Method and System for Predicting Urban Electric Vehicle Charging Demand Based on the Fusion of Heterogeneous Multi-Graph Convolutional Networks

Abstract

The present invention relates to the technical field of charging demand prediction, and discloses a method and system for predicting the charging demand of urban electric vehicles based on the fusion of heterogeneous multi-graph convolutional networks. This method focuses on predicting the charging demands of electric vehicles of multiple different vehicle types. By adopting multi-task learning technology, it effectively extracts the shared features of the charging demand data of electric vehicles of different types, thereby exploring the potential associations between the charging modes of electric vehicles of different types. Compared with the prediction of the charging demand of single-type electric vehicles, the prediction of the charging demand of electric vehicles of multiple vehicle types effectively improves the prediction performance; it can adapt to the scenario where multiple types of electric vehicles coexist and interact with each other in the real world, and improves the scalability and generalization ability.

Description

Technical Field

[0001] The present invention relates to the technical field of charging demand prediction, in particular to a method and system for predicting urban electric vehicle charging demand based on the fusion of heterogeneous multi-graph convolutional networks. Background Art

[0002] With the continuous increase in global carbon emissions and the urgent need for sustainable solutions, reducing fossil fuel consumption and achieving "carbon neutrality" have become common goals of many national governments. Electric vehicles (EVs), as a key alternative to traditional gasoline vehicles, are increasingly favored by the public and widely used. This shift towards a green and intelligent transportation system has prompted the rapid deployment of charging infrastructure in urban areas. With the continuous growth of the number of electric vehicles, accurately predicting urban charging demand has become particularly important, which is not only crucial for understanding regional power load changes and scheduling energy distribution, but also significant for grid operators to take effective measures, such as dynamic pricing.

[0003] Currently, a large number of studies have been carried out in the field of electric vehicle charging demand prediction, covering statistical methods, time series prediction methods, spatio-temporal prediction methods, etc. However, existing studies on electric vehicle charging demand prediction often focus only on a single type of vehicle, such as taxis or buses, while ignoring the relationships between different vehicle types and their overall impact on regional charging demand. In addition, most studies have failed to comprehensively consider the multiple spatio-temporal dependencies of electric vehicle charging demand, including the multiple impacts of historical charging demand, regional points of interest (hereinafter referred to as POIs), and vehicle mobility on charging demand. Therefore, existing charging demand prediction methods have significant limitations in capturing the interaction of charging demands between different vehicle types, and their scalability and generalization ability are also limited in the real urban environment where multiple types of electric vehicles coexist. In addition, due to the insufficient representation ability of existing methods in modeling the multiple spatio-temporal relationships of electric vehicle charging demand, their effectiveness and accuracy in practical applications are also affected. Summary of the Invention

[0004] The present invention provides a method and system for predicting urban electric vehicle charging demand based on the fusion of heterogeneous multi-graph convolutional networks, which is used to solve the problem that the existing methods have insufficient representation ability in modeling the multiple spatio-temporal relationships of electric vehicle charging demand, resulting in the reduction of their effectiveness and accuracy in practical applications.

[0005] To solve the above technical problems, the technical solutions proposed by the present invention are as follows:

[0006] In a first aspect, the present application provides a method for predicting urban electric vehicle charging demand based on the fusion of heterogeneous multi-graph convolutional networks, including:

[0007] S1: Construct a multi-vehicle sharing feature extraction module;

[0008] S2: Input the charging demand data of multiple types of electric vehicles into the multi-vehicle sharing feature extraction module, and obtain the fused features output by the sharing feature extraction module;

[0009] S3: Construct and train a heterogeneous multi-graph spatio-temporal prediction module;

[0010] S4: Input the charging demand data of multiple types of electric vehicles and the fused features into the trained heterogeneous multi-graph spatio-temporal prediction module to obtain the charging demands of multiple types of electric vehicles at the prediction moment.

[0011] In a second aspect, the present application also provides an urban electric vehicle charging demand prediction system based on the fusion of heterogeneous multi-graph convolutional networks, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the steps of the method described in the first aspect above are implemented.

[0012] The present invention has the following beneficial effects:

[0013] The urban electric vehicle charging demand prediction method based on the fusion of heterogeneous multi-graph convolutional networks provided by the present application focuses on the charging demand prediction of electric vehicles of multiple different vehicle types. By adopting multi-task learning technology, it effectively extracts the shared features of the charging demand data of different types of electric vehicles, thereby exploring the potential associations between the charging modes of different types of electric vehicles. Compared with the charging demand prediction of single-type electric vehicles, the charging demand prediction of multiple vehicle types of electric vehicles effectively improves the prediction performance; it can adapt to the scenario where multiple types of electric vehicles coexist and interact in the real world, and improves the scalability and generalization.

[0014] In a further technical solution, the multiple spatio-temporal relationships of electric vehicle charging demands are mined and revealed from multiple dimensions. The impacts of historical charging demands, POIs in different regions, and the movement patterns of vehicles on charging demands are respectively constructed into three types of graph structures, enabling the model to effectively learn various factors affecting charging demands from heterogeneous multi-graphs, thereby improving the accuracy of charging demand prediction.

[0015] In a further technical solution, three different types of graphs, namely the historical charging demand graph, the POI similarity graph, and the dynamic vehicle movement graph, are effectively fused using a graph convolutional network, capturing various spatio-temporal dependence relationships and improving the reliability of electric vehicle charging demand prediction. Description of the Drawings

[0016] The accompanying drawings, which form a part of this application, are used to provide a further understanding of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0017] Figure 1 It is a flowchart of a method for predicting urban electric vehicle charging demand based on the fusion of heterogeneous multi-graph convolutional networks provided for this application;

[0018] Figure 2 It is a framework diagram of a model (EVCha) for predicting the charging demand of multi-type electric vehicles in an embodiment of the present invention.

[0019] Figure 3 It is a comparison diagram of the charging demand prediction results of the EVCha model in the 2021Q2 dataset for different regions with other baseline models in an embodiment of the present invention.

[0020] Figure 4 It is a comparison diagram of the charging demand prediction results of the EVCha model in the 2022Q1 dataset for different regions with other baseline models in an embodiment of the present invention.

[0021] Figure 5 It is a comparison diagram of the charging demand prediction results of the EVCha model in the 2021Q2 dataset for different time periods with other baseline models in an embodiment of the present invention.

[0022] Figure 6 It is a comparison diagram of the charging demand prediction results of the EVCha model in the 2022Q1 dataset for different time periods with other baseline models in an embodiment of the present invention. Detailed implementation manners

[0023] To facilitate the understanding of the present invention, the following will describe the present invention more comprehensively and meticulously in conjunction with the accompanying drawings of the specification and preferred embodiments, but the protection scope of the present invention is not limited to the following specific embodiments.

[0024] Unless otherwise defined, all professional terms used hereinafter have the same meaning as commonly understood by those skilled in the art. The professional terms used herein are only for the purpose of describing specific embodiments and are not intended to limit the protection scope of the present invention.

[0025] Unless otherwise specifically stated, various raw materials, reagents, instruments, and equipment used in the present invention can be obtained through the market or can be prepared by existing methods.

[0026] Unless otherwise defined, the technical terms or scientific terms used in the present invention shall have the ordinary meanings understood by those of ordinary skill in the field to which the present invention pertains. The "first", "second" and similar terms used in the present invention do not denote any order, quantity or importance, but are only used to distinguish different components. Similarly, terms such as "a" or "one" do not denote a quantity limitation, but mean that there is at least one. The terms such as "connected" or "coupled" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The terms such as "upper", "lower", "left" and "right" are only used to represent relative positional relationships, and when the absolute position of the object being described changes, the relative positional relationship also changes accordingly.

[0027] Please refer to Figure 1 , this application provides a method for predicting urban electric vehicle charging demand based on the fusion of heterogeneous multi-graph convolutional networks, including:

[0028] S1: Construct a multi-vehicle shared feature extraction module;

[0029] S2: Input charging demand data of multiple types of electric vehicles into the multi-vehicle shared feature extraction module, and obtain the fused features output by the shared feature extraction module;

[0030] S3: Construct and train a heterogeneous multi-graph spatio-temporal prediction module;

[0031] S4: Input charging demand data of multiple types of electric vehicles and the fused features into the trained heterogeneous multi-graph spatio-temporal prediction module to obtain the charging demands of multiple types of electric vehicles at the prediction moment.

[0032] The above method for predicting urban electric vehicle charging demand based on the fusion of heterogeneous multi-graph convolutional networks focuses on the prediction of charging demands of multiple different types of electric vehicles. By adopting multi-task learning technology, it effectively extracts the shared features of charging demand data of different types of electric vehicles, thereby mining the potential associations of charging modes of different types of electric vehicles. Compared with the prediction of charging demand of a single type of electric vehicle, the prediction of charging demand of multiple types of electric vehicles effectively improves the prediction performance; it can adapt to the scenario where multiple types of electric vehicles coexist and interact in the real world, and improves the scalability and generalization ability.

[0033] Since different types of electric vehicles show similar charging demand patterns in the time dimension, and these similar features are beneficial to improving the model performance, the similar features in the time dimension can be extracted and fused from the charging demand data of multiple types of electric vehicles.

[0034] The processing flow of the multi-vehicle sharing feature extraction module includes three steps: cosine similarity calculation, adaptive similarity threshold learning, and feature extraction. The input of the multi-vehicle sharing feature extraction module is various types of electric vehicle charging demand data where k i represents different types of electric vehicles. Here, taking the charging demands of two types of electric vehicles as an example, namely and the output is the fusion feature F. Then, the processing steps of the multi-vehicle sharing feature extraction module constructed in S1 are as follows:

[0035] S1.1 Cosine similarity calculation: Use cosine similarity to calculate the similarity of two charging demand time series. Take the above-mentioned charging demand data of different types of electric vehicles and as the input, calculate the cosine similarity, and obtain the global similarity value S of the charging demands of these two types of electric vehicles. This value reflects the global similarity of all time steps. In actual electric vehicle charging demand prediction, generally, historical charging demand data of a specific length of time steps is used to predict the charging demand of future time steps. Therefore, it is also necessary to calculate the cosine similarity of the historical charging demand data of a specific length of time steps. Here, use and to represent the historical charging demand data of a specific length of time steps of different types of electric vehicles, and use it as the input to calculate the cosine similarity to obtain the local similarity value s. Among them, the specific length of time steps can refer to selecting 24 hours as the time step. This is only an example here and is not limited

[0036] S1.2 Adaptive similarity threshold learning: The adaptive similarity threshold learning includes two fully connected layers, ReLU activation function, and Sigmod function. Take the global similarity value S obtained in S1.1 as the input of the adaptive similarity threshold learning. First, pass through the first fully connected layer and ReLU activation function, and then take the obtained value as the input of the second fully connected layer, and then pass through the Sigmod function to obtain the threshold T. This threshold is used as the condition for judging whether to perform similarity feature extraction

[0037] S1.3 Feature extraction: The feature extraction includes two strategies, namely the similarity feature extraction strategy and the simple addition strategy. When the local similarity value s in S1.1 is greater than the threshold T in S1.2, the similarity feature extraction strategy is adopted, that is, add the historical charging demand data of two types of electric vehicles and and obtain the fusion feature F through the Softmax function; when the local similarity value s in S1.1 is not greater than the threshold T in S1.2, the simple addition strategy is adopted, that is, add the historical charging demand data of two types of electric vehicles and They are directly added together to obtain the fused feature F. According to the similarity degree of the electric vehicle charging demand data, two different feature extraction strategies are adopted to flexibly extract the fused feature, thereby improving the accuracy and stability of the prediction.

[0038] It should be noted that the heterogeneous multi-graph spatio-temporal prediction module constructed in S3 is used for the multi-graph spatio-temporal prediction of the charging demand of each type of electric vehicle. That is, for each type of electric vehicle, this module is independently initialized and its parameters are not shared. This module can effectively capture the static and dynamic spatio-temporal dependency relationships between regions, including historical charging demand similarity, POI similarity, and vehicle mobility.

[0039] The heterogeneous multi-graph spatio-temporal prediction module includes a spatio-temporal attention block, a heterogeneous multi-graph construction block, a spatial graph convolution block, and a temporal learning block. Each module will be described in detail below:

[0040] The spatio-temporal attention block in S3.1 includes two time attention blocks with the same structure, a fusion module, and a spatial attention block. The input is the electric vehicle charging demand data and the fused feature F obtained by the multi-vehicle shared feature extraction module in S1, and the output is the spatio-temporal attention weight matrix The specific implementation steps are as follows:

[0041] Each time attention block contains a multi-head attention structure, a Concat function, a linear transformation, and layer normalization. Here, taking one of the time attention blocks with the input of the electric vehicle charging demand data as an example to describe the specific process. The multi-head attention structure consists of self-attention blocks of multiple heads, and each self-attention block of a head calculates the attention weight through the scaled dot-product attention mechanism. First, is multiplied by W i Q , W i K and W i V respectively to obtain Q i , K i and V i , where W i Q , W i K and W i V are three different weight matrices. The weight matrices are randomly initialized, continuously learned and updated during model training. Subsequently, after transposing K i and multiplying it with Q i , then dividing by where d k is the dimension of Q i , Ki and V i The common dimension of the three is then obtained through the Softmax function to obtain the attention weight of each head, and then combined with V i Multiply to get the output of each head's self-attention block. Since the multi-head attention structure consists of multiple heads' self-attention blocks, it will contain multiple output results. The Concat function is needed to concatenate the output results of all heads, and then perform a linear transformation to combine the results with the original input. Add and apply layer normalization to get the final result For another temporal attention block with the same structure, the principle is the same as above. The input is the fusion feature F obtained by the multi-vehicle sharing feature extraction module in S1, and the output is Z F This step captures the temporal relationship between sequences through a multi-head self-attention mechanism, thereby effectively identifying temporal trends.

[0042] After the above steps, the output results of the two time attention blocks are and Z F , which is used as the input of the fusion module. The fusion module includes a Concat function and a convolution layer. and Z F After connecting with the Concat function, it is converted through the convolution layer to obtain the time attention tensor The temporal attention tensor It contains both rich temporal features and similar features of charging requirements for different types of vehicles.

[0043] The spatial attention block includes multi-head parallel attention calculation and Softmax function, and the input is the final result of the above fusion module, that is, the temporal attention tensor. The output is the spatiotemporal attention weight matrix The temporal attention tensor The projection is linearly transformed and segmented into multiple heads. Each head calculates the attention score in parallel. The calculation method of the attention score is the same as the above-mentioned time attention block, and the scaled dot product attention mechanism is also adopted. Then the obtained result is normalized by the Softmax function to obtain the spatiotemporal attention weight matrix

[0044] The heterogeneous multi-graph construction block described in S3.2 includes the construction of the historical charging demand graph, the construction of the POI similarity graph, and the construction of the vehicle movement graph, which respectively generate the historical charging demand graph POI Similarity Graph and dynamic vehicle movement diagram By fully modeling the multiple correlations of spatial regions from different angles, various spatial correlations can be effectively mined.

[0045] The construction of the historical charging demand graph is to reveal the potential correlation of charging demands among different regions, because charging demands often exhibit temporal correlation with historical observations. Define as the historical charging demand graph, where the vertex set represents all regions, and ε C represents the set of edges. The value of each edge is either 0 or 1, which depends on the similarity of the historical charging demands of these two points (i.e., two regions). The similarity of the historical charging demands is measured by the Pearson correlation coefficient of the historical charging demand data of these two regions. If the value is greater than the set threshold, the edge value is 1; otherwise, it is 0. The threshold is set according to the charging demand data of the actual vehicle types, such as being set as the value of the 75th percentile of the correlation coefficients of all data.

[0046] The construction of the POI similarity graph is to capture the functional similarity among different regions. When predicting the charging demand of a region, it is reasonable to refer to other regions with similar functional areas. Define as the POI similarity graph, where the vertex set represents all regions, and ε P represents the set of edges. The value of each edge depends on the POI similarity of these two points (i.e., two regions). The value of the POI similarity is obtained by calculating the cosine similarity of the POI vectors of the two regions. The dimension of the POI vector is equal to the number of POI categories in the region, and the value of each dimension represents the number of POIs of this category in the region.

[0047] The construction of the vehicle movement graph is to capture the impact of vehicle mobility on charging demand. For example, areas with a higher vehicle density tend to have greater charging demands due to the increased movement frequency. Define as the vehicle movement graph. Since the movement behavior of vehicles changes over time, each time step should correspond to a dynamic vehicle movement graph where the vertex set represents all regions, and represents the set of all edges at the t-th time step. The value of each edge depends on the number of vehicle transfers between these two points (i.e., two regions), that is, the number of vehicles moving from one region to another region at the t-th time step.

[0048] The spatial graph convolution block described in S3.3 includes three parallel graph convolution network (abbreviated as GCN) layers, whose inputs include electric vehicle charging demand data the spatio-temporal attention weight matrix obtained in the spatio-temporal attention block of S3.1 and the three graphs obtained in the heterogeneous multi-graph construction block of S3.2 (i.e., the historical charging demand graph the POI similarity graph and the dynamic vehicle movement map ), and the output is the multi-dimensional spatio-temporal feature tensor H.

[0049] Among the three parallel GCN layers, each GCN layer combines a weight matrix and a corresponding graph for graph convolution, and each GCN layer corresponds to a kind of graph respectively. First, use the historical charging demand map to calculate the initial Chebyshev polynomial T l . Then, for each GCN layer, add the corresponding graph to the spatio-temporal attention weight matrix respectively, and after passing through the Softmax function, multiply it with the initial Chebyshev polynomial T l to obtain the l-th order Chebyshe polynomial and including each kind of graph and the weight matrix, with a total of L orders. To better combine the dynamic attributes of nodes, graph convolution is performed on the electric vehicle charging demand data at each time step t. In each GCN layer, the L-th order Chebyshev polynomials (which are and ) are used to perform graph convolution aggregation on the charging demand data of electric vehicles at the t-th time step, and then the results of the three GCN layers are added together as the graph convolution result at the t-th time step. The above steps are executed for all time steps, and finally the results of all time steps are concatenated and passed through the ReLU function to obtain the multi-dimensional spatio-temporal feature tensor H.

[0050] The time learning block described in S3.4 includes three gated convolutional neural network (abbreviated as GatedCNN) layers with different convolutional kernel sizes, a Concat concatenation layer, two fully connected layers, and a convolutional layer. The input of this module is the multi-dimensional spatio-temporal feature tensor H of the above S3.3 spatial graph convolution block, and the output is the charging demand of electric vehicles at the prediction moment. By introducing a gating mechanism, this module can better learn and select important features, and multiple convolutional kernels can effectively capture features at different time scales. The specific implementation steps are as follows:

[0051] The convolutional kernel sizes of the three Gated CNN layers are set to 3, 5, and 7 respectively, and the multi-dimensional spatio-temporal feature tensor H of the above spatial graph convolution block is used as the input of each Gated CNN layer to obtain the outputs respectively. Then, these three outputs are processed by the Concat concatenation layer and passed through the first fully connected layer to obtain the output H out . Subsequently, adjust the dimension of H out through the convolutional layer, and finally obtain the charging demand of electric vehicles at the prediction moment through the second fully connected layer.

[0052] Furthermore, the above-mentioned urban electric vehicle charging demand prediction method based on heterogeneous multi-graph convolution network fusion further includes using the charging demand data of multiple types of electric vehicles for model training and prediction.

[0053] Before model training, the charging demand data of multiple types of electric vehicles are divided into a training set, a validation set, and a test set according to the ratio of 6:2:2. In the model training stage, the relevant data are respectively passed through S1 to extract shared features of the charging demands of multiple types of vehicles, and then in S3, the charging demands of each vehicle type are predicted. At this point, the charging demands at the prediction moment of multiple types of vehicles can be obtained. The predicted results are compared with the real results, and the Huber loss function corresponding to each vehicle type is calculated. Then, the loss functions of all vehicle types are summed up according to a certain ratio (for example, the ratio of two vehicles is 1:1) to obtain the total loss. The gradient of the total loss is calculated through the backpropagation algorithm, and the model parameters are updated using gradient descent. After each training cycle ends, the model obtained in the current training cycle is evaluated on the validation set, and the training is stopped when the performance of the validation set no longer improves, so as to obtain the model with the optimal performance.

[0054] In the model prediction stage, the data of the test set are input into the trained model obtained, so as to obtain the charging demands of multiple types of electric vehicles at the prediction moment.

[0055] In summary, in this example, by constructing a multi-vehicle shared feature extraction module and a heterogeneous multi-graph spatio-temporal prediction module, adopting technologies such as multi-task learning, graph convolutional neural network, multi-head attention mechanism, and gated convolutional neural network, comprehensively considering the mutual influence of the charging demands of multiple types of electric vehicles, and combining various influencing factors such as historical charging demands, regional POIs, and vehicle mobility, complex spatio-temporal relationships are effectively captured, so as to realize the joint prediction of the charging demands of multiple types of electric vehicles. This method not only considers various factors affecting the charging demand, but also effectively improves the prediction accuracy of the charging demands of multiple types of electric vehicles by extracting the similarity of the charging demands of different types of vehicles.

[0056] Next, a complete example is used to describe in detail the above-mentioned urban electric vehicle charging demand prediction method based on heterogeneous multi-graph convolution network fusion:

[0057] I. Data preprocessing

[0058] The electric vehicle operation dataset used in this example comes from real data in a large city in China and covers two time periods, namely from April to June 2021 (hereinafter referred to as 2021Q2) and from January to March 2022 (hereinafter referred to as 2022Q1). The dataset contains two different types of data, namely vehicle movement data and vehicle charging data. Among them, the vehicle movement data describes the information at the start and end of each electric vehicle trip, and the vehicle charging data records the charging information of each electric vehicle. This dataset contains 29,717 electric vehicles and 11.48 million records.

[0059] The preprocessing of the data includes the following three aspects:

[0060] First, perform data preprocessing on the vehicle movement data and vehicle charging data. Since there are problems such as duplicate records, incorrect field values, and missing individual fields in the data, it is necessary to perform preprocessing such as data cleaning and filling on the data. The preprocessed vehicle movement data contains 8 fields, namely vehicle ID number, vehicle type (taxi or private car), trip start and end times, and GPS longitude and latitude coordinates of the start and end positions. The preprocessed vehicle charging data includes 8 fields, namely vehicle ID number, charging start and end times, GPS longitude and latitude coordinates of the vehicle charging location, vehicle type (taxi or private car), and state of charge (SOC) before and after charging.

[0061] Second, divide the preprocessed vehicle movement data and vehicle charging data based on regions. In this example, the city is divided into 138 grid regions of equal size, and the size of each grid is 7km×7km. Subsequently, map the vehicle movement data and vehicle charging data to different grid regions and process them. Summarize the charging data of the same vehicle type in the same region to obtain the charging demand of a specific vehicle type (taxi or private car) in that region. Summarize the movement data of the same vehicle type transferred from one region to another to generate the mobility data of a specific vehicle type (taxi or private car) between these two regions.

[0062] Third, perform preprocessing on the POI data. Since the purpose of this example is to predict the electric vehicle charging demand in each region, it is necessary to map the POI data to each region and count the number of different types of POIs in each region to generate the POI vector of that region.

[0063] In this example, the above dataset is divided into a training set, a validation set, and a test set according to a ratio of 6:2:2.

[0064] II. Build a multi-vehicle shared feature extraction module

[0065] Vehicles of different types have differences in usage patterns and operating characteristics, but there are also some similarities. For example, private cars and taxis in the same area tend to charge during off-peak hours with less congestion, showing similar charging demand patterns in the time dimension. Identifying and sharing these similar characteristics helps improve the performance of the model. Therefore, in this example, Figure 2 the "multi-vehicle shared feature extraction module" shown in the figure is constructed, and a multi-task learning-based method is used to capture the similar time features of different vehicles and share features with each other.

[0066] To select representative sequence features, first, the cosine similarity function is used to calculate the similarity between two time series. Intuitively, sequences with similar data distributions can share features to improve the performance of task learning. The cosine similarity can be calculated by the following formula:

[0067]

[0068] The charging demand data of different types of electric vehicles (private cars) and (taxis) are used as inputs to calculate the global similarity value S; the historical charging demand data of different types of electric vehicles at a specific length of time step (24 hours in this example) and are used as inputs to obtain the local similarity value s.

[0069] However, only sequence features with similarity higher than a certain threshold can be shared to promote the learning performance of multi-type vehicles. Therefore, adaptive similarity threshold learning is used to determine this threshold. The adaptive similarity threshold learning includes two fully connected layers, a ReLU activation function, and a Sigmod function. Taking the global similarity value S as the input of the adaptive similarity threshold learning, it first passes through the first fully connected layer and the ReLU activation function, and then the obtained value is used as the input of the second fully connected layer, and then passes through the Sigmod function to obtain the threshold T, which is expressed by the formula as follows:

[0070] T = σ(w2 · (ReLU(w1 · S + b1)) + b2_(2)

[0071] where w1 and b1 are the weight matrix and bias vector of the first fully connected layer respectively, used to convert the global similarity value S into an intermediate feature representation; while w2 and b2 are the weight matrix and bias vector of the second fully connected layer respectively, used to convert the output of the first fully connected layer into the input of the Sigmod function. This threshold T is used as the condition for judging whether to perform similarity feature extraction.

[0072] Next, two strategies are adopted to extract the shared features of two sequences from different vehicles: the similarity feature extraction strategy and the simple addition strategy. When the local similarity value s is higher than the threshold T, the similarity feature extraction strategy is used to extract the correlation of the key time steps in the sequence and enhance the significance of the time steps corresponding to the similar distribution; when the local similarity value s is lower than the threshold T, the two sequences are simply added together. The output of the above two strategies is the fused feature F, which can be expressed as:

[0073]

[0074] where w a is a learnable weight parameter.

[0075] III. Spatio-Temporal Attention Block

[0076] The charging demand of electric vehicles exhibits significant characteristics in both the time and space dimensions. To effectively capture the spatio-temporal dependence of the charging demand of electric vehicles, in this example, temporal attention and spatial attention are combined in sequence in the spatio-temporal attention module to generate spatio-temporal attention weights. These weights will be used as the input to the spatial convolution block in Step V.

[0077] The spatio-temporal attention block includes two time attention blocks with the same structure, a fusion module, and a spatial attention block.

[0078] Each time attention block uses the multi-head self-attention mechanism, which can effectively capture the time trend. For each head of the self-attention block, scaled dot-product attention is used to calculate the attention weights, and the calculation method is as follows:

[0079]

[0080] where X is the input data, and are three different weight matrices respectively, and d k is the common dimension of Q i , K i and V i . Since the multi-head self-attention mechanism contains multiple output results, the output results of all heads are concatenated through the Concat function, and then the output of the multi-head self-attention mechanism is obtained through a linear transformation. Then, it is added to the original input X and layer normalization is applied to obtain the result of this time attention block:

[0081] Z = LayerNorm(X + Concat[head i , …, head h W O )(5)

[0082] where W Ois a linear transformation matrix.

[0083] For this module, the input is and F, where is the charging demand data of each vehicle, and F is the fusion feature obtained through the multi-vehicle sharing feature extraction module. After inputting and F into two temporal attention blocks respectively, and Z F are obtained, which are used as the input of the fusion module. The fusion module includes a Concat function and a convolutional layer. After connecting and Z F with the Concat function and then performing transformation through the convolutional layer, the temporal attention tensor is obtained. The specific formula is as follows:

[0084]

[0085] The spatial attention block includes multi-head parallel attention calculation and the Softmax function. The input is the final result of the above fusion module, i.e., the temporal attention tensor The output is the spatio-temporal attention weight matrix

[0086] IV. Heterogeneous multi-graph construction block

[0087] In charging demand prediction, relying solely on time series data may not be able to fully reflect the complex spatio-temporal dependence relationship of charging demand. Therefore, a multi-graph mechanism is proposed to consider the correlation between regions from different perspectives.

[0088] The heterogeneous multi-graph construction block includes historical charging demand graph construction, POI similarity graph construction, and vehicle movement graph construction, and respectively obtains the historical charging demand graph G C , the POI similarity graph G P and the dynamic vehicle movement graph

[0089] The historical charging demand graph can reveal the potential correlation of charging demand between different regions. Define G C =(V, ε C ) as the historical charging demand graph, where the vertex set V represents all regions, and ε C represents the set of edges. Each element C in ε is defined as the similarity of the historical charging demands of regions r i and r j :

[0090]

[0091] Among them, represents region ri and r j of the historical charging demand sequence, indicating to calculate their Pearson correlation coefficient, and θ is the correlation coefficient threshold.

[0092] The POI similarity graph can capture the functional similarity between different regions. Define G P =(V, ε p ) as the POI similarity graph, and each element in ε p is defined as the POI similarity between regions r and r i and r j :

[0093]

[0094] where represents the POI vectors of regions r i and r j , and calculates the cosine similarity of the POI vectors of these two regions.

[0095] The vehicle movement graph can capture the impact of vehicle mobility on charging demand. For example, areas with a higher vehicle density tend to have a greater charging demand due to the increased movement frequency. Define the dynamic vehicle movement graph for each time interval as where each element in represents the number of vehicles moving from one region to another at the t-th time step.

[0096] V. Spatial Graph Convolution Block

[0097] The graph convolutional network (GCN) can effectively aggregate neighbor node information and capture the complex relationships between nodes. In this example, to enhance the feature learning effect, a multi-graph mechanism and spatio-temporal attention weights are combined to dynamically adjust the aggregation operation in the graph convolutional network. The spatial graph convolution block includes three parallel GCN layers, and each GCN layer combines a weight matrix and a type of graph for graph convolution operations.

[0098] First, use the historical charging demand graph G C to calculate the initial Chebyshev polynomial T1. Then, in each GCN layer, add the corresponding graphs G C , G p , to the spatio-temporal attention weight matrix respectively, and after passing through the Softmax function, multiply with the initial Chebyshev polynomial T l ​Multiply them to obtain the l-th order Chebyshev polynomial containing each graph and the weight matrix and

[0099]

[0100] To better combine the dynamic attributes of nodes, next, graph convolution is performed on the electric vehicle charging demand data at each time step t. In each GCN layer, the Chebyshev polynomials of order L (which are and respectively) are used to aggregate the graph convolution of the electric vehicle charging demand data. Then, the results of the three GCN layers are added together to obtain the graph convolution result at the t-th time step. The formula is as follows:

[0101]

[0102] In the formula, Θ l represents the learnable weight value, is the charging demand at each time step;

[0103] Perform the above steps for all time steps, concatenate the obtained graph convolution results, and after passing through the ReLU activation function, obtain the multi-dimensional spatio-temporal feature tensor H. The formula is as follows:

[0104]

[0105] VI. Temporal Learning Block

[0106] To effectively extract features at different time scales, the temporal learning block adopts multiple gated convolutional neural networks (Gated CNN) with different convolutional kernels. Gated CNN controls the transmission of features through a gating mechanism, which helps the model learn key features and suppress noise. In addition, using Gated CNN with different convolutional kernel sizes can capture features at multiple time scales.

[0107] The temporal learning block shown includes three Gated CNN layers, and the convolutional kernel sizes are set to 3, 5, and 7 respectively. The multi-dimensional spatio-temporal feature tensor H obtained from the spatial graph convolution block is used as the input for each Gated CNN layer. The implementation of each Gated CNN is as follows:

[0108]

[0109] Among them, W i and V i are the weight matrices of the i-th convolutional kernel, b i and c i represent the bias terms of the i-th convolutional kernel, represents element-wise multiplication, * represents the convolution operation, bi and c i is the bias term, W i and V i are the sizes of the convolutional kernels, and σ is the sigmoid activation function. After the above operations, the output results of the Gated CNN with different convolutional kernels are respectively and Then, after concatenating using the Concat function, the aggregated output result is obtained through the fully connected layer:

[0110]

[0111] In the formula, FC represents the fully connected layer.

[0112] Finally, H out After the input is processed by the final convolutional layer, the dimension is adjusted and the output is generated through the fully connected layer to obtain the charging demand of the electric vehicle at the final predicted moment.

[0113] VII. Model Training and Prediction

[0114] In the model training stage, the training set of the electric vehicle charging demand data is input into the model, and the shared feature extraction of the charging demands of multiple types of vehicles, the spatio-temporal attention block, the heterogeneous multi-graph construction block, the spatial convolutional block, and the temporal learning block are processed in sequence to fully capture the complex spatio-temporal dependency relationships of the charging demand sequence, and finally the charging demand result at the predicted moment is generated. Subsequently, the charging demand predicted by the model is compared with the actual charging demand, and the corresponding Huber loss function is calculated for each vehicle type. In the multi-task learning framework, the losses of different types of vehicles are weighted and summed according to a certain weight ratio to obtain the total loss function:

[0115]

[0116] where is the loss function of private cars, is the loss function of taxis, and λ is the weight coefficient. The gradient of the total loss is calculated through the backpropagation algorithm, and the model parameters are updated using gradient descent. After each training epoch ends, the model obtained in the current training epoch is evaluated on the validation set. If the performance on the validation set no longer improves, the training is stopped, thus obtaining the model with the optimal performance.

[0117] In the model prediction stage, the data of the test set is input into the above-trained model, so as to obtain the charging demands of multi-type electric vehicles at the predicted moment.

[0118] VIII. Model Performance Evaluation

[0119] To verify the prediction performance of the multi-type electric vehicle charging demand prediction model (EVCha) proposed in this example, eight methods including historical average (HA), long short-term memory network (LSTM), spatio-temporal multi-graph convolutional network (ST-MGCN), adaptive graph convolutional recurrent network (AGCRN), spatio-temporal fusion graph neural network (STFGNN), decoupled dynamic spatio-temporal graph neural network (D 2 STGNN) are selected as the comparison baselines for the experiment. The comparison baseline models are as follows:

[0120] Historical average (HA): Prediction is made based on the average charging demand at the same time interval in the training set.

[0121] Long short-term memory network (LSTM): By introducing forget, input, and output gating mechanisms, it can capture long-term and short-term dependencies and is a commonly used time series prediction model.

[0122] Spatio-temporal multi-graph convolutional network (ST-MGCN): Encodes the non-Euclidean correlations between regions into multiple graphs and models them through multi-graph convolution and context-gated recurrent neural networks to enhance temporal correlations.

[0123] Adaptive graph convolutional recurrent network (AGCRN): Enhances traditional graph convolution through an adaptive module and combines them into a recurrent network to capture spatio-temporal correlations.

[0124] Spatio-temporal fusion graph neural network (STFGNN): Integrated by the STFGN module and a new type of gated CNN module, it captures hidden spatial dependencies through data-driven graphs and their further fusion with the given spatial graph.

[0125] Decoupled dynamic spatio-temporal graph neural network (D 2 STGNN): After separating diffusion and inherent traffic information, they are processed by the diffusion module and the inherent module respectively, and a dynamic graph learning module is used to learn the dynamic features of the traffic network.

[0126] Spatio-temporal self-supervised learning (ST-SSL): Through the paradigm of spatio-temporal self-supervised learning, it enhances traffic pattern representation to reflect spatio-temporal heterogeneity.

[0127] Self-supervised spatio-temporal bottleneck attention network (SSTBAN): Follows a multi-task framework and aims to solve the long-term traffic prediction problem.

[0128] The two evaluation metrics selected are MAE and EMSE, as follows:

[0129]

[0130] Among them, yi represents the true charging demand value of the i-th sample, represents the predicted charging demand value of the i-th sample, and n represents the total number of samples.

[0131] The performance of the model proposed in this example will be verified and compared from multiple perspectives, including the following experimental results:

[0132] (1) Baseline performance comparison

[0133] The EVCha model proposed in the example was experimented on the above dataset, and the performance of the EVCha model and other baseline models was compared in two evaluation metrics, MAE and RMSE. The experimental results are shown in Tables 1 and 2, where Table 1 shows the overall performance of the 2021Q2 dataset and Table 2 shows the overall performance of the 2022Q1 dataset.

[0134] Table 1 Overall performance of the 2021Q2 dataset

[0135]

[0136]

[0137] Table 2 Overall performance of the 2022Q1 dataset

[0138]

[0139] As can be seen from Tables 1 and 2, compared with other baselines, the EVCha model is superior to other baseline models in both MAE and RMSE, verifying the superiority of the EVCha model in the prediction task of multi-type electric vehicle charging demand.

[0140] (2) Verification of prediction performance in the spatial dimension

[0141] To further verify the prediction performance of the EVCha model in different regions, box plots of the performance of the charging demand in all regions in the RMSE evaluation metric were plotted. Figure 3 and Figure 4 respectively show the RMSE box plots of different vehicle types (taxis and private cars) on the 2021Q2 and 2022Q1 datasets. From Figure 3-4 it can be observed that the RMSE results of the EVCha model for taxis and private cars are significantly lower than those of other models, indicating that the EVCha model has achieved good prediction performance in each region.

[0142] (3) Comparison of prediction performance in the time dimension

[0143] To further verify the prediction performance of the EVCha model in the time dimension, a day was divided into four different time periods (0:00 - 6:00, 7:00 - 12:00, 13:00 - 18:00, 19:00 - 23:00), and a comparative analysis of the RMSE performance of the model was conducted within these time periods. Figure 5 and Figure 6 Figure 5 and Figure 6 respectively show the RMSE performance of each model in different time periods on the datasets of 2021Q2 and 2022Q1. For simplicity of display, only the top 3 baseline models with better performance were selected for comparison with the EVCha model. The results show that EVCha performs better than other baseline models in most time periods, indicating that it can effectively predict the charging demand of electric vehicles in different time periods.

[0144] (4) Ablation experiment

[0145] To evaluate the design effectiveness of different components in the EVCha model, ablation experiments were conducted by sequentially removing certain components to observe the impact on the model performance. The components to be ablated included: 1) removing the vehicle movement graph; 2) removing the historical charging demand graph; 3) removing the POI similarity graph; 4) adopting single-task learning; 5) removing the multi-layer Gated CNN and only using a single-layer Gated CNN layer.

[0146] Table 3 Ablation results of each component on the 2021Q2 and 2022Q1 datasets

[0147]

[0148] Table 3 shows the results of the ablation experiments on the two datasets of 2021Q2 and 2022Q1. It can be seen from the table that removing any one component will lead to a decline in the model performance. Especially removing different heterogeneous graphs significantly reduces the prediction accuracy of the model, indicating that fusing multiple heterogeneous graphs is crucial for improving the prediction performance. In addition, compared with the multi-type vehicle prediction of the EVCha model, single-task learning, that is, separately predicting the charging demand of taxis or private cars, also reduces the performance, proving the effectiveness of shared feature extraction in multi-task learning. In addition, the performance of only using a single-layer Gated CNN is not as good as using a multi-layer Gated CNN, which further verifies the effectiveness of the multi-Gated CNN structure in time feature learning.

[0149] The present application also provides an urban electric vehicle charging demand prediction system based on heterogeneous multi-graph convolutional network fusion, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the steps of the above method are implemented. The urban electric vehicle charging demand prediction system based on heterogeneous multi-graph convolutional network fusion can implement various embodiments of the above-mentioned urban electric vehicle charging demand prediction method based on heterogeneous multi-graph convolutional network fusion, and can achieve the same beneficial effects, which will not be elaborated here.

[0150] The preferred specific embodiments of the present invention have been described in detail above. It should be understood that those of ordinary skill in the art can make many modifications and variations based on the concept of the present invention without creative efforts. Therefore, all technical solutions that can be obtained by those skilled in the art in the technical field of the present invention through logical analysis, reasoning, or limited experiments based on the concept of the present invention on the basis of the prior art should fall within the protection scope determined by the claims.

Claims

1. A method for predicting the charging demand of urban electric vehicles based on the fusion of heterogeneous multi-graph convolutional networks, characterized in that, Including: S1: Construct a multi-vehicle sharing feature extraction module; S2: Input the charging demand data of multiple types of electric vehicles into the multi-vehicle sharing feature extraction module, and obtain the fusion features output by the sharing feature extraction module; S3: Construct and train a heterogeneous multi-graph spatio-temporal prediction module; S4: Input the charging demand data of multiple types of electric vehicles and the fusion features into the trained heterogeneous multi-graph spatio-temporal prediction module to obtain the charging demands of multiple types of electric vehicles at the prediction moment; The S2 includes: S21: Calculate the cosine similarity of N charging demand time series using cosine similarity to obtain the global similarity value and the local similarity value; S22: Calculate the threshold T based on the global similarity value; S23: When the local similarity value is greater than the threshold T, adopt the similarity feature extraction strategy to obtain the fusion features; when the local similarity value is not greater than the threshold T, adopt the addition strategy to obtain the fusion features.

2. The method for predicting the charging demand of urban electric vehicles based on the fusion of heterogeneous multi-graph convolutional networks according to claim 1, wherein The S21 includes: Take the charging demand data of different types of electric vehicles as input, calculate the cosine similarity, and obtain the global similarity value of the charging demands of these two types of electric vehicles; Calculate the cosine similarity of the historical charging demand data of a set length time step to obtain the local similarity value.

3. The method for predicting the charging demand of urban electric vehicles based on the fusion of heterogeneous multi-graph convolutional networks according to claim 1, wherein The S22 includes: Take the global similarity value as the input of the adaptive similarity threshold learning. First, pass through the first fully connected layer and the ReLU activation function, then take the obtained value as the input of the second fully connected layer, and then pass through the Sigmod function to obtain the threshold T. The threshold T is used as the condition for judging whether to perform similarity feature extraction.

4. The method for predicting the charging demand of urban electric vehicles based on the fusion of heterogeneous multi-graph convolutional networks according to claim 1, wherein The similarity feature extraction strategy satisfies: Add the historical charging demand data of N types of electric vehicles, and obtain the fusion features after passing through the Softmax function; The addition strategy satisfies: directly adding the historical charging demand data x of N types of electric vehicles k1 and x k2 to obtain a fused feature.

5. The method for predicting the charging demand of urban electric vehicles based on the fusion of heterogeneous multi-graph convolutional networks according to claim 1, characterized in that The heterogeneous multi-graph spatio-temporal prediction module includes: a spatio-temporal attention block, a heterogeneous multi-graph construction block, a spatial graph convolution block, and a time learning block; The S4 includes: S41: Input the charging demand data of multiple types of electric vehicles and the fusion features into the spatio-temporal attention block, and obtain the spatio-temporal attention weight matrix output by the spatio-temporal attention block; S42: Use the heterogeneous multi-graph construction block to generate a historical charging demand graph, a POI similarity graph, and a dynamic vehicle movement graph; S43: Input the spatio-temporal attention weight matrix, the historical charging demand graph, the POI similarity graph, and the dynamic vehicle movement graph into the spatial graph convolution block, and obtain the multi-dimensional spatio-temporal feature tensor output by the spatial graph convolution block; S44: Input the multi-dimensional spatio-temporal feature tensor into the time learning block, and obtain the charging demand of the electric vehicle at the prediction moment output by the time learning block.

6. The method for predicting the charging demand of urban electric vehicles based on the fusion of heterogeneous multi-graph convolutional networks according to claim 5, wherein The time attention block includes a first time attention block, a second time attention block, a fusion module, and a spatial attention block; The S41 includes: Input the charging demand data of the multiple types of electric vehicles into the first temporal attention block, and obtain the Input the fused feature into the second temporal attention block, and obtain Z output by the second temporal attention block F ; Taking the and the Z F as the inputs of the fusion module, after connecting them through the Concat function of the fusion module, and then converting them through the convolutional layer of the fusion module, a temporal attention tensor is obtained; Input the time attention tensor into the spatial attention block. The spatial attention block projects the time attention tensor into multiple heads for linear transformation and segmentation, and then normalizes the obtained result through the Softmax function of the spatial attention block to obtain the spatio-temporal attention weight matrix.

7. The method for predicting the charging demand of urban electric vehicles based on the fusion of heterogeneous multi-graph convolutional networks according to claim 5, wherein The historical charging demand map is used to indicate the potential correlation of charging demands between different regions and satisfies the following relational expression: Among them, the vertex set represents all regions, and ε C represents the set of edges, and the value of each edge is 0 or 1; The POI similarity map is used to indicate the functional similarity between different regions and satisfies the following relational expression: Among them, the vertex set represents all regions, while ε P represents the set of edges; The dynamic vehicle movement map is used to indicate the impact of vehicle mobility on charging demands and satisfies the following relational expression: Among them, the vertex set represents all regions, represents the set of all edges at the t-th time step; The spatial graph convolution block includes three parallel graph convolution network layers GCN; The S43 includes: Using the historical charging demand graph G C Calculate the initial Chebyshev polynomial T l ; For each GCN layer, the corresponding historical charging demand map, POI similarity map, and dynamic vehicle movement map are respectively added to the spatio-temporal attention weight matrix, and after passing through the Softmax function, they are multiplied by the initial Chebyshev polynomial T l to obtain the l-th order Chebyshev polynomial containing each map and the weight matrix and for a total of L orders; Perform graph convolution on the electric vehicle charging demand data at each time step t. In each GCN layer, use the L-order Chebyshev polynomial to perform graph convolution aggregation on the charging demand data of the electric vehicle at the t-th time step, and then add the results of the three GCN layers as the graph convolution result at the t-th time step; Execute the above steps for all time steps, and finally splice the results of all time steps and obtain a multi-dimensional spatio-temporal feature tensor after passing through the ReLU function.

8. The method for predicting the charging demand of urban electric vehicles based on the fusion of heterogeneous multi-graph convolutional networks according to claim 5, characterized in that, The time learning block includes three gated convolutional neural network layers GatedCNN with different convolutional kernel sizes, a Concat splicing layer, two fully connected layers, and a convolutional layer; The S44 includes: Use the multi-dimensional spatio-temporal feature tensor as the input for each Gated CNN layer to obtain the outputs respectively Process these three outputs through a Concat concatenation layer and obtain output H through the first fully connected layer out ; Apply H out Adjust the dimension through the convolutional layer, and finally obtain the charging demand of the electric vehicle at the prediction time through the second fully connected layer.

9. An urban electric vehicle charging demand prediction system based on the fusion of heterogeneous multi-graph convolutional networks, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of any one of the methods described in claims 1 to 8 above.

Images