Electric vehicle charging prediction method and model fusing space-time node information

Abstract

The invention provides an electric vehicle charging prediction method and model fused with time-space node information. The prediction method comprises the following steps: constructing an adjacent matrix according to traffic position characteristics, and obtaining a historical sub-number set according to historical number set slicing so as to obtain a node physical graph; obtaining a PageRank value according to the node physical graph so as to obtain a time sequence matrix and a spatio-temporal information graph; based on a multi-head attention mechanism, respectively constructing a space attention matrix and a time attention matrix according to the time sequence matrix; obtaining a spatial dynamic correlation matrix according to the spatial attention matrix and the adjacency matrix, and combining the spatial-temporal information graph to obtain a fusion graph; and obtaining a spatial convolutional network based on graph convolution according to the fusion graph, and obtaining a time convolutional network based on causal convolution according to the time attention matrix, so as to obtain a prediction model for electric vehicle charging prediction. According to the electric vehicle charging prediction method fusing the space-time node information, the space-time coupling effect can be comprehensively considered, and the prediction precision is high.

Description

Technical Field

[0001] The present invention relates to the technical field of electric vehicle charging prediction, and in particular to an electric vehicle charging prediction method and model integrating time and space node information. Background Art

[0002] With the increasing attention paid to environmental protection and sustainable development around the world, the electric vehicle (EV) industry has developed rapidly and has become an important direction for future transportation development. However, the popularity of EVs has also brought about a significant increase in charging demand, posing new challenges to the stable operation of the power grid and the balance of the power market. At present, there are the following problems with the prediction of EV charging load: (1) There are differences in charging demand in different regions, which is affected by multiple factors such as the number of EVs, user charging habits, and the distribution of charging facilities. (2) The charging load also fluctuates over time, which is restricted by factors such as the driving patterns of EVs and changes in user charging demand. (3) Traditional EV charging load prediction methods often perform statistical analysis based on historical data, ignoring the importance of spatiotemporal node information and failing to accurately reflect the spatiotemporal distribution characteristics of EV charging load. With the popularization of EV applications, EV charging load prediction is not only related to the stable operation of the power grid, but also directly affects the optimal use of energy. With the increasing number of EVs, disordered charging may lead to peak load on the power grid, affecting the stability and safety of the power grid. Therefore, it is necessary to formulate a reasonable charging strategy to balance the power grid load and avoid problems such as overload and power outages.

[0003] However, most current studies focus on the impact of the time dimension, while when mining spatial dimension information, artificial intelligence methods are mostly used to extract it from load data. Insufficient consideration is given to the impact of actual traffic geographic information on charging load. How to reasonably construct a traffic network model and fully consider the impact of spatial dimension information in the prediction of the spatiotemporal distribution of electric vehicle charging load still requires further research. Summary of the invention

[0004] The present invention aims at the problems existing in the prior art and proposes a prediction model for electric vehicle charging that considers the importance of spatiotemporal nodes. The prediction model considers the actual traffic network information, i.e., the traffic location characteristics, and learns the potential spatiotemporal dynamic characteristics by combining the load information to generate a spatiotemporal information graph that considers the importance of nodes. Then, the spatiotemporal information graph is combined with the spatiotemporal attention layer to further extract the spatiotemporal feature information. In addition, the extracted features are input into the Graph Convolutional Network (GCN) to obtain spatial dependencies from the neighborhood. At the same time, the extended causal convolution is used to obtain the temporal dependencies from the adjacent time, which enhances the model's ability to capture long-term dependencies, captures the spatiotemporal coupling, and improves the accuracy of electric vehicle charging prediction. Finally, considering the randomness of electric vehicle user demand, based on the analysis of electric vehicle charging characteristics, a complete electric vehicle charging prediction method and model that integrates spatiotemporal node information is established.

[0005] To achieve the above purpose, the technical solution adopted by the present invention is as follows: a method for predicting electric vehicle charging by integrating spatiotemporal node information, comprising the following steps:

[0006] Obtaining the total duration and time slice; obtaining the traffic location characteristics of the charging station and the historical data set of the charging station within the total duration;

[0007] Slicing the historical data set according to the time slice to obtain a plurality of historical sub-data sets;

[0008] Based on the time slice, construct an adjacency matrix according to the traffic location characteristics, the adjacency matrix including nodes and edges, the nodes are charging stations, and the edges are road connectivity between charging stations within the time slice;

[0009] Obtaining a node physical graph according to the adjacency matrix and the plurality of historical subsets;

[0010] Using a PageRank algorithm to obtain a PageRank value according to the node physical graph, the PageRank value is the PageRank value of a historical subset of charging stations within the time slice;

[0011] Based on the total duration, obtaining a time series matrix and a spatiotemporal information graph according to the PageRank value;

[0012] Based on the multi-head attention mechanism, a spatial attention matrix and a temporal attention matrix are respectively constructed according to the time series matrix;

[0013] Obtaining a spatial dynamic correlation matrix according to the spatial attention matrix and the adjacency matrix;

[0014] Obtaining a fusion graph according to the spatial dynamic correlation matrix and the spatiotemporal information graph;

[0015] Based on graph convolution, a spatial convolution network is obtained according to the fusion graph;

[0016] Based on causal convolution, a temporal convolutional network is obtained according to the temporal attention matrix;

[0017] Obtaining a prediction model according to the spatial convolutional network and the temporal convolutional network;

[0018] Electric vehicle charging prediction is performed according to the prediction model.

[0019] Furthermore, the historical data set includes at least power data.

[0020] Furthermore, the historical data set also includes at least one of electricity price data, road network data, weather data, travel information data or charging facility data.

[0021] Furthermore, the formula for obtaining the PageRank value according to the node physical graph using the PageRank algorithm is:

[0022]

[0023] In the formula, PR is the PageRank value of the current charging station, d is the damping factor, N is the total number of other charging stations connected to the current charging station, n is the total number of charging stations, i is the serial number of the charging station connected to the current charging station, and j is the number of charging stations connected to T. i The serial number of the connected charging station excluding the current charging station, T i is the i-th charging station connected to the current charging station, PR(T i ) is T i The PageRank value, ω ij is the weight of the edge between the i-th charging station and the j-th charging station, ω ih is the weight of the edge between the i-th charging station and the h-th charging station, and h is the weight of the edge between the i-th charging station and the h-th charging station. i Serial number of the connected charging station.

[0024] Furthermore, the step of obtaining the fusion graph is:

[0025] Constructing a spatial attention matrix according to the time series matrix;

[0026] Obtaining a standardized spatial attention distribution matrix according to the spatial attention matrix;

[0027] Obtaining the spatial dynamic correlation matrix according to the spatial attention distribution matrix and the adjacency matrix;

[0028] The fusion graph is obtained according to the spatial dynamic correlation matrix and the spatiotemporal information graph.

[0029] Furthermore, the steps of obtaining the spatial convolutional network are:

[0030] Obtaining a Laplacian matrix according to the fusion graph;

[0031] Based on the Laplace matrix, a transformation matrix of the fusion graph is obtained by Fourier transform;

[0032] Using Chebyshev polynomials, with Chebyshev polynomials as the kernel, to obtain the eigenfunctions of the Laplace matrix according to the transformation matrix;

[0033] Substituting the eigenfunction into the convolution expression, the spatial convolutional network is obtained.

[0034] Furthermore, the spatial convolutional network is:

[0035]

[0036] In the formula, g is the filter function for L, x 1 is the transformation matrix, k 1 is the size of the convolution kernel in the spatial convolutional network, K 1 is the order of the Chebyshev polynomial, is the Chebyshev coefficient vector, k 1 is the Chebyshev polynomial of order, and L is the Laplace matrix of the fusion graph.

[0037] Furthermore, the steps of obtaining the temporal convolutional network are:

[0038] Constructing a temporal attention matrix according to the time series matrix;

[0039] Obtain a standardized time attention distribution matrix according to the time attention matrix;

[0040] The temporal convolutional network is constructed according to the temporal attention distribution matrix.

[0041] Furthermore, the temporal convolutional network is:

[0042]

[0043] In the formula, y is the output of the temporal convolutional network, t is the time, and k is 2 is the serial number of the convolution kernel in the temporal convolutional network, K 2 is the total number of convolution kernels in the temporal convolutional network, w(k 2 ) is the kth 2 The weight of the convolution kernel, x 2is the temporal attention distribution matrix, and d' is the expansion rate.

[0044] A model for implementing the electric vehicle charging prediction method integrating spatiotemporal node information, comprising a data acquisition unit, a PageRank value calculation unit, a feature extraction unit and a time convolution unit;

[0045] The data acquisition unit:

[0046] Used to obtain the total duration and the time slice;

[0047] Used to obtain the traffic location characteristics of the charging station and the historical data set of the charging station within the total duration;

[0048] The PageRank value calculation unit:

[0049] Used to slice the historical data set according to the time slice to obtain a plurality of historical sub-data sets;

[0050] for constructing an adjacency matrix based on the time slice and the traffic location characteristics;

[0051] Used to obtain a node physical graph according to the adjacency matrix and the several historical subsets;

[0052] Used to obtain a PageRank value according to the node physical graph using a PageRank algorithm, wherein the PageRank value is the PageRank value of a historical subset of charging stations within the time slice;

[0053] Used to obtain a time series matrix and a spatiotemporal information graph based on the total duration and the PageRank value;

[0054] The feature extraction unit:

[0055] Used to construct a spatial attention matrix and a temporal attention matrix respectively according to the time series matrix based on a multi-head attention mechanism;

[0056] Used to obtain a spatial dynamic correlation matrix according to the spatial attention matrix and the adjacency matrix;

[0057] Used to obtain a fusion graph according to the spatial dynamic correlation matrix and the spatiotemporal information graph;

[0058] The temporal convolution unit:

[0059] Used to obtain a spatial convolution network according to the fusion graph based on graph convolution;

[0060] Used to obtain a temporal convolutional network based on the causal convolution and the temporal attention matrix.

[0061] Compared with the prior art, the present invention has the following beneficial effects:

[0062] The electric vehicle charging prediction method provided by the present invention integrates the time-space node information, generates a time-space information graph considering the importance of nodes in combination with the load information, learns the potential time-space dynamic characteristics, and analyzes the charging load according to the system integrating the time-space characteristics, establishes a prediction model that can capture long-term dependencies and fully consider the time-space coupling effect, and improves the accuracy of the prediction;

[0063] The present invention is suitable for alleviating the pressure of charging congestion. Compared with traditional methods, the prediction method has achieved a significant improvement in prediction accuracy, which not only verifies the effectiveness of the proposed method in mining complex spatiotemporal characteristics, but also provides a new research perspective and technical path for electric vehicle charging load prediction.

[0064] The electric vehicle charging prediction model provided by the present invention integrates spatiotemporal node information, which will deepen the understanding of the influence of spatial factors on load prediction accuracy, provide a more accurate and reliable solution for the practical application of electric vehicle charging load prediction technology, and promote research in this field to a higher level. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 It is a schematic diagram of the principle of the electric vehicle charging prediction method integrating spatiotemporal node information of the present invention;

[0066] Figure 2 This is a schematic diagram of the framework of the electric vehicle charging prediction model that integrates spatiotemporal node information of the present invention. DETAILED DESCRIPTION

[0067] The disadvantage of most attention-based methods is that they share all parameters of location and time interval, i.e., the parameters are used for both spatial and temporal features, so the node correlation depends only on their respective features. However, the flow correlation of different nodes at different times is different, and the degree of correlation between the historical data of various charging stations is also different. Therefore, considering various types of spatial correlation, the spatial PageRank node weight value affected by load is used for spatial feature learning to fully explore the spatiotemporal characteristics.

[0068] In order to clearly illustrate the technical features of this solution, this solution is described below through a specific implementation method.

[0069] See also Figure 1 The embodiment of the present invention provides an electric vehicle charging prediction method integrating spatiotemporal node information, comprising the following steps:

[0070] The total duration and time slices are obtained, and the time slices are obtained by dividing the total duration into segments according to the preset duration; the traffic location characteristics of the charging station and the historical data set of the charging station within the total duration are obtained; the historical data set at least includes power data; the historical data set also includes at least one of electricity price data, road network data, weather data, travel information data or charging facility data.

[0071] Slice the historical data set according to the time slice to obtain several historical sub-data sets;

[0072] Based on the time slice, an adjacency matrix is ​​constructed according to the traffic location characteristics. The adjacency matrix includes nodes and edges. The nodes are charging stations, and the edges are the road connectivity between charging stations within the time slice.

[0073] Obtaining a node physical graph based on an adjacency matrix and several historical subsets;

[0074] The PageRank algorithm is used to obtain the PageRank value according to the node physical graph. The PageRank value is the PageRank value of the historical subset of the charging station in the time slice.

[0075] The PageRank algorithm is an algorithm used to rank web pages. Its purpose is to evaluate the importance of each web page by calculating the link relationship between them. The general calculation formula of the PageRank algorithm can be expressed as Formula 1:

[0076]

[0077] Where PR is the PageRank value of the requested web page, d is the damping factor, i is the serial number of other web pages connected to the current web page, n is the total number of other web pages connected to the current web page, PR(T i ) is T i The PageRank value of C(T i ) is T i The PageRank algorithm can effectively identify the most influential nodes in the network, so this study introduces it into the identification of key nodes in charging stations, and improves Formula 1 to obtain Formula 2. Formula 2 is the formula for obtaining the PageRank value based on the node physical graph using the PageRank algorithm. Formula 2 is:

[0078]

[0079] In the formula, PR is the PageRank value of the current charging station, d is the damping factor, N is the total number of other charging stations connected to the current charging station, n is the total number of charging stations, i is the serial number of the charging station connected to the current charging station, and j is the number of charging stations connected to T. i The serial number of the connected charging station excluding the current charging station, T iis the i-th charging station connected to the current charging station, PR(T i ) is T i The PageRank value, ω ij is the weight of the edge between the i-th charging station and the j-th charging station, ω ih is the weight of the edge between the i-th charging station and the h-th charging station, and h is the weight of the edge between the i-th charging station and the h-th charging station. i The serial number of the connected charging station;

[0080] The PageRank algorithm takes into account the neglected spatial features and effectively integrates the spatiotemporal features. It can more comprehensively evaluate the relative importance of charging stations under different spatiotemporal conditions and improve the prediction accuracy.

[0081] Based on the total duration, the time series matrix and spatiotemporal information graph are obtained according to the PageRank value; see Figure 2 , the time series matrix is Figure 1 In the feature matrix, the spatiotemporal information graph is formed in the graph convolution layer;

[0082] By calculating the PageRank value, the neglected spatial features are taken into account, and the spatiotemporal features are effectively integrated, which can more comprehensively evaluate the relative importance of charging stations under different spatiotemporal conditions and improve the prediction accuracy;

[0083] See also Figure 2 In the attention layer, based on the multi-head attention mechanism, the spatial attention matrix and the temporal attention matrix are constructed according to the time series matrix respectively; the introduction of the multi-head attention mechanism enables the prediction model to identify which features should be given more weight in a specific scenario, thereby improving the prediction model's ability to integrate spatial information and temporal information.

[0084] Obtain the spatial dynamic correlation matrix based on the spatial attention matrix and the adjacency matrix;

[0085] The fusion map is obtained according to the spatial dynamic correlation matrix and the spatiotemporal information map; the steps to obtain the fusion map are:

[0086] Construct a spatial attention matrix based on the time series matrix. The spatial attention matrix is:

[0087] S=V s ·σ((χ h W 1 )W 2 (W 3 χ h ) Τ +b s );

[0088] Where S is the spatial attention matrix, χ h is the time series matrix, χ h=(X 1 ,X 2 ,…,X γ ), X γ is the PageRank value of the γth time slice in the time series matrix, γ is the total number of time slices in the total duration, and χ h ∈R N×F×γ , V s , W 1 , W 2 , W 3 and b s are all learnable parameters, V s ,b s ∈R N×N , W 1 ∈R T , W 2 ,W 3 ∈R F×γ , superscript Τ represents transposition, σ is the activation function, N is the total number of PageRank values ​​in the time slice, and F is the total number of charging stations;

[0089] According to the spatial attention matrix, the standardized spatial attention distribution matrix is ​​obtained. The spatial attention distribution is:

[0090]

[0091] In the formula, S i ' ,j is the spatial attention distribution, exp is the expected function, S i,j is the association strength between charging station i and charging station j;

[0092] The spatial dynamic correlation matrix is ​​obtained by multiplying the spatial attention distribution matrix and the adjacency matrix;

[0093] Obtain a fusion map based on the spatial dynamic correlation matrix and the spatiotemporal information map;

[0094] In the time domain convolution layer, based on the graph convolutional neural network, the spatial convolutional network is obtained according to the fusion graph. The steps to obtain the spatial convolutional network are:

[0095] The Laplace matrix is ​​obtained according to the fusion graph. The Laplace matrix is:

[0096] L = DA;

[0097] Where L is the Laplacian matrix of the fused graph, D is the degree matrix of the fused graph, and A is the adjacency matrix of the fused graph;

[0098] Based on the Laplace matrix, the Fourier transform is used to obtain the transformation matrix of the fusion graph:

[0099] GF(x1 )=U Τ x 1 ;

[0100] Where GF is the Fourier transform, U is the matrix in the Laplace transform, and L = U∧U Τ , ∧ is the eigenvector, x 1 is the transformation matrix;

[0101] To simplify the calculation, the Chebyshev polynomial is used. The Chebyshev polynomial is used as the kernel, and the eigenfunction of the Laplace matrix is ​​obtained according to the transformation matrix;

[0102]

[0103] In the formula, g θ is the eigenfunction of L, the subscript of T is the order of the Chebyshev polynomial, is (k 1 -1) order Chebyshev polynomial, is (k 1 -2) Chebyshev polynomial, T 0 is the 0th order Chebyshev polynomial, T 1 is the 1st order Chebyshev polynomial;

[0104] Substituting the eigenfunction into the convolution expression obtains the spatial convolutional network.

[0105] The spatial convolutional network is:

[0106]

[0107] In the formula, g is the filter function for L, k 1 is the size of the convolution kernel in the spatial convolutional network, K 1 is the order of the Chebyshev polynomial, is the Chebyshev coefficient vector, k 1 Chebyshev polynomial of order;

[0108] The Chebyshev graph convolution layer with spatial attention mechanism, namely the spatial convolution network, can dynamically adjust the adjacency of the fusion graph during the convolution operation by introducing the spatial attention matrix, so as to better capture the important information in the fusion graph structure. Compared with the direct graph convolution operation, it can enhance the prediction model's perception of important nodes and improve the effect of the convolution operation.

[0109] Obtain a temporal convolutional network based on the temporal attention matrix;

[0110] The steps to obtain a temporal convolutional network are:

[0111] Construct a temporal attention matrix based on the time series matrix;

[0112]

[0113] Where E is the temporal attention matrix, V e , U 1 , U 2 , U 3 and b e are all learnable parameters, V s ,b s ∈R γ×γ , U 1 ∈R N , W 2 ∈R F×N , W 3 ∈R F ;

[0114] According to the time attention matrix, the standardized time attention distribution matrix is ​​obtained, and the time attention distribution is:

[0115]

[0116] Where, E' α,β is the temporal attention distribution, E α,β is the correlation strength between time slice α and time slice β;

[0117] Also in the time domain convolution layer, based on causal convolution, a temporal convolution network is constructed according to the temporal attention distribution matrix; causal convolution is often used as temporal convolution because it only depends on current and past inputs and does not involve future inputs. In the actual process of capturing long-term dependencies, causal convolution requires more layers when spanning long time intervals, which increases computational costs and memory usage, and multi-layer causal convolution also increases training time and increases the complexity of the model. Therefore, in this study, we adopted dilated causal convolution, which introduces a dilation rate d to increase the receptive field so that the convolution kernel can capture features over a larger time range without increasing the size of the convolution kernel; the temporal convolution network is:

[0118]

[0119] In the formula, y is the output of the temporal convolutional network, t is the time, and k is 2 is the serial number of the convolution kernel in the temporal convolutional network, K 2 is the total number of convolution kernels in the temporal convolutional network, w(k 2 ) is the kth 2 The weight of the convolution kernel, x 2 is the time attention distribution matrix, d' is the expansion rate;

[0120] The temporal convolutional network built based on dilated causal convolution is able to capture dependencies over a long time range without significantly increasing the computational cost. In the processing of long time series forecasting tasks, dilated causal convolution can enhance the ability of the forecasting model to capture long-term dependencies of loads and improve the forecasting accuracy.

[0121] Obtaining a prediction model based on a spatial convolutional network and a temporal convolutional network;

[0122] Electric vehicle charging forecasting based on prediction model.

[0123] When constructing a prediction model, historical data is used for charging load; the prediction model can also capture input information when in use, including charging load (power data), electricity price data, road network data, weather data, travel information data or charging facility data, etc., to enhance the prediction model's ability to understand and capture long-term dependencies, effectively extract the long-term dependencies between electric vehicle charging and input information, and fully consider the spatiotemporal coupling effect by integrating time and space information. It can not only analyze the changing trend of charging load in different time periods, but also evaluate the impact of spatial factors on charging load, which helps the prediction model to more comprehensively understand the spatiotemporal distribution characteristics of electric vehicle charging load.

[0124] A model for realizing an electric vehicle charging prediction method integrating spatiotemporal node information, comprising a data acquisition unit, a PageRank value calculation unit, a feature extraction unit and a time convolution unit;

[0125] Data acquisition unit:

[0126] Used to obtain total duration and time slice;

[0127] Used to obtain the traffic location characteristics of the charging station and the historical data set of the charging station within the total duration;

[0128] PageRank value calculation unit:

[0129] Used to slice the historical data set according to time slices to obtain several historical sub-data sets;

[0130] Used to construct an adjacency matrix based on time slices and traffic location characteristics;

[0131] Used to obtain a node physical graph based on an adjacency matrix and several historical subsets;

[0132] Used to obtain the PageRank value according to the node physical graph using the PageRank algorithm, where the PageRank value is the PageRank value of the historical subset of charging stations within the time slice;

[0133] Used to obtain time series matrix and spatiotemporal information graph based on the total duration and PageRank value;

[0134] Feature extraction unit:

[0135] It is used to construct the spatial attention matrix and the temporal attention matrix according to the time series matrix based on the multi-head attention mechanism;

[0136] Used to obtain the spatial dynamic correlation matrix based on the spatial attention matrix and the adjacency matrix;

[0137] Used to obtain a fusion map based on the spatial dynamic correlation matrix and the spatiotemporal information map;

[0138] Temporal Convolution Unit:

[0139] Used to obtain a spatial convolutional network based on a fusion graph based on graph convolution;

[0140] Used to obtain a temporal convolutional network based on the temporal attention matrix based on causal convolution.

[0141] In order to further verify the superiority of the electric vehicle charging prediction that integrates spatiotemporal node information proposed in this disclosure, four algorithms, namely long short-term memory neural network (LSTM), GCN, TCN, and GCN-TCN, were used as control benchmarks for experiments to verify the applicability of the preset method and prediction model proposed in this disclosure. The mean absolute percentage error (MAPE), root mean square error (RMSE) and goodness of fit (coefficient of determination, R2) of the test results are shown in Table 1. It can be seen that the GCN algorithm is poor in stability and accuracy, and the MAPE of the prediction result is 9.15%, which is mainly due to the difficulty of GCN in capturing the changing laws of node features over time. In contrast, LSTM and TCN perform well in processing sequence data, but lack effective use of graph structure information. The electric vehicle charging prediction method that integrates spatiotemporal node information proposed in this disclosure can take into account both graph structure information and time series dynamic information, and adjust feature weights based on the attention mechanism. The test result MAPE is only 4.06%, which is 33.11% lower than the TCN algorithm that cannot process graph structure information. In summary, the example results verify the improvement effect of the proposed electric vehicle charging load prediction method that considers spatiotemporal coupling effects in processing graph data tasks with time series characteristics. Compared with single models such as LSTM and TCN, it shows more superior training performance and prediction accuracy. The prediction method proposed in this disclosure can take into account both graph structure information and time series dynamic information, and further adjust feature weights based on the attention mechanism. The MAPE of the prediction result is 39.49% lower than that of the LSTM algorithm, with higher prediction accuracy.

[0142] In view of the limitations of existing methods in the field of electric vehicle charging load forecasting in mining complex dynamic spatiotemporal features, this study innovatively integrates the attention mechanism of charging station nodes with spatiotemporal importance evaluation to construct a comprehensive feature representation system with spatiotemporal correlation. This strategy greatly enhances the ability of the prediction model to extract spatiotemporal feature information from the data. Subsequently, we input these features into the graph convolutional neural network (GCN) to capture the spatiotemporal coupling relationship and deeply analyze the spatial interaction effect between charging station nodes and the dynamic evolution of time series. This process not only improves the perception ability of the model, but also significantly improves the accuracy of the prediction results.

[0143] Table 1 Electric vehicle charging prediction results using different methods

[0144]

[0145] The experimental verification results show that compared with the traditional methods, the prediction method proposed in this paper has achieved a significant improvement in prediction accuracy. This not only verifies the effectiveness of the proposed prediction method in mining complex spatiotemporal characteristics, but also provides a new research perspective and technical path for electric vehicle charging load prediction. Looking to the future, the research will deepen the understanding of the impact of spatial factors on load forecasting accuracy, and continue to optimize the model architecture based on this disclosure, aiming to further improve the prediction accuracy. This will provide a more accurate and reliable solution for the practical application of electric vehicle charging load forecasting technology, while promoting research in this field to a higher level.

[0146] Finally, it should be noted that the above content is only used to illustrate the technical solution of the present invention, rather than to limit the scope of protection of the present invention. Simple modifications or equivalent substitutions of the technical solution of the present invention by ordinary technicians in this field do not deviate from the essence and scope of the technical solution of the present invention.

Claims

1. A method for predicting electric vehicle charging by integrating spatiotemporal node information, characterized in that: The following steps are involved: Obtaining the total duration and time slice; obtaining the traffic location characteristics of the charging station and the historical data set of the charging station within the total duration; Slicing the historical data set according to the time slice to obtain a plurality of historical sub-data sets; Based on the time slice, construct an adjacency matrix according to the traffic location characteristics, the adjacency matrix including nodes and edges, the nodes are charging stations, and the edges are road connectivity between charging stations within the time slice; Obtaining a node physical graph according to the adjacency matrix and the plurality of historical subsets; Using a PageRank algorithm to obtain a PageRank value according to the node physical graph, the PageRank value is the PageRank value of a historical subset of charging stations within the time slice; Based on the total duration, obtaining a time series matrix and a spatiotemporal information graph according to the PageRank value; Based on the multi-head attention mechanism, a spatial attention matrix and a temporal attention matrix are respectively constructed according to the time series matrix; Obtaining a spatial dynamic correlation matrix according to the spatial attention matrix and the adjacency matrix; Obtaining a fusion graph according to the spatial dynamic correlation matrix and the spatiotemporal information graph; Based on graph convolution, a spatial convolution network is obtained according to the fusion graph; Based on causal convolution, a temporal convolutional network is obtained according to the temporal attention matrix; Obtaining a prediction model according to the spatial convolutional network and the temporal convolutional network; The electric vehicle charging prediction is performed according to the prediction model.

2. The electric vehicle charging prediction method integrating spatiotemporal node information according to claim 1 is characterized in that: The historical data set includes at least power data.

3. The electric vehicle charging prediction method integrating spatiotemporal node information according to claim 2 is characterized in that: The historical data set also includes at least one of electricity price data, road network data, weather data, travel information data or charging facility data.

4. The electric vehicle charging prediction method integrating spatiotemporal node information according to claim 1 is characterized in that: The formula for obtaining the PageRank value according to the node physical graph using the PageRank algorithm is: In the formula, PR is the PageRank value of the current charging station, d is the damping factor, N is the total number of other charging stations connected to the current charging station, n is the total number of charging stations, i is the serial number of the charging station connected to the current charging station, and j is the number of charging stations connected to T. i The serial number of the connected charging station excluding the current charging station, T i is the i-th charging station connected to the current charging station, PR(Ti) is T i The PageRank value, ω ij is the weight of the edge between the i-th charging station and the j-th charging station, ω ih is the weight of the edge between the i-th charging station and the h-th charging station, and h is the weight of the edge between the i-th charging station and the h-th charging station. i Serial number of the connected charging station.

5. The electric vehicle charging prediction method integrating spatiotemporal node information according to claim 4 is characterized in that: The steps of obtaining the fusion graph are: Constructing a spatial attention matrix according to the time series matrix; Obtaining a standardized spatial attention distribution matrix according to the spatial attention matrix; Obtaining the spatial dynamic correlation matrix according to the spatial attention distribution matrix and the adjacency matrix; The fusion graph is obtained according to the spatial dynamic correlation matrix and the spatiotemporal information graph.

6. The electric vehicle charging prediction method integrating spatiotemporal node information according to claim 5 is characterized in that: The steps of obtaining the spatial convolutional network are: Obtaining a Laplacian matrix according to the fusion graph; Based on the Laplace matrix, a transformation matrix of the fusion graph is obtained by Fourier transform; Using Chebyshev polynomials, with Chebyshev polynomials as the kernel, to obtain the eigenfunctions of the Laplace matrix according to the transformation matrix; Substituting the eigenfunction into the convolution expression, the spatial convolutional network is obtained.

7. The electric vehicle charging prediction method integrating spatiotemporal node information according to claim 6 is characterized in that: The spatial convolutional network is: In the formula, g is the filter function for L, x1 is the transformation matrix, k1 is the size of the convolution kernel in the spatial convolution network, K1 is the order of the Chebyshev polynomial, is the Chebyshev coefficient vector, is the k1-order Chebyshev polynomial, and L is the Laplace matrix.

8. The electric vehicle charging prediction method integrating spatiotemporal node information according to claim 6 is characterized in that: The steps of obtaining the temporal convolutional network are: Constructing a temporal attention matrix according to the time series matrix; Obtain a standardized time attention distribution matrix according to the time attention matrix; The temporal convolutional network is constructed according to the temporal attention distribution matrix.

9. The electric vehicle charging prediction method integrating spatiotemporal node information according to claim 8 is characterized in that: The temporal convolutional network is: Where y is the output of the temporal convolutional network, t is time, k2 is the sequence number of the convolution kernel in the temporal convolutional network, K2 is the total number of convolution kernels in the temporal convolutional network, w(k2) is the weight of the k2th convolution kernel, x2 is the temporal attention distribution matrix, and d' is the expansion rate.

10. A model for implementing the electric vehicle charging prediction method integrating spatiotemporal node information as described in any one of claims 1 to 9, characterized in that: It includes a data acquisition unit, a PageRank value calculation unit, a feature extraction unit and a time convolution unit; The data acquisition unit: Used to obtain the total duration and the time slice; Used to obtain the traffic location characteristics of the charging station and the historical data set of the charging station within the total duration; The PageRank value calculation unit: Used to slice the historical data set according to the time slice to obtain a plurality of historical sub-data sets; for constructing an adjacency matrix based on the time slice and the traffic location characteristics; Used to obtain a node physical graph according to the adjacency matrix and the several historical subsets; Used to obtain a PageRank value according to the node physical graph using a PageRank algorithm, wherein the PageRank value is the PageRank value of a historical subset of charging stations within the time slice; Used to obtain a time series matrix and a spatiotemporal information graph based on the total duration and the PageRank value; The feature extraction unit: Used to construct a spatial attention matrix and a temporal attention matrix respectively according to the time series matrix based on a multi-head attention mechanism; Used to obtain a spatial dynamic correlation matrix according to the spatial attention matrix and the adjacency matrix; Used to obtain a fusion graph according to the spatial dynamic correlation matrix and the spatiotemporal information graph; The temporal convolution unit: Used to obtain a spatial convolution network according to the fusion graph based on graph convolution; Used to obtain a temporal convolutional network based on the causal convolution and the temporal attention matrix.

Images