A physical model driven and deep learning collaborative electric vehicle charging load prediction method

Abstract

The application discloses a kind of physical model driven and deep learning collaborative electric vehicle charging load prediction method, first based on the multi-source data including time, space, user behavior dimension Construction physical model, output reflects the internal law of charging load Physical characteristics, the physical characteristics of physical model output are input as one of key inputs Deep learning model (CNN-Transformer-LSTM);In the loss function of deep learning model, the constraint term reflecting the physical law of charging load is explicitly added, and the model prediction result is forced to comply with the physical law, the model after the fusion of physical model and deep learning model is (Phy-CNN-Transformer-LSTM, PCTL), the hyperparameters of PCTL fusion model are optimized using CPO algorithm.Optimized PCTL fusion model is used for prediction.The application realizes the double fusion of physical modeling and deep learning in feature input and loss function, and is assisted by CPO algorithm, which significantly improves the physical rationality and overall reliability of charging load prediction in complex multi-dimensional scenarios.

Description

Technical Field

[0001] This invention belongs to the field of electric vehicle charging load prediction and energy management, specifically involving a method for predicting electric vehicle charging load driven by a physical model and combined with deep learning. Background Technology

[0002] With increasing global emphasis on environmental protection and sustainable development, the new energy electric vehicle industry is experiencing rapid growth. The widespread application of electric vehicles not only helps reduce carbon emissions and alleviate environmental pollution caused by the large number of traditional gasoline-powered vehicles, but also promotes the transition of the energy structure to clean energy. However, the large-scale adoption of electric vehicles also brings many challenges to the power system, among which accurate prediction of electric vehicle charging load has become a critical issue that urgently needs to be addressed. Currently, existing methods for predicting electric vehicle charging load have certain limitations. Some traditional methods often consider only single-dimensional factors, such as relying solely on time-series data for prediction, failing to comprehensively cover important information such as spatial distribution and user behavior, resulting in predictions that do not accurately reflect the actual charging load. Although some deep learning-based methods can handle complex nonlinear relationships, the lack of effective integration with physical laws may lead to discrepancies between predicted results and actual physical conditions.

[0003] To ensure the stable operation of the power system, rationally plan charging infrastructure, and improve the charging experience for electric vehicle users, it is of great practical significance and urgency to develop a method for predicting electric vehicle charging load that can comprehensively consider multi-dimensional information such as time, space, and user behavior, and effectively integrate physical laws. Summary of the Invention

[0004] To address the aforementioned issues, this invention aims to overcome the limitations of existing electric vehicle charging load prediction methods by providing a physical model-driven and deep learning-based method for predicting electric vehicle charging load. This method aims to improve the accuracy and reliability of predictions and provide a scientific basis for power system planning, charging facility layout, and user charging services.

[0005] Technical Solution: This invention discloses 1. A method for predicting electric vehicle charging load using a physical model-driven and deep learning-integrated approach, characterized by comprising the following steps:

[0006] Step 1: Integrate multi-source data from time, space, and user behavior dimensions to construct a physical model, and use the physical model to output the physical characteristics of the charging load. The time dimension includes differences between weekdays and holidays, and seasonal variations; the spatial dimension includes the distribution of charging stations and urban functional zoning; and the user behavior dimension includes user preferences and travel purposes.

[0007] Step 2: The deep learning model uses CNN-Transformer-LSTM, and its input features include historical charging load data. Environmental data and the physical feature vectors output by the physical model Mapped to the predicted load value By adding a physical constraint term to the loss function of the deep learning model, the model after fusing the physical model and the deep learning model is called the PCTL fusion model.

[0008] Step 3: With the goal of minimizing the loss function, optimize the hyperparameters of the PCTL fusion model using the Crowned Porcupine algorithm, including the learning rate, number of network layers, and attention mechanism weights;

[0009] Step 4: Perform charging load prediction using the optimized PCTL fusion model.

[0010] Furthermore, the specific process of constructing the physical model is as follows:

[0011] (1) Time dimension data processing:

[0012] set up Indicates the characteristics of a workday. Indicates the characteristics of holidays, Seasonal characteristics, weekday and holiday characteristics are represented by binary variables. If the day is a weekday that is neither a holiday nor a weekend, then... ,otherwise If the day is a public holiday or weekend, then ,otherwise ; and Strictly mutually exclusive: That is, when =1 =0, when =1 =0, seasonal characteristics are represented by a periodic function. Considering The function takes values ​​in the range [-1, 1]. After normalization to [0, 1], the formula is as follows: ,in Indicates the first of the year sky; The larger the value, the closer the day is to summer, and the higher the demand for travel and charging; the smaller the value, the closer the day is to winter, and the lower the demand.

[0013] Comprehensive characteristics of the time dimension Represented as:

[0014]

[0015] in, , , The weights are assigned to weekdays, holidays, and seasonal characteristics, respectively.

[0016] (2) Spatial dimension data processing: Combining the distribution of charging stations and urban functional zoning, calculate the charging demand in different areas, and set... The distribution characteristics of charging stations are represented by their density. ,in This indicates the number of charging stations in a given area. Indicates the area of ​​the region; This represents the characteristics of urban functional zoning, which is divided into commercial areas, residential areas, and industrial areas. Different urban functional areas have different charging needs, which is represented by a vector: ;

[0017] Comprehensive characteristics of spatial dimensions Represented as:

[0018]

[0019] in, and The weights are respectively the distribution of charging stations and the characteristics of urban functional areas, and ;

[0020] (3) User behavior dimension data processing: Based on user preferences and travel purpose, analyze users' charging behavior patterns and set... Indicates user preference characteristics, The characteristics of travel purpose are represented by a vector; user preference characteristics are analyzed through historical charging data, which shows that users are more likely to charge during certain time periods, thus assigning higher weight to those time periods. Different travel purposes have different charging needs. ,in Indicates the first The weight of different travel purposes;

[0021] Comprehensive characteristics of user behavior:

[0022]

[0023] in, and The weights are respectively the user preferences and the travel purpose characteristics, and ;

[0024] (4) Data fusion: The data from the three dimensions are fused to obtain the physical characteristics of the charging load output by the final physical model. , is represented as:

[0025]

[0026] in, For energy conservation operators, It is a characteristic coupling function. As a comprehensive feature of the time dimension, As a comprehensive feature of spatial dimensions, These are comprehensive features based on user behavior. , , These are the weighting coefficients for each dimension. ;

[0027] Furthermore, a physical constraint term is added to the loss function, and the loss function is expressed as follows:

[0028]

[0029] in, For the prediction error loss term, For physical constraints, These are the weighting coefficients;

[0030] Prediction error loss function Calculated using mean square error:

[0031]

[0032] in, This represents the actual charging load value. For predicted values, This represents the number of samples.

[0033] Furthermore, the physical constraint term mentioned in step 2 Specifically, this includes energy conservation constraints and power balance constraints:

[0034] Energy conservation constraint:

[0035] in, For the first Predicted charging load at each time step For time step, Total energy demand;

[0036] Power conservation constraint:

[0037] in, This refers to the maximum power limit of the charging station.

[0038] Furthermore, the CPO algorithm for optimizing the hyperparameters of the PCLT model is described in the following steps:

[0039] (1) Initialize the population: Set the population size Maximum number of iterations Search space range Initialize the position of each individual in the population. ,in , The number of hyperparameters;

[0040] (2) Fitness calculation: For each individual in the population, a PCTL fusion model is constructed. The historical data obtained in step 1 is divided into a training set and a validation set in a 7:3 ratio. The model is then trained on the training set and the loss function value is calculated on the validation set. Fitness is For negative numbers, maximizing fitness is equivalent to minimizing loss;

[0041] (3) Update individual positions: Update the position of each individual according to the position update formula of the Crowned Porcupine algorithm: , For the first In the next iteration, individuals The position update is calculated using the following formula:

[0042]

[0043] in, This represents the position of the best individual in the current population. The location of the individual is randomly selected; , The result is a random number, with a value range of [0,1]. , These are control parameters used to balance global and local searches;

[0044] (4) Boundary handling: If the updated individual If the hyperparameter value is outside its range, adjust it to the boundary value.

[0045] (5) Termination condition judgment: Repeat steps (2)-(4) until the maximum number of iterations is reached. ;

[0046] (6) Output the optimal combination of hyperparameters .

[0047] Furthermore, when performing charging load prediction using the optimized PCTL fusion model, the following steps are performed:

[0048] Step 1: Determine the input data: The input data includes historical charging load data. Environmental data and physical model output ;in, This represents the actual charging load sequence n time steps before the predicted time. This represents environmental characteristic data at the time of prediction, including temperature. ,humidity Wind speed Precipitation ; The physical characteristics of the charging load output by the physical model;

[0049] Step 2: Standardize the three types of input data: historical charging load data and environmental data. The standardized input data is a concatenated vector of each standardized feature.

[0050] Step 3: Obtain standardized prediction output: Input the standardized input data into the optimized PCTL fusion model to obtain the standardized prediction output;

[0051] Step 4: Convert the standardized forecast output into a charging load forecast with actual physical meaning through destandardization.

[0052] Beneficial effects:

[0053] This invention constructs a physical model by integrating multi-source data from time dimensions (such as differences between weekdays and holidays, seasonal variations), spatial dimensions (such as charging station distribution, urban functional zoning), and user behavior dimensions (such as user preferences, travel purposes). This model comprehensively reflects the time-varying characteristics of charging load and the diversity of user behavior, solving the problem that traditional methods rely solely on historical data and ignore physical laws and user behavior, thus significantly improving the comprehensiveness and accuracy of prediction.

[0054] This invention uses the output of the physical model as one of the input features of the deep learning model to ensure that the prediction results conform to physical laws (such as energy conservation and power balance). A physical constraint term is added to the loss function of the deep learning model to further constrain the physical rationality of the prediction results, thus solving the problem that traditional prediction methods rely solely on data-driven approaches and ignore physical laws.

[0055] This invention uses the Crowned Porcupine algorithm to optimize the hyperparameters of the fusion model, including the learning rate, the number of network layers, and the attention mechanism weights. By combining global search and local search, the optimal combination of hyperparameters is found. Attached Figure Description

[0056] Figure 1 This is the overall flowchart of the present invention;

[0057] Figure 2 This is a flowchart of the training process for adding physical constraints to the deep learning model in this invention;

[0058] Figure 3 This is a flowchart of the pine hog algorithm optimization fusion model in this invention;

[0059] Figure 4 This is a comparison chart of electric vehicle charging load prediction using the physical model, deep learning model, and fusion model in this invention;

[0060] Figure 5 This is a comparison chart of electric vehicle charging load prediction using the optimized PCTL fusion model and the unoptimized PCTL fusion model based on the Crowned Porcupine Algorithm (CPO) in this invention. Detailed Implementation

[0061] The specific technical solution of the present invention will be further described in detail below with reference to specific examples.

[0062] like Figure 1 As shown, this invention discloses a physical model-driven and deep learning-integrated method for predicting electric vehicle charging load. First, a physical model is constructed based on multi-source data encompassing time, space, and user behavior dimensions. This model outputs physical features reflecting the inherent laws of charging load. These physical features are then used as one of the key inputs to a deep learning model (CNN-Transformer-LSTM). A constraint term reflecting the physical laws of charging load is explicitly added to the loss function of the deep learning model, forcing the model's prediction results to conform to these physical laws. The fused model of the physical model and the deep learning model is (Phy-CNN-Transformer-LSTM, PCTL). The Crown Porcupine algorithm is applied to optimize the hyperparameters of the PCTL fusion model, including the learning rate, network structure, and attention weights, minimizing the loss function with added physical constraints. The optimized PCTL fusion model is then used for prediction. This invention, through the dual fusion of physical modeling and deep learning in feature input and loss function, supplemented by the CPO algorithm, significantly improves the physical rationality and overall reliability of charging load prediction in complex, multi-dimensional scenarios.

[0063] The steps for constructing a physical model are as follows:

[0064] (1) Time dimension data processing:

[0065] set up Indicates the characteristics of a workday. Indicates the characteristics of holidays, Seasonal characteristics, weekday and holiday characteristics are represented by binary variables. If the day is a weekday that is neither a holiday nor a weekend, then... ,otherwise If the day is a public holiday or weekend, then ,otherwise ; and Strictly mutually exclusive: That is, when =1 =0, when =1 =0, seasonal characteristics are represented by a periodic function. Considering The function takes values ​​in the range [-1, 1]. After normalization to [0, 1], the formula is as follows: ,in Indicates the first of the year sky; The larger the value, the closer the day is to summer, and the higher the demand for travel and charging; the smaller the value, the closer the day is to winter, and the lower the demand.

[0066] Comprehensive characteristics of the time dimension Represented as:

[0067]

[0068] in, , , The weights are assigned to weekdays, holidays, and seasonal characteristics, respectively.

[0069] (2) Spatial dimension data processing: Combining the distribution of charging stations and urban functional zoning, calculate the charging demand in different areas, and set... The distribution characteristics of charging stations are represented by their density. ,in This indicates the number of charging stations in a given area. Indicates the area of ​​the region; This represents the characteristics of urban functional zoning, which is divided into commercial areas, residential areas, and industrial areas. Different urban functional areas have different charging needs, which is represented by a vector: ;

[0070] Comprehensive characteristics of spatial dimensions Represented as:

[0071]

[0072] in, and The weights are respectively the distribution of charging stations and the characteristics of urban functional areas, and ;

[0073] (3) User behavior dimension data processing: Based on user preferences and travel purpose, analyze users' charging behavior patterns and set... Indicates user preference characteristics, The characteristics of travel purpose are represented by a vector; user preference characteristics are analyzed through historical charging data, which shows that users are more likely to charge during certain time periods, thus assigning higher weight to those time periods. Different travel purposes have different charging needs. ,in Indicates the first The weight of different travel purposes;

[0074] Comprehensive characteristics of user behavior:

[0075]

[0076] in, and The weights are respectively the user preferences and the travel purpose characteristics, and ;

[0077] (4) Data fusion: The data from the three dimensions are fused to obtain the physical characteristics of the charging load output by the final physical model. , is represented as:

[0078]

[0079] in, For energy conservation operators, It is a characteristic coupling function. As a comprehensive feature of the time dimension, As a comprehensive feature of spatial dimensions, These are comprehensive features based on user behavior. , , These are the weighting coefficients for each dimension. ;

[0080] like Figure 2 As shown, physical constraint terms To constrain the prediction results to conform to physical rules, it is expressed as:

[0081]

[0082] in, For the first A physical constraint function, These are the corresponding physical constraint constants. The number of physical constraints.

[0083] Physical constraints Specifically, this includes energy conservation constraints and power balance constraints.

[0084] Energy conservation constraint:

[0085] in, For the first Predicted charging load at each time step For time step, This represents the total energy demand.

[0086] Power conservation constraint:

[0087] in, This refers to the maximum power limit of the charging station.

[0088] like Figure 3 As shown, the CPO algorithm for optimizing the hyperparameters of the PCLT model follows a specific process:

[0089] (1) Initialize the population: Set the population size Maximum number of iterations Search space range Initialize the position of each individual in the population. ,in , The number of hyperparameters;

[0090] (2) Fitness calculation: For each individual in the population, a PCTL fusion model is constructed. The historical data obtained in step 1 is divided into a training set and a validation set in a 7:3 ratio. The model is then trained on the training set and the loss function value is calculated on the validation set. Fitness is For negative numbers, maximizing fitness is equivalent to minimizing loss;

[0091] (3) Update individual positions: Update the position of each individual according to the position update formula of the Crowned Porcupine algorithm: , For the first In the next iteration, individuals The position update is calculated using the following formula:

[0092]

[0093] in, This represents the position of the best individual in the current population. The location of the individual is randomly selected; , The result is a random number, with a value range of [0,1]. , These are control parameters used to balance global and local searches;

[0094] (4) Boundary handling: If the updated individual If the hyperparameter value is outside its range, adjust it to the boundary value.

[0095] (5) Termination condition judgment: Repeat steps (2)-(4) until the maximum number of iterations is reached. ;

[0096] (6) Output the optimal combination of hyperparameters .

[0097] like Figure 4 As shown in the figure, the physical model, deep learning model, and the fusion of the physical model and deep learning model described in this invention are compared for electric vehicle prediction. The plotted curves clearly show that the prediction results of the fusion model are closest to the actual load data, while the prediction results of the physical model and deep learning model show some deviation from the actual load data. The statistical indicators show that the fusion model has the lowest RMSE and MAE, indicating the highest prediction accuracy; the physical model has the second lowest RMSE and MAE; and the deep learning model has the highest RMSE and MAE, indicating the lowest prediction accuracy.

[0098] like Figure 5 As shown, the Crowned Poor Pig Algorithm (CPO) in this invention optimizes the PCTL fusion model and compares it with the unoptimized PCTL fusion model for electric vehicle charging load prediction. The CPO algorithm optimizes model parameters, simultaneously reducing MSE, RMSE, MAE, and MAPE, demonstrating its ability to effectively handle different types of prediction errors, especially showing significant advantages in load surges and peak scenarios. The substantial improvement in R² indicates that the CPO-optimized model not only fits well on the training set but also maintains high generalization on the test set, adapting to unseen load patterns such as charging demands during sudden holidays and extreme weather conditions.

[0099] The above embodiments are only for illustrating the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent transformations or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting electric vehicle charging load using a physical model-driven and deep learning-integrated approach, characterized in that, Includes the following steps: Step 1: Integrate multi-source data from time, space, and user behavior dimensions to construct a physical model, and use the physical model to output the physical characteristics of the charging load. The time dimension includes differences between weekdays and holidays, and seasonal variations; the spatial dimension includes the distribution of charging stations and urban functional zoning; and the user behavior dimension includes user preferences and travel purposes. The specific process of constructing the physical model is as follows: (1) Time dimension data processing: set up Indicates the characteristics of a workday. Indicates the characteristics of holidays, Seasonal characteristics, weekday and holiday characteristics are represented by binary variables. If the day is a weekday that is neither a holiday nor a weekend, then... ,otherwise If the day is a public holiday or weekend, then ,otherwise ; and Strictly mutually exclusive: That is, when =1 =0, when =1 =0, seasonal characteristics are represented by a periodic function. Considering The function takes values ​​in the range [-1, 1]. After normalization to [0, 1], the formula is as follows: ,in Indicates the first of the year sky; The larger the value, the closer the day is to summer, and the higher the demand for travel and charging; the smaller the value, the closer the day is to winter, and the lower the demand. Comprehensive characteristics of the time dimension Represented as: ； in, , , The weights are for weekdays, holidays, and seasonal characteristics, respectively. (2) Spatial dimension data processing: Combining the distribution of charging stations and urban functional zoning, calculate the charging demand in different areas, and set... The distribution characteristics of charging stations are represented by their density. ,in This indicates the number of charging stations in a given area. Indicates the area of ​​the region; This represents the characteristics of urban functional zoning, which is divided into commercial areas, residential areas, and industrial areas. Different urban functional areas have different charging needs, which is represented by a vector: ; Comprehensive characteristics of spatial dimensions Represented as: ； in, and The weights are respectively the distribution of charging stations and the characteristics of urban functional areas, and ; (3) User behavior dimension data processing: Based on user preferences and travel purpose, analyze users' charging behavior patterns and set... Indicates user preference characteristics, The characteristics of travel purpose are represented by a vector. User preference characteristics are analyzed through historical charging data. The analysis shows that users are more likely to charge during certain time periods, thus assigning higher weight to those time periods. Different travel purposes result in different charging needs. ,in Indicates the first The weight of different travel purposes; Comprehensive characteristics of user behavior: ； in, and The weights are respectively the user preferences and the travel purpose characteristics, and ; (4) Data fusion: The data from the three dimensions are fused to obtain the physical characteristics of the charging load output by the final physical model. , is represented as: ； in, For energy conservation operators, It is a characteristic coupling function. As a comprehensive feature of the time dimension, As a comprehensive feature of spatial dimensions, These are comprehensive features based on user behavior. , , These are the weighting coefficients for each dimension. ; Step 2: The deep learning model uses CNN-Transformer-LSTM, and its input features include historical charging load data. Environmental data and the physical feature vector output by the physical model Mapped to the predicted load value By adding a physical constraint term to the loss function of the deep learning model, the model after fusing the physical model and the deep learning model is called the PCTL fusion model. Step 3: Optimize the hyperparameters of the PCTL fusion model using the Crowned Porcupine algorithm, including the learning rate, number of network layers, and attention mechanism weights; Step 4: Perform charging load prediction using the optimized PCTL fusion model.

2. The electric vehicle charging load prediction method based on physical model-driven and deep learning-assisted approach according to claim 1, characterized in that, The loss function incorporates a physical constraint term, and the loss function is expressed as follows: ； in, For the prediction error loss term, For physical constraint terms, These are the weighting coefficients; Prediction error loss function Calculated using mean square error: ； in, This represents the actual charging load value. For predicted values, This represents the number of samples.

3. The electric vehicle charging load prediction method based on physical model-driven and deep learning-assisted approach according to claim 2, characterized in that, The physical constraint term mentioned in step 2 Specifically, this includes energy conservation constraints and power balance constraints: Energy conservation constraint: ; in, For the first Predicted charging load at each time step For time step, Total energy demand; Power conservation constraint: ; in, This refers to the maximum power limit of the charging station.

4. The electric vehicle charging load prediction method based on physical model-driven and deep learning-assisted approach according to claim 1, characterized in that, The CPO algorithm for optimizing PCLT model hyperparameters, as described above, follows this process: (1) Initialize the population: Set the population size Maximum number of iterations Search space range Initialize the position of each individual in the population. ,in , The number of hyperparameters; (2) Fitness calculation: For each individual in the population, a PCTL fusion model is constructed. The historical data obtained in step 1 is divided into a training set and a validation set in a 7:3 ratio. The model is then trained on the training set and the loss function value is calculated on the validation set. Fitness is For negative numbers, maximizing fitness is equivalent to minimizing loss; (3) Update individual positions: Update the position of each individual according to the position update formula of the Crowned Porcupine algorithm: , For the first In the next iteration, individuals The position update is calculated using the following formula: ； in, This represents the position of the best individual in the current population. The location of the individual is randomly selected; , The result is a random number, with a value range of [0,1]. , These are control parameters used to balance global and local searches; (4) Boundary handling: If the updated individual If the hyperparameter value is outside its range, adjust it to the boundary value. (5) Termination condition judgment: Repeat steps (2)-(4) until the maximum number of iterations is reached. ; (6) Output the optimal combination of hyperparameters .

5. The electric vehicle charging load prediction method based on physical model-driven and deep learning-assisted approach according to claim 1, characterized in that, When performing charging load prediction using the optimized PCTL fusion model, follow these steps: Step 1: Determine the input data: The input data includes historical charging load data. Environmental data and physical model output ;in, This represents the actual charging load sequence n time steps before the predicted time. This represents environmental characteristic data at the time of prediction, including temperature. ,humidity Wind speed Precipitation ; The physical characteristics of the charging load output by the physical model; Step 2: Standardize the three types of input data: historical charging load data and environmental data. The standardized input data is a concatenated vector of each standardized feature. Step 3: Obtain standardized prediction output: Input the standardized input data into the optimized PCTL fusion model to obtain the standardized prediction output; Step 4: Convert the standardized forecast output into a charging load forecast with actual physical meaning through destandardization.

Images