A novel LSTM based deep learning approach for
multi-time scale electric vehicles charging load
prediction
Juncheng Zhu, Zhile Yang, Yan Chang, Yuanjun Guo, Kevin Zhu, Jianhua Zhang
Abstract—Short-term load forecasting is an important issue
in energy management system and a key measure to maintain
the stable and effective operation of power systems, providing
reasonable future load curve feeding to the unit commitment and
economic load dispatch. Specifically, load forecasting at different
time scales has various and corresponding roles in power systems.
In recent years, plug-in electric vehicles have gradually become
popular in major cities, and the number of electric vehicles is
growing rapidly. However, the mass roll out of EVs may cause
severe problems to the power system due to the huge charging
power and stochastic charging behaviours of the EVs drivers. The
accurate model of EVs charging load forecasting is therefore an
emerging topic. In this paper, artificial neural networks and a
long short term memory model based deep learning approaches
are employed and compared in forecasting the EVs charging
load from the charging station perspective. Numerical results
show that the long short term memory model has demonstrated
better performance and provided a model of higher accuracy in
short-term EVs load forecasting comparing with the traditional
artificial neural networks.
I. INTRODUCTION
Power load forecasting has long been an important com-
ponent in power and energy sectors and well studied since
the 1980s [1]. Featured load forecasting is based on the
operating characteristics of the system, capacity-enhancing
decisions, natural conditions as well as the social impacts,
determining load demand at particular time slots in the future
under conditions and providing fundamental reference for
power system scheduling. Accurate load forecasting could
significantly contribute to various aspects including econom-
ical and reasonable arrangement of the start and stop of
internal generator sets of a power grid, maintaining the safety
and stability of the grid operation, reducing the unnecessary
This research is financially supported by China NSFC under grants
61773252, Natural Science Foundation of Guangdong Province under grants
2018A030310671, China Post-doctoral Science Foundation (2018M631005),
and State Key Laboratory of Alternate Electrical Power System with Renew-
able Energy Sources (Grant No. LAPS18020)
Juncheng Zhu is the Zhengzhou University, Zhengzhou, 450001, China (e-
mail: jczhu
siat@163.com)
Zhile Yang and Yuanjun Guo are with Shenzhen Institute of Advanced
Technology, Chinese Academy of Sciences, Shenzhen, Guangdong, 518055,
China (e-mail:zl.yang,yj.guo@siat.ac.cn)
Yan Chang is with University of Science and Technology of China, Hefei
230026, China (e-mail: sa517011@mail.ustc.edu.cn)
Kevin Zhu is with Lynbrook High School, San Jose, CA, USA,
(email:Kevinrzhu123@gmail.com)
Jianhua Zhang is with the State Key Laboratory of Alternate Electrical
Power System with Renewable Energy Sources, North China Electric Power
University, Beijing 102206, China (e-mail: zjh@ncepu.edu.cn)
rotating reserve capacity, and eventually improving economic
and social benefits. Load forecasting could be divided into
ultra-short-term load forecasting, short-term load forecasting,
medium-term load forecasting and long-term load forecasting
depending on the expected time length. The ultra-short-term
load forecasting refers to the load forecast within 1 hour ahead.
In the safety monitoring state, the predicted values of 5 to 10
second or 1 to 5 minute are required, and the preventive control
and emergency state processing require a predicted value of
10 minute to 1 hour. Various research approaches have been
developed in predicting the power demand in multiple time
scales.
The state-of-the-art power load forecasting models can
be categorized into traditional statistical models and artifi-
cial intelligence (AI) models. Traditional forecasting methods
majorly include the time series method [2], autoregressive
integrated moving average [3], regression analysis [4], Kalman
filtering [5], etc. On the other hand, AI methods consist of
artificial neural networks (ANN) [6], support vector machines
(SVM) [7], and deep learning methods [8]. Before the 21st
century, due to the strong adaptive, self-learning and genere-
lization ability, ANN had become an important technique in
load forecasting. Hippert et al. [9] reviewed the application
of ANNs for load forecasting and claimed that ANNs have
effectiveness for load forecasting in terms of the accuracy and
efficiency.
In recent years, the deep learning methods have been
obtained wide attractions and used in image semantic seg-
mentation [10], classification [11], target detection [12], nat-
ural language processing [13] and many other science and
engineering fields. The network structure constructed by the
deep learning methods are more complex with a large number
of hidden layers and/or recurrent structure, which endowing
stronger learning and self-adaptive ability than ANN methods.
Therefore, it has also been paid attention in the field of load
forecasting. In 1998, Vermaak and Botha [14] used recurrent
neural network (RNN) for the first time to establish a short-
term load forecasting model. However, the conventional RNN
would suffer from the gradient vanishing problem and the long
short term memory (LSTM) network is an effective approach
to relief the issue. More recently, Marino et al. [15] proposed a
LSTM architecture to forecast the load of individual residences
[16]. Kong et al. [17] combined the energy consumption of
a residence with the behaviour of a resident, converted the
behaviour patterns of energy consumers into a sequence of
3531
2019 IEEE PES Innovative Smart Grid Technologies Asia
978-1-7281-3520-5/19/$31.00 ©2019 IEEE