasting in hong ergy consumption behaviour and public awareness of renewable energy in Qatar.
kong using arima and artificial neural networks models. Applied Mechanics and Sustainable Cities and Society, 41, 639–646.
Materials, 672–674, 2085–2097. Yao, J., Yang, F., Zhuang, Z., Shao, Y., & Yuan, P. F. (2018). The effect of personal and
Li, W., & Shang, Y. (2008). Prediction of energy production and energy consumption microclimatic variables on outdoor thermal comfort: A field study in a cold season in
based on BP neural networks. IEEE International Symposium on Knowledge Acquisition Lujiazui CBD, Shanghai. Sustainable Cities and Society, 39, 181–188.
& Modeling Workshop. Yousfi, S., Berrani, S. A., & Garcia, C. (2016). Contribution of recurrent connectionist
Li, X. L., Lin, D. M., Wang, R. J., & Wang, X. W. (2011). Hourly heating load prediction of language models in improving lstm-based arabic text recognition in videos. Pattern
radiant floor heating system based on the bp neural network. Advanced Materials Recognition, 64, 245–254.
Research, 243–249, 4913–4917. Yuce, B., Li, H., Rezgui, Y., Petri, L., Jayan, B., & Yang, C. (2014). Utilizing artificial
Liang, C., Zhen, W., & Gang, W. (2017). Application of lstm networks in short-term power neural network to predict energy consumption and thermal comfort level: An indoor
load forecasting under the deep learning framework. Electric Power Information and swimming pool case study. Energy and Buildings, 80, 45–56.
Communication Technology.. Zhu, H., Yu, C. W., & Cao, S. (2019). Ventilation online monitoring and control system
Liu, C., Wu, K., & Tsao, M. (2005). Energy efficient information collection with the from the perspectives of technology application. Indoor and Built Environment, 10(4).
ARIMA model in wireless sensor networks. IEEE Global Telecommunications Conference

10