User group charging load baseline quantification method and system fusing multi-source data

Abstract

The invention discloses a user group charging load baseline quantification method and system fusing multi-source data, and relates to the field of electric vehicle charging, and the method comprises the following steps: S1, building a graph structure based on load similarity; s2, external feature space correction based on GAT; s3, time sequence load prediction based on TimeXer is carried out; and S4, result comparison and evaluation. In order to solve the problem that an existing model is insufficient in precision, the invention aims to improve the precision and robustness of a load prediction model by effectively fusing multi-source data and carrying out space-time mining and correction.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of electric vehicle charging, and particularly relates to a user group charging load baseline quantification method and system fusing multi-source data. BACKGROUND

[0002] At present, electric vehicle charging load prediction methods are mainly divided into physical simulation and data-driven methods.

[0003] (1) Data-driven method: This method is the mainstream, which inputs historical data into the model to output prediction results. Early methods include traditional statistical models (such as ARIMA), but they are difficult to fit nonlinear charging load data. Traditional machine learning models (such as SVM, random forest) have advantages in dealing with nonlinear problems, but it is difficult to capture deep temporal dependence.

[0004] (2) Deep learning method: With the development of artificial intelligence, deep learning models have obvious advantages. Recurrent neural networks (such as LSTM, GRU) can better capture long-term temporal relationships, but they are not efficient when dealing with ultra-long sequences. Models centered on Transformer have become a research hotspot and can effectively solve long-distance dependence problems, but many variants still have deficiencies in capturing complex relationships between multiple variables. SUMMARY

[0005] The purpose of the present application is to provide a user group charging load baseline quantification method and system fusing multi-source data to solve the problems raised in the background.

[0006] To achieve the above purpose, the present application provides the following technical solutions: A user group charging load baseline quantification method fusing multi-source data, comprising the following steps: S1. Data preprocessing is performed on the electric vehicle charging transaction data set, and user-level data is integrated into hourly charging sessions; feature extraction is performed on the hourly charging sessions, charging load is extracted as an endogenous variable, and weather, date, and price features are extracted as original exogenous variables; the Pearson correlation coefficient of each charging station historical load sequence is calculated, and an adjacency matrix is constructed therefrom; S2. The original exogenous variables and the adjacency matrix are jointly input into a GAT model, the original exogenous variables are reconstructed, and the reconstructed exogenous variables are obtained; S3. The endogenous variables and the reconstructed exogenous variables are jointly input into a TimeXer model for training, the TimeXer model can capture the temporal dependence of the charging load itself in the endogenous variables and the influence between the reconstructed exogenous variables and the endogenous variables, and finally generate load prediction values of different time lengths in the future.

[0007] As a further aspect of the present invention, the method for preprocessing the electric vehicle charging transaction dataset in step S1 includes: removing charging sessions that failed to complete the transaction or had abnormal transaction volume due to system failure; and using accumulation technology to aggregate the data, summarizing the second-level transaction data into multiple charging station charging load time series with hourly intervals.

[0008] As a further aspect of the present invention, the method for constructing the adjacency matrix in step S1 specifically involves: using the load time series data of each charging station, and employing the Pearson correlation coefficient to measure the adjacency between any two charging stations. i , j Hourly average power sequence P i , P j The linear correlation between them, Pearson correlation coefficient r ij The calculation formula is as follows: (1) in, P i,t , P j,t These are charging stations i and j At the point of time t average power, , It is the average power of each over the entire time period; T It is the total time step, Pearson correlation coefficient r ij The value range is [-1, 1]; by performing Pearson correlation coefficient analysis on the load data of all charging stations, a charging station adjacency matrix is ​​constructed using the load similarity between charging stations, and a load correlation heatmap of each charging station is generated; the Pearson correlation coefficient... r ij The closer the value is to 1 or -1, the stronger the positive or negative correlation between the load curves of the two charging stations; a value close to 0 indicates that the correlation is not obvious.

[0009] As a further aspect of the present invention, the method for reconstructing the spatial dimension of the adjacency matrix data in step S2 specifically involves: reconstructing the external feature space based on the GAT model, wherein the GAT model includes a graph attention layer, and the input of the graph attention layer is a set of node features. h ={ h 1, h 2, ..., h N},in, It is a node i ofF 3D feature vectors N Given the number of nodes, the graph attention layer outputs a new set of node features. h ={ h ʹ 1, h ʹ 2, ..., h ʹ N},in, It is a node i After transformation F ʹ 3D feature vector; First, from the weight matrix A parameterized shared linear transformation is applied to each node to enhance feature representation. Then, the GAT model transforms the node pairs ( i , j Calculate the attention coefficient e ij , used to represent nodes j Features of nodes i Importance, attention coefficient e ij The calculation method is as follows: (2) in, The attention mechanism is a single-layer feedforward neural network consisting of a weight vector. Parameterization was performed, and LeakyReLU was used as the activation function. The attention coefficients of each node were normalized using the Softmax function. The normalized attention coefficients were then... α ij for: (3) Where, N i It is a node i In the first-order neighbor set of the graph, the model employs a masked attention mechanism, only when a node... j It is a node i The attention coefficient is only calculated when the neighbor is involved. e ij For non-neighbor nodes, their attention coefficients are replaced with a very small negative number to ensure that their weights approach zero in the Softmax calculation; finally, the nodes... i New features The weights are obtained by weighted summation of the features of its neighboring nodes; these weights are the normalized attention coefficients. α ij And use ELU as the activation function: (4).

[0010] As a further aspect of the present invention, the GAT model has two layers: the first layer consists of... K The attention heads are composed of multiple attention heads in parallel. Each head receives the raw features and processes them... K The output features of each element are concatenated to form a wider feature representation; the concatenated features of the first layer output are... As shown below: (5) Where || represents the concatenation operation; The second layer receives the features concatenated from the first layer. As input, the final feature is output by averaging the results from multiple sources. As shown below: (6).

[0011] As a further aspect of the present invention, step S3 specifically involves time-series load forecasting based on TimeXer, and the embedding strategy of the TimeXer model is as follows: Endogenous sequences X en To predict the target column, satisfying For embedding endogenous sequences, TimeXer employs a block embedding strategy, dividing long sequences into... N = T / P Non-overlapping blocks { s 1, s 2, ..., s N Each block contains P At each point in time, each block is mapped to a separate linear layer. D By using a time-series token and overlaying it with positional encoding, the model introduces a learnable global token. G en Endogenous sequences aim to capture macroscopic information about the entire sequence and serve as a bridge for interaction with exogenous variables. X en Ultimately represented as N +1 token H en : (7) Among them, Emb( s i =Linear( s i )+PE( i ); Exogenous sequences Z It is by CComposed of external features, satisfying Z ={ z (1) , ..., z (C)},in, Each exogenous sequence is mapped as a whole through an independent embedding layer. D Dimensional variable tokens, C exogenous sequence Z Represented as C Variable tokens V ex : (8).

[0012] As a further aspect of this invention, the attention mechanism employed by the TimeXer model is as follows: Endogenous self-attention: In each block of the encoder, for all endogenous tokens H en Perform standard multi-head self-attention operations to obtain the processed tokens. H en satisfy: (9) Exogenous-Endogenous Cross-Attention: A cross-attention layer is introduced after the self-attention layer, in which the global tokens of the endogenous sequence are... G en from H en The variable tokens of all exogenous sequences are separated from ʹ and used as the query. V ex The key and value are input into the cross-attention module, resulting in an updated global token. G en ʹ: (10) Updated global token G en Replace ʹ H en The old global tokens in ʹ are then sent to the feedforward network layer together, thereby passing the aggregated external information to all endogenous tokens; go through L After processing by the layer encoder, the model will finally output all the endogenous tokens. Flatten the block dimension and generate the future through a linear prediction head. S Predicted values ​​at each time step : (11).

[0013] As a further aspect of the present invention, the historical load is the hourly charging station load data for the past 7 days.

[0014] As a further aspect of the present invention, step S4, model testing and effect evaluation, is also included: We selected the historical mean method and compared its performance with five other models—Informer, Transformer, PatchTST, LSTM, and Bi-GRU—and the TimeXer model to evaluate their performance. We chose four indicators as the standard for measuring the prediction results. We conducted ablation experiments on three types of charging stations, calculated the percentage decrease in indicators, and verified the effectiveness of the GAT feature reconstruction module and multi-source data. The four indicators are specifically the Mean Absolute Error (MAE), Mean Square Error (MSE), Root Mean Square Error (RMSE), and Coefficient of Determination (R²). 2 The calculation formula is as follows: (12) (13) (14) (15) in, n It is the sample size. y i It is the first i The true value of each sample That is the corresponding predicted value. It is the average of all true values; MAE, MSE, and RMSE reflect the degree to which predicted values ​​deviate from actual values. 2 It measures the extent to which a model explains changes in predicted values, and its value ranges between (−∞, 1]. When R0... 2 When R = 0, it means the model's performance is at the average level; 2 The closer the value is to 1, the better the model's predictive performance.

[0015] Furthermore, this invention also provides a user group charging load baseline quantification system that integrates multi-source data, including a data preprocessing module, a GAT spatial feature reconstruction module, a TimeXer time series prediction module, and an effect evaluation module. The data preprocessing module integrates hourly charging sessions and quantifies the functional similarity between charging stations by calculating the Pearson correlation coefficient of their historical load sequences, thereby constructing an adjacency matrix to capture spatial correlations between stations without precise geographical location. The GAT spatial feature reconstruction module is responsible for reconstructing the spatial dimensions of the original multi-source external data input to the data preprocessing module, treating each charging station as a node on a graph. By learning the attention weights between stations, it transforms macroscopic information that is the same for all stations into differentiated features that reflect the unique spatial relationships of each station, while simultaneously reducing the dimensionality of the vector features. The TimeXer time series prediction module receives the historical load of the target station and the reconstructed features processed by the GAT model, while capturing the time dependence of the historical load itself and the influence of external factors, ultimately generating future load prediction values.

[0016] Compared with the prior art, the beneficial effects of the present invention are: (1) Improve prediction accuracy and robustness: In view of the problem of insufficient accuracy of existing models, this invention aims to improve the accuracy and robustness of load prediction models by effectively integrating multi-source data and performing spatiotemporal mining and correction.

[0017] (2) Realize spatial modeling without geographic information: In view of the shortcomings of existing GNN methods that are highly dependent on geographic information, this invention aims to propose a spatial relationship modeling method based on load function similarity, so as to capture the spatial coupling relationship between stations without obtaining precise location information.

[0018] (3) Achieving differentiated correction of external information: In response to the problem of homogenization of external information, this invention aims to use the GAT model to transform macro-level external information into characteristics that reflect the differences in the characteristics of each site, thereby enhancing the effectiveness of the information. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the data input and framework of the GAT-TimeXer model in this embodiment of the invention. Figure 2 This is a heatmap showing the load correlation of 13 charging stations in an embodiment of the present invention; Figure 3 This is a schematic diagram of the multi-head attention mechanism of the GAT model in an embodiment of the present invention; Figure 4a This is the t-SNE dimensionality reduction diagram before feature correction in this embodiment of the invention; Figure 4bThis is the violin diagram before feature correction in an embodiment of the present invention; Figure 5a This is the t-SNE dimensionality reduction diagram after feature correction in this embodiment of the invention; Figure 5b This is a violin diagram with feature corrections in an embodiment of the present invention; Figure 6 This is a single-layer TimeXer model structure in an embodiment of the present invention; Figure 7 This is a graph showing the 24-hour prediction results for different charging stations in different seasons in an embodiment of the present invention; Figure 8 This is a comparison chart of 72 prediction results from a multi-benchmark model for different charging stations in this embodiment of the invention; Figure 9 This is a graph showing the change of the R2 index in the ablation experiment in an embodiment of the present invention. Detailed Implementation

[0020] 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.

[0021] like Figure 1 As shown, the method embodiment of the present invention mainly includes the following steps: S1. Construction of a graph structure based on load similarity; S2. External feature space correction based on GAT; S3. Time-series load forecasting based on TimeXer; S4. Results Comparison and Evaluation.

[0022] Step S1: Construction of a graph structure based on load similarity The dataset used in this invention originates from the high-resolution electric vehicle charging transaction dataset of Jiaxing City, China. This dataset records 441,077 real charging transaction records from 13 different types of charging stations (including bus stops, highway service areas, industrial parks, shopping malls, parks, etc.) within the three administrative districts of Nanhu, Tongxiang, and Xiuzhou in Jiaxing City from December 31, 2019 to December 31, 2021. The data collection time resolution is accurate to the second. The original dataset includes multi-dimensional features such as transaction information, status information, and external information. Transaction information includes user ID, charging pile ID, order creation / start / end / payment time, transaction electricity (kWh), electricity fee (yuan), and service fee (yuan). Status information includes the reason for charging termination, such as user-initiated stopping, full charge, system failure, etc. External information includes weather data such as temperature, relative humidity, and precipitation, as well as local time-of-use electricity prices. This paper preprocesses the dataset to remove charging sessions that failed to complete or had abnormal transaction volumes due to system failures or other reasons. The paper also uses an accumulation technique to aggregate the data, summarizing the second-level transaction data into a charging load time series of 13 charging stations at 1-hour intervals.

[0023] In this invention, the computer hardware configuration used for the example analysis is: an Intel Xeon Platinum 8352V CPU and an NVIDIA GeForce RTX 4090 GPU. The software environment is: Python 3.11 programming language, PyTorch 2.6 deep learning framework, and the training process utilizes the GPU for CUDA (12.6) parallel computing to accelerate training.

[0024] In the absence of precise geographic latitude and longitude location information, this invention employs a functional similarity-based method to construct an adjacency matrix of charging stations for use as input to a graph attention network. The core idea is that charging stations with similar charging behavior patterns have stronger functional or geographical correlations. To quantify the similarity between different charging stations, this paper utilizes the load time-series data of each charging station and uses the Pearson correlation coefficient to measure the similarity between any two charging stations. i , j Hourly average power sequence P i , P j The linear correlation between them. Pearson correlation coefficient. r ij The calculation formula is as follows: (1) in, P i,t , P j,t These are charging stations i andj At the point of time t average power, , It is the average power of each over the entire time period; T This is the total time step. The value of this coefficient ranges from [-1, 1]. The closer the value is to 1 or -1, the stronger the positive or negative correlation between the load curves of the two charging stations; a value close to 0 indicates that the correlation is not obvious.

[0025] By performing Pearson correlation coefficient analysis on the load data of all charging stations, an adjacency matrix was constructed using the load similarity between charging stations, generating a heatmap of load correlation for 13 charging stations. As shown in Figure 4, the correlation between most charging stations is weak. According to Evans (1996) in statistics, a correlation coefficient with an absolute value greater than 0.6 can be considered a strong correlation. As shown in the figure, only the bus station, shopping center, and park B charging stations have correlation coefficients exceeding 0.6, and their coefficients were set to 1 when constructing the adjacency matrix; the coefficients between the remaining charging stations were set to 0. Notably, by examining the region name field in the dataset, it was found that all three charging stations are located in Tongxiang City, indicating that load similarity can serve as an effective proxy for geographic clustering to a large extent. By analyzing users' charging behavior patterns, a graph structure reflecting real-world spatial relationships can still be constructed.

[0026] Step S2: External Feature Space Correction Based on GAT The Graph Attention Layer (GAL) is the core of the GAT model. The input to this layer is a set of node features. h ={ h 1, h 2, ..., h N},in, It is a node i of F 3D feature vectors N The number of nodes. This layer outputs a new set of node features. h ={ h ʹ 1, h ʹ 2, ..., h ʹ N},in, It is a node i After transformation F ʹ 3D eigenvectors.

[0027] First, from the weight matrix A parameterized shared linear transformation is applied to each node to enhance feature representation. Then, the model is used for node pairs ( i , j Calculate the attention coefficient e ij , used to represent nodes j Features of nodes i Importance, attention coefficient e ij The calculation method is as follows: (2) in, The attention mechanism is a single-layer feedforward neural network consisting of a weight vector. Parameterization is performed, and LeakyReLU is used as the activation function. The attention coefficients of each node are normalized using the Softmax function. The normalized attention coefficients are... α ij for: (3) Where, N i It is a node i The first-order neighbor set in the graph. It is worth noting that this paper employs a masked attention mechanism, meaning that attention is only applied when a node... j It is a node i The attention coefficient is calculated only when the neighbor is a given element (the corresponding element in the adjacency matrix is ​​greater than 0). e ij For non-neighbor nodes, their attention coefficients are replaced with a very small negative number to ensure that their weights approach zero in the Softmax calculation. Finally, the nodes... i New features The weights are obtained by weighted summation of the features of its neighboring nodes; these weights are the normalized attention coefficients. α ij And use ELU as the activation function: (4) The GAT model is characterized by the introduction of a multi-head attention mechanism, allowing the model to execute in parallel. K Each attention mechanism is used, and then the feature vectors obtained from each are concatenated and averaged, such as... Figure 2 As shown. The GAT model in this paper has two layers: the first layer consists of... K The attention heads are composed of multiple heads in parallel. Each head receives the raw features and processes them... K The output features of each element are concatenated to form a wider feature representation; the concatenated features of the first layer output are... As shown below: (5) Here, || represents the concatenation operation. The second layer receives the features concatenated by the first layer. As input, the final feature is output by averaging the results from multiple sources. As shown below: (6) In the prediction task of this invention, exogenous variables are important inputs to the TimeXer model. However, the original exogenous variables, such as temperature, relative humidity, rainfall, and electricity price, have almost identical observed values ​​for multiple neighboring charging stations, failing to reflect the differences between different charging stations; moreover, the large number of features and the strong correlation between individual features can also affect the prediction model's learning of features. To address this issue, the proposed GAT model is used to correct the exogenous variables. The GAT model treats each charging station as a graph node and quantifies and aggregates the influence of neighboring charging stations on the central charging station by learning the attention weights between nodes. This invention constructs a two-layer graph attention network model based on the PyTorch framework, with normalized 9-dimensional original exogenous variables as input and normalized 4-dimensional corrected exogenous variables as output. The specific parameter configuration of this model is shown in Table 1.

[0028] This paper evaluates the effectiveness of the GAT model in feature correction tasks from two dimensions: the t-SNE dimensionality reduction graph reflecting the spatial distribution of exogenous variables and the violin graph reflecting the statistical distribution of features. Since this paper uses a masked attention mechanism, the attention coefficient is calculated only when the adjacency matrix between adjacent nodes is greater than 0. For charging stations other than the three charging stations (bus station, shopping mall, and park B), the correlation coefficients between them are low, and the spatial coupling is almost negligible. To simplify the analysis, this paper only presents the results for these three highly correlated charging stations. The feature visualization before GAT model correction is shown in Figure 4, and the feature visualization after correction is shown in Figure 5. As shown in Figures 4(a) and 4(b), the data points of the three charging stations in the t-SNE dimensionality reduction graph almost overlap, and the distribution of the three charging stations under the same feature in the violin graph shows no significant difference. This indicates that the original features cannot effectively reflect the behavioral patterns of different stations. After GAT model correction, as shown in Figures 5(a) and 5(b), the three colored data points in the t-SNE dimensionality reduction graph cluster towards their respective regions, forming clusters with relatively clear boundaries, and the distribution of each feature of the three charging stations in the violin graph is different. This indicates that the GAT model enhances the ability of features to distinguish between different sites by incorporating the influence of neighboring sites. Simultaneously, the number of features is reduced from 9 to 4, decreasing the complexity of exogenous variables.

[0029] Step S3: Time-series load forecasting based on TimeXer TimeXer is an enhanced Transformer architecture designed for time series forecasting tasks with exogenous variables. Its core lies in introducing an embedding strategy and a dual attention mechanism to effectively distinguish and fuse endogenous and exogenous sequence information. A single-layer TimeXer model structure is as follows: Figure 3 As shown.

[0030] (1) Embedding strategy Endogenous sequences X en To predict the target column, satisfying For embedding endogenous sequences, TimeXer employs a block embedding strategy. It divides long sequences into... N = T / P Non-overlapping patches { s 1, s 2, ..., s N Each block contains P Each time point. Each block is mapped to a separate linear layer. D The model uses a time-series token and overlays positional encoding. Simultaneously, to capture macroscopic information about the entire sequence and serve as a bridge for interaction with exogenous variables, a learnable global token is introduced. G en Therefore, endogenous sequences X en Ultimately represented as N +1 token H en : (7) Among them, Emb( s i =Linear( s i )+PE( i ).

[0031] Exogenous sequences Z It is by C Composed of external features, satisfying Z ={ z (1) , ..., z (C)},in, TimeXer employs a variable-level embedding strategy. Each exogenous sequence is mapped as a whole through an independent embedding layer. D Dimensional variable tokens. Thus, C exogenous sequence Z It is represented as CVariable tokens V ex : (8) (2) Attention mechanism Endogenous self-attention: In each block of the encoder, for all endogenous tokens H en Perform standard multi-head self-attention operations. This mechanism aims to capture complex temporal dependencies within endogenous sequences and obtain processed tokens. H en satisfy: (9) Exogenous-Endogenous Cross-Attention: To effectively incorporate external information into the prediction process, the model introduces a cross-attention layer after the self-attention layer. In this layer, the global tokens of the endogenous sequence... G en from H en The ʹ is separated as a query, while the variable tokens of all exogenous sequences are... V ex The updated global token is input as both key and value to the cross-attention module. G en ʹ: (10) In this way, the global token can selectively aggregate the most relevant information from all external variables. Updated global token. G en ʹ will be replaced H en The old global tokens in ʹ are then fed into the feedforward network layer, thereby passing the aggregated external information to all endogenous tokens.

[0032] go through L After processing by the layer encoder, the model will finally output all the endogenous tokens. Flatten the block dimension and generate the future through a linear prediction head. S Predicted values ​​at each time step : (11) For the selection of endogenous variables, the endogenous variables in this invention are the historical values ​​of the prediction object: hourly charging station load data (168 points) over the past 7 days. The charging station load basically shows a regular daily change, but the load on weekdays and weekends is also different. Selecting 7 days allows the model to learn the changes in load over the past week, thereby making more accurate predictions.

[0033] For the selection of exogenous variables, this invention considers three factors: time characteristics, weather characteristics, and electricity price characteristics. Regarding time characteristics, date types such as holidays, number of hours in a day, and weekday all affect the charging behavior of electric vehicle users. Regarding weather characteristics, temperature and humidity affect the power consumption of electric vehicles, and rainfall also affects the frequency of user charging. Regarding electricity price characteristics, it affects the time shift of load charging. All of the above characteristics can influence the electric vehicle charging load; therefore, nine types of data (168 points) including hourly time, weather, and electricity price over the past seven days were selected as exogenous variables. The time length of each characteristic was aligned with the endogenous variables. For example, daily characteristics such as weekday were copied and extended to 24 hours in a day. The specific multi-source feature fields are shown in Table 2.

[0034] In the prediction task of this invention, the TimeXer model input consists of endogenous variables and corrected exogenous variables at the hourly level over the past 7 days. The endogenous variables are the historical power values ​​of the charging stations, and the corrected exogenous variables are the output of the GAT model. The TimeXer model output is the predicted value of the endogenous sequence over the next S time steps. In the prediction scenario of this invention, the prediction step size is S = 24, 48, 72, and 96. The specific parameter configuration of this model is shown in Table 3.

[0035] like Figure 7 As shown in the figure, the load forecast results for bus stations, shopping centers and Park B charging station are presented in summer, autumn and winter. Input the hourly endogenous and exogenous variable data for the 7 days before the forecast date, and output the load data for the forecast date (24:00). The input in the figure only shows the endogenous variables, i.e. the historical load data. Figure 7 This indicates that the TimeXer model performs excellently in capturing load peaks and valleys, and can predict single-peak, double-peak, and multi-peak curves that are close to the actual values. It can also accurately predict the timing of daily charging peaks and valleys and capture the daily periodicity of charging load well.

[0036] like Figure 7 As shown in (a), (b), and (c), the load curves of bus station charging stations exhibit daily periodic changes, generally showing typical bi-peak or multi-peak loads. By comparing the load curves for the three seasons and the National Day holiday shown in autumn, it can be seen that the load power values ​​do not change significantly across the three seasons. This is because buses operate on a regular schedule with the same departure times each day, and their charging demand is largely unaffected by seasonality. Figure 7As shown in (d)(e)(f), the charging load at shopping mall charging stations exhibits a single-peak charging pattern in summer and multiple peaks in autumn and winter. During the National Day holiday, a significant power surge peak occurs, due to the concentrated time users spend at shopping malls during the holiday. Figure 7 As shown in (g)(h)(i), the daily load of charging station B in park B exhibits a multi-peak charging model, with small power differences between peaks, indicating that the number of people charging and the charging time are relatively fixed each day. It is noteworthy that the park charging station did not experience a power surge during holidays; instead, the power increased significantly during autumn and winter. This may be because users showed a stronger tendency to travel during the autumn and winter of 2021, leading to an increase in the park's charging power.

[0037] In addition to predicting with an output length of 24, the model's output length was changed to 48, 72, and 96 to evaluate its predictive performance under different output lengths. Specific metrics are shown in Table 4. It can be seen that the bus station and shopping mall charging stations have the lowest MSE, MAE, and RMSE metrics and the R² index is closest to 1 when the output length is 48. Park B charging station performs best when the output length is 24. When the prediction length is adjusted between 24 and 96, the changes in various metrics for each charging station are not significant, indicating that the model has good stability.

[0038] Step S4: Result Comparison and Evaluation To comprehensively and objectively evaluate the performance of prediction models, this paper selects four evaluation metrics widely used in prediction tasks: Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and Coefficient of Determination (R-Squared). 2 The calculation formula is as follows: (12) (13) (14) (15) in, n It is the sample size. y i It is the first i The true value of each sample That is the corresponding predicted value. It is the average of all true values.

[0039] MAE, MSE, and RMSE reflect the degree to which predicted values ​​deviate from actual values. Therefore, the closer these three indicators are to 0, the better the model's predictive performance. 2 This measures the model's ability to explain changes in predicted values, and its value ranges from (−∞, 1]. When R... 2 When R = 0, it means the model's performance is at the average level; 2 The closer the value is to 1, the better the model's predictive performance. It's worth noting that this study did not use percentage-based metrics such as Mean Absolute Percentage Error (MAPE). This is primarily because charging load data contains a large number of zero or near-zero values ​​during nighttime and other periods. When the true value... y i When the denominator is zero, the denominator in the MAPE formula is zero, which leads to meaningless calculations or extreme results, making it impossible to stably measure model performance.

[0040] Selecting a scenario with a prediction length of 72 hours, we compared the performance and evaluated metrics of six model methods—Informer, Transformer, PatchTST, LSTM, Bi-GRU, and historical mean method—with the TimeXer model. Specific prediction results can be found in [link to results]. Figure 8 Table 5. (For example...) Figure 8 As shown, predicted data from bus stops (14:00-26:00, November 23, 2021), shopping malls (22:00-18:00, August 15, 2021), and Park B (5:00-31:00, August 28, 2021) are displayed. It can be seen that the TimeXer model can capture the peaks and troughs of the curve and fit the actual values ​​well. MAE, MSE, RMSE, and R... 2 These four indicators serve as the standards for measuring the prediction results. As shown in Table 5, the MAE, MSE, and RMSE of the model used in this paper are closest to 0, and R... 2 The value closest to 1 provides the best fit. It's worth noting that the graph shows significant fluctuations in charging station power throughout the day, with many low power values ​​close to 0. Therefore, most models can only learn the general trend and cannot perfectly fit sudden, dramatic increases in specific loads. Hence, R0... 2 An indicator is considered acceptable if its fit on this dataset exceeds 0.6.

[0041] To verify the effectiveness of the proposed model method, this section conducts ablation experiments, comparing the prediction metrics MAE, MSE, RMSE, and R for three charging stations under three conditions: the proposed model, removal of the GAT feature correction module, and removal of multi-source data. 2 The changes can be observed by adjusting the forecast length to provide a more comprehensive view of the indicators. For example...Figure 9 As shown in the figure, the experiments conducted at three charging stations with different prediction lengths, including removing the GAT feature correction module and removing multi-source data, compared to the experimental metric R of the model presented in this paper. 2 The decrease rate shows that removing the GAT model or removing multi-source data will reduce R. 2 The indicators decreased, and Park B showed the most significant effect in the ablation experiment. This is because users going to the park to charge are greatly affected by factors such as weather and holidays in the multi-source data, so the model's prediction effect improved significantly after multi-source data correction. Buses, on the other hand, have fixed daily schedules, so their charging demand is largely unaffected by external factors, and therefore their effect in the ablation experiment was not significant. The remaining indicators are shown in Tables 6, 7, and 8. It can be seen that the model in this paper exhibits the lowest MAE, MSE, and RMSE values ​​across different charging stations and prediction lengths, with R... 2 The value is the largest, which indicates that the consideration of multi-source data and the GAT feature correction module in this paper have a significant effect.

[0042] Explanation of relevant terms: (1) Exogenous variables: These are external variables that affect the prediction target but are not determined by the model itself, such as weather, electricity price, date type, etc.

[0043] (2) Endogenous variable: refers to the target variable of the prediction task, whose value is determined by the system. In this paper, it specifically refers to the historical charging load of the site to be predicted.

[0044] (3) Graph Attention Network (GAT): A deep learning model for processing graph-structured data. It aggregates information by assigning different attention weights to neighboring nodes, thereby updating the feature representation of the nodes.

[0045] (4) TimeXer: An enhanced Transformer architecture designed for time series forecasting tasks with exogenous variables.

[0046] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0047] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for quantifying the baseline charging load of a user group by integrating multi-source data, characterized in that, The steps include: S1. Preprocessing the electric vehicle charging transaction dataset to integrate user-level data into hourly charging sessions; extracting features from the hourly charging sessions, extracting charging load as an endogenous variable, and extracting weather, date, and electricity price features as original exogenous variables; Calculate the Pearson correlation coefficient of the historical load sequence of each charging station and construct the adjacency matrix accordingly; S2. Input the original exogenous variables and the adjacency matrix into the GAT model to reconstruct the original exogenous variables and obtain the reconstructed exogenous variables; S3. Input the endogenous variables and the reconstructed exogenous variables into the TimeXer model for training. The TimeXer model can capture the time-series dependence of the charging load itself in the endogenous variables and the influence between the reconstructed exogenous variables and the endogenous variables, and finally generate load forecast values ​​for different time lengths in the future.

2. The user group charging load baseline quantification method based on multi-source data as described in claim 1, characterized in that, The method for preprocessing the electric vehicle charging transaction dataset in step S1 includes: removing charging sessions that failed to complete due to system failure or had abnormal transaction volume; and using accumulation technology to aggregate the data, summarizing the second-level transaction data into multiple charging station charging load time series with hourly intervals.

3. The user group charging load baseline quantification method based on multi-source data as described in claim 1, characterized in that, The method for constructing the adjacency matrix in step S1 is as follows: using the load time series data of each charging station, the Pearson correlation coefficient is used to measure the adjacency between any two charging stations. i , j Hourly average power sequence P i , P j The linear correlation between them, Pearson correlation coefficient r ij The calculation formula is as follows: (1) in, P i,t , P j,t These are charging stations i and j At the point of time t average power, , It is the average power of each over the entire time period; T It is the total time step, Pearson correlation coefficient r ij The value range is [-1, 1]; by performing Pearson correlation coefficient analysis on the load data of all charging stations, a charging station adjacency matrix is ​​constructed using the load similarity between charging stations, and a load correlation heatmap of each charging station is generated; the Pearson correlation coefficient... r ij The closer the value is to 1 or -1, the stronger the positive or negative correlation between the load curves of the two charging stations; a value close to 0 indicates that the correlation is not obvious.

4. The method for quantifying the baseline of user group charging load by fusing multi-source data according to claim 1, characterized in that, The method for reconstructing the spatial dimension of the adjacency matrix data in step S2 is as follows: The external feature space is reconstructed based on the GAT model, whereby the GAT model includes a graph attention layer, and the input to the graph attention layer is a set of node features. h ={ h 1, h 2, ..., h N },in, It is a node i of F 3D feature vectors N Given the number of nodes, the graph attention layer outputs a new set of node features. h ={ h ʹ 1, h ʹ 2, ..., h ʹ N },in, It is a node i After transformation F ʹ 3D feature vector; First, from the weight matrix A parameterized shared linear transformation is applied to each node to enhance feature representation. Then, the GAT model transforms the node pairs ( i , j Calculate the attention coefficient e ij , used to represent nodes j Features of nodes i Importance, attention coefficient e ij The calculation method is as follows: (2) in, The attention mechanism is a single-layer feedforward neural network consisting of a weight vector. Parameterization was performed, and LeakyReLU was used as the activation function. The attention coefficients of each node were normalized using the Softmax function. The normalized attention coefficients were then... α ij for: (3) Where, N i It is a node i In the first-order neighbor set of the graph, the model employs a masked attention mechanism, only when a node... j It is a node i The attention coefficient is only calculated when the neighbor is involved. e ij For non-neighbor nodes, their attention coefficients are replaced with a very small negative number to ensure that their weights approach zero in the Softmax calculation; finally, the nodes... i New features The weights are obtained by weighted summation of the features of its neighboring nodes; these weights are the normalized attention coefficients. α ij And use ELU as the activation function: (4)。 5. The user group charging load baseline quantification method based on multi-source data as described in claim 4, characterized in that, The GAT model has two layers: the first layer consists of... K The attention heads are composed of multiple attention heads in parallel. Each head receives the raw features and processes them... K The output features of each element are concatenated to form a wider feature representation; the concatenated features of the first layer output are... As shown below: (5) Where || represents the concatenation operation; The second layer receives the features concatenated from the first layer. As input, the final feature is output by averaging the results from multiple sources. As shown below: (6)。 6. The method for quantifying the baseline of user group charging load by fusing multi-source data according to claim 1, characterized in that, Step S3 specifically involves time-series load forecasting based on TimeXer. The embedding strategy of the TimeXer model is as follows: Endogenous sequences X en To predict the target column, satisfying For embedding endogenous sequences, TimeXer employs a block embedding strategy, dividing long sequences into... N = T / P Non-overlapping blocks { s 1, s 2, ..., s N Each block contains P At each point in time, each block is mapped to a separate linear layer. D By using a time-series token and overlaying it with positional encoding, the model introduces a learnable global token. G en Endogenous sequences aim to capture macroscopic information about the entire sequence and serve as a bridge for interaction with exogenous variables. X en Ultimately represented as N +1 token H en : (7) Among them, Emb( s i =Linear( s i )+PE( i ); Exogenous sequences Z It is by C Composed of external features, satisfying Z ={ z (1) , ..., z (C) },in, Each exogenous sequence is mapped as a whole through an independent embedding layer. D Dimensional variable tokens, C exogenous sequence Z Represented as C Variable tokens V ex : (8)。 7. The user group charging load baseline quantification method based on multi-source data as described in claim 6, characterized in that, The attention mechanism used in the TimeXer model is as follows: Endogenous self-attention: In each block of the encoder, for all endogenous tokens H en Perform standard multi-head self-attention operations to obtain the processed tokens. H en satisfy: (9) Exogenous-Endogenous Cross-Attention: A cross-attention layer is introduced after the self-attention layer, in which the global tokens of the endogenous sequence are... G en from H en The variable tokens of all exogenous sequences are separated from ʹ and used as the query. V ex The key and value are input into the cross-attention module, resulting in an updated global token. G en ʹ: (10) Updated global token G en Replace ʹ H en The old global tokens in ʹ are then sent to the feedforward network layer together, thereby passing the aggregated external information to all endogenous tokens; go through L After processing by the layer encoder, the model will finally output all the endogenous tokens. Flatten the block dimension and generate the future through a linear prediction head. S Predicted values ​​at each time step : (11)。 8. The method for quantifying the baseline of user group charging load by fusing multi-source data according to claim 1, characterized in that, The historical load refers to the hourly charging station load data over the past 7 days.

9. The method for quantifying the baseline of user group charging load by fusing multi-source data according to claim 1, characterized in that, It also includes step S4: model testing and performance evaluation. We selected the historical mean method and compared its performance with five other models—Informer, Transformer, PatchTST, LSTM, and Bi-GRU—and the TimeXer model to evaluate their performance. We chose four indicators as the standard for measuring the prediction results. We conducted ablation experiments on three types of charging stations, calculated the percentage decrease in indicators, and verified the effectiveness of the GAT feature reconstruction module and multi-source data. The four indicators are specifically the Mean Absolute Error (MAE), Mean Square Error (MSE), Root Mean Square Error (RMSE), and Coefficient of Determination (R²). 2 The calculation formula is as follows: (12) (13) (14) (15) in, n It is the sample size. y i It is the first i The true value of each sample That is the corresponding predicted value. It is the average of all true values; MAE, MSE, and RMSE reflect the degree to which predicted values ​​deviate from actual values. 2 It measures the extent to which a model explains changes in predicted values, and its value ranges between (−∞, 1]. When R0... 2 When R = 0, it means the model's performance is at the average level; 2 The closer the value is to 1, the better the model's predictive performance.

10. A system employing the user group charging load baseline quantization method as described in any one of claims 1-9, characterized in that, The system includes a data preprocessing module, a GAT spatial feature reconstruction module, a TimeXer time series prediction module, and an effect evaluation module. The key feature is that the data preprocessing module integrates hourly charging sessions and quantifies the functional similarity between charging stations by calculating the Pearson correlation coefficient of their historical load sequences, thereby constructing an adjacency matrix to capture the spatial correlation between stations even without precise geographical location data. The GAT spatial feature reconstruction module reconstructs the spatial dimensions of the raw multi-source external data input from the data preprocessing module, treating each charging station as a node on a graph. By learning the attention weights between stations, it transforms macroscopic information that is the same for all stations into differentiated features that reflect the unique spatial relationships of each station, while simultaneously reducing the dimensionality of the vector features. The TimeXer time series prediction module receives the historical load of the target station and the reconstructed features processed by the GAT model, capturing both the time dependence of the historical load itself and the influence of external factors, ultimately generating a future load prediction value.

Images