Transportation Research Part C 153 (2023) 104205

Contents lists available at ScienceDirect

Transportation Research Part C

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

Predicting electric vehicle charging demand using a

heterogeneous spatio-temporal graph convolutional network
Shengyou Wang a, f, h, Anthony Chen b, e, g, Pinxi Wang c, Chengxiang Zhuge d, e, f, g, *
a
Key Laboratory of Transport Industry of Big Data Application Technologies for Comprehensive Transport, Beijing Jiaotong University, 3
Shangyuancun, Xizhimenwai, Beijing 100044, China
b
Department of Civil and Environmental Engineering, The Hong Kong Polytechnic University, Hong Kong, China
c
Beijing Transport Institute, No. 9 LiuLiQiao South Lane, Fengtai District, Beijing 100073, China
d
Department of Land Surveying and Geo-Informatics, The Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong, China
e
Research Institute for Sustainable Urban Development, The Hong Kong Polytechnic University, Hong Kong, China
f
The Hong Kong Polytechnic University Shenzhen Research Institute, Shenzhen, China
g
Smart Cities Research Institute, The Hong Kong Polytechnic University, Hong Kong, China
h
School of Traffic Management, People’s Public Security University of China, 100091 Beijing, China

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

Keywords: Short-term Electric Vehicle (EV) charging demand prediction is an essential task in the fields of
Electric Vehicle smart grid and intelligent transportation systems, as understanding the spatiotemporal distribu­
Charging Demand Prediction tion of charging demand over the next few hours could help operators of charging stations and the
Heterogeneous Graph
grid to take measures (e.g., dynamic pricing) in response to varying charging demand. This study
Spatio-temporal Data Mining
Graph Convolutional Network
proposed a heterogeneous spatial–temporal graph convolutional network to predict the EV
charging demand at different spatial and temporal resolutions. Specifically, we first learned the
spatial correlations between charging regions by constructing heterogeneous graphs, i.e., a
geographic graph and a demand graph. Then, we used graph convolutional layers and gated
recurrent units to extract spatio-temporal features in the observations. Further, we designed a
region-specific prediction module that grouped regions based on graph embedding and point of
interest (POI) data for prediction. We used a large real-world GPS dataset which contained over
76,000 private EVs in Beijing for model training and validation. The results showed that,
compared with recently popular spatio-temporal prediction methods, the proposed model had
superior prediction accuracy and steady performance at different scales of regions. In addition,
we conducted ablation studies and hyperparameter sensitivity tests. The results suggested that
incorporating the demand graph and geographic graph could help improve model performance.

1. Introduction

Due to the global energy crisis and climate change, reducing fossil fuel consumption and moving toward carbon neutrality has
become a common objective for many governments (Hu et al., 2018; Ma and Faye, 2022). Electric Vehicles (EVs), as an important
alternative to convolutional vehicles (e.g., petrol cars), have been widely accepted as a solution to reduce global carbon dioxide

* Corresponding author.
E-mail addresses: [email protected] (S. Wang), [email protected] (A. Chen), [email protected] (P. Wang), chengxiang.
[email protected] (C. Zhuge).

https://doi.org/10.1016/j.trc.2023.104205
Received 27 September 2022; Received in revised form 26 April 2023; Accepted 9 June 2023
Available online 23 June 2023
0968-090X/© 2023 Elsevier Ltd. All rights reserved.
S. Wang et al. Transportation Research Part C 153 (2023) 104205

emissions (Engel et al., 2018; Yuan et al., 2019). The EV market has soared in the past decade due to supportive government policies
and technological advancements by EV manufacturers (Yang et al., 2021). The widespread promotion of EVs has resulted in an
increasing demand for electricity (Lin et al., 2019). Although governments and agencies have put lots of effort into deploying charging
infrastructure, the lack of charging infrastructure is still a barrier to the purchase of EVs (Murugan and Marisamynathan, 2022).
Moreover, the imbalance between supply and demand for EV charging remains as a problem. On the one hand, the low-density
charging station network cannot support such increasing charging demand, due to the rising EV adoption rate (Yi et al., 2021). On
the other hand, some charging stations are significantly underutilized due to their difficult access (Yi et al., 2022). Furthermore, the
highly cyclical and volatile nature of EV charging demand (particularly large spikes of charging demand in a super-short-term time
scale) causes intense pressure on the power grid (Boulakhbar et al., 2022; Zhu et al., 2019).
Therefore, it is necessary for the power grid system to know the temporal and spatial distributions of short-term EV charging
demand in advance, so as to adopt appropriate management strategies in the context of the smart grid. Specifically, accurately pre­
dicting charging demand in different regions can play an essential role in addressing the supply–demand imbalance issue, and the
prediction results can be used in the three key aspects of the smart grid. Firstly, it can assist in releasing information on available
charging stations to EV users, so as to better satisfy the heterogeneous charging demand. Second, charging station operators can
properly process charging requests of EV users for economic dispatch at multiple time scales, so as to maintain a high utilization rate of
all charging facilities. Third, the charging demand prediction information benefits the smart grid system from both economic and
security perspectives, particularly in terms of optimal utility operations, power quality and security monitoring, peak shaving and
frequency regulation services (Buzna et al., 2021; Zhu et al., 2019). For example, the dynamic electricity pricing strategy, which
fluctuates according to variations in grid loads (including EV charging demand), would impact EV users’ charging behavior. Cheaper
charging fees may encourage EV users to postpone charging behavior dramatically, shifting charging demand to peak periods (e.g.,
midnight) to off-peak periods (e.g., early morning) (Hu et al., 2019). However, the possible negative influence of dynamic electricity
pricing on the traffic system should also be considered due to the coupled transportation and electricity systems (Ma and He, 2021).
To accurately predict charging demand in the short term, some attempts have been made (Yao et al., 2022). Since the future
charging demand is correlated with its historical demand, some studies developed a series of statistically based time series forecasting
models such as Moving Average (MA), Autoregressive (AR), and Autoregressive Integrated Moving Average (ARIMA) (Kalogirou,
2003). However, these methods generally assumed that forecast and historical data are linearly correlated. Thus they could not reflect
the non-linear characteristics (e.g., weather and date) in real situations (Flammini et al., 2019). Compared with statistical model-based
methods, machine learning methods or deep learning methods can model complex non-linear information in temporal terms with
remarkable results (Amara-Ouali et al., 2022). For example, Soldan et al. (2021) applied a logistic regression model to predict charging
station usage using temporal and historical occupancy information as input features (Ermagun and Levinson, 2019). Ma and Faye
(2022) used a hybrid Long Short-Term Memory (LSTM) network (Hochreiter and Schmidhuber, 1996) to merge heterogeneous fea­
tures, i.e. historical charging state and multiple time-related features, to predict EV charging demand. To sum up, the existing research
of EV charging demand prediction pays more attention to extracting temporal characteristics for EV charging demand, while ignoring
the spatial collection between different regions. However, according to the Tobler’s First Law of Geography (Tobler, 1970), the de­
mand of the target region could be strongly influenced by its spatial neighbors. Meanwhile, there may also be strong demand cor­
relations between those regions that are far apart but have similar built environments (Yu et al., 2020). Therefore, it would be useful
but also challenging to capture and incorporate the spatio-temporal characteristics for EV charging demand prediction.
A common method for characterizing spatial correlations is to create grid-based maps over the study area, and then to apply
convolutional neural networks (CNNs) and their variants to construct spatial correlation patterns (Lee and Rhee, 2022). Furthermore,
to capture spatio-temporal correlation concurrently, some scientists integrated CNN with Recurrent Neural Network (RNN) into a
hybrid network (Guo and Zhang, 2020; Li et al., 2023). However, EV charging demand across regions cannot be extracted and captured
by these types of prediction methods accurately since they ignore the heterogeneity of spatial correlations. For example, those regions
with a similar built environment but far away from each other might still exhibit a high correlation in EV charging demand, while CNN
and RNN-based approaches do not account for this. To extract the heterogeneity of spatial correlation, Graph-based data structures and
Graph Convolutional Networks (GCNs) have received considerable attention in demand prediction due to their high prediction ac­
curacy (Jin et al., 2020; Yu et al., 2020). Graphs are a type of data that can quantify the connections between various entities, which
can be used to obtain the spatial correlation of its neighborhood by GCNs (Wu et al., 2020).
Inspired by the successful application of GCN for graph data, we attempt to construct graphs to represent the spatial correlation of
charging demand, and propose a heterogeneous spatio-temporal graph convolutional network for predicting short-term charging
demand at the region level. Specifically, the prediction model will have two advanced modules: First, we will construct heterogeneous
graphs (i.e., a geographic graph and a demand graph) to model the spatial correlations between charging regions for spatio-temporal
EV charging demand prediction; Second, we will design a region-specific prediction module that clusters charging regions into groups
and trains region-specific prediction layers. Compared to unified ones, this module could help improve prediction accuracy. In
addition, we will use a unique one-month EV trajectory dataset in Beijing in January 2018, which contains rich charging information of
over 76,000 actual EV users. This will serve as an important dataset for model training, and validation.

2. Literature review

2.1. EV charging demand estimation and prediction

Many attempts have been made to estimate and predict EV charging demand. In this study, the EV charging demand prediction

2
S. Wang et al. Transportation Research Part C 153 (2023) 104205

refers to an accurate prediction on charging demand for a short period (e.g., in the next few minutes or hours), which is useful for real-
time scheduling and management of charging facilities and vehicles; while the EV charging demand estimation refers to the estimation
of EV charging demand within various scenarios, usually from a long-term perspective. The estimation results are useful for EV policy
making (e.g., EV subsidies) and facility planning (e.g., charging station location) (Yao and Moawad, 2019). Our study falls into the
former group, namely short-term EV charging demand prediction.
The EV charging demand estimation has received considerable attention in previous studies. For example, Valsera-Naranjo et al.
(2011) estimated the EV charging demand in winter, and applied both deterministic and probabilistic methods to analyze the impact of
the charging demand on the power grid. Tu et al. (2016) estimated the electric taxi charging demand using GPS trajectory data and the
transportation network in Shenzhen. Dong et al. (2019) delineated the area into a discrete grid with a resolution of 1 km × 1 km, which
served as the analysis unit for estimating the EV charging demand. From a long-term perspective, Luo et al. (2011) applied the Monte
Carlo simulation method to estimate the EV charging demand in China for the years of 2015, 2020, and 2030. Salah et al. (2015) used
the substation load data in Switzerland in 2010 to fit a charging demand curve, so as to estimate the long-term EV charging demand in
2040.
Compared to EV charging demand estimation, short-term charging demand prediction has received much less attention, in part
because such short-term predictions need proper datasets for model training and validation, but the privacy-sensitive character of this
data makes it extremely difficult to acquire (Butt et al., 2019; Flammini et al., 2019). Few attempts have been made to develop short-
term EV charging demand prediction models. For example, Ma and Faye, (2022) developed a hybrid LSTM neural network with EV
charging session data in the city of Dundee, which could predict charging demand for each charging pile in 10 min and 1 h. The results
showed that the proposed model performed better than LSTM and GRU models. However, this study only considered the temporal
characteristic and ignored the spatial feature of the EV charging demand. Yi et al. (2021) presented a sequence-to-sequence neural
network (Shen et al., 2020) to predict the EV charging demand in one month and validated it with real-world datasets. The case study
showed that the proposed approach can effectively infer the charging demand. However, the time interval of prediction was a month.
Hüttel et al. (2021) proposed Temporal Graph Convolutional Networks to predict 1-day, 7-day and 30-day EV charging demand, where
the time interval of the data is one day. The proposed model results outperformed have superior prediction performance compared to
other methods (e.g., CNN, CNN + LSTM). However, The time intervals of these charging demand predictions are too long (e.g., days,
months), which is not conducive to real-time applications for scheduling and management of charging facilities and vehicles.

2.2. Graph-based Spatio-temporal modeling

Graph neural network has shown excellent performance in a range of demand prediction problems, due to its powerful ability to
represent the spatial structure and learn contextual information (Jiang and Luo, 2022; Salah et al., 2015; Wu et al., 2021). However,
such modeling approaches have been seldom applied in EV charging demand prediction. Therefore, we will review how graph-based
spatio-temporal modeling approaches have been applied to the other demand prediction problems in relevant fields, including ride-
hailing demand prediction, public transit demand prediction, human mobility demand prediction, taxi demand prediction, and traffic-
related demand prediction.
For ride-hailing demand prediction, Huang et al. (2022) applied dynamic Origin-Destination (OD) graphs to character the ride-
hailing demand, and introduced a Graph Convolutional Network (GCN) to predict short-term (20 min) ride-hailing demand be­
tween origin and destination. The results showed that the proposed model performed better than the baseline models, i.e. Multiple
Layer Perceptron (MLP), LSTM. Feng et al. (2022) developed a GCN model to predict OD based ride-hailing demand. The results
showed that the proposed model outperformed Random Forest and MLP model significantly. For public transit demand prediction, Li
et al. (2022) treated the OD as a graph, and developed a probabilistic graph convolution model to predict robust public transit demand.
The results demonstrated that the proposed model got higher accuracy than several established tools, i.e., Linear Regression (LR), and
ARIMA. For human mobility demand prediction, Wang et al. (2021) applied a GCN model to merge the feature of metro system to­
pology and pedestrian travel patterns of multiple time intervals for node-level passenger flow prediction. The prediction accuracy of
the experiment showed the superior performance of the proposed method compared with the Locally Weighted Linear Regression
(LWR), RF model, and ARIMA models. For taxi demand prediction, Chen et al. (2020) presented a model based on GCN, and LSTM to
merge the spatial feature of taxi trips and temporal features of traffic flow to achieve an accurate prediction of taxi demand. The
experimental results showed that the proposed model performed a higher accuracy of the short-term taxi demands than MLP and
Support Vector Regression (SVR).
For traffic-related demand prediction, a road network is usually treated as a graph, in which road intersections as treated as nodes.
Several attempts have been made to develop graph-based spatio-temporal modeling methods, and finally use the feature information
of each node to make predictions (Rodrigues, 2022; Yin et al., 2021). Li et al. (2018) proposed a diffusion convolutional recurrent
neural network (DCRNN) for traffic prediction. DCRNN employs bi-directional random walks on a graph to capture spatial correlations
in traffic flows and an encoder-decoder structure to gather temporal correlations. In addition, considering the capability of a GCN to
represent spatial features, Zhao et al. (2019) adopted GCN and combined it with Gate Recurrent Unit (GRU) to obtain spatio-temporal
patterns in traffic. Yu et al. (2018) developed a spatio-temporal graph convolutional structure model that enables to capture spatio-
temporal features simultaneously with fewer parameters. Furthermore, Guo et al. (2019) introduced the attention mechanism into
a spatio-temporal graph convolutional network and used multi-graphs to model three temporal attributes of traffic flow, namely,
recent, daily-periodic, and weekly-periodic dependencies. To overcome the spatial dependence of the above work on the static graph
structure, Wu et al. (2019) proposed the Graph WaveNet for spatio-temporal graph modeling. Specifically, Graph WaveNet can learn
the adjacency matrix adaptively by node embedding and then combine it with a one-dimensional convolutional network to achieve

3
S. Wang et al. Transportation Research Part C 153 (2023) 104205

end-to-end traffic prediction.

2.3. Comments on previous work

As reviewed above, we identified two research gaps: first, short-term charging demand prediction has received much less attention
than EV charging demand estimation, in part because the short-term prediction generally needs a large charging dataset for model
training and validation. In response, this paper will use a unique large EV dataset, which contains charging information of over 76,000
private EVs in Beijing. Second, graph-based spatio-temporal modeling approaches have shown good performance in a wide range of
demand prediction problems, but their performance in EV charging demand prediction remains unclear. In response, this paper will
propose a heterogeneous spatial–temporal graph convolutional network for short-term electric vehicle charging demand prediction. In
particular, to mine the spatial features of EV charging demand, we treat the geographical region distribution and EV charging demand
pattern as a geographic graph and a demand graph, respectively, so as to implement the spatio-temporal EV charging demand pre­
diction with heterogeneous graphs. Furthermore, we will design a region-specific prediction module that clusters regions into groups
and trains region-specific prediction layers, which is expected to further improve prediction accuracy.

3. Preliminaries

In this paper, EV charging demand is defined as the number of EVs served within a certain time interval (e.g., 1 h) and a specific
area or region. The EV charging demand prediction aims to predict charging demand over a short-term period (i.e., 3 h) in the near
future, primarily using the observed historical charging demand.
The EV charging demand in all regions is defined as X = (X1 , X2 , ..., Xt , ...) ∈ RN×n , where t = (1, 2, ..., n), n is the number of
observed charging demand records for a specific region; N is the number of regions. Xt = (x1,t , x2,t , ..., xN,t )T , xN,t denotes the EV
charging demand in the N-th observation region at t-th time series. The EV charging demand prediction problem is formulated by
Equation (1).

(1)
f (⋅)
[Xt− m+1 , ..., Xt ] ̅̅→[Xt+1 , ..., Xt+p ]

where m (1⩽m < n) is the length of historical EV charging demand records (i.e., the length of input data); p (0 < p < t +p⩽n) is the
length of charging demand records to be predicted (i.e., the length of output data). Essentially, the objective of an EV charging demand
prediction model is to learn a function f(.) that can well predict the EV charging demand for the next p steps based on the information
extracted from the previous m steps.
The above function f(.) mainly consider the temporal characteristics. To improve the prediction accuracy, we further extract the
spatial characteristics of the region. An EV charging region network is defined as an undirected graph G = (V,E,A), where V is a set of
vertexes, i.e., a set of regions in the study area |V| = N, where N is the number of regions. Edges e(vi , vj ) ∈ E represent the connection
between a pair of charging regions. The adjacent matrix of graph G is represented by A ∈ RN×N , in which aij = 1 indicates the regions i
and j are connected and aij = 0 otherwise. Therefore, the EV charging demand prediction problem for a network of charging regions can
be formulated by Equation (2).

Fig. 1. The structure of the heterogeneous spatio-temporal graph convolutional network.

4
S. Wang et al. Transportation Research Part C 153 (2023) 104205

(2)
f (⋅)
[Xt− m+1 , ..., Xt , G] ̅̅→[Xt+1 , ..., Xt+p ]

where the function f(.) in Eq. (2) contains both temporal and spatial features.

4. Methodology

As reviewed, the temporal and spatial characteristics of EV charging demand are critical but have received little attention. To
obtain accurate prediction, we attend to capture the spatio-temporal correlation of EV charging demand at the region level. Specially,
we propose a novel heterogeneous spatio-temporal deep learning model for the EV charging demand prediction. We will start with an
introduction to the overall model framework and the basic functions of each module in Section 4.1. Then, we will elaborate on the
design of each module and present the detailed formulation in Sections 4.2, and 4.3. Finally, we will provide the region-specific
prediction scheme in Section 4.4.

4.1. Model overview

Fig. 1 shows the overall structure of the proposed model. The model comprises two main modules, namely, the spatio-temporal
learning module and the region-specific prediction module.
Module 1: the spatio-temporal learning module. This module is designed for extracting the spatial correlation with two different
types of graphs, i.e., a geographic graph and a demand graph (see Section 4.2). Graph Convolutional Network (GCN) layers and Gated
Recurrent Units (GRU) are employed in the module to extract spatio-temporal features. After the historical observation with two types
of graphs going through the network, to fuse the features of the graphs, we design a heterogeneous hidden representation ensembled
by a linear combination (see Section 4.3).
Module 2: the region-specific prediction module (see Section 4.4). To further improve the prediction accuracy, boost computational
effectiveness, and avoid falling into the local optimum, this module clusters the regions into different groups for region-specific
prediction by using the K-means clustering algorithm. This module considers the region correlation which is determined by the
geographical Graph Embedding and Point of Interest (POI). Finally, the region-specific output layers are trained and a Fully Connected
(FC) layer is used to obtain the final prediction.

4.2. Spatial correlation

According to the Tobler’s First Law of Geography (Tobler, 1970), spatial correlation is highly related to geographic distance (Yu
et al., 2018; Zhao et al., 2019). In other words, the closer the geographic distance between the two regions, the stronger the EV
charging demand correlation. To represent this spatial characteristic, we introduce a geographic graph and set the geographic weight
to be higher when the geographic distance is closer (see section 4.2.1). In addition, the two regions may also have strong charging
demand correlations despite the geographic distance (Yu et al., 2020). We introduce a demand graph and set the demand weight to be
higher when the demand correlation between regions is closer (see section 4.2.2). Therefore, we construct a heterogeneous graph
ensemble to incorporate both the geographic and demand graphs to characterize the spatial correlation scenario with a close
geographic distance and a strong charging demand correlation (see section 4.3.3). This is expected to improve prediction accuracy and
make the model more realistic.

4.2.1. Geographic graph

We use the Euclidean distance (Morin et al., 2020) to measure the adjacency of charging region. A pair of charging regions are
considered adjacent only when their Euclidean distance is smaller than a pre-defined threshold value. The geographic graph is
mathematically formulated by Equation (3).
{
1, dgeo (vi , vj ) < εg
Agij = (3)
0, otherwise

g
where Aij is the geographic adjacent relationship of region i and region j; εg is the threshold; dgeo (vi , vj ) is the Euclidean distance be­
tween region i and region j, which can be calculated based on their locations.

4.2.2. Demand graph

Aggregating observations from regions with similar charging demand patterns can generally help to improve prediction accuracy.
However, these regions are not always located close to each other (Lei et al., 2022). As a result, the geographic graph above could not
capture this information. In response, we construct a demand graph based on the extent to which a pair of regions is similar to each
other in terms of their time-series of charging demand. Specifically, we utilize Dynamic Time Warping (DTW) (Herrmann and Webb,
2023) to measure the similarity of two charging region demand sequences. Compared to the Euclidean distance, DTW incorporates a
temporal alignment of the series through dynamic programming and further assesses their similarity. In other words, DTW is applied to
find the minimum distance between objects based on their time series data. Here, the minimum distance is used to represent the
similarity of two charging regions in charging demand. The calculation method of DTW is given in Appendix 1 of Supplementary
Materials. The demand graph is presented by Equation (4).

5
S. Wang et al. Transportation Research Part C 153 (2023) 104205

{
1, ddem (vi , vj ) < εd
Adij = (4)
0, otherwise

where Adij is the adjacent relationship of charging demand for region i and region j; εf is the threshold used to decide the sparsity of the
demand graph; ddem (vi , vj ) is the DTW distance between region i and region j.

4.3. Spatio-temporal learning module

With the heterogeneous graph above, we design a spatio-temporal learning module which extracts spatio-temporal correlations
from the observations. In this module, we apply graph convolutional layers for extracting spatial correlations and gated recurrent units
(GRU) for learning temporal relations. As shown in Fig. 1, the module takes the observation with two graphs as input, i.e., a geographic
graph and a demand graph. Then, the graphs are fed into two groups of graph convolutional layers and gated recurrent units,
respectively. Finally, the hidden representations of heterogeneous graphs are ensembled by linear combination.

4.3.1. Graph convolutional layer

Due to the irregularity of regional relations (see the established adjacency (i.e., geographic and demand graphs) of the charging
region in Section 4.2), the distribution of the EV charging demand at the region level is a non-Euclidean structure. The approaches,
including convolutional Neural Networks (CNN) (Bouwmans et al., 2019) and ConvLSTM, do not meet the requirements of our study,
because they are capable of aggregating data with Euclidean structure but show limitations when processing data in non-Euclidean
space, such as graphs. However, as an extension of CNN on the graph domain, a Graph Convolutional Network (GCN), which
learns spatial correlations from non-Euclidean data, has been widely applied in various spatio-temporal learning models (Jiang and
Luo, 2022; Wu et al., 2021). Therefore, GCN is chosen to address the challenge of extracting the spatial characteristics of EV charging
demand.
Inherits from convolutional convolutional kernels, the idea of graph convolutional is to aggregate node features from neighboring
nodes with a filter in the Fourier domain (Kipf and Welling, 2016). Given a graph (V, A, X), where X ∈ RN×F is the node feature, and F is
the number of features. A GCN layer can be formulated by Equation (5).
1 1
̃− 2A
Xgcn = σ (D ̃D̃ − 2 XW) (5)

̃ = A +I ∈ RN×N is the normalized adjacent matrix with self-connection; I is the identity matrix; D ∑̃
where A ̃=
j Aij s the degree matrix.

W = RF×F is learnable parameter; Xgcn = RN×F is the output node feature; σ(.) is the sigmoid function for nonlinear activating.
′ ′

As a single layer of GCN only aggregates information from one-hop neighbors (directly adjacent), we stack two GCN layers for a
broader view of spatial dependence in this research. The input of the first layer X0 = RN×m is the historical observation from the
previous h steps.

4.3.2. Gated recurrent unit (GRU)

After learning spatial correlations by GCN layers, we focus on extracting the temporal characteristics, which are another key factor
in predicting EV charging demand. Based on the temporal propagation structure, recurrent layers are widely applied in various time-
series modeling tasks. GRU is a variant of recurrent neural networks (RNN) and is able to learn the temporal characteristics from the
observations. GRU solves the gradient vanishing problem in convolutional RNN and has a more efficient structure than the classical
long short-term memory (LSTM) (Ahuja et al., 2022). It consists of reset gates r and update gates z, which decide whether to pass the
hidden information from the previous step to the next one. Given a sequence of hidden representations from GCN layers [X1 , X2 , ...,
Xt , ...], where t = (1, 2, ..., m) , the recursion can be formulated by Equations (6)-(9).
rt = σ (Wr ⋅ (ht− 1 , Xt ) + br ) (6)

zt = σ (Wz ⋅ (ht− 1 , Xt ) + bz ) (7)

̃
ht = tanh (Wh ⋅ (rt ∗ ht− 1 , Xt ) + bh ) (8)

ht = zt ∗ ̃
ht + (1 − zt ) ∗ ht− 1 (9)

where ht− 1 is the hidden state of the previous time step; ̃ ht is the candidate state that stores temporary information from reset and
update gates; σ is the sigmoid activation function; Wr , Wz , and Wh are learnable weights; br , bz , and bh are learnable biases that can be
updated by backpropagation.

4.3.3. Heterogeneous graph ensemble

To represent the two graphs, simple fusion processes, such as addition, are not appropriate because the different graphs were
unlikely to contribute equally to EV charging demand. Therefore, we attempt to ensemble the two graphs by a linear combination to
better represent the heterogeneous spatial correlation.

6
S. Wang et al. Transportation Research Part C 153 (2023) 104205

As aforementioned, we first model the spatial correlations with two different graphs, and then feed the two graphs along with the
historical observations into two groups of GCN layers and GRU cells. This can be formulated by Equations (10)-(11).
H g = ST([Xt− m+1 , ..., Xt ], A
g
) (10)

H d
= ST([Xt− m+1 , ..., Xt ], A
d
) (11)

where ST represents GCN layers and GRU cells (see Section 4.3); [Xt− m+1 , ..., Xt ] ∈RN×m is the historical observation of the past m steps;
Ag and Ad are the geographic adjacent matrix and demand adjacent matrix (see Section 4.2), respectively. Hg ∈ RN×K and Hd ∈ RN×K
are the outputs from the two groups of GCN layers and GRU cells, respectively, where K is the hidden dimension of GRU cells.
Therefore, Hg and Hd are assembled by a linear combination to represent the heterogeneous spatial correlation, as shown by Equation
(12).

H = α × H g + (1 − α) × H d (12)

where α is a self-adaptive weight that is updated in the training stage along with other network parameters via backpropagation; H ∈
RN×K is the hidden representation of heterogeneous graphs, which is utilized as the output of the spatio-temporal learning module.

Fig. 2. The EV charging demand in Beijing aggregated with different grid scales.

7
S. Wang et al. Transportation Research Part C 153 (2023) 104205

4.4. Region-specific prediction module

After learning spatio-temporal features by GCN layers and GRU units, a common practice in the existing spatio-temporal prediction
model is utilizing a unified prediction module to map the hidden representation to the output, e.g., linear layer and convolutional layer
(Wu et al., 2019; Zhao et al., 2019). However, as charging demand patterns vary across charging stations/regions, using a unified
model ignores these differences and may further compromise the prediction accuracy. To address this issue, we design a novel region-
specific prediction module that provides specialized prediction modules for regions with different charging demand patterns. The
module has the advantage of further improving the prediction accuracy, boosting computational effectiveness, and avoiding falling
into the local optimum. The module starts by clustering regions into different groups based on region correlations, i.e., graph
embedding and point of interest (POI). Then, a group of linear layers is trained for the region-specific prediction. We will introduce
region correlation modeling and region-specific prediction in Sections 4.4.1 and 4.4.2, respectively.

4.4.1. Region correlation modeling

Due to the mobility of EV users, the charging demand in one region may be affected by its neighbors. For example, high EV charging
demand is concentrated in those regions in central areas, while those regions in suburban areas tend to have a low charging demand
(see Fig. 2). Therefore, the geographic adjacent relationship is a crucial characteristic that needs to be considered in the model. In
attention, the EV charging demand in a region is closely associated with EV users’ travel demand, and thus should also be associated
with land use patterns (or the built environment). For example, residential regions may have high EV charging demand on weekdays,
while restaurants and shopping malls may attract more EV users on weekends; The region with a small number of POIs may have low
charging demand. Thereby, some regions may have similar EV charging demand patterns even though they are far away from each
other, which indicates that POIs can be applied to explore the spatial interaction of charging demand.

(1). Graph Embedding (GE).

Graph Embedding (GE) is to vectorize the adjacency relations of geographic graphs and, at the same time, reduce the dimen­
sionality of the geographic adjacency matrix. A geographic graph has more than hundreds of dimensions (see Fig. 4), while POIs only
have 12 dimensions in our study (representing 12 POI types). Their dimensions are greatly different. Therefore, we use GE to transform
the geographic characteristics, so as to better integrate geographic information and POI information.
Recall that we have defined a geographic graph G = (V, E, Ag ) in Section 4.2. Graph embedding methods aim to compress a sparse
high-dimensional graph into a low-dimensional latent representation, which can be utilized in downstream tasks (Cai et al., 2018). The
mathematical expression can be learned by a mapping function f(⋅) : RN×N →RN×d that represents the graph topology RN×N as a
low-dimensional matrix representation RN×d , where d is much smaller than N. In this paper, we employ DeepWalk, which has been
widely adopted in various graph-based tasks, such as node classification, and link prediction (Perozzi et al., 2014). DeepWalk starts by
generating random walk sequences W = {v1 , v2 , ..., vn } from a certain node v with a transition probability between nodes. We define the
transition probability from node vi to node vj by Equation (13).
⃒ wij
Pr(vi ⃒vj ) = ∑ (13)
vi ∈N(vj ) wij

where N(vj ) is the set of adjacent nodes of vj ; wij is the weight between nodes which is represented by the geographic distance in this
study. In this way, random walks rooted from neighboring nodes are similar. Then, the skip-gram algorithm is adopted to learn the
node embedding (Mikolov et al., 2013), as shown in Appendix 2 of Supplementary Materials. After many experiments, to maximize the
performance of the proposed model, the window size and the dimension of embedding are set to 10 and 16, respectively. Thus, the
graph topology is embedded into GE ∈ RN×16 , where those neighboring nodes have similar embedding vectors.

(2). Point of Interest (POI).

Unlike graph embedding (GE), which represents graph topology, Point of Interests (POIs), which are a series of attributes that
counts the number of objects in a specific region with different functions, such as residential and leisure functions, have been widely
used to characterize land use patterns (Bian et al., 2022; Liang et al., 2023), and thus can be potentially used to characterize similarity
of EV charging demand pattern. In this study, we utilize different types of POI to describe the functions of a region. Specifically, we
consider 12 different POI types (see Section 5.1), POI ∈ RN×12 . Such POI data can be collected from map service providers, such as
OpenStreetMap, Google Map, and Baidu Map. In this study, we used POIs from the Baidu Map.

4.4.2. Region-specific prediction

With the GE and POI features above, we can train region-specific prediction layers. We start by concatenating GE and POI and
employ K-means clustering to group regions into a specific number of clusters (denoted as C here). Then, for each cluster, we train fully
connected layers. As depicted in Fig. 1, nodes are coded with different colors after clustering and then fed into corresponding Fully
Connected (FC) layers. The region-specific prediction can be formulated by Equation (14).
̂t
[X + 1 , ...,
̂ t+p ] = σ (Wc × Hc + bc ) c ∈ [1, ..., C]
X (14)

8
S. Wang et al. Transportation Research Part C 153 (2023) 104205

where p is the prediction length; Hc is the hidden representation of cluster c from the previous spatial–temporal learning module; Wc
and bc are corresponding learnable parameters; σ is the sigmoid activation function.

4.5. Model performance evaluation

To evaluate the performance of the proposed model, we will compare it against the following spatio-temporal prediction methods
(see Table 1 for a comparison between these methods):

• History Average 1 (HA1) is a traditional statistical method that generates predictions by simply using the arithmetic average of
previous observations (Kamarianakis and Prastacos, 2003). In this study, we average EV charging demands of the past 12 steps as
the prediction.
• History Average 2 (HA2) is another format of History Average with a different data structure (compared to HA1). It predicts
charging demand at a specific time of day by averaging the charging demands at the same time of day over the past few days. In this
study, for a specific time of day, we average EV charging demands of the past 12 days as the prediction.
• Support Vector Regression (SVR) is a machine learning method designed for solving regression problems. It is an extension of the
Support Vector Machine (SVM) (Cheng et al., 2017). SVR attempts to identify a non-linear regression equation in high-dimensional
feature space by tackling a quadratic programming problem (Chen and Yu, 2014). Essentially, SVR is a typical regression-based
method.
• Random Forest (RF) is a machine learning method and can be used to deal with classification and regression problems. Specif­
ically, RF applies the bootstrap sampling and classification and regression tree (CART) method to build multiple independent
decision trees (Liu et al., 2019). The outcomes are then combined by RF to improve prediction performance. In this study, we use RF
as a representation of the classification- and regression-based methods.
• Gated Recurrent Unit (GRU) is a deep learning method for time-series modeling, and it has shown its advantages in various
prediction tasks (Seng et al., 2021; Sun et al., 2022; Zhao et al., 2022). An introduction to GRU is presented in Section 4.3. In this
study, we use GRU as a representation of the time-series deep learning methods.
• Temporal Graph Convolutional Network (T-GCN) is a deep learning method and it incorporates graph convolutional layers and
GRU (Zhao et al., 2019). It has the same structure as the spatio-temporal modeling module (see Section 4.3). In this study, we use T-
GCN as a representation for extracting spatio-temporal relationships, where the spatial feature is contributed by a single geographic
graph.
• Spatio-temporal Graph Neural Controlled Differential Equation (STG-NCDE) is a novel deep learning model which particu­
larly introduced Neural controlled differential equations (NCDEs) for time series prediction (Choi et al., 2021). NCDEs were
recently proposed as ground-breaking approaches for processing sequential data, and it was considered as a continuous analog to

Table 1
Comparing our model against typical spatio-temporal prediction models.
Model Names Model Type Feature Extraction Structure

Statistical Machine Deep Temporal Spatial

Learning Learning

History Average 1 (HA1) √ √ • Average charging demands over the

previous time steps
History Average 2 (HA2) √ √ • Average charging demands at the same
time of day over the past few days
Support Vector Regression (SVR) √ √ • Regression
Random Forest (RF) √ √ • Classification and regression
Gated Recurrent Unit (GRU) √ √ • Temporal recurrent network
Temporal Graph Convolutional √ √ √ • Temporal recurrent layer
Network (T-GCN) • Spatial graph convolutional layer
STG-NCDE √ √ √ • NCDE for the temporal and spatial
processing
• Graph convolutional layer
Spatio-Temporal Graph √ √ √ • Temporal convolutional layer
Convolutional Network • Spatial graph convolutional layer
(STGCN)
Graph WaveNet (GWNet) √ √ √ • Self-adaptive Adjacency Matrix
• Temporal convolutional layer
• Spatial graph convolutional layer
MTGNN √ √ √ • Graph learning layer
• Graph convolution module
• Temporal convolution module
• Output module
Our Model √ √ √ • Geographic graphs and demand graphs
• Graph convolutional layers
• Gated recurrent units layers
• Region-specific prediction module

9
S. Wang et al. Transportation Research Part C 153 (2023) 104205

RNN. STG-NCDE applied one NCDE for the temporal extraction, and another for spatial processing. Then the two NCDEs and GCN
were combined into a framework for the sequential data prediction.
• Spatio-Temporal Graph Convolutional Network (ST-GCN) is a deep learning method and has recently received more attention
in prediction studies (Yu et al., 2018). Unlike other methods that apply recurrent units on time series, it saves computational effort
by applying convolutional operation on spatial and temporal dimensions (Yu et al., 2021). In this study, we use STGCN as a
representation to efficiently extract spatio-temporal relationships, where ST-GCN utilizes the geographic graph.
• Graph WaveNet (GWNet) is a novel spatio-temporal graph modeling method based on the deep learning approach, and it does not
rely on any explicit graph structure (Wu et al., 2019). It utilizes an adaptive dependency matrix which is trained through node
embedding. We use GWNet as a representation of prediction methods that do not need a predefined adjacent matrix.
• Multivariate Time series forecasting with Graph Neural Networks (MTGNN) is a general framework for graph neural networks
created especially for multivariate sequential data prediction (Wu et al., 2020). Firstly, it can automatically capture the uni-
directed relationships between variables through a graph learning module; then it uses the novel mix-hop propagation layer
and the dilated inception layer to extract the spatial and temporal dependencies within the sequential data.

5. Case study of Beijing

5.1. Dataset and model settings

In this study, we used a unique real-world EV trajectory dataset in Beijing, China, which contained 76,774 private EVs in January of
2018: see the work by Sun et al. (2021a) for a detailed introduction to the dataset. Each record in the dataset contains several key fields,
including time, location, state of charge, and instantaneous speed. Through an analytical framework proposed by Sun et al. (2021b),
we extracted charging events of EVs as charging demand, which was further aggregated at the region level. The EV charging events
occurred at 1,128 public charging stations with both slow and fast chargers available. Their power rates ranged from 7 to 30 kW. This
means that the proposed heterogeneous spatial–temporal graph convolutional network could be applied to real-world scenarios with
different charger types available. Although it is generally more useful to examine the performance of a new model using public datasets
(Amara-Ouali et al., 2022). However, our unique EV trajectory dataset would be a more appropriate datasets for fully testing the
performance of our proposed model. The reasons are twofold: First, our case study tested the model performance on a large-scale
charging network with over 1,000 charging stations (comprising both slow and fast charging posts); while the public datasets usu­
ally only contained a few charging stations. Consequently, the model performance on a large-scale charging network could not be well
examined with these datasets. Second, our model particularly considered spatial correlations between different regions. Those public
datasets with a small number of charging stations would not be qualified for testing this unique feature of our model. In addition, we
used POI data in Beijing, whose data source was Baidu Map, one of the most-used map service providers. The POI data contains 12
different types, namely Transport Infrastructure, Educational Institutions, Financial Institutions, Scenic Sports, Media Facilities,
Health Care Facilities, Gymnasia, Government Agencies, Companies, Service Facilities, Commercial Establishments, and Residential
Buildings. In total, there are 117,901 POIs. A detailed introduction to the POI dataset can be found in work by (Yang et al., 2022).
Moreover, to evaluate the model performance at different scales, we generated three networks of charging regions with different
grid scales, i.e., 1 km × 1 km, 2 km × 2 km, and 4 km × 4 km. The spatial distributions of charging events at different grid scales are
shown in Fig. 2 (note: the blank pixels indicate there are no charging events in this region). It can be found from Fig. 2 that the charging
demand is concentrated in the central areas. In addition, we can find the spatial patterns of EV charging demand were different when
different spatial analysis units were used, which is known as the modifiable areal unit problem (MAUP) (Dark and Bram, 2007). The
area of Beijing is 16,410 km2. There will be 6,253 grids under the scale of 1 km × 1 km, and 721 grids under the scale of 4 km × 4 km.
Understanding the model performance under different spatial scales would also be helpful for different stakeholders involved: for EV

Fig. 3. Temporal charging patterns in Beijing (note: we here use 1 × 1 grid as spatial analysis unit).

10
S. Wang et al. Transportation Research Part C 153 (2023) 104205

users and operators of charging stations, the predicted charging demand at a higher spatial resolution (e.g., 1 km × 1 km grid) would be
more helpful; while for the power grid system, the predicted charging demand at a lower spatial resolution would be enough.
Fig. 3-(a) and (b) show the temporal distributions of charging demand by the hour and day, respectively. It can be found that the
peak period in a day is from 11 am to 5 pm, and there are no significant differences between the seven days in a week.
We further visualize geographic and demand graphs at different grid scales in Fig. 4, The black pixels indicate that the two regions
are connected while the white ones indicate no connection. It can be observed that the connections in the geographic graph are
gathered on the diagonal, while the connections in the demand graph are scattered. The significant difference between the geographic
graph and demand graph suggests that using heterogeneous graphs (instead of a single graph type) can provide thorough spatial
correlations, which is expected to improve model performance. In addition, we can also find that the same type of adjacent matrix has a
similar pattern at different grid scales.
The proposed model utilizes the historical EV charging demands in the previous 12 h (or 12 steps) to predict charging demands in
the next 3 h and 6 h, which is a common practice in similar studies (Bogaerts et al., 2020; Fukuda et al., 2020; Ma and Faye, 2022). For
cross-validation, we sequentially divide the dataset into training, validation, and testing sets with a ratio of 3: 1: 1. The model pa­
rameters are trained by the Adam optimizer (Kingma and Ba, 2015) with an initial learning rate of 0.001. The model is trained for 200

Fig. 4. Visualization of different types of adjacent matrix.

11
S. Wang et al. Transportation Research Part C 153 (2023) 104205

Fig. 4. (continued).

epochs with a batch size of 32. The initial weight value of two TGCN modules in Eq. (12) is set to 0.5. All models are implemented with
PyTorch (Li et al., 2020) and run on a server with 12th Gen Intel (R) Core (TM) I7-12700KF@ 3.60 GHz, 64-bit operating system, 16 GB
RAM and NVIDIA GeForce RTX 3070 Ti.
In this study, the model’s prediction accuracy is evaluated by three commonly used metrics, i.e., Symmetric Mean Absolute Per­
centage Error (SMAPE), Root Mean Square Error (RMSE), and Mean Absolute Error (MAE), which are defined by Equations (15)-(17),
respectively.
N ⃒ ⃒
1 ∑ ⃒ yi − ̂y i ⃒
SMAPE(y, ̂ y) = ⃒ ⃒ × 100% (15)
N i ⃒(yi + ̂y i )/2 ⃒
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
√
√1 ∑ N
y) = √
RMSE(y, ̂ (yi − ̂y i )2 (16)
N i

12
S. Wang et al. Transportation Research Part C 153 (2023) 104205

Table 2
Comparing the short-term performance of different prediction models across spatial scales.
(a) Prediction results for the length of 3 h

Methods Metrics

1 km £ 1 km 2 km £ 2 km 4 km £ 4 km

SMAPE(%) RMSE MAE SMAPE(%) RMSE MAE SMAPE(%) RMSE MAE

(std) (std) (std) (std) (std) (std) (std) (std) (std)

HA1 50.85 1.29 0.63 48.4 1.65 0.75 47.68 2.83 1.06
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
HA2 34.78 1.16 0.64 38.01 1.41 0.72 42.96 2.191 0.950
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
SVR 31.28 0.80 0.35 30.98 0.98 0.42 32.24 1.43 0.55
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
RF 30.85 0.87 0.40 31.45 1.09 0.49 32.30 1.70 0.66
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
GRU 27.81 0.86 0.37 27.74 1.04 0.45 28.82 1.47 0.59
(0.005) (0.011) (0.014) (0.022) (0.049) (0.035) (0.048) (0.103) (0.078)
TGCN 29.01 0.89 0.40 34.55 1.14 0.56 41.99 1.70 0.80
(0.005) (0.002) (0.003) (0.011) (0.037) (0.022) (0.016) (0.043) (0.025)
STG- 35.82 0.80 0.41 40.93 1.03 0.53 35.52 1.51 0.77
NCDE (0.012) (0.013) (0.020) (0.015) (0.026) (0.030) (0.010) (0.016) (0.022)
STGCN 32.96 0.90 0.44 37.83 1.13 0.56 42.98 1.61 0.74
(0.015) (0.257) (0.067) (0.031) (0.076) (0.045) (0.069) (0.149) (0.061)
GWNet 26.48 0.80 0.35 25.82 0.97 0.41 26.89 1.41 0.57
(0.023) (0.029) (0.210) (0.109) (0.212) (0.122) (0.186) (0.080) (0.304)
MTGNN 26.68 0.74 0.33 28.61 0.89 0.41 36.65 1.38 0.59
(0.024) (0.205) (0.103) (0.037) (0.126) (0.304) (0.031) (0.017) (0.029)
Ours 24.28 0.73 0.32 24.87 0.88 0.39 26.35 1.23 0.51
(0.019) (0.053) (0.039) (0.027) (0.069) (0.047) (0.025) (0.016) (0.050)

(b) Prediction results for the length of 6 h

Methods Metrics

1 km£1 km 2 km£2 km 4 km£4 km

SMAPE(%) RMSE MAE SMAPE(%) RMSE MAE SMAPE(%) RMSE MAE

(std) (std) (std) (std) (std) (std) (std) (std) (std)

HA1 54.61 1.38 0.68 51.83 1.77 0.81 50.93 3.05 1.14
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
HA2 36.90 1.16 0.64 42.70 1.41 0.72 41.42 2.18 0.94
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
SVR 39.95 0.98 0.47 39.62 1.22 0.56 41.04 1.85 0.73
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
RF 39.57 1.06 0.52 40.01 1.35 0.63 40.94 2.17 0.86
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
GRU 34.89 1.03 0.49 34.72 1.25 0.58 36.27 1.80 0.76
(0.005) (0.011) (0.014) (0.015) (0.035) (0.040) (0.031) (0.039) (0.053)
TGCN 37.53 1.07 0.54 42.65 1.35 0.69 47.27 1.95 0.95
(0.005) (0.002) (0.003) (0.010) (0.006) (0.004) (0.008) (0.018) (0.013)
STG- 43.07 0.93 0.54 49.88 1.25 0.66 47.84 1.81 0.8
NCDE (0.302) (0.201) (0.054) (0.030) (0.063) (0.129) (0.023) (0.071) (0.103)
STGCN 40.30 1.05 0.54 42.35 1.29 0.65 55.32 1.83 0.84
(0.029) (0.059) (0.024) (0.034) (0.254) (0.086) (0.039) (0.050) (0.016)
GWNet 33.06 0.96 0.46 34.68 1.31 0.58 34.90 1.72 0.70
(0.031) (0.019) (0.027) (0.201) (0.103) (0.035) (0.219) (0.037) (0.501)
MTGNN 33.77 0.88 0.43 33.72 1.08 0.50 42.90 1.50 0.66
(0.025) (0.102) (0.213) (0.021) (0.177) (0.042) (0.038) (0.050) (0.101)
Ours 34.44 0.93 0.45 34.48 1.12 0.54 34.19 1.58 0.68
(0.026) (0.012) (0.018) (0.031) (0.080) (0.026) (0.021) (0.013) (0.011)

Note: “std” in the table stands for standard deviation, which is used to quantify the variances of the ten-run results).

13
S. Wang et al. Transportation Research Part C 153 (2023) 104205

1 ∑
N
MAE(y, ̂y ) = yi|
|yi − ̂ (17)
N i

where ̂
y i and yi denote the predicted and observed values, respectively; N is the number of samples.

5.2. Prediction performance

Table 2 shows the results with two different prediction lengths (i.e., 3 and 6 h) at three different levels of the spatial analysis unit.
To summarize, the proposed method in general outperforms other typical models, as evident from the three lower metrics in all the 3-
hour prediction experiments, i.e., SPMAE, RMSE, and MAE. Meanwhile, the performance of the proposed model is very close to the
best performance of the baseline models (e.g., MTGNN) in the 6-hour prediction experiments. However, the model structures are more
complex, which means a much longer training time is needed (see Table 3). In brief, compared with the baseline models, the proposed
model tended to have less model parameters, a shorter training time, and a higher prediction accuracy. In general, the computational
cost (e.g., training time, computer power) of training a deep learning model can be estimated by multiplying the number of model
parameters by the number of data points (Shen et al., 2022). This implies that given the same amount of input data, the model effi­
ciency is inversely proportional to the number of model parameters. Compared with the baseline deep learning models (i.e., MTGNN,
GWNet, STG-NCDE), the proposed model has a more concise structure (see section 4.5). Specifically, MTGNN designed an additional
module (namely the mix-hop propagation layer and the dilated inception layer) to extract the characteristics; GWNet additionally
constructed an adaptive dependency matrix through node embedding for prediction; and STG-NCDE extra added a module, i.e., neural
controlled differential equations, for time series prediction. These additional modules in the baseline approaches need more param­
eters and training time for fitting.
In addition, as the grid (or spatial analysis unit) size increases, RMSE and MAE increase because the average number of charging
demands in each grid grows (see Fig. 3).
Furthermore, it can be found that those convolutional deep learning prediction methods that incorporate the geographic graph tend
to have relatively worse performance in the EV charging demand prediction. For example, the prediction accuracies of TGCN, STG-
NCDE and STGCN are even lower than that of GRU. This suggests that using geographic spatial correlations may introduce noise
into GCN layers, which could lower prediction accuracy. Also, this is consistent with the hypothesis that the EV charging demand in
regions is not highly associated with the geographical location of charging stations. In other words, geographic adjacent regions may
have different charging demand patterns. In addition, it can be observed that GraphWave (GW) Net and MTGNN performed better than
GRU, TGCN, STG-NCDE, and STGCN. This indicates that the adaptive dependency matrix in GraphWave Net and the mix-hop prop­
agation layer in MTGNN can learn spatial correlations from the observations and further improves prediction accuracy.
Besides, we further evaluate the performance of the proposed model from a spatial perspective, using RMSE as an indicator. As
shown in Fig. 5, the model performance remains almost the same across grid scales: the regions with a high prediction error are mostly
located in the central areas. This might be because the influential factors to EV charging demand in these regions tended to be more
complicated. We might need to additionally consider other factors (or incorporate these factors into the prediction models) to further
improve model accuracy in these regions.

5.3. Ablation study

In this section, we conduct ablation studies on the following variants to quantify the contribution of each module to the overall
performance. For a fair comparison, all the variants are trained with the same configuration as detailed in Section 5.1.

• No specific region: The region-specific prediction module in Section 4.4 is removed.

• No demand graph: The graph convolutional layers that utilize the demand graph in the spatio-temporal modeling module is
removed.
• No geographic graph: The graph convolutional layers that utilize geographic graph in the spatio-temporal modeling module is
removed.

Table 3
The method parameters and epoch time in the 3-hour prediction.
Methods Number of Model Parameters Epoch Time
(Million) (s)

GRU 0.0129 0.4453

TGCN 0.0503 1.1210
STG-NCDE 2.5591 70.5172
STGCN 0.0993 0.6302
GWNet 0.3002 5.0665
MTGNN 0.6025 9.4932
Ours 0.2082 3.7312

14
S. Wang et al. Transportation Research Part C 153 (2023) 104205

Fig. 5. Heat maps of RMSE with different grid scales for 3 h/6h-ahead prediction.

Fig. 6 shows the prediction performance of ablation methods for both the 3 h- and 6 h-ahead predictions. Generally, compared to
the three variants, the proposed method in the baseline achieves the best performance, indicating that each module could help to
improve model performance. Among the three variants, the model without the demand graph has the worst performance, especially in
those scenarios with a larger spatial analysis unit (or grid) used. This suggests that the demand graph is an essential element in the
model and could heavily influence model performance. It also indicates regions that are correlated even when they are spatially distant
from each other. Besides, removing the geographic graph could lower the prediction accuracy as well. This suggests that geographic
correlation is also an important factor and should be included in the EV charging demand prediction model. Meanwhile, this indicates
our model assumption is realistic: the closer the geographic distance between the two regions, the stronger the EV charging demand
correlation that can be expected. Based on the prediction performance of “no demand graph” and “no geographic graph” variants, we
can find that it would be better to jointly consider the spatial correlation of regions characterized by the geographic distance and the
demand correlation of regions that are spatially distant from each other. Furthermore, the region-specific prediction module con­
tributes little to the improvement of prediction accuracy. However, the improvement for the 4 km × 4 km grid scale is not as significant
as that for the 1 km × 1 km grid scale by comparing the SMAPE values of “No specific region” and “Ours” at different grid scales,
indicating that the number of regions (or the region size) could be an influential factor. In summary, the ablation studies indicate the
efficiency of each designed module. Especially, the heterogeneous graph ensemble that linearly combined the geographic graph and
demand graph is an advantage for characterizing the spatial correlation compared with other baseline models.

15
S. Wang et al. Transportation Research Part C 153 (2023) 104205

Fig. 5. (continued).

5.4. Hyperparameter sensitivity test

In the heterogeneous graph ensemble part (see Section 4.3.3), the parameter α for representing the relationship between
geographic and demand graphs (see Eq.(12)) is tested (see Fig. 7). Since the parameter α is an adaptive weight, it is updated through
backpropagation along with other parameters during the training phase. Therefore, it can be seen from Fig. 7 that as the training epoch
increased, α decreased. The final values of α for the 1x1 km, 2x2 km, and 4x4 km scales are 0.33, 0.34, and 0.30, respectively.
We further investigate hyperparameters that may have a great influence on prediction accuracy, i.e., the number of clusters C and
the hidden dimension K of GRU cells. We empirically set C ∈ {4, 16, 64} and K ∈ \ { 32, 128, 256 \ } , respectively. Fig. 8 and Fig. 9
show model performance with different numbers of cluster C and different hidden dims K for 3 h-ahead prediction, respectively. To
summarize, the default setting (C = 16 and K = 128) of hyperparameters achieves the best prediction accuracy among all the settings.
These two hyperparameters exhibit approximately the same trend for the three-evaluation metrics and different grid scales: as the
value of the hyperparameter increases, the prediction accuracy first improves and then drops. The reason may be that both parameters
are positively correlated with the complexity of the model: a model with an oversimplified structure is not able to learn complex spatio-
temporal features. However, an over complicated model may lead to overfitting. Besides, we can observe that at the same grid scale,
changing the hidden dim K tends to more heavily influence prediction accuracy than changing the number of cluster C, indicating that
the model performance is more sensitive to K.

16
S. Wang et al. Transportation Research Part C 153 (2023) 104205

Fig. 6. Contribution of key modules to Model Performance.

Fig. 7. The values of parameter α under each epoch.

5.5. Model performance with a longer-term prediction

To further investigate the generalization of the proposed model, we set up another two groups of experiments in which the model
was used to predict charging demand over a longer period (i.e., 12 and 24 h). It can be seen from Table 4 that the proposed model still
has a satisfactory performance in these long-term prediction scenarios. Almost all metrics of our model are better than those of most the
baseline models, and are very close to the state-of-the-art models (e.g., MTGNN). However, MTGNN has much more parameters and
thus a longer training time than our model (see Table 3). Therefore, our model is not only suitable for short-term (e.g., 3 and 6 h)
predictions, but also has a relatively good performance in those longer-term (e.g., 12 and 24 h) predictions.

17
S. Wang et al. Transportation Research Part C 153 (2023) 104205

Fig. 8. Model performance with different numbers of Cluster C for the 3 h-ahead prediction.

Fig. 9. Model performance with different hidden dim K for the 3 h-ahead prediction.

6. Evaluating model performance with a public dataset

6.1. Data description

We additionally used one public charging session dataset to further evaluate performance of the proposed model with more time
steps. The public dataset contains one-year charging sessions from 15 charging stations in the City of Boulder, USA in 2020 (Boulder
Open Data., 2021). We firstly converted the raw data into time series charging demand data (see Section 3), and then removed those
charging stations with much less charging demand (specifically, those charging station only severing two EVs per month, on average).
Ultimately, we obtained 8,784 (=366 days × 24 h) time steps for each charging station. Since the number of charging stations of
Boulder is much smaller than that in Beijing (which is 1,128), it is difficult to generate grids for the case of Boulder. In response, we
treated a charging station as a grid. In addition, the charging stations in the case of Beijing is grouped into 16 categories according to
the POIs around the stations. While Boulder only had 15 stations in total, making it infeasible to use a region-specific prediction
module. Therefore, the region-specific prediction module was not included in the model. The spatial and temporal distributions of
charging demand for Boulder are shown in Fig. 10 and Fig. 11, respectively. It can be found that the charging stations had quite
different charging demands, ranging from 0.84 to 19.03 EVs served per day. From the temporal perspective, the peak period is from 11
am to 5 pm, and there is no significant difference in the EV charging demand between the seven days in a week.

6.2. Model performance

Table 5 shows the model performance with one-year charging session dataset in Boulder. Here, two discrete prediction horizons of
3 and 6 h are employed. In summary, the proposed model exhibits superior performance compared to other conventional models in the
Boulder scenario, as evident from the lower RMSE and MAE in the 3-hour prediction experiments, and lower RMSE in the 6-hour
prediction experiments.
It’s worth noting that the performance of the proposed model with the Boulder dataset may not be as good as that with the Beijing
dataset. There are two possible reasons. Firstly, the distance between charging stations in Boulder is relatively far, and the model could

18
S. Wang et al. Transportation Research Part C 153 (2023) 104205

Table 4
Model performance with a longer-term prediction.
(a) Model performance for the prediction length of 12 h

Methods Metrics

1 km £ 1 km 2 km £ 2 km 4 km £ 4 km

SMAPE(%) RMSE MAE (std) SMAPE(%) RMSE MAE SMAPE(%) RMSE MAE (std)
(std) (std) (std) (std) (std) (std) (std)

HA1 78.43 1.82 1.06 74.76 2.34 1.19 74.89 4.13 1.67
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
HA2 67.78 1.54 0.81 68.31 1.90 0.84 69.43 3.51 1.20
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
SVR 44.01 1.12 0.56 45.01 1.35 0.79 47.90 2.34 0.89
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
RF 45.89 1.36 0.76 47.61 1.52 0.81 50.01 2.54 1.32
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
GRU 43.43 1.20 0.60 43.40 1.47 0.71 45.28 2.16 0.93
(0.023) (0.014) (0.051) (0.007) (0.020) (0.015) (0.034) (0.028) (0.019)
TGCN 43.66 1.16 0.62 46.79 1.45 0.75 47.10 2.06 0.94
(0.104) (0.211) (0.020) (0.013) (0.027) (0.041) (0.035) (0.061) (0.035)
STG- 49.77 1.06 0.62 43.97 1.24 0.69 54.99 1.91 0.91
NCDE (0.132) (0.033) (0.040) (0.026) (0.057) (0.022) (0.130) (0.027) (0.032)
STGCN 44.72 1.23 0.65 48.48 1.56 0.81 53.90 2.25 1.11
(0.121) (0.052) (0.024) (0.036) (0.037) (0.042) (0.027) (0.031) (0.041)
GWNet 41.08 1.21 0.60 43.82 1.35 0.65 43.61 1.89 0.80
(0.045) (0.025) (0.150) (0.057) (0.051) (0.027) (0.043) (0.062) (0.063)
MTGNN 41.83 1.03 0.53 41.44 1.26 0.61 41.65 1.69 0.74
(0.023) (0.042) (0.029) (0.130) (0.032) (0.041) (0.025) (0.048) (0.0623)
Ours 41.52 1.06 0.55 40.16 1.28 0.64 40.67 1.83 0.79
(0.0140) (0.023) (0.0152) (0.032) (0.021) (0.015) (0.017) (0.026) (0.034)

(b) Model performance for the prediction length of 24 h

Methods Metrics

1km£1km 2km£2km 4km£4km

SMAPE(%) RMSE MAE SMAPE(%) RMSE MAE SMAPE(%) RMSE MAE

(std) (std) (std) (std) (std) (std) (std) (std) (std)

HA1 76.89 1.73 0.96 72.15 2.32 1.13 68.45 3.98 1.76
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
HA2 71.32 1.59 0.87 69.15 1.98 0.91 69.48 3.76 1.35
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
SVR 55.52 1.24 0.59 58.69 1.57 0.89 52.01 2.71 0.91
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
RF 51.34 1.32 0.68 50.83 1.69 0.93 51.98 2.69 1.32
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
GRU 49.14 1.28 0.68 48.28 1.57 0.79 48.59 2.34 1.04
(0.005) (0.027) (0.012) (0.006) (0.038) (0.024) (0.019) (0.012) (0.023)
TGCN 52.29 1.24 0.67 51.81 1.49 0.77 56.50 2.15 0.96
(0.024) (0.031) (0.015) (0.026) (0.041) (0.013) (0.052) (0.061) (0.030)
STG- 54.56 1.14 0.66 50.54 1.35 0.74 51.59 2.01 0.95
NCDE (0.052) (0.121) (0.039) (0.070) (0.046) (0.032) (0.019) (0.218) (0.000)
STGCN 49.27 1.30 0.71 49.21 1.61 0.87 52.55 2.33 1.14
(0.041) (0.069) (0.104) (0.0703) (0.041) (0.039) (0.124) (0.033) (0.052)
GWNet 48.19 1.23 0.65 46.92 1.55 0.76 50.67 2.11 0.91
(0.026) (0.031) (0.069) (0.033) (0.028) (0.130) (0.293) (0.146) (0.079)
MTGNN 48.73 1.17 0.62 46.11 1.39 0.69 46.15 1.93 0.84
(0.136) (0.061) (0.079) (0.240) (0.061) (0.130) (0.038) (0.256) (0.091)
Ours 47.67 1.16 0.63 47.22 1.38 0.71 46.97 2.01 0.89
(0.023) (0.054) (0.018) (0.040) (0.019) (0.021) (0.011) (0.138) (0.049)

Note: “std” in the table stands for standard deviation, which is used to quantify the variances of the ten-run results).

not adopt the grid-level charging demand approach which was used in the Beijing case. Instead, it treated each charging station as a
grid. This simplification might have influenced the model performance. Secondly, the Boulder dataset contained only 15 charging
stations, making it difficult to use POIs for classification (see Section 4.4). As a result, the region-specific prediction module, which is
found useful in the Beijing case (see Section 5.3), cannot be considered in the model.

19
S. Wang et al. Transportation Research Part C 153 (2023) 104205

Fig. 10. The spatial distribution of charging stations and their charging demand in Boulder.

Fig. 11. Temporal charging patterns in Boulder.

7. Conclusion

In this paper, we propose a heterogeneous spatio-temporal graph convolutional network for predicting electric vehicle charging
demand. Specifically, we first construct heterogeneous graphs to represent the spatial correlations between different charging regions,
i.e., a geographic graph and a demand graph. Subsequently, we employ graph convolutional layers and gated recurrent units to extract
spatial and temporal features and ensemble the hidden representations of heterogeneous graphs by linear combination. In addition, as
different regions have different charging demand patterns, we improve prediction accuracy by designing a region-specific prediction
module, which first clusters regions into different groups and then trains region-specific layers for the prediction. To evaluate the

20
S. Wang et al. Transportation Research Part C 153 (2023) 104205

Table 5
Comparing the prediction performance of different prediction models with the public one-year charging session dataset.
Methods Prediction for the length of 3 h Prediction for the length of 6 h

SMAPE(%) (std) RMSE (std) MAE (std) SMAPE(%) (std) RMSE (std) MAE (std)

HA1 98.72 0.24 0.06 100.37 0.25 0.06

(0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
HA2 187.90 0.30 0.08 187.45 0.30 0.08
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
SVR 74.29 0.21 0.05 100.76 0.23 0.06
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
RF 69.21 0.21 0.05 99.17 0.22 0.05
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
GRU 61.68 0.22 0.05 74.32 0.23 0.06
(0.012) (0.008) (0.004) (0.049) (0.065) (0.031)
TGCN 0.61 0.21 0.05 78.65 0.23 0.06
(0.012) (0.008) (0.004) (0.032) (0.023) (0.003)
STG-NCDE 66.82 0.21 0.11 72.17 0.25 0.16
(0.035) (0.011) (0.007) (0.052) (0.036) (0.081)
STGCN 120.85 0.30 0.09 126.00 0.34 0.11
(0.028) (0.017) (0.009) (0.043) (0.016) (0.030)
GWNet 62.32 0.21 0.05 74.72 0.23 0.06
(0.019) (0.015) (0.008) (0.019) (0.026) (0.011)
MTGNN 59.65 0.21 0.05 75.19 0.23 0.05
(0.018) (0.023) (0.006) (0.034) (0.018) (0.009)
Ours 63.29 0.21 0.05 76.45 0.23 0.06
(0.013) (0.010) (0.005) (0.013) (0.015) (0.004)

Note: “std” in the table stands for standard deviation, which is used to quantify the variances of the ten-run results).

prediction accuracy of the proposed method, we use a unique charging dataset containing real-world GPS trajectories of over 76,000
EVs in Beijing, China, and further divide them into regions by different grid scales.
We found that compared to the traditional temporal models and recently popular spatio-temporal modeling methods (HA1, HA2,
SVR, RF, GRU, T-GCN, STGCN, GWNet), the proposed method can achieve remarkable prediction accuracy with different prediction
lengths (3 h and 6 h) and on different grid scales (i.e., 1 km × 1 km, 2 km × 2 km, 4 km × 4 km). Furthermore, the charging demand
regions with a high prediction error are mostly located in the central areas. This might be because the influential factors to EV charging
demand in these regions tended to be more complicated. The ablation studies on the model variants show that the demand graph and
the geographic graph, as well as the region-specific prediction module, could significantly improve model performance. Besides, the
number of regions could also influence the model performance.
In our future work, we will try to integrate external factors (such as real-time traffic status and weather conditions) into the model,
which may influence the charging demand pattern. Incorporating these factors may not only help to improve model performance but
also explain the complex charging demand pattern in the central areas. In addition, we can continue to examine performance of the
proposed short-term charging demand prediction model with a longer time-series dataset. Currently, we use one-month dataset in our
study for model training and validation. A longer time-series dataset would allow us to further explore how more time steps may
influence model performance.

CRediT authorship contribution statement

Shengyou Wang: Methodology, Conceptualization, Software, Formal analysis, Visualization, Writing – original draft. Anthony
Chen: Methodology, Funding acquisition, Writing – review & editing. Pinxi Wang: Methodology, Validation, Writing – review &
editing. Chengxiang Zhuge: Methodology, Funding acquisition, Writing – review & editing.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to
influence the work reported in this paper.

Data availability

The data that has been used is confidential.

Acknowledgement

The work described in this paper was jointly supported by the National Natural Science Foundation of China (52002345), the
Fundamental Research Funds for the Central Universities (2023JKF02ZK08), the research grants from the Research Institute for

21
S. Wang et al. Transportation Research Part C 153 (2023) 104205

Sustainable Urban Development (1-BBWF and 1-BBWR), the Smart Cities Research Institute (CDAR and CDA9) and the funding for
Project of Strategic Importance provided by The Hong Kong Polytechnic University (1-ZE0A).

Appendix A. Supplementary material

Supplementary data to this article can be found online at https://doi.org/10.1016/j.trc.2023.104205.

References

Ahuja, S., Shelke, N.A., Singh, P.K., 2022. A deep learning framework using CNN and stacked Bi-GRU for COVID-19 predictions in India. SIViP 16 (3), 579–586.
https://doi.org/10.1007/s11760-021-01988-1.
Amara-Ouali, Y., Goude, Y., Hamrouche, B., & Bishara, M., (2022). A benchmark of electric vehicle load and occupancy models for day-ahead forecasting on open
charging session data, in: Proceedings of the Thirteenth ACM International Conference on Future Energy Systems, 193-207. https://doi.org/10.1145/
3538637.3538850.
Bian, H., Zhou, C., Guo, Z., Wang, X., He, Y., Peng, S., 2022. Planning of electric vehicle fast-charging station based on POI interest point division, functional area, and
multiple temporal and spatial characteristics. Energy Rep. 8, 831–840. https://doi.org/10.1016/j.egyr. (2022).10.161.
Bogaerts, T., Masegosa, A.D., Angarita-Zapata, J.S., Onieva, E., Hellinckx, P., 2020. A graph CNN-LSTM neural network for short and long-term traffic forecasting
based on trajectory data. Trans. Res. Part C: Emerging Technologies 112, 62–77. https://doi.org/10.1016/j.trc. (2020).01.010.
Boulakhbar, M., Farag, M., Benabdelaziz, K., Kousksou, T., Zazi, M., 2022. A deep learning approach for prediction of electrical vehicle charging stations power
demand in regulated electricity markets: the case of Morocco. Cleaner Energy Systems 3, 100039. https://doi.org/10.1016/j.cles.(2022).100039.
Boulder Open Data. 2021. “Boulder Police Department Arrests Dataset”. City of Boulder. https://open-data.bouldercolorado.gov/datasets/
39288b03f8d54b39848a2df9f1c5fca2_0/explore. Accessed April 22, 2023.
Bouwmans, T., Jayed, S., Sultana, M., Jung, S.K., 2019. Deep neural network concepts for background subtraction: a systematic review and comparative evaluation.
Neural Netw. 117, 8–66.
Butt, T.A., Iqbal, R., Salah, K., Aloqaily, M., Jararweh, Y., 2019. Privacy management in social internet of vehicles: review, challenges and blockchain based solutions.
IEEE Access 7, 79694–79713.
Buzna, L., De Falco, P., Ferruzzi, G., Khormali, S., Proto, D., Refa, N., Straka, M., van der Poel, G., 2021. An ensemble methodology for hierarchical probabilistic
electric vehicle load forecasting at regular charging stations. Appl. Energy 283, 116337. https://doi.org/10.1016/j.apenergy.(2020).116337.
Cai, H., Zheng, V.W., Chang, K.C.-C., 2018. A comprehensive survey of graph embedding: problems, techniques, and applications. IEEE Transactions on Knowledge
and Data Engineering 30, 1616–1637. https://doi.org/10.1109/tkde.2018.2807452.
Chen, K., Yu, J., 2014. Short-term wind speed prediction using an unscented kalman filter based state-space support vector regression approach. Appl. Energy 113,
690–705. https://doi.org/10.1016/j.apenergy.2013.08.025.
Chen, Z., Zhao, B., Wang, Y., Duan, Z., Zhao, X., 2020. Multitask learning and GCN-Based taxi demand prediction for a traffic road network. Sensors 20 (13), 3776.
https://doi.org/10.3390/s20133776.
Cheng, A., Jiang, X., Li, Y., Zhang, C., Zhu, H., 2017. Multiple sources and multiple measures based traffic flow prediction using the chaos theory and support vector
regression method. Physica A 466, 422–434. https://doi.org/10.1016/j.physa.2016.09.041.
Choi, J., Choi, H., Hwang, J., Park, N., 2021. Graph neural controlled differential equations for traffic forecasting. Association for the Advancement of Artificial
Intelligence (AAAI) 1–13. https://doi.org/10.48550/arXiv.2112.03558.
Dark, S.J., Bram, D., 2007. The modifiable areal unit problem (MAUP) in physical geography. Progress in Phys. Geography: Earth and Environ. 31, 471–479. https://
doi.org/10.1177/0309133307083294.
Dong, G., Ma, J., Wei, R., Haycox, J., 2019. Electric vehicle charging point placement optimisation by exploiting spatial statistics and maximal coverage location
models. Transp. Res. Part D: Transp. Environ. 67, 77–88.
Engel, H., Hensley, R., Knupfer, S., & Sahdev, S. (2018). Charging ahead: Electric-vehicle infrastructure demand. McKinsey Center for Future Mobility, 8. https://
www.mckinsey.com/industries/automotive-and-assembly/our-insights/charging-ahead-electric-vehicle-infrastructure-demand.
Ermagun, A., Levinson, D., 2019. Spatiotemporal short-term traffic forecasting using the network weight matrix and systematic detrending. Trans. Res. Part C:
Emerging Technologies 104, 38–52. https://doi.org/10.1016/j.trc.2019.04.014.
Feng, S., Ke, J., Yang, H., Ye, J., 2022. A multi-task matrix factorized graph neural network for co-prediction of zone-based and OD-based ride-hailing demand. IEEE
Trans. Intelligent 23, 5704–5716. https://doi.org/10.1109/TITS.(2021).3056415.
Flammini, M.G., Prettico, G., Julea, A., Fulli, G., Mazza, A., Chicco, G., 2019. Statistical characterisation of the real transaction data gathered from electric vehicle
charging stations. Electr. Pow. Syst. Res. 166, 136–150.
Fukuda, S., Uchida, H., Fujii, H., Yamada, T., 2020. Short-term prediction of traffic flow under incident conditions using graph convolutional recurrent neural network
and traffic simulation. IET Intel. Transport Syst. 14 (8), 936–946. https://doi.org/10.1049/iet-its.2019.0778.
Guo, S., Lin, Y., Feng, N., Song, C., & Wan, H. (2019). Attention based spatial-temporal graph convolutional networks for traffic flow forecasting. National Conference
on Artificial Intelligence. Association for the Advancement of Artificial Intelligence (AAAI), 922-929. https://doi.org/10.1609/aaai.v33i01.3301922.
Guo, G., Zhang, T., 2020. A residual spatio-temporal architecture for travel demand forecasting. Trans. Res. Part C: Emerging Technologies 115, 102639. https://doi.
org/10.1016/j.trc.2020.102639.
Herrmann, M., Webb, G.I., 2023. Amercing: an intuitive and effective constraint for dynamic time warping. Pattern Recogn., 109333 https://doi.org/10.1016/j.
patcog.2023.109333.
Hochreiter, S., Schmidhuber, J., 1996. LSTM can solve hard long time lag problems. Adv. Neural Inf. Proces. Syst. 473–479 https://doi.org/10.1162/
neco.1997.9.8.1735.
Hu, L., Dong, J., Lin, Z., Yang, J., 2018. Analyzing battery electric vehicle feasibility from taxi travel patterns: the case study of New York city. Trans. Res. Part C:
Emerging Technologies 87, 91–104.
Hu, L., Dong, J., Lin, Z., 2019. Modeling charging behavior of battery electric vehicle drivers: a cumulative prospect theory based approach. Trans. Res. Part C:
Emerging Technologies 102, 474–489. https://doi.org/10.1016/j.trc.2019.03.027.
Huang, Z., Zhang, W., Wang, D., Yin, Y., 2022. A GAN framework-based dynamic multi-graph convolutional network for origin-destination-based ride-hailing demand
prediction. Inf. Sci. 601, 129–146. https://doi.org/10.1016/j.ins. (2022).04.024.
Hüttel, F.B., Peled, I., Rodrigues, F., & Pereira, F. C., (2021). Deep spatio-temporal forecasting of electrical vehicle charging demand. https://doi.org/10.48550/
arXiv.2106.10940.
Jiang, W., Luo, J., 2022. Graph neural network for traffic forecasting: a survey. Expert Syst. Appl. 207 (30), 117921 https://doi.org/10.1016/j.eswa. (2022).117921.
Jin, G., Cui, Y., Zeng, L., Tang, H., Feng, Y., Huang, J., 2020. Urban ride-hailing demand prediction with multiple spatio-temporal information fusion network. Trans.
Res. Part C: Emerging Technologies 117, 102665. https://doi.org/10.1016/j.trc.2020.102665.
Kalogirou, S.A., 2003. Artificial intelligence for the modeling and control of combustion processes: a review. Prog. Energy Combust. Sci. 29 (6), 515–566.
Kamarianakis, Y., Prastacos, P., 2003. Forecasting traffic flow conditions in an urban network: comparison of multivariate and univariate approaches. Trans. Res.
Record: J. Trans. Res. Board 1857 (1), 74–84. https://doi.org/10.3141/1857-09.

22
S. Wang et al. Transportation Research Part C 153 (2023) 104205

Kingma, D. P., & Ba, J. (2015). Adam: A method for stochastic optimization. 3rd International Conference on Learning Representations (ICLR), 1-15. https://doi.org/
10.48550/arXiv.1412.6980.
Kipf, T.N., Welling, M., 2016. Semi-supervised classification with graph convolutional networks. Computer Sci. 3, 1–13. https://doi.org/10.48550/arXiv.1609.02907.
Lee, K., Rhee, W., 2022. DDP-GCN: multi-graph convolutional network for spatiotemporal traffic forecasting. Trans. Res. Part C: Emerging Technologies 134, 103466.
https://doi.org/10.1016/j.trc.2021.103466.
Lei, T., Guo, S., Qian, X., Gong, L., 2022. Understanding charging dynamics of fully-electrified taxi services using large-scale trajectory data. Trans. Res. Part C:
Emerging Technologies 143, 103822. https://doi.org/10.1016/j.trc. (2022).103822.
Li, C., Bai, L., Liu, W., Yao, L., Waller, S.T., 2022. Graph neural network for robust public transit demand prediction. IEEE Trans. Intell. Transp. Syst. 23, 4086–4098.
https://doi.org/10.1109/TITS. (2020).3041234.
Li, Y., Yu, R., Shahabi, C., & Liu, Y. (2018). Diffusion convolutional recurrent neural network: Data-driven traffic forecasting. 6th International Conference on
Learning Representations (ICLR), 1-16. https://doi.org/10.48550/arXiv.1707.01926.
Li, S., Zhao, Y., Varma, R., & Salpekar, O. (2020). PyTorch distributed: Experiences on accelerating data parallel training. Proceedings of the VLDB Endowment, 13:
3005-3018. https://doi.org/10/gpxj23.
Li, X., Xu, Y., Zhang, X., Shi, W., Yue, Y., Li, Q., 2023. Improving short-term bike sharing demand forecast through an irregular convolutional neural network. Trans.
Res. Part C: Emerging Technologies 147, 103984. https://doi.org/10.1016/j.trc.2022.103984.
Liang, J., Tang, J., Gao, F., Wang, Z., Huang, H., 2023. On region-level travel demand forecasting using multi-task adaptive graph attention network. Inf. Sci. 622,
161–177. https://doi.org/10.1016/j.ins. (2022).11.138.
Lin, Y., Zhang, K., Shen, Z.-J.-M., Ye, B., Miao, L., 2019. Multistage large-scale charging station planning for electric buses considering transportation network and
power grid. Trans. Res. Part C: Emerging Technologies 107, 423–443. https://doi.org/10.1016/j.trc.2019.08.009.
Liu, L., Chen, R.C., Zhao, Q., Zhu, S., 2019. Applying a multistage of input feature combination to random forest for improving MRT passenger flow prediction.
J. Ambient Intell. Hum. Comput. 10, 4515–4532. https://doi.org/10.1007/s12652-018-1135-2.
Luo, Z., Song, Y., Hu, Z., Xu, Z., Yang, X., Zhan, K., 2011. Forecasting charging load of plug-in electric vehicles in China. IEEE Power and Energy Society General
Meeting 1–8. https://doi.org/10.1109/PES.2011.6039317.
Ma, T.Y., Faye, S., 2022. Multistep electric vehicle charging station occupancy prediction using hybrid LSTM neural networks. Energy 244, 1–13.
Ma, X., He, X., 2021. Period-to-period toll adjustment schemes for mixed traffic with time-varying electric vehicle penetration. Trans. Res. Part C: Emerging
Technologies 129, 103237. https://doi.org/10.1016/j.trc. (2021).103237.
Mikolov, T., Chen, K., Corrado, G., & Dean, J. (2013). Efficient estimation of word representations in vector space. 1st International Conference on Learning
Representations (ICLR), 1-12. 10.48550/arXiv.1301.3781.
Morin, L., Gilormini, P., Derrien, K., 2020. Generalized Euclidean distances for elasticity tensors. J. Elast. 138, 221–232.
Murugan, M., Marisamynathan, S., 2022. Elucidating the Indian customers requirements for electric vehicle adoption: an integrated analytical hierarchy process -
Quality function deployment approach. Case Studies on Transport Policy 10, 1045–1057. https://doi.org/10.1016/j.cstp.2022.03.017.
Perozzi, B., Al-Rfou, R., & Skiena, S. (2014). Deep walk: Online learning of social representations. 20th ACM SIGKDD International Conference on Knowledge
Discovery and Data Mining, USA, 701-710. https://doi.org/10/gfkpqt.
Rodrigues, F., 2022. On the importance of stationarity, strong baselines and benchmarks in transport prediction problems, 1-6. 10.48550/ARXIV.2203.02954.
Salah, F., Ilg, J.P., Flath, C.M., Basse, H., Dinther, C., 2015. Impact of electric vehicles on distribution substations: a swiss case study. Appl. Energy 137, 88–96.
Seng, D., Lv, F., Liang, Z., Shi, X., Fang, Q., 2021. Forecasting traffic flows in irregular regions with multi-graph convolutional network and gated recurrent unit.
Frontiers of Information Technol. & Electronic Eng. 22, 1179–1193. https://doi.org/10.1631/FITEE.2000243.
Shen, X., Gao, J., Wu, W., Li, M., Zhou, C., Zhuang, W., 2022. Holistic network virtualization and pervasive network intelligence for 6G. IEEE Commun. Surv. Tutorials
24, 1–30. https://doi.org/10.1109/COMST.2021.3135829.
Shen, G., Kurths, J., Yuan, Y., 2020. Sequence-to-sequence prediction of spatiotemporal systems. Chaos 30, 023102. https://doi.org/10.1063/1.5133405.
Soldan, F., Bionda, E., Mauri, G., Celaschi, S., 2021. Short-term forecast of EV charging stations occupancy probability using big data streaming analysis. Computer
Sci. 1–7. https://doi.org/10.48550/arXiv.2104.12503.
Sun, Z., Hu, Y., Li, W., Feng, S., Pei, L., 2022. Prediction model for short-term traffic flow based on a K-means-gated recurrent unit combination. IET Intel. Transport
Syst. 16, 675–690. https://doi.org/10.1049/itr2.12165.
Sun, M., Shao, C., Zhuge, C., Wang, P., Yang, X., Wang, S., 2021a. Exploring the potential of rental electric vehicles for vehicle-to-grid: a data-driven approach. Resour.
Conserv. Recycl. 175, 1–19. https://doi.org/10.1016/j.resconrec.(2021).105841.
Sun, M., Shao, C., Zhuge, C., Wang, P., Yang, X., Wang, S., 2021b. Uncovering travel and charging patterns of private electric vehicles with trajectory data: evidence
and policy implications. Transportation. https://doi.org/10.1007/s11116-021-10216-1.
Tobler, W.R., 1970. A computer movie simulating urban growth in the detroit region. Econ. Geogr. 46, 234–240. https://doi.org/10.2307/143141.
Tu, W., Li, Q., Fang, Z., Shaw, S., Zhou, B., Chang, X., 2016. Optimizing the locations of electric taxi charging stations: a spatial-temporal demand coverage approach.
transportation research Board 95th Annual Meeting. United States 65, 172–189.
Valsera-Naranjo, E., Martínez-Vicente, D., Sumper, A., Villafàfila-Robles, R., & Sudrià-Andreu, A. (2011). Deterministic and probabilistic assessment of the impact of
the electrical vehicles on the power grid. 2011 IEEE Power and Energy Society General Meeting, 1-8. https://doi.org/10/c326r4.
Wang, J., Zhang, Y., Wei, Y., Hu, Y., Piao, X., Yin, B., 2021. Metro passenger flow prediction via dynamic hypergraph convolution networks. IEEE Trans. Intell. Transp.
Syst. 22, 7891–7903. https://doi.org/10.1109/TITS.(2021).3072743.
Wu, Z., Pan, S., Long, G., Jiang, J., & Zhang, C. (2019). Graph wavenet for deep spatial-temporal graph modeling. 28th International Joint Conference on Artificial
Intelligence, 1907-1913. 10.48550/arXiv.1906.00121.
Wu, Z., Pan, S., Long, G., Jiang, J., Chang, X., Zhang, C., 2020. Connecting the dots: multivariate time series forecasting with graph neural networks. 10.48550/
arXiv.2005.11650.
Wu, Z., Pan, S., Chen, F., Long, G., Zhang, C., & Yu, P. S. (2021). A comprehensive survey on graph neural networks. IEEE Transactions on Neural Networks and
Learning Systems, 32: 4-24. https://doi.org/10/ggrj8p.
Yang, D., Sarma, N.J.S., Hyland, M.F., Jayakrishnan, R., 2021. Dynamic modeling and real-time management of a system of EV fast-charging stations. Trans. Res. Part
C: Emerging Technologies 128, 103186. https://doi.org/10.1016/j.trc.(2021).103186.
Yang, X., Zhuge, C., Shao, C., Huang, Y., Hayse, C.G., Tang, J., Sun, M., Wang, P., Wang, S., 2022. Characterizing mobility patterns of private electric vehicle users
with trajectory data. Appl. Energy 321, 1–17.
Yao, Z., Gendreau, M., Li, M., Ran, L., Wang, Z., 2022. Service operations of electric vehicle carsharing systems from the perspectives of supply and demand: a
literature review. Trans. Res. Part C: Emerging Technologies 140, 103702. https://doi.org/10.1016/j.trc.(2022).103702.
Yao, J., Moawad, A., 2019. Vehicle energy consumption estimation using large scale simulations and machine learning methods. Trans. Res. Part C: Emerging
Technologies 101, 276–296. https://doi.org/10.1016/j.trc.2019.02.012.
Yi, Z., Liu, X.C., Wei, R., Chen, X., Dai, J., 2022. Electric vehicle charging demand forecasting using deep learning model. Journal of Intelligent Transportation Systems
26, 690–703. https://doi.org/10.1080/15472450.2021.1966627.
Yi, Z., Liu, X.C., Wei, R., Chen, X., Dai, J., 2021. Electric vehicle charging demand forecasting using deep learning model. J. Intell. Transp. Syst. 1–14. https://doi.org/
10.1080/15472450.(2021).1966627.
Yin, X., Wu, G., Wei, J., Shen, Y., Qi, H., Yin, B., 2021. Deep learning on traffic prediction: methods, analysis and future directions. IEEE Trans. Intell. Transp. Syst. 23,
1–17. https://doi.org/10.1109/TITS.(2021).3054840.
Yu, L., Du, B., Hu, X., Sun, L., Han, L., Lv, W., 2021. Deep spatio-temporal graph convolutional network for traffic accident prediction. Neurocomputing 423, 135–147.
https://doi.org/10.1016/j.neucom.(2020).09.043.
Yu, B., Yin, H., & Zhu, Z. (2018). Spatio-temporal graph convolutional networks: A deep learning framework for traffic forecasting. 27th International Joint
Conference on Artificial Intelligence, 3634-3640. https://doi.org/10/gftc76.

23
S. Wang et al. Transportation Research Part C 153 (2023) 104205

Yu, B., Lee, Y., Sohn, K., 2020. Forecasting road traffic speeds by considering area-wide spatio-temporal dependencies based on a graph convolutional neural network
(GCN). Trans. Res. Part C: Emerging Technologies 114, 189–204. https://doi.org/10.1016/j.trc.(2020).02.013.
Yuan, Y., Zhang, D., Miao, F., Chen, J., He, T., & Lin, S. (2019). p^2 charging: Proactive partial charging for electric taxi systems. 2019 IEEE 39th International
Conference on Distributed Computing Systems (ICDCS), 688-699. https://doi.org/10.1109/ICDCS.2019.00074.
Zhao, Y., Han, X., Xu, X., 2022. Traffic flow prediction model based on the combination of improved gated recurrent unit and graph convolutional network. Front.
Bioeng. Biotechnol. 10, 1–17. https://doi.org/10.3389/fbioe. (2022).804454.
Zhao, L., Song, Y., Zhang, C., Liu, Y., Wang, P., Lin, T., Deng, M., Li, H., 2019. T-GCN: a temporal graph convolutional network for traffic prediction. IEEE Trans. Intell.
Transp. Syst. 21, 1–11. https://doi.org/10.1109/TITS.2019.2935152.
Zhu, J., Yang, Z., Mourshed, M., Guo, Y., Zhou, Y., Chang, Y., Wei, Y., Feng, S., 2019. Electric vehicle charging load forecasting: a comparative study of deep learning
approaches. Energies 12, 2692. https://doi.org/10.3390/en12142692.

24