Regional electric vehicle charging load prediction method based on GAT, GRU and gravity model

Abstract

The invention relates to a regional electric vehicle charging load prediction method based on a GAT, a GRU and a gravity model, and the method comprises the steps: obtaining a charging network, and constructing a topological graph through the charging network; the method comprises the following steps: acquiring charging data, carrying out feature engineering processing on the charging data according to a topological graph, inputting the processed data into a gating cycle unit network model, acquiring initial time features, constructing multiple views based on a node relationship in the topological graph, taking the multiple views as a basic structure, inputting the initial time features into a graph attention network model, and constructing the gating cycle unit network model. Obtaining deep-level spatio-temporal features; the gating circulation unit network model and the graph attention network model are obtained by training through a training set; and performing weighted combination on the spatial-temporal features and the initial time features to obtain node features, and obtaining a charging load prediction result through the node features. According to the method, the accuracy and robustness of charging load prediction are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical fields of deep learning, energy management and intelligent transportation, in particular to a regional electric vehicle charging load prediction method based on GAT and GRU and gravity model. BACKGROUND

[0002] With the increasing emphasis on environmental protection and greenhouse gas emission reduction worldwide, electric vehicles (EV) as a key technology to replace traditional fuel vehicles are being promoted and deployed at a rapid pace globally. According to the International Energy Agency (IEA) report in 2025, the surge in the number of electric vehicles has created a huge demand for charging infrastructure, driving the large-scale construction of charging piles and other supporting facilities. Although the number of global public charging piles doubled between 2022 and 2024, the deployment speed and rationality of the spatial layout have not fully matched the actual needs of the rapidly expanding electric vehicle fleet. The irrational layout of charging facilities and the low utilization rate of some sites, among other issues, in turn limit the further popularization of electric vehicles and the improvement of user experience.

[0003] In order to address the above challenges and achieve the optimal configuration of charging infrastructure and the intelligent scheduling of power grid resources, it is crucial to accurately predict the charging demand of electric vehicles. Precise and short-term charging demand prediction can provide a scientific basis for power grid operators to alleviate the impact of large-scale and disorderly charging behavior on the power grid, ensuring the stability and safety of the power grid.

[0004] Although the existing technical solutions have developed prediction methods that integrate dynamic and static features based on traffic simulation and machine learning technology, these methods often fail to effectively capture the correlation of charging demand between nodes, especially in the case of long-distance traffic nodes and heterogeneous data sources. SUMMARY

[0005] To solve the above problems existing in the prior art, the purpose of the present application is to provide a regional electric vehicle charging load prediction method based on GAT and GRU and gravity model, which can overcome the shortcomings of existing methods, effectively integrate the spatio-temporal relationship and traffic distribution characteristics contained in the charging demand data without the need for high-precision and all-around traffic data, thereby establishing a more accurate and robust regional short-term charging demand prediction model, significantly improving the accuracy and robustness of charging load prediction.

[0006] To achieve the above purpose, the present application provides the following solutions:

[0007] A regional electric vehicle charging load prediction method based on GAT and GRU and gravity model, comprising:

[0008] Obtain the charging network and construct a topology graph using the charging network;

[0009] The charging data is acquired, and feature engineering is performed on the charging data according to the topology graph. The processed data is then input into a gated recurrent unit network model to obtain initial temporal features. Based on the node relationships in the topology graph, a multi-view is constructed. Using the multi-view as the basic structure, the initial temporal features are input into a graph attention network model to obtain deep spatiotemporal features. Both the gated recurrent unit network model and the graph attention network model are trained using a training set.

[0010] The spatiotemporal features are weighted and combined with the initial time features to obtain node features, and the charging load prediction results are obtained through the node features.

[0011] Optionally, feature engineering processing of the charging data includes:

[0012] The charging load data in the charging data is transformed by log1p, and external features related to charging behavior and timestamp information are introduced to normalize the transformed charging load data to the target interval.

[0013] Extract the inherent attributes of each charging station in the charging data that do not change over time, and generate static features based on the inherent attributes; the static features include: the total number of charging piles, Boolean features with DC piles as the main type, and the ratio of DC piles to AC piles. Broadcast the static features along the time dimension and concatenate them with the time series features to obtain the static information of the nodes in each time step.

[0014] Calculate the Pearson correlation between the charging quantity sequences of each node in the topology graph, and introduce the charging quantity of the most correlated node of each node as an enhancement feature into the corresponding node;

[0015] For each time series, reversible instance normalization is performed, that is, the mean and standard deviation of the time series are calculated and normalized.

[0016] Optionally, obtaining the initial time feature includes:

[0017] The processed data is decomposed into recent series, daily periodic series, and weekly periodic series;

[0018] The recent sequence, the daily cycle sequence, and the weekly cycle sequence are respectively input into the gated recurrent unit network model to obtain multiple time-dependent features. These time-dependent features are then concatenated to obtain the initial time features.

[0019] H (0) =[h reccent |hday |h week ]

[0020] Among them, H (0) For the initial time feature, h reccent h day and h week These are recent sequences, daily cycle sequences, and weekly cycle sequences, respectively.

[0021] Optionally, the multiple views include: a static physical map, a dynamic gravity map, and a dynamic correlation map;

[0022] The adjacency matrix of the static physical graph is constructed based on physical distance:

[0023]

[0024] Where, N k (i) represents the k nearest neighbors of node i in terms of physical distance, A phy [i,j] is the adjacency matrix of the static physical graph;

[0025] The adjacency matrix of the dynamic gravity graph:

[0026] The average value of the data within the dynamic backtracking window of each node is calculated based on the normalized time series data to obtain the dynamic potential energy. Based on this dynamic potential energy, the gravitational pull between the node and its neighboring nodes is determined.

[0027]

[0028] Among them, W grav [i,j] represents the gravitational force between nodes i and j, d(i,j) represents the physical distance between the nodes, and σ dist To control the hyperparameter that affects distance decay, P i and P j Let i and j be the dynamic potential energies of nodes i and j, respectively.

[0029] Select the strongest gravitational force corresponding to the target number of nodes, and construct the adjacency matrix of the dynamic gravity graph;

[0030] The adjacency matrix of the dynamic correlation graph:

[0031] Calculate the Pearson correlation coefficient between the time series of any two nodes within the dynamic window, and construct the adjacency matrix and edge weights of the dynamic correlation graph based on the Pearson correlation coefficient:

[0032]

[0033] Where, θ corr For the initially set correlation threshold, ρ ij Let A be the Pearson correlation coefficient.corr [i,j] is the adjacency matrix of the dynamic correlation graph, W corr [i,j] represents the edge weights.

[0034] Optionally, calculating the average value based on the standardized time series data within the dynamic backtracking window of each node includes:

[0035]

[0036] Where, x i,t' For node i within the dynamic window t ' The value at time, μ(X) dyn,i ) and σ(X dyn,i ) represent node i within the entire window X dyn,i The mean and standard deviation of T dyn For the length of the dynamic backtracking window sequence, Softplus(x) = ln(1+e x ) is the activation function.

[0037] Optionally, obtaining the deep spatiotemporal features includes:

[0038] Using the multi-view as the basic structure, the initial temporal features are input into the graph attention network model. Each GAT layer in the graph attention network model is used to calculate the attention coefficient between a node and its neighboring nodes, and to update the node representation by aggregating neighbor information through attention weighting, thereby obtaining the deep spatiotemporal features.

[0039] Optionally, obtaining the node features includes:

[0040] A gating fusion mechanism is set up, which uses a fully connected layer and a sigmoid activation function to calculate the gating vector;

[0041] Based on the gate vector, the spatiotemporal features and the initial time features are weighted and combined to obtain node features:

[0042] H final =(1-g)⊙H (0) +g⊙H refined

[0043] Among them, H final Let g be the node feature, and H be the gating vector. refined As a spatiotemporal feature, H (0) For the initial time characteristics, ⊙ represents element-wise multiplication.

[0044] Optionally, calculating the gate vector includes:

[0045] g = sigmoid(W g [H spatial‖H (0) ]+b g )

[0046] Among them, H spatial This is a preliminary refined spatial feature representation, where || denotes the concatenation operation, g is the gating vector, sigmoid is the activation function, and W... g b is a learnable weight matrix. g It is a learnable bias vector.

[0047] Optionally, obtaining the charging load prediction result includes:

[0048] The node features are input into a linear fully connected output layer to obtain normalized prediction values.

[0049] The predicted values ​​are inversely normalized using the mean and standard deviation stored for each sequence to obtain charging load prediction results with a true physical scale.

[0050]

[0051] in, The normalized predicted value is represented by RevIN_denorm, which is the inverse normalization function.

[0052] The beneficial effects of this invention are as follows:

[0053] This invention replaces traditional static geographic maps by constructing dynamic gravity maps and dynamic correlation maps that update in real time with the data. This allows the model to capture the real node relationships determined by load size and functional similarity, transcending fixed geographical limitations and thus more accurately reflecting complex interactions, significantly improving prediction accuracy. It also enhances model robustness and adapts to changes in data distribution.

[0054] This invention employs Reversible Instance Normalization (RevIN) technology to effectively address the distribution shift (non-stationarity) problem of charging load data caused by factors such as holidays and seasons. This mechanism enables the model to adapt to dynamic changes in data patterns, greatly enhancing the model's stability and generalization ability in different scenarios.

[0055] This invention optimizes prediction logic through intelligent fusion of multi-dimensional features and incorporates a gated fusion mechanism that adaptively balances the node's own time-series features with complex spatiotemporal features learned from its neighborhood. This intelligent information filtering ensures that the model prioritizes the most effective information sources based on specific circumstances, achieving complementary feature advantages and further optimizing prediction results. Attached Figure Description

[0056] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0057] Figure 1 This is a flowchart of a regional electric vehicle charging load prediction method based on GAT, GRU and gravity model according to an embodiment of the present invention.

[0058] Figure 2 This is a schematic diagram of a dynamic gravity diagram according to an embodiment of the present invention. Detailed Implementation

[0059] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0061] like Figure 1 As shown in the figure, this embodiment discloses a method for predicting regional electric vehicle charging load based on GAT, GRU, and a gravity model, including: acquiring a charging network and constructing a topology graph using the charging network; acquiring charging data, performing feature engineering processing on the charging data according to the topology graph, inputting the processed data into a gated recurrent unit network model to obtain initial temporal features, constructing a multi-view based on the node relationships in the topology graph, using the multi-view as the basic structure, inputting the initial temporal features into a graph attention network model to obtain deep spatiotemporal features; both the gated recurrent unit network model and the graph attention network model are trained using a training set; weighting and combining the spatiotemporal features with the initial temporal features to obtain node features, and obtaining the charging load prediction result through the node features.

[0062] Furthermore, feature engineering processing of the charging data includes: performing a log1p transformation on the charging load data and introducing external features related to charging behavior and timestamp information; normalizing the transformed charging load data to the target interval; extracting the inherent attributes of each charging station in the charging data that do not change over time, and generating static features based on these inherent attributes; the static features include: the total number of charging piles, Boolean features with DC piles as the main component, and the ratio of DC piles to AC piles; broadcasting the static features along the time dimension and concatenating them with the time series features to obtain the static information of the nodes at each time step; calculating the Pearson correlation between the charging quantity sequences of each node in the topology graph, and introducing the most correlated node of each node as an enhancement feature into the corresponding node; performing reversible instance normalization on each time series, i.e., calculating the mean and standard deviation of the time series respectively, and normalizing the mean and standard deviation.

[0063] Specifically, step one, feature engineering and feature normalization:

[0064] Comprehensive feature engineering is performed on the acquired raw data, which is divided into three parts: time-series features, static features, and correlation-based feature enhancement.

[0065] First, the original charging load data y was transformed using a log1p method to mitigate its highly skewed distribution. Then, external features related to charging behavior, including temperature, electricity price, and current service fee, were introduced. Finally, timestamp information, including hour, weekday, and month, was extracted and normalized to the [0,1] interval.

[0066] For static features, inherent attributes of each charging station that do not change over time are extracted, namely the charging station coordinates, the number of DC charging piles, and the number of AC charging piles. Based on this, three new static features are generated: the total number of charging piles, a Boolean feature indicating whether DC charging piles are dominant, and the ratio of DC charging piles to AC charging piles. These static features are broadcast along the time dimension and concatenated with time series features, enabling the model to obtain static information of nodes at each time step.

[0067] To capture the functional similarity between nodes that the topology graph may not fully reflect, this invention introduces an independent feature enhancement method. This method calculates the Pearson correlation of each node's historical load sequences only on the training set, thereby identifying the Top-K most functionally similar 'neighbors' for each node i. Subsequently, this invention performs mean aggregation on the features of these K neighbors and then concatenates the aggregation result with the node i's own features, thereby generating more informative enhanced features for the model to use.

[0068] To address the data distribution offset issue, before inputting the model, this invention applies Reversible Instance Normalization (RevIN) to each sequence to mitigate the potential distribution offset caused by stabilizing the charging load time series of each charging station (or region). For an input sequence containing S samples... RevIN for each sample (i.e., each time series) (where i = 1, ..., S) Instance normalization is performed independently. First, calculate x for each sample. i The mean μ(x) i ) and standard deviation σ(x) i ):

[0069]

[0070] Then the sample is normalized:

[0071]

[0072] Where, x i,j This represents the value of the i-th sample at the j-th time step; This represents the input batch, which contains S independent samples (e.g., historical load sequences of S charging stations); each sample x i It is a time series of length F, where ε is a very small constant added to prevent the denominator from being zero, and x i,norm It is sample x i The result after normalization.

[0073] Save the calculated μ(x) and σ(x) for subsequent inverse normalization steps.

[0074] Furthermore, obtaining the initial time features includes: decomposing the processed data into recent sequences, daily periodic sequences, and weekly periodic sequences; inputting the recent sequences, daily periodic sequences, and weekly periodic sequences into a gated recurrent unit network model to obtain multiple time-dependent features; and concatenating the time-dependent features to obtain the initial time features.

[0075] Specifically, step two involves encoding multi-cycle event sequences:

[0076] After feature engineering, for any historical time t, this invention constructs a complete feature vector containing all information. This vector is composed of three parts: the dynamic temporal features at that time, the static features that do not change over time, and the correlation features between nodes. To comprehensively encode time dependencies at different time scales, this invention employs different historical backtracking methods, utilizing the aforementioned complete feature vector to construct the following three independent input sequences for each node to capture complex time dependencies:

[0077] (1) Recent sequence X reccent Includes the nearest T before the predicted point r Feature data from consecutive time steps are used to capture short-term dependencies.

[0078] (2) Daily cycle sequence X day Extracted past T d The characteristic data of the hour corresponding to the prediction time point are used to model daily periodicity.

[0079] (3) Periodic sequence X week Extracted past T w During the week, feature data from the same week and hour as the prediction time point are used to model weekly periodicity.

[0080] The three normalized sequences are fed into independent gated recurrent unit (GRU) networks. In this invention, the final hidden state vector output by each GRU after processing the corresponding sequence is used as the feature representation of that sequence, denoted as h. reccent h day h week Subsequently, these three vectors are concatenated to obtain an initial time representation that incorporates multi-period time information. N is the total number of nodes, where D h It is the feature dimension after splicing.

[0081] H (0) =[h reccent |h day |h week ]

[0082] Among them, H final Let g be the node feature, and H be the gating vector. refined As a spatiotemporal feature, H (0) For the initial time characteristics, ⊙ represents element-wise multiplication.

[0083] Furthermore, the multi-view includes: a static physical graph, a dynamic gravity graph, and a dynamic correlation graph; the adjacency matrix of the static physical graph is constructed based on physical distance; the average value of the time series data within the dynamic backtracking window of each node is calculated to obtain the dynamic potential energy; based on the dynamic potential energy, the gravity between the node and its neighboring nodes is determined; the strongest gravity corresponding to the target number of nodes is selected to construct the adjacency matrix of the dynamic gravity graph; the Pearson correlation coefficient between the time series of any two nodes within the dynamic window is calculated; based on the Pearson correlation coefficient, the adjacency matrix and edge weights of the dynamic correlation graph are constructed.

[0084] Specifically, step three, multi-stream GAT construction:

[0085] To comprehensively describe the complex relationships between nodes, a multi-view graph structure containing static and dynamic views is constructed:

[0086] (1) Static physical diagram:

[0087] This graph captures static proximity relationships determined by geographical location and that do not change over time. Its adjacency matrix A phy Based on physical distance:

[0088]

[0089] Where, N k (i) represents the k nearest neighbors of node i in terms of physical distance. This invention ensures the symmetry of the graph by taking the union of the sets.

[0090] (2) Dynamic gravity diagram, such as Figure 2 As shown:

[0091] Based on the gravity model, it is assumed that the greater the recent average charge of a node, the stronger its attraction to other nodes, and this attraction will decrease with increasing distance.

[0092] First, this invention acquires the historical charging load data of each node within the dynamic backtracking window, denoted as a matrix. Based on this data, the dynamic potential energy P of each node is calculated. i This potential energy is obtained by averaging the standardized time series of node i within the window:

[0093]

[0094] Where, x i,t' Is node i within the dynamic window t? ' The value at time, and μ(X) dyn,i and σ(X) dyn,i ) are node i in the entire window X dyn,i The mean and standard deviation of T dyn N represents the length (time steps) of the dynamic backtracking window, and N represents the total number of charging nodes in the system.

[0095] Then, the gravitational force W between node i and node j grav [i,j] is defined as:

[0096]

[0097] Where d(i,j) is the physical distance between nodes, σ dist This is a hyperparameter that controls the effect of distance decay. The above formula defines the gravitational weight between any two nodes, forming a weight matrix W. gravBased on the weight matrix, this invention employs a Top-K method to filter the weights: specifically, for each node i, this invention selects the k node with the largest weight value from all other nodes. attract A dynamic adjacency matrix A is constructed using neighboring nodes. grav And retain the weights of these connections, while resetting the weights of other connections to 0.

[0098] (3) Dynamic correlation diagram:

[0099] The flow focuses on uncovering data-driven functional connections. Its graph construction process is completely independent of any prior knowledge based on physical space. The graph's topology and edge weights are dynamically determined solely by the Pearson correlation between nodes over time. This mechanism aims to capture spatial dependencies orthogonal to the first two flows based on physical proximity and gravity models, which are purely based on behavioral similarity, thus forming a key complement in the information dimension.

[0100] Calculate the Pearson correlation coefficient ρ between the time series of any two nodes i and j within the dynamic window. ij :

[0101]

[0102] The adjacency matrix A of the graph corr And edge weight W corr Based on correlation:

[0103]

[0104] Where, θ corr This is the initially set relevance threshold. Simultaneously, to ensure model stability and prevent isolated nodes in GAT calculations, this invention forcibly adds a self-loop, A', to the dynamically generated adjacency matrix. corr =A corr ∪I, where I is the identity matrix.

[0105] Furthermore, obtaining deeper spatiotemporal features includes: using multiple views as the basic structure, inputting initial temporal features into a graph attention network model, where each GAT layer is used to calculate the attention coefficient between a node and its neighboring nodes, and updating the node representation by aggregating neighbor information through attention weighting, thereby obtaining deeper spatiotemporal features.

[0106] Specifically, step four involves the extraction of spatiotemporal features using a multi-graph attention network:

[0107] Using the above three graphs (or combinations thereof) as the basic structure, the initial time feature is represented as H. (0)As input, the data is fed into a graph attention network (GAT) with L layers for deep extraction of spatiotemporal features. In each GAT layer, node i and its neighbor j ∈ N are considered. i Attention coefficient e ij The importance of different neighbors is calculated and used to learn the significance of each neighbor. Finally, the node representation is updated by aggregating neighbor information through attention-weighted aggregation. This process allows the model to learn highly complex, non-linear spatiotemporal dependencies, resulting in a deep spatiotemporal feature representation H. refined .

[0108] Furthermore, obtaining node features includes: setting a gating fusion mechanism, which uses a fully connected layer and a sigmoid activation function to calculate a gating vector; weighting and combining spatiotemporal features with initial time features based on the gating vector to obtain node features; inputting the node features into a linear fully connected output layer to obtain normalized prediction values; and calling the mean and standard deviation stored in each sequence to perform inverse normalization on the prediction values ​​to obtain charging load prediction results with a real physical scale.

[0109] Specifically, step five involves gating fusion and predicting the output:

[0110] To dynamically weigh and fuse information from the time dimension (step two) and the spatiotemporal dimension (step four), this invention designs a gated fusion mechanism. This mechanism first calculates a gate vector g using a fully connected layer and a sigmoid activation function:

[0111] g = sigmoid(W g [H spatial ‖H (0) ]+b g )

[0112] Among them, H spatial This is a preliminary spatial feature representation, where || denotes the concatenation operation, g is the gating vector, and sigmoid is the activation function. W g b is a learnable weight matrix. g This is a learnable bias vector. Then, the initial temporal feature representation H is obtained by using this gating vector. (0) and deep spatiotemporal feature representation H refined By performing weighted combination, the final node feature representation is obtained:

[0113] H final =(1-g)⊙H (0) +g⊙H refined

[0114] Here, ⊙ represents element-wise multiplication. This mechanism allows the model to adaptively decide whether to rely more on the node's own temporal pattern or on the influence of its spatiotemporal neighborhood in the current prediction. Finally, H... final Input a fully connected linear output layer to obtain normalized predicted values.

[0115]

[0116] Then, the mean and standard deviation saved for each sequence in steps one and three are used to perform inverse normalization to obtain the final charging load prediction result with true physical scale.

[0117]

[0118] in, These are the normalized predicted values.

[0119] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for predicting regional electric vehicle charging load based on GAT, GRU, and a gravity model, characterized in that, include: Obtain the charging network and construct a topology graph using the charging network; The charging data is acquired, and feature engineering is performed on the charging data according to the topology graph. The processed data is then input into a gated recurrent unit network model to obtain initial temporal features. Based on the node relationships in the topology graph, a multi-view is constructed. Using the multi-view as the basic structure, the initial temporal features are input into a graph attention network model to obtain deep spatiotemporal features. Both the gated recurrent unit network model and the graph attention network model are trained using a training set. The spatiotemporal features are weighted and combined with the initial time features to obtain node features, and the charging load prediction results are obtained through the node features.

2. The regional electric vehicle charging load prediction method based on GAT, GRU, and gravity model according to claim 1, characterized in that, Feature engineering processing of the charging data includes: The charging load data in the charging data is transformed by log1p, and external features related to charging behavior and timestamp information are introduced to normalize the transformed charging load data to the target interval. Extract the inherent attributes of each charging station in the charging data that do not change over time, and generate static features based on the inherent attributes; the static features include: the total number of charging piles, Boolean features with DC piles as the main type, and the ratio of DC piles to AC piles. Broadcast the static features along the time dimension and concatenate them with the time series features to obtain the static information of the nodes in each time step. Calculate the Pearson correlation between the charging quantity sequences of each node in the topology graph, and introduce the charging quantity of the most correlated node of each node as an enhancement feature into the corresponding node; For each time series, reversible instance normalization is performed, that is, the mean and standard deviation of the time series are calculated and normalized.

3. The regional electric vehicle charging load prediction method based on GAT, GRU, and gravity model according to claim 1, characterized in that, Obtaining the initial time features includes: The processed data is decomposed into recent series, daily periodic series, and weekly periodic series; The recent sequence, the daily cycle sequence, and the weekly cycle sequence are respectively input into the gated recurrent unit network model to obtain multiple time-dependent features. These time-dependent features are then concatenated to obtain the initial time features. H (0) =[h reccent |h day |h week ] Among them, H (0) For the initial time feature, h reccent h day and h week These are recent sequences, daily cycle sequences, and weekly cycle sequences, respectively.

4. The regional electric vehicle charging load prediction method based on GAT, GRU, and gravity model according to claim 1, characterized in that, The multi-view includes: a static physics diagram, a dynamic gravity diagram, and a dynamic correlation diagram; The adjacency matrix of the static physical graph is constructed based on physical distance: Where, N k (i) represents the k nearest neighbors of node i in terms of physical distance, A phy [i,j] is the adjacency matrix of the static physical graph; The adjacency matrix of the dynamic gravity graph: The average value of the data within the dynamic backtracking window of each node is calculated based on the normalized time series data to obtain the dynamic potential energy. Based on this dynamic potential energy, the gravitational pull between the node and its neighboring nodes is determined. Among them, W grav [i,j] represents the gravitational force between nodes i and j, d(i,j) represents the physical distance between the nodes, and σ dist To control the hyperparameter that affects distance decay, P i and P j Let i and j be the dynamic potential energies of nodes i and j, respectively. Select the strongest gravitational force corresponding to the target number of nodes, and construct the adjacency matrix of the dynamic gravity graph; The adjacency matrix of the dynamic correlation graph: Calculate the Pearson correlation coefficient between the time series of any two nodes within the dynamic window, and construct the adjacency matrix and edge weights of the dynamic correlation graph based on the Pearson correlation coefficient: Where, θ corr For the initially set correlation threshold, ρ ij Let A be the Pearson correlation coefficient. corr [i,j] is the adjacency matrix of the dynamic correlation graph, W corr [i,j] represents the edge weights.

5. The regional electric vehicle charging load prediction method based on GAT, GRU, and gravity model according to claim 4, characterized in that, The average value is calculated based on the standardized time series data within the dynamic backtracking window of each node, including: Where, x i,t' Let μ(X) be the value of node i at time t' within the dynamic window. dyn,i ) and σ(X dyn,i ) represent node i within the entire window X dyn,i The mean and standard deviation of T dyn For the length of the dynamic backtracking window sequence, Softplus(x) = ln(1+e x ) is the activation function.

6. The regional electric vehicle charging load prediction method based on GAT, GRU, and gravity model according to claim 1, characterized in that, Obtaining the deep spatiotemporal features includes: Using the multi-view as the basic structure, the initial temporal features are input into the graph attention network model. Each GAT layer in the graph attention network model is used to calculate the attention coefficient between a node and its neighboring nodes, and to update the node representation by aggregating neighbor information through attention weighting, thereby obtaining the deep spatiotemporal features.

7. The regional electric vehicle charging load prediction method based on GAT, GRU, and gravity model according to claim 1, characterized in that, Obtaining the node features includes: A gating fusion mechanism is set up, which uses a fully connected layer and a sigmoid activation function to calculate the gating vector; Based on the gate vector, the spatiotemporal features and the initial time features are weighted and combined to obtain node features: H final =(1-g)⊙H (0) +g⊙H refined Among them, H final Let g be the node feature, and H be the gating vector. refined As a spatiotemporal feature, H (0) For the initial time characteristics, ⊙ represents element-wise multiplication.

8. The regional electric vehicle charging load prediction method based on GAT, GRU, and gravity model according to claim 7, characterized in that, Calculating the gate vector includes: g=sigmoid(W g [H spatial ‖H (0) ]+b g ) Among them, H spatial This is a preliminary refined spatial feature representation, where || denotes the concatenation operation, g is the gating vector, sigmoid is the activation function, and W... g b is a learnable weight matrix. g It is a learnable bias vector.

9. The regional electric vehicle charging load prediction method based on GAT, GRU, and gravity model according to claim 1, characterized in that, Obtaining the charging load prediction results includes: The node features are input into a linear fully connected output layer to obtain normalized prediction values. The predicted values ​​are inversely normalized using the mean and standard deviation stored for each sequence to obtain charging load prediction results with a true physical scale. in, The normalized predicted value is represented by RevIN_denorm, which is the inverse normalization function.

Images