Intelligent automobile charging load probability prediction method based on Informer model improvement
By introducing SMAN module and DDI module in the Informer model, the charging load prediction method is improved, and the problems of insufficient prediction accuracy and inability to effectively deal with complex load changes in the prior art are solved, thereby achieving higher accuracy and stability charging load prediction.
Patent Information
- Application Number
- CN202510218247.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-26
- Publication Date
- 2025-06-24
AI Technical Summary
When the existing charging load prediction methods deal with complex charging load fluctuations, it is difficult to capture the nonlinear characteristics and long-term dependencies of the data, and the prediction accuracy is insufficient in complex environments, so they cannot effectively deal with complex load changes.
A charging load probability prediction method based on an improved Informer model is proposed. By introducing SMAN module and DDI module, the model's spatio-temporal feature extraction capability and spatial relationship capture capability are enhanced, and the model performance is optimized to improve prediction accuracy.
Through the improved Informer model, the spatial and temporal changes of charging load can be captured more accurately, which significantly improves the accuracy and stability of charging load prediction, effectively deals with complex load changes, and provides more accurate prediction support.
Smart Images
Figure CN120197967A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of power system load forecasting and new energy charging facility optimization, and particularly relates to a charging load forecasting method and its system implementation that integrates time series feature encoding and probability density estimation. More specifically, the present invention proposes a probabilistic forecasting scheme based on the optimization of the Informer neural network architecture to achieve the quantitative characterization of multi-dimensional uncertainties during the operation of charging infrastructure. Background Art
[0002] With the popularization of electric vehicles and the continuous construction of charging facilities, charging load forecasting has become an important task for power grid dispatching and optimization. Accurate charging load forecasting not only helps improve the operation efficiency of the power grid, but also effectively reduces the burden on the power grid and ensures the stable operation of charging stations under high load conditions. Therefore, accurate forecasting of charging load is of great significance for the planning and dispatching of power systems.
[0003] Most traditional charging load forecasting methods rely on statistical time series models such as ARIMA models, linear regression models, etc. Although these methods have achieved certain results in some scenarios, they often have difficulty capturing the non-linear characteristics and long-term dependence relationships of data when dealing with complex charging load fluctuations. In addition, the change of charging load is affected not only by historical data, but also by various factors such as weather, traffic, and user behavior, and traditional methods have certain limitations in dealing with these factors.
[0004] In recent years, deep learning models, especially the Informer model based on the Transformer architecture, have gradually become an important tool in the field of time series forecasting due to their excellent sequence modeling ability and long-term dependence capturing ability. The Informer model effectively solves the problem of information loss in long sequences through the self-attention mechanism and has achieved excellent performance in multiple time series forecasting tasks. Therefore, the charging load forecasting method based on the Informer model, with its powerful modeling ability, can more accurately predict the change trend of charging load, improve the forecasting accuracy, and provide reliable decision-making support for power grid dispatching and optimization.
[0005] However, the existing charging load forecasting methods still have certain deficiencies. Especially in complex environments, how to improve the forecasting accuracy, reduce errors, and accurately reflect the fluctuation characteristics of charging load is an important challenge in current research. For this reason, the present invention proposes a charging load probability forecasting method based on an improved Informer model, aiming to improve the accuracy and stability of charging load forecasting by introducing new model structures and optimization strategies, so as to provide more accurate forecasting support for the efficient operation and energy management of the power grid. Summary of the Invention
[0006] The object of the present invention is to provide an intelligent vehicle charging load probability prediction method improved based on the Informer model, so as to solve the problems of insufficient prediction accuracy, large prediction fluctuations and inability to effectively cope with complex load changes in traditional charging load prediction.
[0007] To achieve the above object, the present invention provides an intelligent vehicle charging load probability prediction method improved based on the Informer model, including:
[0008] Step 1, collect and sort out historical charging data, and preprocess the historical charging load data;
[0009] Step 2, optimize the Informer model, and propose the SMAN module and the DDI module to enhance the model performance;
[0010] Step 3, train the optimized model to obtain a high-precision training model, and calculate the prediction error;
[0011] Step 4, conduct ablation experiments to evaluate the influence of each module on the model performance;
[0012] Step 5, realize the probability prediction of the intelligent vehicle charging load, so as to improve the prediction accuracy of the charging load and assist in power grid management.
[0013] Preferably, in the above technical solution, the charging load data obtained in Step 1 covers the changes in different time periods and the load of charging stations, so as to ensure that the model can adapt to diverse charging demands.
[0014] Preferably, in the above technical solution, in Step 1, the historical charging data is preprocessed, and each data point is strictly reviewed to ensure that it can accurately reflect the change trend of the charging load.
[0015] Preferably, in the above technical solution, the improvement of the Informer model in Step 2 includes:
[0016] (1) Embed the SMAN module into the model architecture of the Informer. The SMAN module can efficiently extract spatio-temporal features and successfully overcome the challenges faced by traditional methods in capturing the changes in load patterns in different time periods and spatial regions. Different from the traditional attention mechanism, the SMAN module generates queries (Q), keys (K) and values (V) through convolutional layers respectively, and ensures that each query only interacts with its local region. This design avoids the high computational overhead generated when performing global dot product calculations. The specific ways for the SMAN module to generate queries, keys and values are as follows:
[0017] Q = Conv Q (X), K = Conv K (X), V = ConvV (X)
[0018] The SMAN module introduces a multi-scale processing strategy. Since the dot product scores between queries and keys usually exhibit a long-tailed distribution and only a few queries contribute significantly to the scores, the SMAN module thus designs a sparse mechanism that focuses on computing those "active" queries - i.e., the queries that have a greater impact on the scores - to achieve computational optimization. In this process, the SMAN module uses the Kullback-Leibler divergence to quantify the sparsity S(q l , K), and its specific formula is as follows:
[0019]
[0020] where q l and k m represent the l-th row and the m-th row of matrices Q and K respectively;
[0021] (2) Embed the DDI module into the model architecture of Informer, behind the SMAN module. The DDI module combines convolutional operations with a bidirectional attention mechanism, significantly enhancing the ability to capture spatial relationships in the feature map and improving the accuracy of charging load prediction. The DDI module generates queries (Q), keys (K), and values (V) within the convolutional layer, achieving local attention to specific spatial regions. This local attention mechanism enables the spatio-temporal variations of the charging load to be more accurately modeled, effectively capturing the dependencies between regions, thereby further improving the prediction accuracy. Let the input feature tensor be X, containing C channels. The generation process of queries, keys, and values in the DDI module can be expressed as:
[0022] Q, K, V = Conv(X)
[0023] where the feature tensor is X, containing C channels;
[0024] The DDI module adopts a bidirectional computing strategy. Under this strategy, the attention mechanism performs calculations in the horizontal and vertical directions respectively to fully utilize the information at each position in the feature map. In this way, the model can capture context information from different directions and adapt to features of different scales, thereby enhancing the adaptability to various input data. In addition, the DDI module also introduces a dual-channel processing strategy. Under this strategy, the attention calculation is performed separately in two independent channels - one channel for the horizontal direction and the other for the vertical direction; the calculated results are then fused, enabling the model to focus on important features in different directions. This method significantly enhances the model's ability to capture complex spatial contexts. The attention calculation in the horizontal direction can be expressed as:
[0025]
[0026] The calculation of vertical attention can be expressed as:
[0027]
[0028] Where Q h 、K h and V h represent the query, key, and value in the horizontal channel respectively, and Q v 、K v and V v represent the query, key, and value in the vertical channel respectively, and d k is the dimension of the key for scaling;
[0029] The DDI module also introduces a residual connection mechanism to combine the output of the attention mechanism with the original input X, thereby promoting the flow of information and stabilizing the training process of the model. The final output can be expressed as:
[0030] Final Output = X + Output h + Output v
[0031] Where Output h and Output v represent the results of the attention mechanism applied in the horizontal and vertical directions respectively.
[0032] Preferably, in the above technical solution, step 3 trains the optimized model to obtain a training model with high accuracy.
[0033] Preferably, in the above technical solution, step 3 inputs real-time charging load data into the trained model to obtain the prediction result of the model and calculates the prediction error.
[0034] Preferably, in the above technical solution, step 4 conducts ablation experiments on the charging load dataset, including combining and training each module to obtain the RMSE, MSE, and MAE of each model, and comparing and analyzing the specific results.
[0035] Preferably, in the above technical solution, step 5 realizes the probability prediction of the charging load of intelligent vehicles, improves the accuracy of charging load prediction, and helps grid management.
[0036] (1) The present invention proposes an intelligent vehicle charging load probability prediction method improved based on the Informer model, which makes innovative improvements to the accurate prediction of charging load. Based on the original Informer model, the SMAN module and the DDI module are proposed. By adding the SMAN module, the spatio-temporal dependence in the charging load can be better captured, thus effectively improving the prediction accuracy; while the DDI module enhances the prediction ability for dynamically changing charging demands and further optimizes the performance of the model.
[0037] (2) The present invention proposes an intelligent vehicle charging load probability prediction method improved based on the Informer model. After applying this method to the power grid management system, it can significantly improve the prediction accuracy of power grid management and optimize the scheduling of charging resources, thereby more effectively supporting the operation and resource allocation of the power grid. Brief Description of the Drawings
[0038] Figure 1 Shows the overall process of the intelligent vehicle charging load probability prediction method improved based on the Informer model proposed by the present invention;
[0039] Figure 2 Is the framework diagram of the improved Informer model;
[0040] Figure 3 Is the structural diagram of the proposed SMAN module;
[0041] Figure 4 Is the structural diagram of the proposed DDI module;
[0042] Figure 5 Is the comparison diagram of the predicted values and the true values of the improved model;
[0043] Figure 6 Is the comparison diagram of the predicted values and the true values of different module combinations in the improved method;
[0044] Figure 7 Is the comparison diagram of the RMSE, MSE and MAE indexes of different models; Detailed Description of the Preferred Embodiments
[0045] The following will describe in detail the specific embodiments of the present invention in conjunction with the accompanying drawings, but it should be understood that the protection scope of the present invention is not limited by the specific embodiments.
[0046] Unless otherwise explicitly stated, in the entire specification and claims, the term "comprising" or its variations such as "including" or "having" etc. will be understood to include the stated elements or components, without excluding other elements or other components.
[0047] Such as Figures 1 to 7As shown in the figure, a probability prediction method for the charging load of intelligent vehicles improved based on the Informer model according to a specific embodiment of the present invention includes the following steps:
[0048] Step 1: Collect and organize historical charging data, and preprocess the historical charging load data
[0049] (1) Considering the challenges that may be encountered in the charging load prediction task, our data covers different charging time periods and takes into account factors such as grid load fluctuations. These factors fully reflect the problems such as the diversity of load patterns and time volatility faced in the charging load prediction task.
[0050] (2) For the preprocessing of the historical charging load data, we first cleaned the data and removed incomplete or abnormal data points. Then, through normalization and standardization processing, the influence brought by different data dimensions and ranges was eliminated to ensure the effectiveness of model training. To improve the prediction accuracy, we also performed feature engineering for time characteristics, including extracting periodic features, holiday effects, etc., and filled in the missing values to ensure the continuity and integrity of the data, thus laying a solid foundation for the training of the subsequent prediction model.
[0051] Step 2: Improve the Informer model and propose the SMAN module and the DDI module
[0052] (1) Embed the SMAN module into the model architecture of the Informer. The SMAN module can efficiently extract spatio-temporal features and successfully overcome the challenges faced by traditional methods in capturing the changes in load patterns in different time periods and spatial regions. The SMAN module generates queries (Q), keys (K), and values (V) through convolutional layers respectively, and ensures that each query only interacts with its local area. This design avoids the high computational overhead generated when performing global dot product calculations. The specific ways for the SMAN module to generate queries, keys, and values are as follows:
[0053] Q = Conv Q (X), K = Conv K (X), V = Conv V (X)
[0054] Among them, the feature tensor is X, which contains C channels;
[0055] The SMAN module introduces a multi-scale processing strategy. Since the dot product scores between queries and keys usually follow a long-tailed distribution and only a few queries contribute significantly to the scores, the SMAN module designs a sparse mechanism that focuses on computing those "active" queries - i.e., the queries that have a greater impact on the scores, thus achieving computational optimization. In this process, the SMAN module uses the Kullback-Leibler divergence to quantify the sparsity S(q l ,K), and its specific formula is as follows:
[0056]
[0057] where q l and k m represent the l-th row and the m-th row of matrices Q and K respectively;
[0058] This multi-scale method ensures that the attention mechanism can simultaneously focus on broader patterns and finer details.
[0059] (2) Add the DDI module to the architecture of the Informer model to improve the model's ability to capture complex spatial contexts, thereby further improving the model's prediction accuracy. The DDI module combines convolutional operations with a bidirectional attention mechanism, significantly enhancing the ability to capture spatial relationships in the feature map and improving the accuracy of charging load prediction. The DDI module generates queries (Q), keys (K), and values (V) within the convolutional layer, achieving local attention to specific spatial regions. This local attention mechanism enables the spatio-temporal changes of the charging load to be more accurately modeled, effectively capturing the dependencies between regions, and further enhancing the prediction accuracy. Let the input feature tensor be X, which contains C channels. The generation process of queries, keys, and values in the DDI module can be expressed as:
[0060] Q, K, V = Conv(X)
[0061] where the feature tensor is X, which contains C channels;
[0062] The DDI module adopts a bidirectional calculation strategy. Under this strategy, the attention mechanism calculates separately in the horizontal and vertical directions to make full use of the information at each position in the feature map. In this way, the model can capture context information from different directions, adapt to features of different scales, and enhance its adaptability to various input data.
[0063] In addition, the DDI module also introduces a dual-channel processing strategy. Under this strategy, the attention calculation is performed separately in two independent channels - one channel for the horizontal direction and the other for the vertical direction. The calculated results are then fused, enabling the model to focus on important features in different directions. This method significantly enhances the model's ability to capture complex spatial contexts. The attention calculation in the horizontal direction can be expressed as:
[0064]
[0065] The attention calculation in the vertical direction can be expressed as:
[0066]
[0067] where Q h 、K h and V h represent the query, key, and value in the horizontal direction channel respectively, Q v 、K v and V v represent the query, key, and value in the vertical direction channel respectively, and d k is the dimension of the key for scaling;
[0068] The DDI module also introduces a residual connection mechanism to combine the output of the attention mechanism with the original input X, thereby facilitating the flow of information and stabilizing the model training process. The final output can be expressed as:
[0069] Final Output = X + Output h + Output v
[0070] where Output h and Output v represent the results of the attention mechanisms applied in the horizontal and vertical directions respectively.
[0071] Step 3. Train the model to obtain an accurate trained model
[0072] During the model training process, we first use the training set in the charging load dataset for training, and at the same time use the validation set to perform real-time verification on the training process and compare it with the prediction accuracy of other existing models, as Figure 7 shown. Through this process, a high-precision USDT charging load probability prediction model is finally obtained. Ensure that the model can maintain a high prediction accuracy when performing the intelligent vehicle charging load probability prediction task, thereby improving its performance in practical applications.
[0073] Step 4. Conduct ablation experiments
[0074] Through ablation experiments on the improved method, we verified the effectiveness of each module and deeply analyzed the impact of each module on the model performance. The experimental results show that when only the SMAN module is added, that is, the UST model, the RMSE, MSE, and MAE indicators of the model all show a slight decrease; when only the DDI module is added, that is, the UDT model, the RMSE, MSE, and MAE indicators of the model also decrease. However, when the SMAN module and the DDI module are applied together, that is, the USDT model, the RMSE, MSE, and MAE indicators of the model are significantly reduced, and the prediction effect is significantly improved, as Figure 6 shown. This result fully proves that the synergistic effect of the two significantly enhances the overall effectiveness of the model.
[0075] Step 5: Implement the deployment and visualize the actual application effect of this method.
[0076] Apply the improved method to the intelligent vehicle charging load probability prediction task and evaluate its performance in actual charging load prediction. We conducted experimental predictions on this method on the EVnetNL dataset and compared and analyzed the prediction results with the real values through visualization means, as Figure 5 shown. This process not only shows the difference between the predicted value and the actual value but also provides an intuitive basis for evaluating the accuracy and effectiveness of this method in actual applications. Through this comparison, the actual application effect of the improved method in the charging load probability prediction task is further verified.
Claims
1. A method for predicting the probability of charging load of intelligent vehicles based on an improved Informer model, characterized in that: include: Step 1: Collect and organize historical charging data, and pre-process the charging load historical data; Step 2: Optimize the Informer model and propose the SMAN module and DDI module to enhance the model performance; Step 3: Train the optimized model to obtain a high-precision training model and calculate the prediction error; Step 4: Conduct ablation experiments to evaluate the impact of each module on model performance; Step 5: Realize the probabilistic prediction of electric vehicle charging load to improve the prediction accuracy.
2. The intelligent vehicle charging load probability prediction method based on the improved Informer model according to claim 1 is characterized in that: Step 1: Obtain the changing trends of the charging history data in different periods of the corresponding year, month, and day.
3. The intelligent vehicle charging load probability prediction method based on the improved Informer model according to claim 1 is characterized in that: Step 2 improves the Informer model by: (1) Embedding the SMAN module into the Informer model architecture can efficiently extract spatiotemporal features and successfully overcome the challenges faced by traditional methods in capturing changes in load patterns in different time periods and spatial regions. At the same time, unlike the traditional attention mechanism, the SMAN module generates queries (Q), keys (K), and values (V) through convolutional layers, and ensures that each query only interacts with its local area, avoiding the high computational overhead generated by global dot product calculations. The specific way in which the SMAN module generates queries, keys, and values is as follows: Q=Conv Q (X),K=Conv K (X),V=Conv V (X) The SMAN module introduces a multi-scale processing strategy. Since the dot product scores between queries and keys are usually distributed in a long-tail distribution and only a few queries contribute significantly to the scores, the SMAN module designs a sparse mechanism to focus on calculating those "active" queries, that is, queries that have a greater impact on the scores, thereby achieving computational optimization. In this process, the SMAN module uses the Kullback-Leibler divergence to quantify the sparsity of the query S(q l ,K), the specific formula is as follows: Among them, q l and k m denote the lth and mth rows of matrices Q and K respectively; (2) The DDI module is embedded in the Informer model architecture, located after the SMAN module. The DDI module combines convolution operations with a bidirectional attention mechanism, significantly improving the ability to capture spatial relationships in feature graphs and enhancing the accuracy of charging load prediction. At the same time, the DDI module generates queries (Q), keys (K), and values (V) in the convolution layer, achieving local attention to specific spatial regions, so that the spatiotemporal changes of charging load can be modeled more accurately and the dependencies between regions can be effectively captured, thereby further improving the accuracy of prediction. The generation process of queries, keys, and values in the DDI module can be expressed as: Q,K,V=Conv(X) The feature tensor is X, which contains C channels; The DDI module adopts a bidirectional calculation strategy, under which the attention mechanism is calculated in the horizontal and vertical directions respectively, so as to make full use of the information of each position in the feature map, so that the model can capture contextual information from different directions and adapt to features of different scales, thereby enhancing the adaptability to a variety of input data. In addition, the DDI module also introduces a dual-channel processing strategy, under which the attention calculation is performed in two independent channels, one for the horizontal direction and the other for the vertical direction. The results calculated by this strategy are then fused, so that the model can focus on important features in different directions, significantly enhancing the model's ability to capture complex spatial contexts. The attention calculation in the horizontal direction can be expressed as: The vertical attention calculation can be expressed as: Among them, Q h , K h and V h Represents the query, key and value in the horizontal channel, Q v , K v and V v Represents the query, key and value in the vertical channel, d k is the dimension of the key used for scaling; The DDI module also introduces a residual connection mechanism to combine the output of the attention mechanism with the original input X, thereby promoting the flow of information and stabilizing the training process of the model. The final output can be expressed as: Final Output=X+Output h +Output v Among them, Output h and Output v Represents the results of the attention mechanism applied in the horizontal and vertical directions, respectively.
4. The intelligent vehicle charging load probability prediction method based on the improved Informer model according to claim 1 is characterized in that: Step 3 obtains an accurate prediction model by inputting historical charging data into the model for training, and calculates its prediction error.
5. The intelligent vehicle charging load probability prediction method based on the improved Informer model according to claim 1 is characterized in that: Step 4 conducts an ablation experiment on the charging dataset, which specifically includes training each module combination, calculating the RMSE, MSE, and MAE of each model, and conducting a detailed comparative analysis of the experimental results.
6. The intelligent vehicle charging load probability prediction method based on the improved Informer model according to claim 1 is characterized in that: Step 5 applies this method to charging load forecasting, thereby improving forecasting accuracy and providing strong support for power grid management.