Charging Stations Selection Using a Graph Convolutional Network from Geographic Grid

Abstract

Electric vehicles (EVs) have attracted considerable attention because of their clean and high-energy efficiency. Reasonably planning a charging station network has become a vital issue for the popularization of EVs. Current research on optimizing charging station networks focuses on the role of stations in a local scope. However, spatial features between charging stations are not considered. This paper proposes a charging station selection method based on the graph convolutional network (GCN) and establishes a charging station selection method considering traffic information and investment cost. The method uses the GCN to extract charging stations. The charging demand of each candidate station is calculated through the traffic flow information to optimize the location of charging stations. Finally, the cost of the charging station network is evaluated. A case study on charging station selection shows that the method can solve the EV charging station location problem.

1. Introduction

Using EVs can reduce the burning of fossil fuels, thus effectively reducing greenhouse gas emissions and easing environmental pressures. However, the development of EVs is constrained by energy technologies and support, including battery storage services [1,2] and charging service support [3,4,5,6]. The main reason is that charging station network planning is not justified. Current charging facilities significantly limit the convenience of the charging station network. The user’s charging demand involves a complete charging station network rather than a single one. Therefore, practical design of charging station networks is necessary to facilitate EVs.

1.1. Motivations and Related Work

Energy consumption by the transportation industry has increased greenhouse gas emissions, posing a significant challenge to environmental protection. Unlike fuel vehicles, EVs can reduce carbon dioxide emissions [7]. However, due to the limited number of charging stations, the lack of charging facilities has become an essential factor limiting EVs development [8]. Therefore, proper charging station planning will help promote more sustainable smart transportation.

In recent years, researchers have shown increased interest in properly selecting charging stations and links in networks. Location method is the primary means of facility location planning to meet its spatial distribution demand. The location planning of charging station facilities is usually considered from the perspective of demand [9,10,11,12,13,14] and economy [15,16]. Analyzing the charging demand in the region is an essential step in planning charging stations. Some studies allocate charging stations based on regional charging demand. These studies can be further divided into node-based methods and flow-based methods. The planning and location model based on traffic capture assumes that the demand is given in origin–destination (O–D) flow aiming to cover as many traffic flows as possible. According to the real-time traffic flow, the influence of human activities on the selection of charging stations is dynamically presented [17]. A two-tier optimization framework was constructed to optimize the configuration of fast charging station facilities with various realities, such as traffic flow as a critical factor [18,19,20,21]. Wang proposed a charging strategy suitable for intercity travel [22]. Depending on the flow rate or the location of the flow capture facility, it is more appropriate to deal with optimizing charging stations between distant regions. Node- or point-based demand models can be helpful if demand is concentrated, such as in urban areas. The size of stations installed to meet the slow charging needs of customers in a highly populated area is determined based on a maximum coverage model [23]. Hu constructed an analytical model for the carrying capacity of the charging demand of the charging station [24]. The article [25] verified the planning results of the designed charging station by dividing the research area into multiple regular grids and calculating the power demand in the grid area. Wang presented a four-step method for electric vehicle charging facility deployment [26]. The optimal planning of charging stations was investigated by analyzing users’ travel patterns [27], calculating regional heat values [24], and analyzing the charging demand of commuter EVs [28].

Additionally, the layout of the charging station is planned from an economic point of view to minimize the investment and charging costs. For example, charging duration and cost [22] are essential factors that influence the location of the charging station planning. Considering the cost of charging station construction, Dong and Li established the location and scale model [29]. Considering the ease of planning and maintenance costs, it is reasonable to assume that most cities will soon plan enough charging stations for users. Chen considered the EV length and charging duration to assign the correct number of cars to the stations in the demand area [30]. In [31], a multiobjective optimization model was proposed for layout of charging stations to address the profitability needs of charging station owners and operators. The article [32] evaluated regional segmentation at both macro and micro levels and analyzed the potential locations of charging stations in the region based on land use. Yang [33] built a stochastic simulation modeling framework by considering the charging station price and waiting time, which can allocate the capacity of the charging station in real time. The location and capacity models were established by using the improved particle swarm optimization algorithm [29]. The key challenges of rational planning of charging stations in real cities are as follows:

• Describing users’ charging behaviors and habits.
• Aggregating multisource geospatial data into reliable characteristic information of charging stations.
• Providing the distribution of charging stations to users with on-demand needs.

We aim to plan reasonable charging stations by considering the charging demand of users, the temporal distribution of charging stations, and the road network. The key is to obtain the data of interest points and relevant data of existing charging stations and consider the rationality of the charging station service network from a spatial perspective. However, it is only sometimes obvious how to effectively integrate geographic resources to achieve rational deployment of charging stations because most charging station planning approaches mainly evaluate subjective metrics without using objective models to make judgments, including population density, traffic flow, construction cost, and latency cost. They only consider the regional supply–demand relationship but do not consider the geospatial correlation between nodes. The deep learning method can help reveal spatial correlations of geographical elements. In the age of big data, deep learning has also become a vital research direction [34], which can perform automatic feature learning on crowdsourced data to discover distributed feature representations of the data [35,36]. As a result, GCN-based classification studies have received much attention [20,37]. Ref. [38] proposed a GCN-based semisupervised classification method using social network data. However, this method mainly starts from the demand and does not consider the spatial correlation between regions. In summary, this paper will obtain the potential spatiotemporal features from space using the node classification method of GCN to plan an appropriate charging station location.

1.2. Our Approach

In this paper, we propose a GCN-based EV charging station planning approach that reliably aggregates geographic information from multiple sources and plans rational charging stations. The proposed method is designed to address the charging needs of EV users traveling to improve the convenience and accessibility of charging for drivers.

To better aggregate spatiotemporal feature information, this paper introduces the geospatial feature attribute to improve the rationality of the charging station network based on considering the charging demand and investment cost. It focuses on the connection between charging demand in a region and the spatiotemporal features of the geographic space. Specifically, the method uses the supervised classification of nodes to extract candidate sites for charging stations according to points of interest (POI), traffic information, and costs. The spatial information of each node is considered comprehensively so that reasonable charging stations can be planned in the real road network.

The main contributions of this paper are as follows:

• Propose a hybrid charging stations selection method. The GCN is combined with traditional judgment methods to comprehensively evaluate the optimal layout of the charging station network.
• Construct the graph structure tensor of the charging station network. We propose the graph structure module to obtain the potential spatiotemporal environment attributes to construct the graph structure tensor.
• Optimize potential charging stations based on preliminary classification results. Demand analysis compares the EV charging station’s data in the grid with the actual station to screen out reasonable charging stations further.

2. Methodology

This paper presents a hybrid method for selecting EV charging stations. As shown in Figure 1, EV charging station planning includes data collection, candidate site extraction, and candidate site optimization. The candidate sites of charging stations are extracted based on the GCN node supervised classification. Candidate sites are evaluated by using traffic flow and construction cost to optimize further the location planning results of charging stations [39]. The results of the data collection stage will provide a well-prepared dataset for building the GCN algorithm. The extract candidate charging station phase includes constructing tensor, GCN-based node classification, and choosing candidate sites. The last stage optimizes the charging station and evaluates the cost of the charging station network. As a result, the EV charging infrastructure will be optimized and planned to meet the increasing number of EVs in the future.

2.1. Study Area and Data

The research selects some areas of Yuhua district, Furong district, and Tianxin district of Changsha City, Hunan Province, as the research area (as shown in Figure 2). Changsha is one of the major cities in Central China and is densely populated, covering many residential areas, shopping malls, cultural and educational places, etc. In 2021, Changsha had a population of 10.0479 million and a jurisdiction area of 11,816 square kilometers. Areas with high population densities tend to be areas with high demand for charging stations. Therefore, some urban characteristics in this area may affect the layout of EV charging facilities. By collecting and preprocessing the corresponding traffic data, POI data, and land price data in the study area, we can provide a fully prepared dataset for the construction of a deep learning algorithm.

2.1.1. Road Network

The road network is freely available through the OpenStreetMap platform, which is easy to access, has more diverse road types, and is more current. After obtaining the original road network, we modified the data and removed the wrong road network data to ensure the correctness of the road network data. Four primary types of roads were extracted in this study: trunk, primary, secondary, and tertiary.

2.1.2. Traffic Data

Traffic data are the real-time traffic situation data obtained through the Amap platform. We obtain the corresponding traffic flow by deduplication and cleaning the original traffic situation data.

2.1.3. POI Data

This study extracts POI as the characteristic of nodes because of its high accuracy, wide coverage, and real-time solidity.POI refers to all geographical objects that can be abstracted as points in geographical scenes in general, such as schools, shopping malls, parks, and other geographical entities. The characteristics can expose various socioeconomic phenomena and provide specific geographic references when classifying nodes. The POI dataset comes from Baidu Map, and each POI record has the name, level classification, address, and geographic coordinates of the POI. This study will extract the POI related to EV charging demand from the original POI.

2.1.4. Land Price Data

The land price data used in this paper are based on the benchmark land price. Urban benchmark land price refers to the average price of each price segment at a certain time point in a city, which is divided into land price regions according to the principles of similar use, land necklace, and land price within a certain region of the city.

2.2. Extracting Candidate Charging Stations

There is a classification and extraction model for EV charging stations. In order to preliminarily extract appropriate EV charging stations, important features of representative geographic regions, including POI, road networks, and construction maps, are extracted in this paper [40].

2.2.1. Constructing the Model Tensor

As shown in Figure 3, this study seeks to construct a graph structure which describes various complex data features and the relationship between structural features. The study area is first divided into regular grids and abstracted into nodes, and the adjacency relationship between nodes represents the proximity relationship between grids. As a complex data structure, the graph contains both attributes of node and attributes of edge. The attributes of nodes contain POIs and roads; the attributes of edges are the connection relationships between nodes. It can represent many feature structures with rich information, so it can be used to describe various complex data features and the relationship between structural features. Set the graph structure G = (V, E), where V = $\{v_1,v_2,v_3,...,v_n\}$ denotes the set of nodes, n is the number of nodes, and E = $\{e_1,e_2,e_3,...,e_m\}$ denotes the set of edges with $e_k$ = <$v_i$, $v_j$ >$\in$ E denotes an edge between $v_i$ and $v_j$.

Since the input of GCN is the features of each node, the feature’s attributes associated with each node type need to be determined to better capture the spatial information. The layout of charging facilities is closely related to the traffic network conditions and the needs of users. Therefore, the distribution of POI and road network is used as the characteristic attribute of each node in this study (as shown in Figure 4).

To classify nodes, it is necessary to determine the characteristic attributes associated with each node type. There are eight POI attributes of node characteristics available, which are the number of gas stations, shopping malls, attractions, parking lots, industrial parks, business offices, residential communities, and government institutions; the road network attributes are the trunk, primary, secondary, and tertiary. Eight POI attributes of 400 nodes and four road network attributes were selected as input features to form a two-dimensional tensor of size $400\times 12$. The characteristics of the nodes in the network topology can be represented by an adjacency matrix $A=a_{i,j}\in R^{N\times N}$, and $a_{i,j}$ is defined as shown in Equation (1).

$$a_{i,j}=\left\{\begin{array}{cc}1,\hfill & \mathrm{if} nu{m}_{poi}>0\hfill \\ 0,\hfill & \mathrm{else}\hfill \end{array}\right.$$

(1)

where the $a_{i,j}$ denotes the input of matrix A at the i-th row and the j-th column, where the $A_{i,j}$ represents node characteristic matrix. The $nu{m}_{poi}$ represents the number of each POI in each node. Considering the denseness of residential communities and business offices, it is assumed by experience that when POI is residential communities and $nu{m}_{poi}$ is greater than 3, $A_{i,j}$ is equal to 1, otherwise it is equal to 0; when POI is a business office and $nu{m}_{poi}$ is greater than 5, $A_{i,j}$ is equal to 1, otherwise it is equal to 0. Typically, the maximum numbers of residential communities and business offices in a grid are around 3 and 5, respectively.

2.2.2. GCN-Based Node Classification

This paper uses a GCN-based method to define graph convolution as the aggregation of feature information from neighbors according to the node supervised method [41]. Edge aggregates node information to generate a new feature representation. GCN is a graph-based neural network. By directly performing convolution-like operations on graphs, the problem that conventional depth learning methods cannot now process irregular graph data is solved [42]. The input of GCN is as follows: (1) an $n\times r$ dimensional input feature matrix X composed of the input feature vectors of all nodes in the graph, with n and r being the number of nodes and the dimensionality of the node input feature vectors, respectively; (2) the adjacency matrix of the graph generally represents the structural feature description of the graph, and the output feature matrix Z of GCN is obtained after the convolution layer of each graph in the middle of k. In the middle k graph convolution layers, the features of the nodes flow in the graph structure in the form of hidden features. The GCN is also a neural network layer that propagates from layer to layer in the following manner [42]:

$$H^{(l+1)}=\sigma ({\widehat{D}}^{-\frac{1}{2}}\widehat{A}{\widehat{D}}^{-\frac{1}{2}}{H}^{\left(l\right)}{W}^{\left(l\right)})$$

(2)

where $H^{l+1}$ and $H^l$ are the output and input matrices of layer l ($l=1,2,...,L$). The number of matrix rows is the number of nodes. The matrix columns are features of a node corresponding to the POI and road level. $\widehat{A}=A+I$ is the graph’s adjacency matrix after adding self-connected edges to the undirected graph G. Matrix A is the graph’s adjacency matrix, reflecting the presence or absence of connections between nodes. Matrix I is the unit matrix. $\widehat{D}$ is the degree matrix corresponding to $\widehat{A}$. $W^l$ is a node-negotiable matrix. $W^l$ is a trainable parameter matrix that represents a linear transformation of the mapped feature space. $\sigma$ denotes an activation function. The nodes contain different categories, and the labels of some of them are known, so we use the nodes with given labels for training based on the GCN model. The trained model is used to predict other nodes without labels.

With the topological information of the input graph and the node feature information, the GCN-based method will learn the potential connections between node features and finally obtain the type of each node. Steps of GCN-based node classification include data import and preprocessing, initializing the GCN structure and parameters, and training the GCN. First, the network topology, the characteristic attributes, and the classification labels of each node are imported. The processed data samples are randomly divided into training, validation, and test sets. The training and validation sets are used to determine the various hyperparameters and weights of the GCN. The test set is used to evaluate the performance of the GCN after the training is completed. Before starting to train the GCN, we need to set up the network structure and parameters of the GCN, including the number of graph convolution layers, optimizer, loss function, etc. Then, in training GCN, the edge set, point set, and each node feature of the graph are the first input to obtain the adjacency matrix A and node feature matrix X of the graph. The collection and standardization of node feature information are realized through multiple graph convolution and normalization operations. We extract and learn the features between nodes and output the classified label vector by the softmax layer. Each output vector corresponds to the classification label of the node. Finally, the observed and actual values are used to calculate all nodes’ total loss function values through the loss function. The weights of the GCN are updated using the backpropagation algorithm. The trained model is used to predict the classification of node types in the study area, and the classification label of each node is obtained.

2.2.3. Choosing Candidate Sites

In the study area, 400 grids were selected as candidate areas for constructing EV charging stations. These 400 candidate areas form a rectangular array of 20 rows and 20 columns. The acquired POIs and road networks were cleaned. The extracted regional road network and POI distribution features are taken as the node attributes of the graph. The distribution of the actual charging stations is used as the nodes’ label for the supervised classification of nodes. The candidate charging stations are preliminarily selected. The label of each node is divided into three categories: empty, sparse, and dense, according to the number of node charging piles. The label division basis is as follows:

$$Label=\left\{\begin{array}{cc}empty,\hfill & \mathrm{if} num=0\hfill \\ sparse,\hfill & \mathrm{if} 1\le num\le 30\hfill \\ dense,\hfill & \mathrm{if} num>30\hfill \end{array}\right.$$

(3)

where $Label$ is the label of each node and $num$ is the number of charging piles.

2.3. Optimal Location of EV Charge Stations

The location of the charging station is further optimized according to the preliminary extraction of the candidate sites of the charging station. Firstly, the charging demand is evaluated to merge the nodes based on the traffic flow information, which is used to optimize the location of charge stations. Then, the investment cost of the charging station network is evaluated.

2.3.1. Traffic Demand Assessment

The traffic flow of each node is obtained by using the traffic situation data corresponding to the roads at all levels. This study mainly focuses on the traffic situation data of trunk, primary, secondary, and tertiary urban roads. Based on the traffic situation data processing method proposed in [43], the corresponding traffic flow of roads at all levels under smooth, slow, and congested roads is obtained, as shown in

Table 1. Statistical table of maximum flow of urban roads under different road conditions.

Road Class	Free Speed		Maximum Flow
		Smooth	Slow Down	Congestion
1	80	1636	2340	1076
2	65	1278	1901	861
3	55	1073	1609	646
4	45	913	1317	430

. The obtained real-time traffic situation data are counted to obtain the average value of traffic flow in the same period. According to the traffic flow in each period, the hourly average traffic flow corresponding to each node is calculated. According to the data from the China Automobile Industry Association, the sales volume of new energy vehicles in China exceeded 3.5 million in 2021, and the market share increased to 13.4%. Therefore, the average daily flow of EVs is shown in the following formula:

$$Q=a\ast q\ast 24$$

(4)

where a represents the occupancy of EVs, and q represents the time average of traffic flow.

According to the study area’s traffic flow information, each node’s corresponding charging demand is analyzed. The demand of the charging point is related to the charging probability of the user. The calculation of the demand of the charging point is shown in Formulas (5) and (6). The method compares the charging demand with the node classification label to screen candidate sites. The node is deleted when the demand is consistent with the forecast label result or is lower than the labeling threshold. If the demand exceeds the labeling threshold, the nodes must be reserved for further optimization. When the demand for charging piles exceeds the threshold of the label, it indicates that more charging piles need to be built to meet their needs.

$$Z=Q\ast P÷T_e$$

(5)

where Z represents the demand for charging piles, Q represents the average daily EV flow, and P represents the user’s charging probability.

$$P=n\ast t÷T_e$$

(6)

where n represents the average daily charging times, and n assumes that the user charges once a day. The t refers to the duration of each charge. This study will set different charging duration for three different scenarios:

1.

When the construction of charging stations is all fast charging, the duration of each charging is set to 0.5 h.

2.

When the construction of the charging station is a mixture of fast and slow charging, the duration of each charge is set to be 2 h.

3.

If the construction of the charging station is slow, each charge’s duration is set to be 6 h.

After filtering and deleting nodes that do not need to build charging stations, the nodes are consolidated. The mathematical description is as follows:

$$p_{ij}\le p_i+p_j$$

(7)

where $p_{ij}$ represents the number of charging points required by the merged nodes; $p_i$ represents the number of charging points required by node i; $p_j$ represents the number of charging points needed by node j. To reasonably merge the candidate nodes according to the demand, we assume that the maximum Manhattan distance the nodes can move is 1 km. When the demand of the merged nodes meets the above formula and the moving distance is less than the distance threshold, the node is the final charging station construction area.

2.3.2. Construction Cost Assessment

The optimized site is used to evaluate the cost of the charging station network. The network planning of the charging station should consider the traffic information within the service range and the cost of the charging station. Therefore, the investment cost of the charging station service network is as follows:

$$F=C_1+C_2+C_3$$

(8)

where F is the total cost of charging station planning; $C_1$ is the fixed construction cost of EV charging station converted to each year; $C_2$ is the operation and maintenance cost of the charging station; $C_3$ is the annual charge cost of the charging station.

1.

Annual construction cost of charging station

$$C_1=\frac{r_0{(1+r_0)}^{n_k}}{{(1+r_0)}^{n_k}-1}[A_r{C}_A+M_r{C}_M+C_j+e_i]$$

(9)

where the infrastructure of the EV charging station mainly includes a charger, battery maintenance equipment, charging monitoring, and safety monitoring equipment; distribution facilities mainly include a distribution cabinet, transformer, cable, active filter device, etc. The $n_k$ is the design service life of the charging station; the $r_0$ is the fund recovery rate; $A_r$ is the land area of charging station r, and the number of chargers installed in charging stations of different sizes is different, so the occupied area is also different, as shown in

Table 2. Average floor area of charger.

Number of Chargers	Average Floor Area
30	23
15	22.5
8	20

; $C_A$ is the unit price of land acquisition; $M_r$ is the number of chargers in the charging station; $C_M$ is the unit price of the charger; $C_j$ is the cost of distribution facilities; $e_{ij}$ refers to other auxiliary expenses under different scenarios.

2.

The operation and maintenance costs of the charging station

$$C_2=\mu C_1$$

(10)

where the costs mainly include equipment depreciation costs, overhaul and maintenance costs, wages, etc. As the cost values are not clear, it is considered to calculate the annual operation and maintenance cost as a percentage of the initial investment; the scale factor is u.

3.

Annual charge cost of charging station

$$C_3=365C_e{T}_e[T_r(C_{Cu}+C_{Fe})+M_r(C_L+C_W)]$$

(11)

where $C_e$ is the charging price; $T_e$ is the average effective working time of the charging station every day; $T_r$ is the number of transformers; $M_r$ is the number of chargers in the charging station; $C_{Cu}$ is the copper loss converted to a single distribution transformer; $C_{Fe}$ is the iron loss converted to a single distribution transformer; $C_L$ is the circuit loss converted to a single distribution transformer; $C_W$ is the charging loss converted to a single distribution transformer. The values of the relevant paramaters are given in

Table 3. The values of the relevant parameters.

Parameter	Value	Parameter	Value
$C_M$	USD 27,780	$e_{1j}$	USD 111,120
$C_j$	USD 266,688	$e_{2j}$	USD 83,340
$C_e$	USD 0.11112/kwh	$e_{3j}$	USD 55,560
$C_{Cu}$/$C_{Fe}$	USD 0.005556/kwh	$n_k$	20
$C_L$/$C_W$	USD 0.006945/kwh	$\mu$	3%
$T_e$	20 (h)	$r_0$	0.12
$T_r$	2

; it mainly includes various parameters required in the cost calculation.

3. Case Study

In the case study, our goal is to optimize the layout of charging stations to support the planning and development of local charging facilities. The research first determines a group of candidate locations, then further optimizes charging stations’ selection in different scenarios. We implemented our algorithm in PyTorch and optimized it using the Adam algorithm. Processing was run on an Intel Core i5 computer running Windows with 16 GB RAM.

3.1. Data Preprocessing

Data preprocessing is used to extract node classification labels and analyze the charging demand of each node. It is necessary to preprocess the raw data before constructing the graph structure tensor. First, various POIs are extracted and their number distributions are counted. Secondly, the topology of the road network is checked, and the trunk, primary, secondary, and tertiary roads are extracted. The collected traffic data are real-time traffic data from 21 July 2021 to 27 July 2021, as one week is a complete description of the minimum period of people’s daily travel. The time interval for data sampling is 15 min. The average value of traffic flow in each period is obtained according to the data statistics. Then, the time average traffic flow corresponding to each node is calculated. In the experiment, we train the model by using complete batch processing during the training period. The maximum number of iterations is set to 300. The GCN learners in the method have 16 hidden neuron layers. In addition, the learning rate is set to 0.01, and the weight-decay rate is set to $5\times {10}^{-4}$.

3.2. Candidate Charging Stations

Based on the actual road network, this section preliminarily extracts the candidate sites of EV charging stations in the study area. The 400 grids are used as candidate nodes to quantify and analyze the interest points in the region. A unique ID is created for each node, and label extraction is carried out based on the GCN supervised classification model. After the dataset is run 85 times, we obtain the training loss and average classification accuracy of 400 nodes, as shown in Figure 5. Figure 5a shows the loss of the training set and the verification set, and Figure 5b shows the accuracy of the training set and the verification set. At this time, the loss of the training set and verification set approaches 0.3, the accuracies of the training set and verification set approach 0.7 and 0.6, respectively, and the accuracy of the test set reaches 0.57. Based on the GCN supervised classification training model, this study predicts the classification labels in the study area. In the label prediction results, the number of correctly predicted labels is 255, and the correct rate is 63.75%.

The predicted category labels are compared with the actual label values to filter out the nodes that need further optimization. The node will be deleted when the predicted label value is the same as the actual label. The demand for charging piles at each node is obtained according to the traffic flow information corresponding to each node. The node will be deleted when the predicted label value is consistent with the actual tag value. The remaining nodes need to be further optimized. According to the charging habits of EV users, we divide the research into three scenarios: the charging time is 0.5 h, the charging time is 2 h, and the charging time is 6 h.

Figure 6 shows the location distribution of candidate sites in three scenarios. It can be seen from the figure that when the distribution of charging demand is different, the number of charging facilities configured in the planning area of each charging station is also different. In the figure, the number of charging piles is divided into five levels: the number of charging piles is less than or equal to 8, and the level is 1; the number of charging points is greater than 8, the number of charging points is less than or equal to 15, and the grade is 2; the number of charging piles is greater than 15, the number of charging piles is less than or equal to 30, and the grade is 3; the number of charging piles is greater than 30, the number of charging piles is less than or equal to 45, and the grade is 4; the number of charging points is greater than 45, and the level is 5. In areas with dense POI and road network, it is more attractive for users to travel to generate greater charging demand. Different charging modes require a different number of planned charging stations. As shown in Figure 6a, when the charging mode is mainly fast charging, i.e., the average charging duration is 0.5 h, there are 270 nodes in total that need to be planned for the charging station; as shown in Figure 6b, when the charging mode is mainly fast charging, i.e., the average charging duration is 2 h, there are 273 nodes in total that need to be planned for the charging station; as shown in Figure 6c, when the charging mode is mainly fast charging, i.e., the average charging duration is 6 h, there are 309 nodes in total that need to be planned for the charging station. It can be seen that when the charging mode is mainly slow, the scope of construction is more extensive, and the number of charging stations is higher. In particular, some main roads are distributed within the black rectangle, which may be why the charging demand increases. More charging stations need to be built in this area. To sum up, the spatial distribution of the road network is an essential factor affecting the network planning of the charging station.

3.3. Optimal Location of EV Charge Stations

Candidate charging stations are obtained based on the GCN method and considering the extraction of traffic information constraints. After the nodes that need to be planned for the charging station are preliminarily extracted, the nodes are merged and optimized according to the demand of the charging pile. Therefore, the optimal charging stations under three different scenarios are obtained.

As shown in Figure 7a, there are 101 charging station nodes to be planned, of which the number of charging piles is 1031; as shown in Figure 7b, there are 115 charging station nodes to be planned, and the number of charging piles is 4295; as shown in Figure 7c, there are 100 charging station nodes to be planned, and the number of charging piles is 14,280. The construction costs corresponding to the three scenarios are, respectively, USD 7.2851 million, USD 13.3229 million, and USD 31.1220 million.

In Scenario I, fewer nodes are planned to charge stations and the scenario hasthe lowest investment and construction costs. The population near the urban business district is more mobile and needs to meet the fast charging demand of users. Users can be charged in the shortest time and save waiting time. For the planner, cost minimization is the ultimate goal. Therefore, this scenario is suitable for city center areas where demand is concentrated.

In Scenario II, the nodes requiring the most planning of charging stations are in the middle of the price range. A charging station network combining fast and slow charging would be more reasonable in mixed urban functional areas, such as commercial and residential intersection areas.

In Scenario III, the lowest number of nodes is needed to plan charging stations, but it has the highest investment and construction costs. Users in residential areas have more time for charging at night, and a scenario with a longer charging time can be considered for planning. For planners, the investment cost of this scenario is too high to be fully applicable.

The obtained optimal charging stations are compared with existing stations’ distribution to validate our proposed approach’s effectiveness. Figure 8 shows the distribution of charging stations in the study area with local enlargement (take Scenario I as an example). Yellow points indicate charging stations obtained by the proposed method, black points indicate existing charging stations, and purple area indicates the range of charging stations planned by the proposed method. When an optimal charging station is in the same purple area as an existing station, we treat them as the same charging station. The results are shown in

Table 4. Comparison of planned charging stations and existing charging stations.

Scenarios	Number of Charging Stations	Probability of Similarity to Existing Charging Stations
Scenario I	101	67.3%
Scenario II	115	60.9%
Scenario III	100	77%

, comparing the similarity probabilities for the charging stations planned by the proposed method and the existing ones. From the table, we can see that all three scenarios have a similarity probability of more than 60%, and Scenario III has a similarity probability of nearly 80% with the existing charging stations. This indicates that our proposed method is effective.

This section compares the proposed method with the conventional charging station locating and sizing method [44]. The method uses a grid partitioning method to partition the planning area and plan the charging stations considering the traffic density and capacity constraints, which require 75 charging stations to be built. Compared with our proposed method, the number of charging stations obtained based on locating and sizing methods is far from meeting the increasing demand for electric vehicle charging in the region year by year. Geographical connections between regions are essential in the deployment of charging stations. This approach considers the traffic flow but ignores the spatiotemporal correlation between charging stations, so it deviates somewhat from our proposed approach in optimizing charging station layout, and also has a large discrepancy with existing charging stations.

4. Discussion

In this section, we further discuss how to enrich our proposed research methodology to suit several practical factors in a real-world charging facility deployment setting.

First, the spatial distribution of charging stations under different scenarios is fully demonstrated. The experimental results show that the proposed method is feasible. Figure 7 shows the number of candidate charging stations and the spatial distribution, which provides the basis for optimal planning of charging stations. By integrating adjacent points of candidate charging stations, the economic scale of charging facilities is increased, and construction costs are directly reduced by saving on land use and other expenses.

Second, in practice, decisions on EV charging locations and infrastructure implementation require continuous and long-term planning efforts and changes in charging demand over time. Therefore, the establishment and improvement of charging infrastructure networks are particularly important. In contrast to the charging-demand-based charging station planning approach, this paper focuses more on the influence of spatial and temporal features on the distribution of charging infrastructure to build user-friendly charging station networks and enhance their convenience and accessibility.

Finally, we propose an approach for deploying charging stations that fully considers the connectivity between geographic spaces and can meet the demands of different users, and also considers investment/budgets. The results for charging station deployment are obtained using the proposed method. The actual total construction cost was estimated using these results. Here, we discuss how the budget constraints of building EV charging stations and the possibility of connecting to the grid can be combined with our proposed approach and its implications. We can incorporate budget constraints and charge loading capacity into our proposed approach in two steps. In the first step, the total construction cost is estimated using the proposed method to obtain the deployment results of charging stations, irrespective of budget constraints. In the second step, the number of charging stations can be constrained by the charging capacity to adjust the construction cost. It is important to note that, regardless of the budget, too many charging facilities may burden the grid; with limited budgets, the deployed charging facilities may not meet the charging demand and may lead to a poor charging service experience.

5. Conclusions

This paper proposes a charging station selection method using GCN from the geographic grid to facilitate EV drivers to charge their vehicles. The method aims to address the charging demand of EV users. It proposes three charging station planning scenarios for different charging behaviors of users, taking into account spatial and temporal geographic features to provide a convenient and efficient charging station network for users. Experimental results show that the proposed charging station placement method can offer an effective network space optimization scheme for various scenarios.