Electric vehicle space-time load prediction method and system based on multi-graph dynamic fusion

Abstract

The invention provides an electric vehicle space-time load prediction method and system based on multi-graph dynamic fusion, and relates to the field of electric vehicles, and the method comprises the steps: obtaining the traffic flow and a traffic network graph of a target region; based on the traffic flow and the traffic network map of the target area, the target area is divided into a plurality of sub-areas, based on the plurality of sub-areas, a plurality of nodes are constructed, and one node corresponds to one sub-area; constructing a distance graph, an adjacent graph and a function similarity graph based on the traffic network graph and the plurality of nodes; acquiring historical charging load data of the target area; through the space-time load prediction model, electric vehicle load prediction is carried out based on the historical charging load data of the target area, the distance map, the adjacency map and the function similarity map, and the method has the advantage of improving the accuracy of the electric vehicle load.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of electric vehicles, in particular to an electric vehicle space-time load prediction method and system based on multi-graph dynamic fusion. BACKGROUND

[0002] The large-scale concentrated charging behavior of electric vehicles brings unprecedented challenges to the safe and stable operation of the power grid. Electric vehicle charging load has the characteristics of randomness, intermittency and uneven space-time distribution. Large-scale concentrated charging will cause the sudden increase of local load of the power grid, causing voltage fluctuation, frequency deviation and other power quality problems. In severe cases, it may even cause the overload damage of power grid equipment, threatening the safe and stable operation of the power grid. Therefore, in-depth study of the space-time distribution law of electric vehicle charging load and accurate prediction of its future trend are of great practical significance for ensuring the safe operation of the power grid, optimizing the allocation of electric power resources and promoting the coordinated development of electric vehicles and the power grid.

[0003] The existing technology has made certain progress in the space-time prediction of electric vehicle charging load, and can model the load in the time-space dimension to a certain extent, but mainly relies on a single graph structure to describe the spatial relationship between regions, ignoring the dynamic fusion between multiple graphs. In actual application, the spatial correlation between urban regions is often multi-faceted and dynamically changing, and a single graph structure is difficult to fully and accurately depict such complex spatial relationships. For example, due to changes in traffic flow, user travel patterns and other factors, the charging load transfer relationship between regions may change significantly over different time periods; at the same time, the spatial correlation reflected by different types of data (such as geographic information data, traffic flow data, user behavior data, etc.) may also differ. Therefore, the prediction method based on a single graph structure cannot fully utilize the rich spatial information contained in these multi-source heterogeneous data, and is difficult to adapt to the complex and changing space-time characteristics in actual scenarios, thereby limiting the further improvement of prediction accuracy.

[0004] Therefore, it is necessary to provide an electric vehicle space-time load prediction method and system based on multi-graph dynamic fusion to improve the accuracy of electric vehicle load. SUMMARY

[0005] The application provides a kind of based on multi-graph dynamic fusion electric vehicle space-time load prediction method, comprising: obtaining the traffic flow and traffic road network graph of target area;Based on the traffic flow and traffic road network graph of target area, the target area is divided into multiple sub-regions, based on multiple sub-regions, multiple nodes are constructed, wherein a node corresponds to a sub-region;Based on traffic road network graph and multiple nodes, distance graph, adjacency graph and function similarity graph are constructed;Obtain the historical charging load data of target area;Through space-time load prediction model, based on the historical charging load data of target area and distance graph, adjacency graph and function similarity graph, electric vehicle load prediction is carried out.

[0006] Further, based on traffic road network graph and multiple nodes, distance graph is constructed, comprising: for any two nodes, based on traffic road network graph, the latitude and longitude coordinates of two nodes are obtained, based on the latitude and longitude coordinates of two nodes, the straight line distance of two nodes is calculated;Based on the straight line distance of any two nodes, distance graph and adjacency matrix are constructed.

[0007] Further, based on traffic road network graph and multiple nodes, adjacency graph is constructed, comprising: for any two nodes, based on traffic road network graph, the adjacency relationship of two nodes is determined;According to the adjacency relationship of any two nodes, adjacency graph is constructed.

[0008] Further, based on traffic road network graph and multiple nodes, function similarity graph is constructed, comprising: for each node, based on traffic road network graph, the functional characteristics of the node are determined;For any two nodes, the similarity of the functional characteristics of two nodes is calculated;Based on the similarity of the functional characteristics of any two nodes, function similarity graph is constructed.

[0009] Further, the space-time load prediction model comprises an input layer, a space-time attention unit, a multi-graph dynamic convolution unit and an output layer, wherein the input layer is used to input the historical charging load data of target area, the space-time attention unit is used to capture the dynamic correlation of multiple nodes at different time points based on historical charging load data, the multi-graph dynamic convolution unit is used to generate multi-graph fusion features based on distance graph, adjacency graph and function similarity graph, and the output layer is used to perform electric vehicle load prediction based on multi-graph fusion features.

[0010] Further, the space-time attention unit comprises a time attention group and a space attention group, wherein the time attention group is used to calculate the time attention weight of any two time steps in historical charging load data, generate time attention features based on historical charging load data and the time attention weight of any two time steps, and the space attention group is used to calculate the space attention of any two nodes based on historical charging load data, and generate space attention features based on the space attention of any two nodes and adjacency matrix.

[0011] Further, the multi-graph dynamic convolution unit comprises a spatial convolution group and a time convolution group, wherein the spatial convolution group comprises a first spatial convolution branch, a second spatial convolution branch, a third spatial convolution branch, and a spatial fusion layer, wherein the first spatial convolution branch is configured to extract distance graph spatial features based on the time attention features and the spatial attention features, the second spatial convolution branch is configured to extract adjacency graph spatial features based on the time attention features and the spatial attention features, and the third spatial convolution branch is configured to extract functionally similar graph spatial features based on the time attention features and the spatial attention features, and the spatial fusion layer is configured to fuse the distance graph spatial features, the adjacency graph spatial features, and the functionally similar graph spatial features to generate multi-graph fusion spatial features.

[0012] Further, the spatial fusion layer fuses the distance graph spatial features, the adjacency graph spatial features, and the functionally similar graph spatial features to generate multi-graph fusion features, comprising: determining attention weights of the distance graph, the adjacency graph, and the functionally similar graph; and fusing the distance graph spatial features, the adjacency graph spatial features, and the functionally similar graph spatial features based on the attention weights of the distance graph, the adjacency graph, and the functionally similar graph to generate the multi-graph fusion features.

[0013] Further, the time convolution unit comprises a first time convolution branch, a second time convolution branch, a third time convolution branch, and a time fusion layer, wherein the first time convolution branch is configured to extract distance graph time features based on the time attention features and the time attention features, the second time convolution branch is configured to extract adjacency graph time features based on the time attention features and the time attention features, and the third time convolution branch is configured to extract functionally similar graph time features based on the time attention features and the time attention features, and the time fusion layer is configured to fuse the distance graph time features, the adjacency graph time features, and the functionally similar graph time features to generate multi-graph fusion time features, wherein the multi-graph fusion features comprise the multi-graph fusion spatial features and the multi-graph fusion time features.

[0014] The application provides a multi-graph dynamic fusion-based electric vehicle space-time load prediction system, which applies the multi-graph dynamic fusion-based electric vehicle space-time load prediction method, a data acquisition module configured to acquire traffic flow and a traffic network graph of a target region, a multi-dimensional graphing module configured to divide the target region into a plurality of sub-regions based on the traffic flow and the traffic network graph of the target region, construct a plurality of nodes based on the plurality of sub-regions, wherein one node corresponds to one sub-region, and construct a distance graph, an adjacency graph, and a functionally similar graph based on the traffic network graph and the plurality of nodes, the data acquisition module is further configured to acquire historical charging load data of the target region, and a load prediction module configured to perform electric vehicle load prediction based on the historical charging load data of the target region and the distance graph, the adjacency graph, and the functionally similar graph through a space-time load prediction model.

[0015] Compared with the prior art, the electric vehicle space-time load prediction method and system based on multi-graph dynamic fusion provided by the present application at least have the following beneficial effects:

[0016] 1. By constructing distance graph, adjacency graph and functional similarity graph, the spatial relationship between sub-regions is described from multiple dimensions such as geographical proximity, traffic connectivity and functional similarity, breaking through the limitation of traditional methods which only rely on single topological structure. The multi-graph dynamic convolution module combined with the space-time attention mechanism can adaptively capture the space-time dynamic correlation between nodes in different graphs, effectively integrate multi-source heterogeneous spatial information, provide richer feature input for the model, and thus more accurately describe the spatial distribution rule of charging load.

[0017] 2. The space-time attention mechanism is combined with the multi-graph dynamic convolution to realize joint modeling of dynamic trends in the time dimension and complex dependencies in the space dimension. Through dynamic allocation of attention weights, the model can focus on key time points and spatial nodes, suppress noise interference, and capture the periodicity, burstiness and spatial spillover effect of charging load. Compared with traditional time series models or static graph convolution methods, this framework can significantly improve the prediction accuracy and generalization ability when dealing with complex scenarios such as holidays and extreme weather.

[0018] 3. The interaction mechanism of charging load between different sub-regions (such as coordinated fluctuation of functionally similar regions and space-time conduction of adjacent regions) can be intuitively revealed, providing quantitative basis for charging facility planning and power grid scheduling. At the same time, the visualization of space-time attention weights can locate key influencing factors, helping operators understand the prediction result generation logic and optimize resource allocation strategies. BRIEF DESCRIPTION OF DRAWINGS

[0019] The present specification will be further illustrated in the form of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting, and in these embodiments, the same reference numbers represent the same structures, wherein:

[0020] Figure 1 is a flowchart of an electric vehicle space-time load prediction method based on multi-graph dynamic fusion according to some embodiments of the present specification;

[0021] Figure 2 is a schematic diagram of a plurality of sub-regions according to some embodiments of the present specification;

[0022] Figure 3 is a schematic diagram of a space-time load prediction model according to some embodiments of the present specification;

[0023] Figure 4 is a module schematic diagram of an electric vehicle space-time load prediction system based on multi-graph dynamic fusion according to some embodiments of the present specification. Detailed Implementation

[0024] To more clearly illustrate the technical solutions of the embodiments in this specification, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are merely some examples or embodiments of this specification. For those skilled in the art, these drawings can be applied to other similar scenarios without creative effort. Unless obvious from the context or otherwise specified, the same reference numerals in the drawings represent the same structures or operations.

[0025] Figure 1 This is a flowchart illustrating a multi-graph dynamic fusion-based method for predicting the spatiotemporal load of electric vehicles, as shown in some embodiments of this specification. Figure 1 As shown, a method for predicting the spatiotemporal load of electric vehicles based on multi-graph dynamic fusion may include the following process.

[0026] Step 110: Obtain traffic flow and traffic network map of the target area.

[0027] For example, traffic flow and road network maps of the target area can be obtained from external data sources.

[0028] Step 120: Based on the traffic flow and traffic network map of the target area, divide the target area into multiple sub-areas, and construct multiple nodes based on the multiple sub-areas.

[0029] Each node corresponds to a sub-region.

[0030] Specifically, such as Figure 2 As shown, locations that are close together and have similar traffic flow can be grouped into the same sub-region. Figure 2 In this context, a closed region is called a subregion.

[0031] Step 130: Based on the traffic network map and multiple nodes, construct a distance graph, an adjacency graph, and a functional similarity graph.

[0032] In some embodiments, a distance graph is constructed based on a traffic network map and multiple nodes, including:

[0033] For any two nodes, based on the traffic network map, obtain the latitude and longitude coordinates of the two nodes, and calculate the straight-line distance between the two nodes based on the latitude and longitude coordinates of the two nodes.

[0034] Construct a distance graph and an adjacency matrix based on the straight-line distance between any two nodes.

[0035] Specifically, using Geographic Information System (GIS) technology, the latitude and longitude coordinates corresponding to these two nodes are accurately extracted from the traffic network map. For example, the latitude and longitude coordinates of the center of a sub-region are selected as the latitude and longitude coordinates corresponding to the nodes. After obtaining the latitude and longitude coordinates, the straight-line distance between the two nodes is calculated. This distance is measured in units of length such as kilometers or meters, accurately reflecting the straight-line interval between the two nodes in space.

[0036] When constructing a distance graph, a weighting mechanism is introduced. The reciprocal of the straight-line distance between two nodes is used as their weight value. In a transportation network, nodes that are closer together usually have stronger traffic connections and flow interactions. By taking the reciprocal of the distance as the weight, higher weight values ​​can be assigned to closer node pairs, highlighting their close connection.

[0037] An adjacency matrix is ​​used as a data structure to systematically describe the distance relationships between multiple nodes. The adjacency matrix is ​​a square matrix whose rows and columns correspond to nodes in the transportation network, and the element values ​​are the weights between corresponding nodes. Through the adjacency matrix... For example, adjacency matrix The weight value between the i-th node and the j-th node for:

[0038]

[0039] in, The Euclidean distance formula is used to calculate the path length between two nodes in a transportation network. That is, the square root of the sum of the squares of the differences in coordinates across all dimensions. If the i-th node and the j-th node are not connected, then... Infinity.

[0040] In some embodiments, an adjacency graph is constructed based on a traffic network map and multiple nodes, including:

[0041] For any two nodes, determine their adjacency relationship based on the traffic network map;

[0042] Construct an adjacency graph based on the adjacency relationship between any two nodes.

[0043] Specifically, the adjacency relationship between nodes is abstracted: a value of 1 represents nodes that are directly connected, and a value of 0 represents nodes that are not directly connected. The specific formula is as follows:

[0044]

[0045] in, This represents the abstracted adjacency relationship between the i-th node and the j-th node.

[0046] In some embodiments, a functional similarity graph is constructed based on a traffic network map and multiple nodes, including:

[0047] For each node, the functional characteristics of the node are determined based on the traffic network map;

[0048] For any two nodes, calculate the similarity of their functional features;

[0049] Construct a functional similarity graph based on the similarity of functional features between any two nodes.

[0050] Specifically, to more accurately reflect the functional characteristics of different sub-regions, a functional similarity map is constructed based on the Points of Interest (POIs) corresponding to the regions of each node. The POIs mainly include five categories: residential, working, entertainment, commercial, and other. Each node is associated with a POI vector, representing the distribution of that region across different POI categories.

[0051] A functional similarity adjacency graph is constructed by calculating the similarity of POI vectors between any two nodes. Functional similarity is calculated using cosine similarity.

[0052] If the similarity between two nodes is lower than a set threshold If the two functions are not significantly similar, then the corresponding weights in the adjacency matrix are reset to 0. The specific expression is as follows:

[0053]

[0054] in, The weights are the adjacency matrices of the functionally similar graphs of the two nodes. Let be the similarity between the i-th node and the j-th node.

[0055] Step 140: Obtain historical charging load data for the target area.

[0056] The historical charging load data of the target area can include the charging load of multiple sub-areas at multiple time steps.

[0057] Step 150: Using a spatiotemporal load prediction model, electric vehicle load is predicted based on historical charging load data, distance map, adjacency map, and functional similarity map of the target area.

[0058] Figure 3 These are schematic diagrams of spatiotemporal load prediction models based on some embodiments of this specification, such as... Figure 3As shown, in some embodiments, the spatiotemporal load prediction model includes an input layer, a spatiotemporal attention unit, a multi-graph dynamic convolution unit, and an output layer. The input layer is used to input historical charging load data of the target area. The spatiotemporal attention unit is used to capture the dynamic correlation of multiple nodes at different time points based on the historical charging load data. The multi-graph dynamic convolution unit is used to generate multi-graph fusion features based on distance graphs, adjacency graphs, and functional similarity graphs. The output layer is used to predict electric vehicle load based on the multi-graph fusion features.

[0059] In some embodiments, the spatiotemporal attention unit includes a temporal attention group and a spatial attention group. The temporal attention group is used to calculate the temporal attention weights of any two time steps in the historical charging load data and generate temporal attention features based on the historical charging load data and the temporal attention weights of any two time steps. The spatial attention group is used to calculate the spatial attention of any two nodes based on the historical charging load data and generate spatial attention features based on the spatial attention of any two nodes and the adjacency matrix.

[0060] Specifically, when processing time-series data, different correlation degrees are assigned to charging load information at different time steps. Attention weights are adaptively allocated to historical loads relative to time steps, and these weights reflect the importance of that time step to the current task. The functional equation for time attention can be expressed as:

[0061]

[0062]

[0063] in, It is the first The input data for a layer ST (Spatio-Temporal) spatiotemporal block contains N nodes. Dimension A historical step, These are the learnable parameters. The temporal attention weight matrix maps intermediate features to temporal step correlation. For time attention bias, This is a node-dimensional projection vector, which compresses node information into a scalar (or low-dimensional vector) for temporal attention computation. This is a channel-node interaction matrix that fuses feature channel and node information. The channel-dimensional projection vector is used to extract global information (such as mean and maximum value) of the feature channels. Give each Characterizing the first and The correlation between time steps, followed by attention weights The result is obtained by normalization using the sofmax function. Ensure that the sum of the attention weights is 1. Finally, multiply the input data by the attention weights to obtain a new output: This means dynamically adjusting the input by calculating the importance weights of the time dimension.

[0064] When processing spatial data, different correlations are assigned to the charging load information between different nodes, i.e., attention weights are calculated between nodes. This adaptively captures the dynamic correlations of nodes in the spatial dimension, and adjusts the importance of each node to the current task based on the calculated weights. The functional equation of spatial attention can be expressed as:

[0065]

[0066]

[0067] in, It is the first Input data for the ST spatiotemporal block, For learnable parameters, The spatial attention weight matrix maps intermediate features to node correlation. For spatial attention bias, This is a temporal projection vector, which compresses temporal information into a scalar (or low-dimensional vector) for spatial attention computation. This is a channel-time interaction matrix that fuses feature channels and time information. The channel projection vector is used to extract global information (such as mean and maximum value) of the feature channels. Give each Characterizing the first and The degree of correlation between nodes, followed by attention weights pass The function is normalized to obtain Ensure that the sum of attention weights is 1.

[0068] In some embodiments, the multi-graph dynamic convolutional unit includes a spatial convolutional group and a temporal convolutional group. The spatial convolutional group includes a first spatial convolutional branch, a second spatial convolutional branch, a third spatial convolutional branch, and a spatial fusion layer. The first spatial convolutional branch is used to extract distance graph spatial features based on temporal attention features and spatial attention features. The second spatial convolutional branch is used to extract adjacency graph spatial features based on temporal attention features and spatial attention features, and to extract functionally similar graph spatial features based on temporal attention features and spatial attention features. The spatial fusion layer is used to fuse distance graph spatial features, adjacency graph spatial features, and functionally similar graph spatial features to generate multi-graph fused spatial features.

[0069] Specifically, a graph convolutional neural network is used to capture the spatial dependencies between node neighborhoods. When using spectral graph convolution for computation, the computational complexity is relatively high at O(N²). However, when using Chebyshev polynomials, the number of parameters can be significantly reduced and the computational complexity is also lowered. Therefore, Chebyshev polynomials are directly used to approximate the correlation of charging loads between different nodes. The specific formula is as follows:

[0070]

[0071] in, This represents the graph convolution operation. Represents the graph convolution operator. Represents the convolution kernel. Indicates input features, Represents the polynomial coefficients. The largest eigenvalue of the Laplace matrix is ​​defined according to the Chebyshev polynomial function definition. , , This represents the order of the neighboring nodes around each node extracted by the convolution kernel.

[0072] Spatial attention is used to match Chebyshev polynomials to extract spatial feature information of charging load between different nodes, and the correlation between nodes is dynamically adjusted. The specific calculation formula is as follows:

[0073]

[0074] in, This represents the Hadamard product, which is element-wise multiplication. Graph nodes can be updated using data from themselves and their neighbors. The form of graph convolution is:

[0075]

[0076] in, Let h be the node embedding matrix of the h-th feature channel after convolution of the i-th layer graph. This is the node feature matrix of the h-th feature channel input to the (r-1)-th layer.

[0077] After obtaining the multi-dimensional spatiotemporal features of the traffic network map, spatial convolution operations are performed on the spatial features of different dimensions to obtain the spatial features of the distance graph. Neighbor graph spatial features and functional similarity graph space features To express to it:

[0078]

[0079]

[0080] in, Let be the node feature matrix output at layer r-1 of the i-th graph structure. For graph convolution operations, σ is the activation function.

[0081] In some embodiments, the spatial fusion layer fuses spatial features of distance graphs, spatial features of adjacency graphs, and spatial features of functionally similar graphs to generate multi-graph fusion features, including:

[0082] Determine the attention weights for distance graphs, adjacency graphs, and functionally similar graphs;

[0083] Based on the attention weights of distance graph, adjacency graph, and functionally similar graph, multi-graph fusion features are generated by fusing spatial features of distance graph, adjacency graph, and functionally similar graph.

[0084] Specifically, different graphs will have different degrees of impact on the final spatiotemporal load prediction results. In order to avoid the information redundancy problem caused by using static weighting to fuse different graphs, an attention mechanism is introduced to realize the dynamic fusion between multiple graphs. The node features are aggregated in the spatial and temporal dimensions, and attention weights for each graph are generated using a multilayer perceptron (MLP). These weights will be adaptively adjusted as the input changes, thereby realizing the dynamic fusion of information from multiple graphs.

[0085] Furthermore, the distance graph space features Neighbor graph spatial features and functional similarity graph space features Multi-graph fusion is performed. After uniformly modeling the three different spatial relationship features represented by physical distance, adjacency, and regional functional similarity in each graph, independent and adaptive weights are assigned to each feature dimension. These weighted sums then generate a new, dynamically integrated graph feature representation that combines spatial dependencies from these three perspectives. Multi-angle spatiotemporal load prediction is achieved by fusing features from all graphs. Because different graphs have varying degrees of influence on the final spatiotemporal load prediction results, the extracted graph features need to be fused based on different weight matrices.

[0086]

[0087]

[0088] in, Let G represent the attention weight of the g-th graph, where G represents the number of graphs. Represented as distance graphs Neighbor Map and functional similarity diagram The attention weight matrix, given initial values, can adapt to the degree of influence of different graphs on future results as the model learns. The global context vector represents the state of all nodes after average pooling at the current time. MLP(·) is a shared multilayer perceptron used to calculate the unnormalized attention score for each graph. This is the node feature matrix generated after multi-graph fusion.

[0089] Finally, by stacking the spatiotemporal blocks, we can obtain the spatiotemporal characteristics of each node at the current moment. This indicates that the output of the node channel enters the next part.

[0090] In some embodiments, the temporal convolutional unit includes a first temporal convolutional branch, a second temporal convolutional branch, a third temporal convolutional branch, and a temporal fusion layer. The first temporal convolutional branch is used to extract temporal features of the distance graph based on temporal attention features and temporal attention features. The second temporal convolutional branch is used to extract temporal features of the adjacency graph based on temporal attention features and temporal attention features. It is also used to extract temporal features of the functionally similar graph based on temporal attention features and temporal attention features. The temporal fusion layer is used to fuse the temporal features of the distance graph, the adjacency graph, and the functionally similar graph to generate multi-graph fusion temporal features. The multi-graph fusion features include multi-graph fusion spatial features and multi-graph fusion temporal features.

[0091] Specifically, after capturing the adjacency information between different nodes in the spatial dimension, a Temporal Convolutional Network (TCN) is used to perform convolution from a temporal perspective. This combines causal convolution and dilated convolution to capture long-term dependencies in the time series. Specifically, causal convolution ensures that the model's output depends only on the current and past inputs, unaffected by future information, while dilated convolution expands the receptive field of the convolutional kernel by sampling the input data at intervals. Finally, gated activation units combined with dilated causal convolution are used to further capture temporal dependencies. The specific formula is as follows:

[0092]

[0093] in, The output of the gating activation unit, To expand the output of causal convolution, This represents the Hadamard product. , For different activation functions, For learnable model parameters, For learnable bias parameters, It is represented as a gated activation unit module.

[0094]

[0095] in, Indicates the first Temporal convolution is performed on each spatiotemporal block. It is a tensor aligned with the time series of the previous layer. This represents the standard convolution operation. The parameters representing the convolution kernel in the time dimension, and the choice of activation function. Nonlinear activation. That is, inner layer. Implement one-dimensional convolution in the time dimension, outer layer This means that information aggregated in the time dimension is then convolved in the spatial neighborhood in both the time and space dimensions.

[0096] The temporal fusion layer fuses temporal features from distance graphs, adjacency graphs, and functionally similar graphs in a similar way to the spatial fusion layer, and will not be elaborated here.

[0097] The output layer employs a fully connected layer or a dedicated temporal prediction architecture (such as LSTM, GRU, or Transformer decoder) to map the integrated multi-dimensional features to the prediction target space, generating predicted electric vehicle charging load values ​​for future timeframes. To improve prediction robustness, the output layer can introduce residual connections or attention mechanisms to weightedly fuse key features and suppress noise interference. The load prediction time series can be output through recursive or sequential generation.

[0098] Figure 4 This is a schematic diagram of a multi-graph dynamic fusion-based spatiotemporal load prediction system for electric vehicles, as shown in some embodiments of this specification. Figure 4 As shown, a spatiotemporal load forecasting system for electric vehicles based on multi-graph dynamic fusion may include a data acquisition module, a multi-dimensional mapping module, and a load forecasting module.

[0099] The data acquisition module is used to acquire traffic flow and traffic network maps of the target area;

[0100] The multi-dimensional mapping module is used to divide the target area into multiple sub-regions based on the traffic flow and traffic network map of the target area, construct multiple nodes based on the multiple sub-regions, where one node corresponds to one sub-region, and construct distance graph, adjacency graph and functional similarity graph based on the traffic network map and multiple nodes.

[0101] The data acquisition module is also used to acquire historical charging load data for the target area;

[0102] The load forecasting module is used to forecast the load of electric vehicles based on historical charging load data, distance map, adjacency map and functional similarity map of the target area through a spatiotemporal load forecasting model.

[0103] For a more detailed description of an electric vehicle spatiotemporal load forecasting system based on multi-graph dynamic fusion, please refer to the relevant description of an electric vehicle spatiotemporal load forecasting method based on multi-graph dynamic fusion, which will not be repeated here.

[0104] Finally, it should be understood that the embodiments described in this specification are merely illustrative of the principles of the embodiments described herein. Other variations may also fall within the scope of this specification. Therefore, alternative configurations of the embodiments described herein are intended to be illustrative rather than limiting, and should be considered consistent with the teachings of this specification. Accordingly, the embodiments described herein are not limited to those explicitly introduced and described herein.

Claims

1. A multi-graph dynamic fusion-based electric vehicle space-time load prediction method and system, characterized in that, The method comprises the following steps: acquiring traffic flow and a traffic road network map of a target region; based on the traffic flow and the traffic road network map of the target region, dividing the target region into a plurality of sub-regions, and based on the plurality of sub-regions, constructing a plurality of nodes, wherein one node corresponds to one sub-region; based on the traffic road network map and the plurality of nodes, constructing a distance map, an adjacency map and a functional similarity map; acquiring historical charging load data of the target region; based on the historical charging load data of the target region and the distance map, the adjacency map and the functional similarity map, performing electric vehicle load prediction through a space-time load prediction model. 2.The method of claim 1, wherein, The method of constructing the distance map based on the traffic road network map and the plurality of nodes comprises the following steps: for any two nodes, based on the traffic road network map, acquiring the latitude and longitude coordinates of the two nodes, and based on the latitude and longitude coordinates of the two nodes, calculating the straight-line distance between the two nodes; based on the straight-line distance between any two nodes, constructing a distance map and an adjacency matrix. 3.The method of claim 1, wherein, The method of constructing the adjacency map based on the traffic road network map and the plurality of nodes comprises the following steps: for any two nodes, based on the traffic road network map, determining the adjacency relationship between the two nodes; based on the adjacency relationship between any two nodes, constructing an adjacency map. 4.The method of claim 1, wherein, The method of constructing the functional similarity map based on the traffic road network map and the plurality of nodes comprises the following steps: for each node, based on the traffic road network map, determining the functional characteristics of the node; for any two nodes, calculating the similarity of the functional characteristics of the two nodes; based on the similarity of the functional characteristics of any two nodes, constructing a functional similarity map.

5. The multi-graph dynamic fusion-based spatiotemporal load forecasting method for electric vehicles according to any one of claims 2-4, characterized in that, The space-time load prediction model comprises an input layer, a space-time attention unit, a multi-graph dynamic convolution unit and an output layer, wherein the input layer is used to input the historical charging load data of the target region, the space-time attention unit is used to capture the dynamic correlation of the plurality of nodes at different time points based on the historical charging load data, the multi-graph dynamic convolution unit is used to generate multi-graph fusion features based on the distance map, the adjacency map and the functional similarity map, and the output layer is used to perform electric vehicle load prediction based on the multi-graph fusion features.

6. The method according to claim 5, wherein, The space-time attention unit comprises a time attention group and a space attention group, wherein the time attention group is used to calculate the time attention weight of any two time steps in the historical charging load data, generate time attention features based on the historical charging load data and the time attention weight of any two time steps, and the space attention group is used to calculate the space attention of any two nodes based on the historical charging load data, and generate space attention features based on the space attention of any two nodes and the adjacency matrix.

7. The method according to claim 6, wherein, The multi-graph dynamic convolution unit comprises a spatial convolution group and a time convolution group, wherein the spatial convolution group comprises a first spatial convolution branch, a second spatial convolution branch, a third spatial convolution branch, and a spatial fusion layer, the first spatial convolution branch is configured to extract distance graph spatial features based on the time attention features and the spatial attention features, the second spatial convolution branch is configured to extract adjacency graph spatial features based on the time attention features and the spatial attention features, extract functionally similar graph spatial features based on the time attention features and the spatial attention features, and the spatial fusion layer is configured to fuse the distance graph spatial features, the adjacency graph spatial features, and the functionally similar graph spatial features to generate multi-graph fusion spatial features. 8.The method of claim 7, wherein, The spatial fusion layer fuses the distance graph spatial features, the adjacency graph spatial features, and the functionally similar graph spatial features to generate multi-graph fusion features, comprising: determining attention weights of the distance graph, the adjacency graph, and the functionally similar graph; fusing the distance graph spatial features, the adjacency graph spatial features, and the functionally similar graph spatial features based on the attention weights of the distance graph, the adjacency graph, and the functionally similar graph to generate multi-graph fusion features. 9.The method of claim 7, wherein, The time convolution unit comprises a first time convolution branch, a second time convolution branch, a third time convolution branch, and a time fusion layer, wherein the first time convolution branch is configured to extract distance graph time features based on the time attention features and the time attention features, the second time convolution branch is configured to extract adjacency graph time features based on the time attention features and the time attention features, extract functionally similar graph time features based on the time attention features and the time attention features, and the time fusion layer is configured to fuse the distance graph time features, the adjacency graph time features, and the functionally similar graph time features to generate multi-graph fusion time features, wherein the multi-graph fusion features comprise multi-graph fusion spatial features and multi-graph fusion time features.

10. A multi-graph dynamic fusion-based electric vehicle space-time load prediction system, characterized in that, The multi-graph dynamic fusion-based electric vehicle spatio-temporal load prediction method of any one of claims 1-9 comprises: a data acquisition module configured to acquire traffic flow and a traffic road network graph of a target region; a multi-dimensional graphing module configured to divide the target region into a plurality of sub-regions based on the traffic flow and the traffic road network graph of the target region, construct a plurality of nodes based on the plurality of sub-regions, wherein one node corresponds to one sub-region, and construct a distance graph, an adjacency graph, and a functionally similar graph based on the traffic road network graph and the plurality of nodes; the data acquisition module is further configured to acquire historical charging load data of the target region; a load prediction module configured to perform electric vehicle load prediction based on the historical charging load data of the target region and the distance graph, the adjacency graph, and the functionally similar graph through a spatio-temporal load prediction model.

Images