Multi-region electric vehicle charging load prediction method based on astgcn

Abstract

The present application belongs to the field of new energy automobile charging load prediction, and discloses a multi-region electric vehicle charging load prediction method based on ASTGCN, aiming at the problems of severe time sequence fluctuation, close spatial correlation, insufficient accuracy of traditional prediction methods and lack of regional collaborative prediction ability of urban area electric vehicle charging load, historical operation data of multiple charging stations in the urban area is collected, and a charging station topology structure diagram and a standardized space-time data set are constructed; an ASTGCN model is used to learn the time dependence law and spatial correlation characteristics of the charging load, and an end-to-end prediction model with multi-feature fusion and multi-time scale output is established; through time sequence division, training, verification and testing are completed, which is suitable for short-term and medium-term load high-precision prediction of urban multi-site, distributed and wide-area distributed charging station groups. The present application improves the charging load prediction accuracy in complex scenarios, and provides reliable data support and decision basis for power grid dispatching, station planning and orderly charging control.

Description

Technical Field

[0001] This invention belongs to the field of load forecasting for power systems of new energy vehicles, and involves smart grids, new energy vehicles, and deep learning time series forecasting technology. Specifically, it involves a multi-region electric vehicle charging load forecasting method based on ASTGCN, which is particularly suitable for short-term and medium-term load forecasting of urban multi-site, distributed, and wide-area distributed charging station groups. Background Technology

[0002] With the rapid growth of electric vehicle ownership, urban charging load is characterized by its large scale, strong fluctuations, high spatiotemporal coupling, and regional mutual influence. Accurate forecasting of charging load is the core foundation for power grid safe dispatching, charging station planning, and orderly charging control.

[0003] Current charging load forecasting methods have the following significant shortcomings: Traditional methods such as ARIMA (Autoregressive Summing Moving Average), LSTM (Long Short-Term Memory), and Transformer (a deep learning architecture based on self-attention mechanisms) only focus on time series patterns, ignoring the spatial correlation and mutual influence between charging stations, resulting in low overall regional forecast accuracy; most methods only target single-station forecasts, failing to meet the power grid's forecasting needs for regional aggregated loads; they do not fully integrate multi-dimensional features such as meteorology, time, and spatial location, leading to weak model generalization ability; and they exhibit significant forecast deviations under complex scenarios such as peak loads, extreme weather, and holidays, making them unsuitable for practical engineering applications.

[0004] Therefore, developing a high-precision method for predicting charging load that can simultaneously uncover temporal patterns and spatial correlations and is applicable to multi-station collaboration in a region has significant engineering value and practical implications. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a multi-region electric vehicle charging load prediction method based on ASTGCN (Attention Based Spatial-Temporal Graph Convolutional Network, also known as "spatial-temporal graph convolutional large model"). This method can achieve multi-station collaboration, spatiotemporal feature fusion, multi-scale output, and high-precision prediction, providing reliable support for the safe and economical operation of the power grid and intelligent scheduling of charging stations.

[0006] The technical solution to achieve the above objective is: a multi-regional electric vehicle charging load prediction method based on ASTGCN, comprising the following steps: S1, Multi-dimensional data collection: Collects historical operating data of multiple electric vehicle charging stations in the area. The data includes real-time total charging load, single pile power, voltage, current, ambient temperature, humidity, wind speed, light intensity and standard timestamp information. S2, Data Preprocessing: Cleaning the raw data, including missing value imputation, outlier detection and removal, data standardization and normalization, and constructing a standardized spatiotemporal dataset with a unified format; S3, Spatial Topology Graph Construction: Using charging stations as nodes and geographical distance, load correlation, and electrical coupling degree as weights, an undirected weighted spatial topology graph is constructed to characterize the spatial influence relationship between charging stations and generate an adjacency matrix and a weight matrix. (1) In equation (1), w ij For charging station sites For the site Spatial impact intensity (charging station sites) With charging station sites Spatial correlation weights between them); For charging station sites With the site The straight-line distance between them; σ is the kernel function width, taken as the regional average inter-station distance; S4, Multidimensional Time Feature Extraction: Extract hour, weekday, month, season, whether it is a weekday, whether it is a holiday, and whether it is a peak period from the timestamp to construct a high-dimensional time series feature vector, which enhances the model's ability to capture time series patterns; S5, Construction of a large adaptive spatiotemporal graph convolutional prediction model: Building an ASTGCN prediction model, including spatial graph convolutional layers, temporal convolutional layers, feature fusion layers, Dropout regularization layers, fully connected layers and output layers, and simultaneously learning temporal dependencies and spatial relationships; (2) In formula (2): Let A be the adjacency matrix with self-connection, where A is the original adjacency matrix and I is the identity matrix. It is a degree matrix; This is the weight matrix. X （l） For the first l The node feature matrix of the layer, X （l+1） For the first l+1 The node feature matrix of the layer; S6, Dataset Partitioning and Model Training: The training set, validation set, and test set are divided into three parts in a time sequence at an 8:1:1 ratio; the Adam (Adaptive Moment Estimation) optimizer and MSE (mean squared error) loss function are used to complete the model training, and the hyperparameters are optimized through the validation set; S7, Multi-scale charging load prediction: Input real-time historical feature sequence, the model outputs multi-scale single-station and regional aggregated charging load prediction values ​​for the next 15min, 30min, 1h, 2h, and 4h, and outputs the prediction results; S8 Model Comprehensive Evaluation and Saving: Using MAE (Mean Absolute Error), RMSE (Root Mean Square Error), MAPE (Mean Absolute Percentage Error), and R... 2 (Coefficient of Determination) Completes the comprehensive evaluation of the model, fully measures the model's prediction accuracy, stability and generalization ability, and saves the best model for online prediction and engineering applications.

[0007] In the above-mentioned multi-region electric vehicle charging load prediction method based on ASTGCN, in step S2, the missing value filling adopts linear interpolation, Lagrange interpolation or temporal proximity filling; the outlier detection and removal adopts the 3σ principle, quartile method or isolated forest algorithm; the data standardization and normalization process adopts normal distribution standard normalization to make the data meet the standard distribution with a mean of 0 and a variance of 1.

[0008] In the above-mentioned multi-region electric vehicle charging load prediction method based on ASTGCN, in step S3, the spatial topology diagram is constructed with charging stations as graph nodes and geographical distance between stations, load correlation coefficient, and electrical coupling degree as edge weights to represent the spatial influence relationship between charging stations.

[0009] In the above-mentioned multi-region electric vehicle charging load prediction method based on ASTGCN, the multi-dimensional time features in step S4 include hour, weekday, month, season, whether it is a weekday, whether it is a holiday, and whether it is a peak period. The time features are used to characterize the impact of morning and evening peaks, daytime fluctuations, weekday-weekend differences, seasonal changes, and special events on the charging load distribution.

[0010] In the above-mentioned multi-region electric vehicle charging load prediction method based on ASTGCN, in step S5, the ASTGCN prediction model integrates temporal features, spatial topological features, electrical features and meteorological features as input, captures spatial correlation characteristics through a graph convolution module, and captures long-term temporal dependency patterns through a temporal convolution module.

[0011] In the above-mentioned multi-region electric vehicle charging load prediction method based on ASTGCN, step S6 divides the dataset into a training set, a validation set, and a test set according to the time sequence. The training set is used to learn the model parameters, the validation set is used to perform hyperparameter tuning and model selection, and the test set is used to evaluate the model's generalization ability.

[0012] In the above-mentioned multi-region electric vehicle charging load prediction method based on ASTGCN, in step S6, the dataset is divided into a training set of 80%, a validation set of 10%, and a test set of 10%, strictly following the chronological order to ensure that the temporal continuity is not disrupted and that the model evaluation conforms to the real prediction scenario.

[0013] The above-mentioned multi-region electric vehicle charging load prediction method based on ASTGCN, in step S7, outputs prediction results including single-station charging load prediction value, regional overall aggregated load prediction value and load confidence interval, supporting grid dispatch and station coordinated control.

[0014] The aforementioned multi-region electric vehicle charging load forecasting method based on ASTGCN is applicable to short- and medium-term load forecasting for urban public charging station clusters, park charging stations, skid-mounted charging stations, community distributed charging piles, and multi-energy complementary charging systems.

[0015] The multi-region electric vehicle charging load prediction method based on ASTGCN of the present invention can realize multi-station collaboration, spatiotemporal feature fusion, multi-scale output, and high-precision prediction, effectively improving the accuracy of charging load prediction in complex scenarios, and providing reliable data support and decision-making basis for power grid dispatching, station planning, and orderly charging control. Attached Figure Description

[0016] Figure 1 This is a flowchart of the multi-region electric vehicle charging load prediction method based on ASTGCN of the present invention. Figure 2 This is a schematic diagram comparing the predicted value and the actual value of a single region obtained by the method of the present invention (region 3). Figure 3 This is a schematic diagram comparing the predicted value and the actual value of a single region obtained by the method of the present invention (region 266). Figure 4This is a multi-region electric vehicle charging load prediction map (spatiotemporal prediction of load in 10 regions) obtained using the method of the present invention. Detailed Implementation

[0017] To enable those skilled in the art to better understand the technical solution of the present invention, its specific embodiments will be described in detail below with reference to the accompanying drawings.

[0018] Please see Figure 1 , Figure 2 , Figure 3 and Figure 4 An embodiment of the present invention provides a multi-region electric vehicle charging load prediction method based on ASTGCN, as detailed below.

[0019] 1. Study Subjects and Data Overview: This study focuses on 15 public electric vehicle charging stations in the main urban area of ​​a provincial capital city in southern my country. The stations cover core commercial areas, residential areas, transportation hubs, and industrial parks, exhibiting clear regional clustering and spatial correlation characteristics. Six stations are located in the core area, five in the near suburbs, and four in the far suburbs. These stations are interconnected via the city's power distribution network, resulting in significant spatiotemporal coupling characteristics in the charging load. Operational data for each charging station was collected over six consecutive months at 15-minute intervals, resulting in a total of 17,280 samples per station and a total of 259,200 samples across all 15 stations.

[0020] The fields of the collected raw data strictly correspond to the dimensions of the code input, as shown in Table 1.

[0021] Table 1, Original Data Table: The prediction targets of this embodiment are: the charging load of each charging station in the next 15 minutes, 30 minutes, 1 hour, 2 hours and 4 hours, as well as the overall aggregated charging load of the entire area.

[0022] 2. Data preprocessing: (1) Missing value handling: Missing values ​​in the original data caused by device disconnection or communication interruption are filled using linear interpolation. The calculation formula is as follows: (1) In equation (1), t1 and t2 are the effective times before and after the missing segment, x(t1) and x(t2) are the load values ​​at the corresponding times, and t is the time to be filled. For long sequence missing data with more than 3 consecutive time steps, weighted filling is performed using data from adjacent stations in the same region during the same period.

[0023] (2) Outlier removal: Calculate the mean μ and standard deviation σ, and use the 3σ principle to identify outliers: (2) In equation (2), μ is the mean of the load time series of a single station, and σ is the standard deviation.

[0024] For values ​​identified as abnormal, a time-series sliding window mean is generated to replace them, with the window size set to 8 (corresponding to 2 hours).

[0025] (3) Standardization: The normal distribution standard is used to normalize all input features to zero mean. (3) In equation (3), x These are the original eigenvalues; The mean of this feature on the training set; The standard deviation of this feature on the training set; x * These are the standardized feature values.

[0026] This maps all features to a distribution interval with a mean of 0 and a variance of 1, eliminating the impact of dimensional differences on model training.

[0027] 3. Construction of the spatial topology of charging stations: Using 15 charging stations as graph nodes, an undirected weighted spatial topology graph is constructed based on the geographical distance between stations, historical load correlation, and electrical path distance to characterize the spatial influence relationship between charging stations.

[0028] The spatial influence weights between charging stations are calculated using a Gaussian kernel function, resulting in a 15×15 dimensional adjacency matrix. The matrix elements represent the intensity of spatial influence between different charging stations. (4) In equation (4), For charging station sites and The straight-line distance between stations is represented by σ, where σ is the kernel function width, and the average inter-station distance within the region is taken. The adjacency matrix is ​​then constructed. Matrix elements Indicates charging station site For the site The spatial influence intensity is used to transmit spatial features in graph convolution.

[0029] 4. Construction of Multidimensional Temporal Features: Extracting and constructing high-dimensional temporal features from the original timestamps, including: (1) Cyclical characteristics: hours (0~23), days of the week (0~6), months (1~12); (2) Type characteristics: whether it is a working day (0 / 1), whether it is a statutory holiday (0 / 1); (3) Peak and valley characteristics: morning peak 7:00-9:00, evening peak 17:00-19:00, flat section, valley section marking; (4) Seasonal characteristics: Spring, summer, autumn and winter are the four seasons.

[0030] Temporal features, spatial features, and meteorological features together constitute an 18-dimensional input feature vector, providing the model with complete spatiotemporal prior information.

[0031] 5. Construction of a large adaptive spatiotemporal graph convolutional prediction model The adaptive spatiotemporal graph convolutional prediction model constructed in this embodiment is based on joint learning of spatiotemporal features. Its overall structure consists of an input layer, a spatial graph convolutional layer, a temporal convolutional layer, a feature fusion layer, a regularization layer, and a multi-scale output layer. The model receives an 18-dimensional spatiotemporal feature sequence as input, with a sequence length of 48 steps, corresponding to 12 hours of operational data, which can fully cover the intraday fluctuations and short-term dependencies of charging load. The spatial graph convolutional layer consists of three layers, each with 64 output channels. It uses a Gaussian kernel-weighted adjacency matrix to complete spatial information propagation, enabling the model to learn the spatial relationships and load transfer patterns between different charging stations, effectively capturing the mutual influence between stations within the region. The temporal convolutional layer adopts a causal dilated convolutional structure with a kernel size of 3 and a dilation rate of 2. This expands the receptive range without increasing the number of model parameters, enabling the model to capture temporal dependencies over longer periods, while ensuring that the prediction process does not introduce future information, conforming to the logic of real-world time-series prediction. The feature fusion layer concatenates spatial and temporal features to form a 128-dimensional high-dimensional fusion feature, providing ample information support for the final output.

[0032] (5) In equation (5), Let A be the adjacency matrix with self-connection, where A is the original adjacency matrix and I is the identity matrix. It is a degree matrix; This is the weight matrix. X （l） For the first l The node feature matrix of the layer, X （l+1） For the first l+1 The node feature matrix of the layer.

[0033] To avoid overfitting during training, a Dropout regularization layer with a dropout rate of 0.2 is added after the feature fusion layer to randomly deactivate some neurons during training, thereby improving the model's generalization ability. The model ultimately achieves multi-timescale output through a multi-branch fully connected layer, simultaneously outputting charging load predictions for the next 15 minutes, 30 minutes, 1 hour, 2 hours, and 4 hours, meeting the scheduling and planning needs of different scenarios.

[0034] 6. Implementation steps for spatial correlation characteristics: A complete characterization of the spatial characteristics of charging load is achieved through step-by-step modeling. The specific process is as follows: (1) Construct a spatial topology of charging stations, treat each charging station in the region as an independent spatial node, and establish a multi-site spatial association topology based on geographical location, electrical connection relationship and load correlation.

[0035] (2) Calculate spatial weights and generate adjacency matrices. Based on the geographical distance and coupling strength between stations, use the Gaussian kernel function to quantify the degree of spatial influence between stations and transform the spatial distribution into a computable mathematical matrix.

[0036] (3) Spatial feature extraction and information propagation are achieved through graph convolution. The adjacency matrix is ​​embedded in the model, and the weighted aggregation and transmission of features of adjacent stations are completed by using graph convolution layers, so that the model can learn the spatial dependence and load transmission law of charging stations in the region.

[0037] (4) Perform deep fusion of spatiotemporal features, and fuse spatial distribution features with temporal change features at the feature layer so that the model can simultaneously grasp the temporal patterns and spatial differences.

[0038] (5) Based on the spatiotemporal joint characteristics, spatial load forecasting is completed, and the output can simultaneously reflect the temporal fluctuations and spatial distribution of multi-station collaborative forecasting results, realizing true spatiotemporal integrated forecasting.

[0039] Prediction results are as follows Figure 4 The model effectively reflects the load differences and mutual influences in the spatial distribution of different regions. Compared with methods that only consider time-series characteristics, spatial modeling makes the prediction curve more closely match the real value, significantly improving the prediction accuracy of charging station clusters in complex areas, and fully demonstrating the superiority of this invention in spatiotemporal joint prediction.

[0040] 7. Dataset Partitioning: In this embodiment, the dataset is divided into training, validation, and test sets in an 8:1:1 ratio according to chronological order, without disrupting the chronological order. 80% is used for model parameter training, 10% for validation and tuning, and 10% for final testing, ensuring that the model training and evaluation scenarios are consistent with actual predictions.

[0041] 8. Model Training Settings: The model is trained using the Adam optimizer with an initial learning rate of 0.001, weight decay of 0.00001, batch size of 64, and a maximum training epoch of 100. Mean squared error is used as the loss function, and an early stopping strategy is employed: training automatically stops when the validation set loss fails to decrease for 15 consecutive epochs, preserving the optimal model.

[0042] 9. Prediction Results and Comparative Analysis: MAE, RMSE, MAPE, and R were used on the test set. 2 A comprehensive evaluation of the model was conducted. The prediction performance of the proposed method was compared with that of the three traditional prediction models: GRU (Gated Recurrent Unit), LSTM (Long Short-Term Memory Network), and Transformer (a deep learning architecture based on self-attention mechanism). The results are shown in Table 2. The proposed method performs excellently in the task of predicting the charging load of regional charging station clusters, and all indicators are superior to traditional time series prediction methods.

[0043] The detailed explanations of each evaluation indicator are as follows: (6) In equation (6), N is the total number of samples. y i For the first i The true value of each sample For the first i The predicted value for each sample.

[0044] (7) In equation (7), N is the total number of samples. y i For the first i The true value of each sample For the first i The predicted value for each sample.

[0045] (8) In equation (8), N is the total number of samples. y i For the first i The true value of each sample For the first i The predicted value for each sample.

[0046] (9) In equation (9), N is the total number of samples. y i For the first i The true value of each sample For the first iThe predicted value for each sample.

[0047] Table 2, Evaluation Results Table: In summary, the multi-region electric vehicle charging load prediction method based on ASTGCN of the present invention can achieve multi-station collaboration, spatiotemporal feature fusion, multi-scale output, and high-precision prediction, effectively improving the accuracy of charging load prediction in complex scenarios, and providing reliable data support and decision-making basis for power grid dispatching, station planning, and orderly charging control.

[0048] Those skilled in the art should recognize that the above embodiments are merely illustrative of the present invention and are not intended to limit the present invention. Any variations or modifications to the above embodiments that are within the spirit and essence of the present invention will fall within the scope of the claims of the present invention.

Claims

1. A multi-regional electric vehicle charging load prediction method based on ASTGCN, characterized in that, Includes the following steps: S1. Collect historical operational data from multiple electric vehicle charging stations within the urban area; S2. Preprocess the collected raw data, including missing value imputation, outlier detection and removal, data standardization and normalization, to construct a standardized spatiotemporal dataset with a unified format; S3. Construct a spatial topology map of charging stations based on their geographical location, electrical distance, line correlation strength, and historical load correlation; S4. Extract multi-dimensional time features based on timestamp information to form a high-dimensional time-series feature vector; S5. Build a large-scale ASTGCN prediction model; S6. Complete the training, validation, and testing of the large-scale ASTGCN prediction model through time-series partitioning; S7. Input real-time historical feature sequences and output charging load prediction results through the trained optimal model; S8. Conduct a comprehensive evaluation of the model's prediction performance and save the optimal prediction model for subsequent online predictions.

2. The method for predicting electric vehicle charging load in multiple regions based on ASTGCN according to claim 1, characterized in that, In step S1, the data includes the real-time total charging power of each charging station, power per charging pile, voltage, current, ambient temperature, humidity, wind speed, light intensity, and standard timestamp information.

3. The method for predicting electric vehicle charging load in multiple regions based on ASTGCN according to claim 1, characterized in that, In step S2, the missing value filling adopts linear interpolation, Lagrange interpolation, or temporal neighbor filling; the outlier detection and removal adopts the 3σ principle, quartile method, or isolated forest algorithm; the data standardization and normalization process adopts normal distribution standard normalization to make the data meet the standard distribution with a mean of 0 and a variance of 1.

4. The method for predicting electric vehicle charging load in multiple regions based on ASTGCN according to claim 1, characterized in that, In step S3, the spatial topology diagram of the charging station is used to characterize the spatial influence relationship between charging stations, and finally generates the corresponding adjacency matrix and weight matrix. (1) In equation (1), w ij For charging station sites For the site Spatial influence intensity; For charging station sites With the site The straight-line distance between them; σ is the kernel function width, which is the average distance between stations in the region.

5. A multi-region electric vehicle charging load prediction method based on ASTGCN according to claim 1 or 4, characterized in that, In step S3, the spatial topology diagram of the charging station is constructed with the charging station as the graph node and the geographical distance between stations, load correlation coefficient, and electrical coupling degree as the edge weights to form an undirected weighted graph structure, which is used to characterize the spatial influence relationship between charging stations.

6. The method for predicting electric vehicle charging load in multiple regions based on ASTGCN according to claim 1, characterized in that, In step S4, the multidimensional time features include hour, weekday, month, season, whether it is a weekday, whether it is a holiday, and whether it is a peak period. The time features are used to characterize the impact of morning and evening peaks, daytime fluctuations, weekday-weekend differences, seasonal changes, and special events on the distribution of charging load.

7. The method for predicting electric vehicle charging load in multiple regions based on ASTGCN according to claim 1, characterized in that, In step S5, the ASTGCN prediction model integrates temporal features, spatial topological features, electrical features, and meteorological features as inputs. It captures spatial correlation characteristics through a graph convolution module and captures long-term temporal dependency patterns through a temporal convolution module.

8. A multi-region electric vehicle charging load prediction method based on ASTGCN according to claim 1 or 7, characterized in that, In step S5, the ASTGCN prediction model includes a spatial graph convolutional layer, a temporal convolutional layer, a feature fusion layer, a Dropout regularization layer, a fully connected layer, and an output layer; this model simultaneously learns long- and short-term temporal dependencies and multi-site spatial correlation characteristics. (2) In formula (2): Let A be the adjacency matrix with self-connection, where A is the original adjacency matrix and I is the identity matrix. It is a degree matrix; This is the weight matrix. X （l） For the first l The node feature matrix of the layer, X （l+1） For the first l+1 Layer node feature matrix.

9. The multi-region electric vehicle charging load prediction method based on ASTGCN according to claim 1, characterized in that, In step S6, the dataset is divided into training set, validation set, and test set in chronological order. The training set is used to learn model parameters, the validation set is used to tune hyperparameters and select a model, and the test set is used to evaluate the model's generalization ability.

10. A multi-regional electric vehicle charging load prediction method based on ASTGCN according to claim 9, characterized in that, In step S6, the dataset is divided into 80% training set, 10% validation set, and 10% test set, strictly following the chronological order to maintain temporal continuity and ensure that the model evaluation matches the real prediction scenario.

11. The method for predicting electric vehicle charging load in multiple regions based on ASTGCN according to claim 1, characterized in that, In step S7, the real-time historical feature sequence is input, and the optimal model that has been trained is used to predict the charging load at multiple time scales for the next 15 minutes, 30 minutes, 1 hour, 2 hours and 4 hours, and the prediction results are output.

12. The method for predicting electric vehicle charging load in multiple regions based on ASTGCN according to claim 11, characterized in that, The output prediction results include the predicted value of charging load at a single station, the predicted value of the overall aggregated load of the region, and the load confidence interval, which supports grid dispatch and station coordinated control.

13. The method for predicting electric vehicle charging load in multiple regions based on ASTGCN according to claim 1, characterized in that, In step S8, the evaluation indicators for comprehensively evaluating the model prediction results include mean absolute error, root mean square error, mean absolute percentage error, and coefficient of determination, which comprehensively measure the model's prediction accuracy, stability, and generalization ability.

14. The method for predicting electric vehicle charging load in multiple regions based on ASTGCN according to claim 1, characterized in that, The multi-region electric vehicle charging load forecasting method is applicable to short- and medium-term load forecasting for urban public charging station clusters, park charging stations, skid-mounted charging stations, community distributed charging piles, and multi-energy complementary charging systems.

Images