Electric vehicle group charging and discharging optimization control method and system based on deep reinforcement learning

Abstract

The invention discloses an electric vehicle group charging and discharging optimization control method and system based on deep reinforcement learning. The system comprises the following modules: step 1, constructing a state input vector; 2, adopting Granger causal relationship test to generate a compression state vector; 3, inputting the compressed state vector into a conditional diffusion model, generating a plurality of target electric quantity tracks corresponding to each electric vehicle, and forming an intention track candidate set; 4, inputting the intention trajectory candidate set into an improved dream optimization algorithm to perform trajectory simulation evaluation, performing belief attribution and sampling correction on the intention trajectory of which the performance does not reach the standard through a self-supervised multi-resolution belief error correction stage, and outputting an optimal intention trajectory; 5, constructing a tension energy field based on the optimal intention trajectory; and step 6, based on the equipotential propagation optimization controller, generating a charging and discharging action plan of each electric vehicle. According to the method, causal modeling and an improved dream optimization algorithm are fused, and intelligent charging and discharging control optimization of the electric vehicle group is realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent charging and discharging control technology, and in particular to a method and system for optimizing the charging and discharging control of electric vehicle groups based on deep reinforcement learning. Background Technology

[0002] With the continuous growth of new energy vehicle ownership, the impact of electric vehicle group charging and discharging on power grid operation is becoming increasingly significant. How to achieve intelligent and orderly scheduling of large-scale electric vehicles has become a key technological challenge for the integrated development of power and transportation systems. Existing electric vehicle group charging and discharging control methods mostly rely on centralized predictive scheduling or rule-based distributed control mechanisms, but these generally suffer from the following problems in practical applications:

[0003] Existing methods lack a systematic approach to processing multi-source heterogeneous information during state perception. Key variable extraction relies on empirical selection, and there is a lack of causal analysis and variable compression mechanisms, resulting in insufficient model generalization ability. In trajectory planning, most methods use single trajectory prediction or static optimal path generation, which is difficult to cover diverse user behavioral intentions under uncertain environments. The trajectory schemes lack diversity and adjustability, affecting the robustness of the control strategy. Most current optimization algorithms are guided by single-scale evaluation indicators and lack fine-grained attribution and error correction mechanisms for trajectory quality. They cannot effectively correct substandard charging and discharging strategies, affecting the final control effect. At the control execution level, conventional optimization methods have high computational complexity and slow convergence speed, making it difficult to meet the actual needs of real-time scheduling of electric vehicle groups. They also lack continuous differentiable modeling and efficient solution mechanisms for control objectives.

[0004] Therefore, how to provide a charging and discharging optimization control system for electric vehicle groups based on deep reinforcement learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a charging and discharging optimization control method and system for electric vehicle groups based on deep reinforcement learning. This invention integrates causal modeling and an improved dream optimization algorithm to achieve intelligent charging and discharging control optimization for electric vehicle groups. It constructs a compressed state vector based on Granger causality test, generates a candidate set of intention trajectories through a conditional diffusion model, corrects low-performance intention trajectories by combining a self-supervised multi-resolution belief error correction stage, and introduces a tension energy field and an equipotential propagation optimization controller. Ultimately, it achieves adaptive optimization scheduling oriented towards the triple objectives of user demand, electricity price response, and grid balance, and has significant advantages such as high modeling accuracy, strong optimization capability, and good operational stability.

[0006] Technical solution:

[0007] Step 1: Collect state information and environmental information of the electric vehicle group to construct a state input vector;

[0008] Step 2: Using the Granger causality test, extract the key causal variables from the state input vector to generate a compressed state vector;

[0009] Step 3: Input the compressed state vector into the conditional diffusion model to generate multiple target battery trajectories for each electric vehicle, forming an intention trajectory candidate set;

[0010] Step 4: Input the candidate set of intent trajectories into the improved dream optimization algorithm for trajectory simulation evaluation. Through the self-supervised multi-resolution belief correction stage, the intention trajectories that do not meet the performance standards are corrected by belief attribution and sampling, and the optimal intention trajectory is output.

[0011] Step 5: Construct a tension energy field based on the optimal intention trajectory. The tension energy field is a continuously differentiable energy function defined in the vehicle and time dimensions.

[0012] Step Six: Based on the current state of the electric vehicle, the optimal intended trajectory, and the tension energy field input equipotential propagation optimization controller, generate a charging and discharging action plan for each electric vehicle according to the minimum energy value criterion.

[0013] Optionally, the electric vehicle group status information specifically includes the battery status, remaining driving range, charging power limit, estimated departure time, and user travel information for each electric vehicle; the environmental information specifically includes grid load data, electricity price forecast information, and road traffic flow.

[0014] Optionally, the constructed state input vector is specifically:

[0015] The electric vehicle group status information and environmental information are respectively time-aligned and normalized.

[0016] The processed electric vehicle group state information and environmental information are arranged according to vehicle identification and time series. The state features corresponding to each time step are extracted and stacked in time order to construct a multi-dimensional time series feature matrix.

[0017] A one-dimensional convolutional encoder is used to compress the features of the multi-dimensional temporal feature matrix to obtain the state input vector.

[0018] Optionally, step two specifically includes:

[0019] Each input variable in the state input vector is divided into sliding windows according to a set time step to construct a time series for each input variable.

[0020] Perform a pairwise Granger causality test on the time series of each input variable, specifically including: using one variable as the response variable and the other variable as the predictor variable, construct an autoregressive model containing only the lagged terms of the response variable and an extended regression model containing the lagged terms of the predictor variable, respectively; and calculate the statistic for significance testing by comparing the sum of squared residuals of the autoregressive model and the extended regression model.

[0021] Based on the F-distribution corresponding to the statistic under a given degree of freedom, determine the significance p-value;

[0022] Variable pairs with a significance p-value less than a preset significance threshold are considered to have statistically significant Granger causal relationships. An initial causal connection graph is constructed based on these variable pairs, where each directed edge represents a significant causal influence between the variables.

[0023] In the initial causal connection graph, a gated weight parameter is introduced for each edge. The gated weight parameter is constrained by the Sigmoid activation function to be a continuous value between 0 and 1. An L1 regularization constraint is introduced during the training process to obtain a sparse causal adjacency matrix.

[0024] The sparse causal adjacency matrix and the node embedding vector of each variable are input into the graph attention encoder, which generates causal weight scores for each variable based on a multi-head attention mechanism.

[0025] Based on the causal weight scores, all variables are sorted from largest to smallest. The top K variables are selected to form a set of key causal variables, and linear dimensionality reduction is performed to generate a compressed state vector.

[0026] Optionally, the conditional diffusion model includes a forward diffusion process and a reverse reconstruction process;

[0027] The forward diffusion process uses the historical statistical average trajectory of each electric vehicle as the base trajectory, divides the base trajectory according to a preset time step, and performs numerical normalization processing on the trajectory data in each time step.

[0028] A Gaussian noise vector is generated based on the noise intensity parameters corresponding to the diffusion stage, and the noise vector is superimposed element by element onto the normalized trajectory data to obtain the noisy trajectory at the current diffusion time step.

[0029] Repeat the noise addition operation until all diffusion time steps are completed, and record the noise intensity parameters and trajectory perturbation amplitude at each diffusion time step to construct a noisy trajectory sequence.

[0030] The reverse reconstruction process concatenates the compressed state vector of each electric vehicle with the noisy trajectory sequence in chronological order as a conditional input, and feeds the conditional input into the denoising neural network based on the encoder-decoder structure step by step in time.

[0031] The denoising neural network uses a convolutional encoder to extract the temporal features of the noisy trajectory sequence and uses the compressed state vector to provide conditional guidance in each layer of the decoder network to generate corrected trajectory segments in sequence.

[0032] The trajectory segments output from all denoising steps of each electric vehicle are spliced ​​together in chronological order to form a complete target trajectory. Multiple different target battery trajectories are generated by repeating the reverse reconstruction process, which constitutes the intention trajectory candidate set of the electric vehicle.

[0033] Optionally, the improved dream optimization algorithm specifically includes a simulation evaluation stage, a self-supervised multi-resolution belief correction stage, and a trajectory selection stage;

[0034] In the simulation evaluation phase, for each intent trajectory in the intent trajectory candidate set, a control vector sequence is generated based on the charging and discharging actions at each time step, which serves as the control command for the interaction between the vehicle and the power grid.

[0035] The initial state information of the vehicle and the control vector sequence are input into the simulation state update model, and the predicted state variables of the vehicle are updated iteratively according to the time step. The predicted state variables include the remaining battery power, charging and discharging power and the estimated departure time.

[0036] A three-dimensional performance index vector is constructed based on the predicted state variables. The three-dimensional performance index vector includes the power supply satisfaction rate index, the electricity price response index, and the load smoothing index.

[0037] Intent trajectories whose dimensions in the three-dimensional performance index vector are below the corresponding set performance threshold are defined as substandard intent trajectories. The self-supervised multi-resolution belief correction stage performs belief attribution and sampling correction on substandard intent trajectories, specifically as follows:

[0038] The intent trajectory is divided into several trajectory segments by a time sliding window. An attribution graph of trajectory segments and performance indicators is constructed by combining the changes in performance indicators. The trajectory segments and performance indicators are used as nodes, and the weight of the edge represents the strength of the causal relationship between performance indicator degradation and trajectory segment changes.

[0039] The attribution graph is input into the attention network to extract the attention weights of trajectory segments at different time scales and calculate the belief bias score of each trajectory segment.

[0040] For trajectory segments whose belief bias scores are higher than a set score threshold, the time and location are locked and local perturbation resampling is performed, and a corrected segment is generated by changing the initial random noise.

[0041] The corrected segment is concatenated with the uncorrected portion of the intent trajectory to generate a supplementary intent trajectory candidate set.

[0042] During the trajectory selection phase, simulation evaluation is performed on the supplementary intention trajectory candidate set. Weighting coefficients are set for each performance index dimension in the three-dimensional performance index vector. The comprehensive performance score is calculated according to the weighted summation method. The intention trajectory with the highest comprehensive performance score is selected as the optimal intention trajectory of the electric vehicle.

[0043] Optionally, the step of inputting the vehicle's initial state information and control vector sequence into the simulation state update model, and iteratively updating the vehicle's predicted state variables step by step, specifically involves:

[0044] A state update rule base is constructed for individual electric vehicles. For each electric vehicle, a state transition calculation process is executed based on the input initial state information and the control vector at the current time step.

[0045] The state transition calculation process is as follows: calculate the change in battery level at the current time step based on the current remaining battery level and the charging / discharging power in the control vector; add the change in battery level to the remaining battery level at the previous time step to obtain the current remaining battery level; determine whether the expected departure conditions have been met by combining the current vehicle state and control commands, and update the expected departure time accordingly.

[0046] After each time step, the updated remaining battery power, charging and discharging power, and estimated departure time are recorded as the vehicle's predicted state for the current time step, and the vehicle's predicted state variables are updated iteratively according to the time step.

[0047] Optionally, step five: constructing a tension energy field based on the optimal intention trajectory, wherein the tension energy field is a continuously differentiable energy function defined in the vehicle and time dimensions, specifically:

[0048] The optimal intended trajectory of each electric vehicle is divided into a series of target battery values ​​according to time steps, and the difference between these values ​​and the current position of the vehicle is calculated to form a time-series error signal.

[0049] Based on the timing error signal, the energy value of each vehicle at each time step is defined as the squared weighted value of the timing error, and the energy is stacked in the two dimensions of vehicle number and time to form an initial energy tensor.

[0050] The initial energy tensor is processed to be continuous using a bilinear interpolation method, and an energy function that is continuously differentiable in both the vehicle number and time dimensions is constructed. This energy function is then used as a tension energy field to reflect the degree of deviation between the vehicle trajectory state and the target trajectory.

[0051] Optionally, the equipotential propagation optimization controller takes the current state of the electric vehicle, the optimal intended trajectory, and the tension energy field as inputs;

[0052] Using the energy value of each time step as a reference, the equipotential propagation algorithm is used to backtrack from the last time step and select the action with the minimum energy value at each time step.

[0053] The charging and discharging actions of each electric vehicle throughout the entire scheduling cycle are generated sequentially, forming a charging and discharging action plan for the electric vehicles.

[0054] An electric vehicle group charging and discharging optimization control system based on deep reinforcement learning according to an embodiment of the present invention includes the following modules:

[0055] The state perception module is used to collect state information of electric vehicle groups and environmental information, and construct state input vectors;

[0056] The causal extraction module is used to extract key causal variables from the state input vector based on the Granger causality test and generate a compressed state vector.

[0057] The trajectory generation module is used to input the compressed state vector into the conditional diffusion model to generate multiple target battery trajectories for each electric vehicle, forming an intention trajectory candidate set;

[0058] The trajectory optimization module is used to perform trajectory simulation evaluation on the candidate set of intention trajectories based on the improved dream optimization algorithm. Through the self-supervised multi-resolution belief correction stage, the intention trajectories with substandard performance are corrected by belief attribution and sampling, and the optimal intention trajectory is output.

[0059] The tension energy field construction module is used to construct a tension energy field defined in the vehicle and time dimensions based on the optimal intention trajectory.

[0060] The optimization control module is used to optimize the controller by propagating the current state of the electric vehicle, the optimal intended trajectory, and the tension energy field input equipotential. Based on the minimum energy value criterion, it generates a charging and discharging action plan for each electric vehicle.

[0061] The beneficial effects of this invention are:

[0062] This invention addresses the problems of inaccurate state perception, unstable trajectory control, incomplete performance optimization, and low control execution efficiency in existing scheduling methods through multi-module collaborative design and algorithm integration. By introducing Granger causality test and graph attention mechanism to extract compressed state vectors, it achieves the screening of key causal variables in high-dimensional multi-source state information, thereby enhancing the system's interpretability and robustness to behavioral decision variables.

[0063] This invention combines a conditional diffusion model to generate diverse target charge trajectories, thereby improving the controllability of the subsequent optimization space with a candidate trajectory set;

[0064] This invention introduces an improved dream optimization algorithm, combines simulation evaluation and attribution graph structure, identifies key trajectory segments with degraded performance, and improves trajectory quality by performing belief attribution and sampling correction on intention trajectories with substandard performance through a self-supervised multi-resolution belief correction stage.

[0065] This invention uses a tension energy field to define a continuously differentiable controllable target, and through an equipotential propagation algorithm, efficiently generates real-time action sequences under the minimum energy criterion. Ultimately, it achieves intelligent scheduling for multi-objective performance indicators, improving the stability, responsiveness, and user satisfaction of electric vehicle fleets in grid interaction. Attached Figure Description

[0066] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0067] Figure 1 This is an overall flowchart of the method of the present invention;

[0068] Figure 2 This is a schematic diagram of the system of the present invention. Detailed Implementation

[0069] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0070] refer to Figure 1 A charging and discharging optimization control method for electric vehicle groups based on deep reinforcement learning includes the following steps:

[0071] Step 1: Collect state information and environmental information of the electric vehicle group to construct a state input vector;

[0072] Step 2: Using the Granger causality test, extract the key causal variables from the state input vector to generate a compressed state vector;

[0073] Step 3: Input the compressed state vector into the conditional diffusion model to generate multiple target battery trajectories for each electric vehicle, forming an intention trajectory candidate set;

[0074] Step 4: Input the candidate set of intent trajectories into the improved dream optimization algorithm for trajectory simulation evaluation. Through the self-supervised multi-resolution belief correction stage, the intention trajectories that do not meet the performance standards are corrected by belief attribution and sampling, and the optimal intention trajectory is output.

[0075] Step 5: Construct a tension energy field based on the optimal intention trajectory. The tension energy field is a continuously differentiable energy function defined in the vehicle and time dimensions.

[0076] Step Six: Based on the current state of the electric vehicle, the optimal intended trajectory, and the tension energy field input equipotential propagation optimization controller, generate a charging and discharging action plan for each electric vehicle according to the minimum energy value criterion.

[0077] In this embodiment, the electric vehicle group status information specifically includes the battery status, remaining driving range, charging power limit, estimated departure time, and user travel information for each electric vehicle; the environmental information specifically includes power grid load data, electricity price forecast information, and road traffic flow.

[0078] In this embodiment, the constructed state input vector is specifically:

[0079] The electric vehicle group status information and environmental information are respectively time-aligned and normalized.

[0080] The processed electric vehicle group state information and environmental information are arranged according to vehicle identification and time series. The state features corresponding to each time step are extracted and stacked in time order to construct a multi-dimensional time series feature matrix.

[0081] A one-dimensional convolutional encoder is used to compress the features of the multi-dimensional temporal feature matrix to obtain the state input vector.

[0082] To achieve intelligent decision-making and control of the charging and discharging behavior of electric vehicles (EVs) driven by deep reinforcement learning, this invention first requires constructing a state input vector that accurately reflects the state of individual EVs and their external environment. Therefore, basic operational information for each EV is collected, including battery status, remaining driving range, maximum charging power, estimated departure time, and user travel information. The user travel information includes the estimated departure time, typical driving trajectory, and destination preference parameters. Simultaneously, to achieve coordinated operation between EV behavior and power grid resources, the system synchronously collects power grid operating environment information, specifically real-time power grid load data, electricity price forecasts, and road traffic flow, reflecting data sources that reflect load-side and traffic-side constraints.

[0083] The collected multi-source heterogeneous data is first time-aligned to ensure synchronization of all data at the same time granularity. Then, normalization preprocessing is performed to unify the data dimensions. The processed data is organized according to vehicle identification and a unified time series. The state features of each time step are stacked sequentially to construct a multi-dimensional temporal feature matrix that reflects the evolution of group behavior. The multi-dimensional temporal feature matrix is ​​then input into a one-dimensional convolutional encoder. The convolution kernel extracts dynamic features within the local time window, achieving dimensionality reduction and compression of state information and abstraction of behavioral features. Finally, a state input vector with high expressive power is output.

[0084] In this embodiment, step two specifically includes:

[0085] Each input variable in the state input vector is divided into sliding windows according to a set time step to construct a time series for each input variable.

[0086] Perform a pairwise Granger causality test on the time series of each input variable, specifically including: using one variable as the response variable and the other variable as the predictor variable, construct an autoregressive model containing only the lagged terms of the response variable and an extended regression model containing the lagged terms of the predictor variable, respectively; and calculate the statistic for significance testing by comparing the sum of squared residuals of the autoregressive model and the extended regression model.

[0087] Based on the F-distribution corresponding to the statistic under a given degree of freedom, determine the significance p-value;

[0088] Variable pairs with a significance p-value less than a preset significance threshold are considered to have statistically significant Granger causal relationships. An initial causal connection graph is constructed based on these variable pairs, where each directed edge represents a significant causal influence between the variables.

[0089] To improve the statistical stability and significance discrimination capability of causal relationship testing, this invention performs a sliding window partitioning operation on each input variable in the original state input vector according to a set time step, constructing multiple input segment samples with a temporal order. Each segment represents the variable change process within a fixed-length time interval, and multiple segments together constitute a sample set that can be used for statistical testing. Taking each pair of variables as a unit, an autoregressive model and an extended regression model containing lagged terms of predictor variables are constructed in each sliding window segment. By calculating the difference between the sum of squared residuals of the two models, a statistic for significance testing is obtained, where the variables refer to each item in the aforementioned electric vehicle group state information and environmental information.

[0090] The basic idea of ​​Granger causality testing is to determine whether a variable X has predictive power for another variable Y. Specifically, two regression models are constructed:

[0091] Autoregressive model (AR): Predicts variable Y using only its own past values;

[0092] Extended Real-Time Model (ARX): Predicts Y using past values ​​of both variable Y and variable X;

[0093] Compare the differences in prediction errors between the two models:

[0094] If adding variable X improves the model performance (significantly reduces error), then we say that variable X Granger causes variable Y.

[0095] Among them, the lag term represents the value of the variable in the first few time steps; the residual sum of squares represents the sum of the squared differences between the predicted and actual values;

[0096] The statistic follows an F-distribution, and its degrees of freedom can be further determined based on the set lag order and sample length. Given the degrees of freedom, the significance p-value is looked up based on the cumulative distribution function value of the statistic under the F-distribution. If the p-value is less than the preset significance threshold, then a statistically significant Granger causal relationship is considered to exist between the variables.

[0097] In the initial causal connection graph, a gated weight parameter is introduced for each edge. The gated weight parameter is constrained by the Sigmoid activation function to be a continuous value between 0 and 1. An L1 regularization constraint is introduced during the training process to obtain a sparse causal adjacency matrix.

[0098] The sparse causal adjacency matrix and the node embedding vector of each variable are input into the graph attention encoder, which generates causal weight scores for each variable based on a multi-head attention mechanism.

[0099] After constructing the initial causal connection graph, to achieve learnable modeling and structural optimization of causal edges, continuously differentiable gated weight parameters are introduced onto the directed edges between each variable. These weights are applied to the neural network output through a sigmoid activation function, strictly limiting their values ​​to between 0 and 1. This preserves gradient propagation capability while providing probabilistic interpretation for the gated weights. To further enhance the interpretability and sparsity of the causal structure, L1 regularization constraints are applied to all gated weights during model training to suppress redundant edge weights and improve the ability to identify core causal pathways. The resulting sparse causal adjacency matrix retains only causal paths with significant contributions, enhancing the robustness and generalization performance of the downstream graph neural network.

[0100] In the causal feature encoding stage, the sparse causal adjacency matrix and the node embedding vector corresponding to each input variable are input into the graph attention encoder. This encoder is based on a multi-head attention mechanism and combines causal structure to perform weighted aggregation operations, making the contextual representation perceived by different variables in the graph structure richer and more directional. The learning of attention weights and the gating of causal edge weights work together to enhance the model's ability to capture long-term dependencies. The causal weight score obtained by each variable at the end not only reflects its own historical influence ability, but also reflects its position and structural contribution in the causal graph, serving as a key reference for subsequent variable selection and state compression.

[0101] Based on the causal weight scores, all variables are sorted from largest to smallest. The top K variables are selected to form a set of key causal variables, and linear dimensionality reduction is performed to generate a compressed state vector.

[0102] In this embodiment, the conditional diffusion model includes a forward diffusion process and a reverse reconstruction process;

[0103] The forward diffusion process uses the historical statistical average trajectory of each electric vehicle as the base trajectory, divides the base trajectory according to a preset time step, and performs numerical normalization processing on the trajectory data in each time step.

[0104] A Gaussian noise vector is generated based on the noise intensity parameters corresponding to the diffusion stage, and the noise vector is superimposed element by element onto the normalized trajectory data to obtain the noisy trajectory at the current diffusion time step.

[0105] Repeat the noise addition operation until all diffusion time steps are completed, and record the noise intensity parameters and trajectory perturbation amplitude at each diffusion time step to construct a noisy trajectory sequence.

[0106] In this invention, a fixed number of diffusion steps is set during the forward diffusion process to control the gradual increase in noise intensity; the number of steps is set to 200. Each diffusion time step corresponds to a noise intensity parameter, which is generated using a cosine scheduling strategy. Cosine scheduling uses a periodic function to generate a non-linearly varying noise intensity, thereby improving the sampling diversity of intermediate time steps. In specific implementation, the Gaussian noise vector of each time step is independently sampled under a standard normal distribution using a pseudo-random number generator, ensuring the controllability and reproducibility of the noise.

[0107] The reverse reconstruction process concatenates the compressed state vector of each electric vehicle with the noisy trajectory sequence in chronological order as a conditional input, and feeds the conditional input into the denoising neural network based on the encoder-decoder structure step by step in time.

[0108] The denoising neural network uses a convolutional encoder to extract the temporal features of the noisy trajectory sequence and uses the compressed state vector to provide conditional guidance in each layer of the decoder network to generate corrected trajectory segments in sequence.

[0109] In the reverse reconstruction process of this invention, the denoising neural network adopts a conditional generative neural network model based on an encoder-decoder structure. Its overall structure includes three parts: a convolutional encoder, a conditional guidance module, and a convolutional decoder. The input noisy trajectory sequence is first processed by the convolutional encoder to extract multi-level features in the time dimension, generating a latent spatial representation. Before entering the decoding stage, the compressed state vector is embedded into a guidance vector consistent with the latent feature dimension through a fully connected mapping layer. This guidance vector is then combined with the latent representation in each layer of the decoder through feature concatenation, enabling conditional information to regulate the generation process. The decoder upsamples and reconstructs the trajectory signal layer by layer, finally outputting the corrected trajectory segment. The encoder and decoder adopt a symmetrical structure, with the kernel size and stride of each convolutional layer set to 3×1 and 1 step, respectively. The ReLU function is used for nonlinear activation.

[0110] The trajectory segments output from all denoising steps of each electric vehicle are spliced ​​together in chronological order to form a complete target trajectory. Multiple different target battery trajectories are generated by repeating the reverse reconstruction process, which constitutes the intention trajectory candidate set of the electric vehicle.

[0111] In this embodiment, the improved dream optimization algorithm specifically includes a simulation evaluation stage, a self-supervised multi-resolution belief correction stage, and a trajectory selection stage;

[0112] In the simulation evaluation phase, for each intent trajectory in the intent trajectory candidate set, a control vector sequence is generated based on the charging and discharging actions at each time step, which serves as the control command for the interaction between the vehicle and the power grid.

[0113] The initial state information of the vehicle and the control vector sequence are input into the simulation state update model, and the predicted state variables of the vehicle are updated iteratively according to the time step. The predicted state variables include the remaining battery power, charging and discharging power and the estimated departure time.

[0114] A three-dimensional performance index vector is constructed based on the predicted state variables. The three-dimensional performance index vector includes the power supply satisfaction rate index, the electricity price response index, and the load smoothing index.

[0115] This invention uses the predicted state variable sequence as the basic data source for constructing performance indicators. The power satisfaction rate indicator depends on whether the remaining power can meet the user's range demand when leaving the station; the electricity price response indicator depends on the degree of matching between charging and discharging behavior and the electricity price curve; and the load smoothing indicator depends on the degree of impact of the power output of all vehicles on the power grid. By iteratively updating the vehicle state, the three-dimensional performance indicator vector of the intended trajectory is accurately calculated.

[0116] Intent trajectories whose dimensions in the three-dimensional performance index vector are below the corresponding set performance threshold are defined as substandard intent trajectories. The self-supervised multi-resolution belief correction stage performs belief attribution and sampling correction on substandard intent trajectories, specifically as follows:

[0117] The intent trajectory is divided into several trajectory segments by a time sliding window. An attribution graph of trajectory segments and performance indicators is constructed by combining the changes in performance indicators. The trajectory segments and performance indicators are used as nodes, and the weight of the edge represents the strength of the causal relationship between performance indicator degradation and trajectory segment changes.

[0118] In the graph construction process, each trajectory segment is treated as a trajectory node in the attribution graph, and each performance index is treated as an index node. The connection between trajectory nodes and index nodes is established by evaluating the contribution of each trajectory segment to the deviation of a specific performance index. Specifically, a multivariate linear sensitivity analysis method is used to model the changes in control behavior of each trajectory segment and the response of the index value through linear regression. In the regression model, the attribute vector of the trajectory segment is used as the independent variable, and the deviation of the index value is used as the dependent variable; the regression coefficient reflects the sensitivity of the control behavior to the index deviation.

[0119] The edge weights are calculated from the standardized regression coefficients of the regression model described above, and then corrected using the residual goodness-of-fit index. Specifically, the weights are obtained by averaging the absolute values ​​of each standardized regression coefficient and multiplying this average by the residual goodness-of-fit index. This ensures that the weights reflect both the strength of the linear causal relationship between variables and their statistical significance. To suppress spurious correlations, only edges with a goodness-of-fit higher than a set correlation strength threshold are retained, and their weights are normalized. The resulting attribution graph structure is a heterogeneous directed graph, with node types including trajectory segment nodes and performance indicator nodes. The edges point from trajectory segments to performance indicator nodes, and the edge weights quantify the causal relationship between performance indicator degradation and trajectory segment changes, providing a causal basis for belief bias localization and error-correcting sampling.

[0120] The attribution graph is input into the attention network to extract the attention weights of trajectory segments at different time scales and calculate the belief bias score of each trajectory segment.

[0121] Specifically, the feature vectors of all trajectory segment nodes in the attribution graph (including changes in control behavior within the segment, trajectory perturbation amplitude, and contextual state information) are first input into the attention network. The attention network is based on a multi-head attention mechanism, which uses time windows of different sizes within each head to aggregate the adjacency structure of the trajectory segments, thereby capturing the changes in the impact of trajectory segments on performance indicators in the short and long term. The aggregated representation vectors of each trajectory segment node at different time scales are weighted and fused to generate a unified attention embedding representation.

[0122] Subsequently, the attention embedding representation of each trajectory segment node is weighted and combined with the weights of its corresponding edges to calculate its response strength to the performance degradation metric. To quantify the uncertainty of the trajectory segment, the response strength is compared with the expected value of the preset target metric response; a larger deviation indicates a higher risk of the trajectory deviating from the ideal state.

[0123] Finally, the belief bias score is obtained by taking the weighted product of the prediction bias value of the trajectory segment node on the index response and the weight of the attribution edge of all the performance index nodes connected in the attribution graph, and then normalizing the result. It reflects the potential influence of each trajectory segment on performance degradation in the group behavior.

[0124] For trajectory segments whose belief bias scores are higher than a set score threshold, the time and location are locked and local perturbation resampling is performed, and a corrected segment is generated by changing the initial random noise.

[0125] The corrected segment is concatenated with the uncorrected portion of the intent trajectory to generate a supplementary intent trajectory candidate set.

[0126] During the trajectory selection phase, simulation evaluation is performed on the supplementary intention trajectory candidate set. Weighting coefficients are set for each performance index dimension in the three-dimensional performance index vector. The comprehensive performance score is calculated according to the weighted summation method. The intention trajectory with the highest comprehensive performance score is selected as the optimal intention trajectory of the electric vehicle.

[0127] The improved dream optimization algorithm proposed in this invention introduces a self-supervised multi-resolution belief correction stage and causal attribution graph modeling on the basis of traditional dream optimization algorithms. This significantly enhances the ability to identify, locate, and correct unsatisfactory intention trajectories during trajectory optimization. Traditional dream optimization algorithms usually rely on heuristic trajectory perturbation and policy sampling mechanisms to generate diverse trajectories, but lack accurate tracing and structured processing of the causes of trajectory failure, resulting in low efficiency of the optimization process, random correction direction, and limited performance improvement.

[0128] In contrast, the improved dream optimization algorithm first constructs an attribution graph of trajectory segments and performance indicators, structuring the performance degradation problem into a causal path in a graph model, quantifying the contribution of different trajectory segments to changes in performance indicators, forming a refined basis for impact weight evaluation, and further using a multi-scale attention network to extract belief bias scores of trajectory segments at different time scales from the graph structure, effectively identifying key trajectory behaviors that cause the decline in indicators under the current power grid and vehicle conditions.

[0129] Through the aforementioned self-supervised multi-resolution belief correction stage, the improved dream optimization algorithm can achieve automated and quantitative bias attribution and error correction sampling, exhibiting stronger directionality and convergence efficiency compared to the original method which relies solely on sampling exploration. Furthermore, in the trajectory selection stage, to achieve a comprehensive trade-off between multiple performance objectives, this invention performs simulation evaluation on the supplementary intention trajectory candidate set. Based on a three-dimensional performance index vector composed of the power satisfaction rate index, electricity price response index, and load smoothing index, weighting coefficients of 0.4, 0.35, and 0.25 are set to reflect the relative importance of user range requirements, electricity price response capability, and grid stability in the scheduling strategy. For each intention trajectory, the three performance indices are weighted and summed using set weighting coefficients to obtain a comprehensive performance score. Finally, the trajectory with the highest comprehensive performance score is selected as the optimal intention trajectory for the electric vehicle, achieving robust screening of the optimal performance solution.

[0130] In this embodiment, the step of inputting the vehicle's initial state information and control vector sequence into the simulation state update model, and iteratively updating the vehicle's predicted state variables step by step, specifically involves:

[0131] A state update rule base is constructed for individual electric vehicles. For each electric vehicle, a state transition calculation process is executed based on the input initial state information and the control vector at the current time step.

[0132] The state transition calculation process is as follows: calculate the change in battery level at the current time step based on the current remaining battery level and the charging / discharging power in the control vector; add the change in battery level to the remaining battery level at the previous time step to obtain the current remaining battery level; determine whether the expected departure conditions have been met by combining the current vehicle state and control commands, and update the expected departure time accordingly.

[0133] After each time step, the updated remaining battery power, charging and discharging power, and estimated departure time are recorded as the vehicle's predicted state for the current time step, and the vehicle's predicted state variables are updated iteratively according to the time step.

[0134] The simulation state update model in this invention is used to simulate the dynamic evolution of the state variables of an electric vehicle over time under a given charging and discharging control strategy. This model uses a single vehicle as the basic unit, constructs a state update rule base, and iteratively executes state transition operations step by step by inputting the initial vehicle state information and control vector sequence into the model to predict the future state of the vehicle. The inputs include: the electric vehicle's current remaining battery power, estimated departure time, initial state parameters of maximum charging power, and the control vector (i.e., charging and discharging power plan) for each time step; the outputs are the predicted vehicle state variables at each time step, including remaining battery power, actual charging and discharging power, and estimated departure time.

[0135] The core function of this model is to provide accurate data support for trajectory simulation evaluation. By introducing state transition calculation rules based on energy conservation and behavioral constraints, the model can realistically reflect the impact of control strategies on the state evolution of electric vehicles. By calculating the superposition of charging and discharging power at the current time step and the battery level at the previous time step, the model can infer the remaining battery level of the vehicle. Combined with the expected departure time and user needs, it can determine whether the current plan is reasonable. This process not only makes the evaluation of the intended trajectory more accurate, but also provides a stable and reliable foundation for the subsequent construction of performance indicators and the execution of optimized control.

[0136] In this embodiment, step five: constructing a tension energy field based on the optimal intention trajectory, wherein the tension energy field is a continuously differentiable energy function defined in the vehicle and time dimensions, specifically:

[0137] The optimal intended trajectory of each electric vehicle is divided into a series of target battery values ​​according to time steps, and the difference between these values ​​and the current position of the vehicle is calculated to form a time-series error signal.

[0138] Based on the timing error signal, the energy value of each vehicle at each time step is defined as the squared weighted value of the timing error, and the energy is stacked in the two dimensions of vehicle number and time to form an initial energy tensor.

[0139] In this embodiment, the system divides the optimal intended trajectory of each electric vehicle—that is, the target battery change path the system expects the vehicle to reach within the current scheduling cycle—into a series of target battery values ​​in chronological order. These target values ​​are then compared with the actual battery value of the vehicle in the current simulated state, and the difference between the two is calculated. This difference represents the degree of deviation between the vehicle's current state and its target trajectory at a specific time step. This deviation is recorded to form a timing error signal. These error signals are subsequently used to construct an energy field, thereby guiding the system to optimize the vehicle's charging and discharging behavior to reduce this deviation and allow the vehicle to operate as close as possible to its ideal trajectory. This invention calculates the square of the error at each time step and multiplies it by a set weighting coefficient of 1.5 to obtain the energy value. The energy values ​​of each vehicle throughout the entire scheduling cycle are stacked according to both vehicle number and time step dimensions to construct an initial energy tensor.

[0140] The initial energy tensor is processed to be continuous using a bilinear interpolation method, and an energy function that is continuously differentiable in both the vehicle number and time dimensions is constructed. This energy function is then used as a tension energy field to reflect the degree of deviation between the vehicle trajectory state and the target trajectory.

[0141] In this embodiment, the equipotential propagation optimization controller takes the current state of the electric vehicle, the optimal intended trajectory, and the tension energy field as inputs;

[0142] Using the energy value of each time step as a reference, the equipotential propagation algorithm is used to backtrack from the last time step and select the action with the minimum energy value at each time step.

[0143] The charging and discharging actions of each electric vehicle throughout the entire scheduling cycle are generated sequentially, serving as the charging and discharging action plan for the electric vehicles.

[0144] In this invention, the core idea of ​​the equipotential propagation algorithm is to plan the charging and discharging action sequence for electric vehicles by minimizing the overall energy function in a tension energy field through a backward recursive approach. First, at the end time step of the scheduling cycle, the energy function value corresponding to each optional action is recorded as the initial cost. Then, the algorithm backtracks progressively from the end time step, traversing all optional charging and discharging actions within each time step, and calculating the sum of the cumulative energy values ​​after the current action is executed, referencing the cost value recorded in the next time step. Among all actions, the one that minimizes the cumulative energy value is selected as the optimal action for the current time step, and its corresponding cumulative energy value is recorded for the next round of backtracking. In this way, the equipotential energy value of the optimal path is propagated progressively along the time axis, ultimately allowing the backtracking of a charging and discharging action path with the minimum energy at the starting time step. This achieves global approximate optimization through local optimum propagation, making it suitable for group scheduling scenarios with high real-time requirements and a large number of electric vehicles. Finally, the optimal charging and discharging action path for each electric vehicle throughout the entire scheduling cycle is the action plan result output by the equipotential propagation algorithm.

[0145] refer to Figure 2 A charging and discharging optimization control system for electric vehicle groups based on deep reinforcement learning includes the following modules:

[0146] The state perception module is used to collect state information of electric vehicle groups and environmental information, and construct state input vectors;

[0147] The causal extraction module is used to extract key causal variables from the state input vector based on the Granger causality test and generate a compressed state vector.

[0148] The trajectory generation module is used to input the compressed state vector into the conditional diffusion model to generate multiple target battery trajectories for each electric vehicle, forming an intention trajectory candidate set;

[0149] The trajectory optimization module is used to perform trajectory simulation evaluation on the candidate set of intention trajectories based on the improved dream optimization algorithm. Through the self-supervised multi-resolution belief correction stage, the intention trajectories with substandard performance are corrected by belief attribution and sampling, and the optimal intention trajectory is output.

[0150] The tension energy field construction module is used to construct a tension energy field defined in the vehicle and time dimensions based on the optimal intention trajectory.

[0151] The optimization control module is used to optimize the controller by propagating the current state of the electric vehicle, the optimal intended trajectory, and the tension energy field input equipotential. Based on the minimum energy value criterion, it generates a charging and discharging action plan for each electric vehicle.

[0152] Example 1:

[0153] To verify the feasibility of this invention in practice, it was applied to a pilot area for electric vehicle aggregation and dispatching in a city. Thirty centralized charging stations and approximately 1,000 registered shared electric vehicles were selected within this area to form an electric vehicle fleet, covering typical business scenarios such as peak commuting, distributed fast charging, and on-site maintenance. During system deployment, the electric vehicles were distributed daily across different areas, exchanging status information with the central server via edge gateways.

[0154] In traditional dispatching systems, vehicles typically formulate charging and discharging plans based on static rules, lacking the ability to dynamically predict future trajectories and adapt to environmental changes. This often leads to problems such as a large number of vehicles charging during peak electricity price periods, severe fluctuations in grid load, and even insufficient power supply to meet users' scheduled departure needs. Therefore, the system of this invention is introduced to solve the problems that occur in traditional dispatching systems.

[0155] After system deployment, the state perception module first collects the state information of the electric vehicle group and environmental information in real time, constructs the state input vector, and through Granger causality test, the system effectively screens out the key variables that have a significant impact on the control strategy, forming a compressed state vector.

[0156] In the conditional diffusion model, each vehicle generates multiple candidate power trajectories that meet user needs and grid constraints based on its current state, forming a diverse set of intention trajectory candidates. Subsequently, an improved dream optimization algorithm performs simulation evaluation on the intention trajectory candidate set, iteratively optimizing low-performance trajectories and resampling and supplementing the optimized path using belief bias, thereby forming the optimal intention trajectory that meets the requirements of power demand, optimal electricity price, and load smoothing.

[0157] Based on the optimized trajectory, a tension energy field is constructed to represent the deviation function between the vehicle's current state and the target state. The equipotential propagation controller, based on the principle of minimizing energy value, backtracks to generate the optimal control path and formulates an adaptive charging and discharging action plan for each vehicle.

[0158] The actual operating cycle is 7 consecutive days, covering the entire cycle of weekdays and weekends. Table 1 reflects the optimization effect of the system on user charging success rate and scheduling balance.

[0159] Table 1 Comparison of User Power Supply Guarantee and Grid Load Response Before and After Optimization

[0160]

[0161] As can be seen from the data in Table 1, the electric vehicle group charging and discharging optimization control system based on deep reinforcement learning proposed in this invention significantly improves group scheduling performance and user experience in practical applications.

[0162] The average charging completion rate improved from 82.4% to 96.7%, an increase of 17.4%. This metric reflects the system's ability to meet users' charging needs. The high completion rate after optimization indicates that the personalized target trajectory generated by the conditional diffusion model and the subsequent intent trajectory optimization mechanism can more accurately predict and schedule charging behavior, reduce charging failures caused by scheduling conflicts or resource waste, and significantly enhance service reliability.

[0163] Regarding user complaint rate, it decreased from 3.2 per 100 users before optimization to 0.7, a reduction of 78.1%. This significant decrease in complaint rate indicates that the system has made substantial improvements in ensuring users' time planning, meeting travel needs, and providing convenient charging. In particular, the introduction of a self-supervised multi-discriminatory belief correction stage to proactively correct substandard performance trajectories effectively avoids deviations between user expectations and system behavior.

[0164] Regarding the average load shift during peak electricity price periods, it decreased from 0.45MW to 0.18MW, a reduction of up to 60%. This indicates that the present invention not only optimizes the charging experience at the user level but also significantly enhances the system's responsiveness to electricity price signals. It can effectively guide users to avoid charging during high electricity price periods, reduce the impact on peak loads of the power grid, and demonstrate the system's advantages in power system friendliness.

[0165] The electricity price response rate increased from 49.8% before optimization to 86.2%, which shows that during the optimization process, the system can more sensitively identify the trend of electricity price changes and achieve efficient and economical energy dispatch by adjusting the charging and discharging strategies of electric vehicles, thereby significantly reducing overall operating costs.

[0166] The proportion of passengers meeting the expected departure time also increased from 88.6% to 98.4%, reflecting the timeliness guarantee of users' travel needs. The optimized high guarantee rate shows that the system can ensure that users' departure time is not delayed while taking into account the power grid dispatch objectives and achieving multi-objective coordinated control.

[0167] This embodiment verifies the application value and technical advantages of the system of the present invention in real-world complex environments. By constructing a state perception module, a causal extraction module, a trajectory generation module, a trajectory optimization module, a tension energy field construction module, and an optimization control module, the system achieves closed-loop control throughout the entire process, from electric vehicle state perception to intention trajectory generation, trajectory error correction, and control optimization driven by the energy field. This not only improves the charging completion rate and electricity price response rate but also effectively reduces peak load fluctuations and user complaint rates, ensuring the dual satisfaction of user travel time and vehicle charging needs. By introducing an improved dream optimization algorithm and an equipotential propagation optimization controller, the system demonstrates good intelligence and stability in multi-source heterogeneous information modeling, trajectory prediction correction, and action decision-making. It provides an efficient and feasible technical path for intelligent scheduling of electric vehicle groups in high-load scenarios, proving that the present invention has achieved significant optimization in charging efficiency, grid response capability, and user satisfaction, and has good prospects for promotion and application.

[0168] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1.A method for optimizing control of charging and discharging of a group of electric vehicles based on deep reinforcement learning, characterized in that, The method comprises the following steps: Step 1: collecting state information and environment information of the electric vehicle group, and constructing a state input vector; Step 2: performing Granger causality test on the state input vector to extract causal key variables and generate a compressed state vector; Step 3: inputting the compressed state vector into a conditional diffusion model to generate multiple target electric quantity trajectories corresponding to each electric vehicle, thereby forming an intention trajectory candidate set; Step 4: inputting the intention trajectory candidate set into an improved dream optimization algorithm for trajectory simulation evaluation, and performing belief attribution and sampling correction on intention trajectories with substandard performance through a self-supervised multi-resolution belief error correction stage, thereby outputting an optimal intention trajectory; Step 5: constructing a tension energy field based on the optimal intention trajectory, wherein the tension energy field is a continuous and differentiable energy function defined in the vehicle and time dimensions; Step 6: inputting the current state of the electric vehicle, the optimal intention trajectory and the tension energy field into an equipotential propagation optimization controller, and generating a charging and discharging action plan for each electric vehicle according to the minimum energy value criterion. 2.The method of claim 1, wherein, The state information of the electric vehicle group specifically includes the electric quantity state, the remaining range, the upper limit of the charging power, the estimated departure time and the user's trip information of each electric vehicle; and the environment information specifically includes the power grid load data, the electricity price prediction information and the road traffic flow. 3.The method of claim 1, wherein, The state input vector is constructed as follows: The electric vehicle group state information and environment information are respectively time-aligned and normalized; The processed electric vehicle group state information and environment information are arranged according to the vehicle identifier and the time sequence, the state features corresponding to each time step are extracted, and a multi-dimensional time sequence feature matrix is constructed by stacking in time sequence; The multi-dimensional time sequence feature matrix is compressed by a one-dimensional convolutional encoder to obtain the state input vector. 4.The method of claim 1, wherein, Step 2 is specifically as follows: The input variables in the state input vector are divided into sliding windows according to the set time step, and the time sequence of each input variable is constructed; The time sequences of the input variables are tested for pairwise Granger causality, which specifically includes: taking one variable as a response variable and another variable as a prediction variable, respectively constructing an autoregressive model containing only the lagged items of the response variable and an extended regression model containing the lagged items of the prediction variable, and calculating a statistic for significance test by comparing the residual sum of squares of the autoregressive model and the extended regression model; Based on the F distribution corresponding to the statistic at a given degree of freedom, the significance p value is determined; The variable pairs with a significance p value less than a preset significance threshold are considered to have statistically significant Granger causality, and an initial causal connection graph is constructed based on the variable pairs, wherein each directed edge represents a significant causal effect between variables; A gating weight parameter is introduced for each edge in the initial causal connection graph, which is constrained by a Sigmoid activation function to be a continuous value between 0 and 1, and L1 regularization constraint is introduced in the training process to obtain a sparse causal adjacency matrix. The sparse causal adjacency matrix is input into a graph attention encoder together with the node embedding vector of each variable, and the graph attention encoder generates a causal weight score of each variable based on a multi-head attention mechanism; All variables are sorted in descending order according to the causal weight score, and the top K variables are selected to form a causal key variable set, and linear dimension reduction processing is performed to generate a compressed state vector. 5.The method of claim 1, wherein, The conditional diffusion model includes a forward diffusion process and a backward reconstruction process; The forward diffusion process takes the historical statistical average trajectory of each electric vehicle as a basis trajectory, divides the basis trajectory according to a preset time step, and performs numerical normalization processing on the trajectory data in each time step; According to the noise intensity parameter corresponding to the diffusion stage, a Gaussian noise vector is generated, and the noise vector is element-wise superimposed on the normalized trajectory data to obtain a noisy trajectory at the current diffusion time step; The noise operation is repeated until all diffusion time steps are completed, and the noise intensity parameter and the trajectory disturbance amplitude used in each diffusion time step are recorded to construct a noisy trajectory sequence; The backward reconstruction process concatenates the compressed state vector of each electric vehicle and the noisy trajectory sequence in time sequence as a conditional input, and inputs the conditional input into a denoising neural network based on an encoder-decoder structure; The denoising neural network uses a convolutional encoder to extract the time features of the noisy trajectory sequence, and uses the compressed state vector to provide conditional guidance in each layer of the decoder network to generate corrected trajectory segments in sequence; The trajectory segments output by each electric vehicle in all denoising steps are concatenated in time sequence to form a complete target trajectory, and multiple different target electric quantity trajectories are generated by repeating the backward reconstruction process to form an intention trajectory candidate set of the electric vehicle. 6.The method of claim 1, wherein, The improved dream optimization algorithm specifically includes a simulation evaluation stage, a self-supervised multi-resolution belief correction stage, and a trajectory selection stage; For each intention trajectory in the intention trajectory candidate set, the simulation evaluation stage generates a control vector sequence according to the charging and discharging actions at each time step, which is used as a control instruction for vehicle and grid interaction; The initial state information of the vehicle and the control vector sequence are input into a simulation state update model, and the predicted state variables of the vehicle are updated iteratively according to the time step, including the remaining electric quantity, the charging and discharging power, and the estimated departure time. A three-dimensional performance index vector is constructed according to the predicted state variables, including the electric quantity satisfaction rate index, the electricity price response index, and the load smoothing index. If any dimension of the three-dimensional performance index vector is lower than the corresponding set performance threshold, the intention trajectory is defined as a performance substandard intention trajectory, and the self-supervised multi-resolution belief correction stage performs belief attribution and sampling correction on the performance substandard intention trajectory, specifically: The intention trajectory is divided into several trajectory segments according to a time sliding window, and an attribution graph of the trajectory segment and the performance index is constructed by combining the performance index changes, taking the trajectory segment and the performance index as nodes, and the weight of the edge representing the causal correlation strength between the performance index degradation and the trajectory segment change. input the attribution graph into an attention network, extract attention weights of the trajectory segments at different time scales, and calculate a belief bias score of each trajectory segment; for the trajectory segment with the belief bias score higher than a set score threshold, lock the time position and perform local perturbation resampling to generate a corrected segment by changing the initial random noise; splice the corrected segment with the uncorrected part of the intended trajectory to generate a candidate set of supplemented intended trajectories; in the trajectory selection stage, perform simulation evaluation on the candidate set of supplemented intended trajectories, set a weighting coefficient for each performance indicator dimension in a three-dimensional performance indicator vector, calculate a comprehensive performance score in a weighted summation manner, and select the intended trajectory with the highest comprehensive performance score as the optimal intended trajectory of the electric vehicle. 7.The method of claim 6, wherein, The simulation state update model is inputted with the initial state information of the vehicle and the control vector sequence, and the predicted state variables of the vehicle are iteratively updated at each time step, specifically: a state update rule library is constructed for each electric vehicle, and a state transition calculation process is performed according to the input initial state information and the control vector at the current time step; the state transition calculation process specifically includes: calculating the power change value at the current time step according to the current remaining power and the charging and discharging power in the control vector; adding the power change value to the remaining power at the last time step to obtain the current remaining power; combining the current state of the vehicle with the control instruction to determine whether the expected departure condition is reached, and updating the expected departure time accordingly; after each time step, the updated remaining power, charging and discharging power, and expected departure time are recorded as the predicted state of the vehicle at the current time step, and the predicted state variables of the vehicle are iteratively updated at each time step. 8.The method of claim 1, wherein, Step five: constructing a tension energy field based on the optimal intended trajectory, the tension energy field being a continuous and differentiable energy function defined in the vehicle and time dimensions, specifically: dividing the optimal intended trajectory of each electric vehicle into a series of target power values according to the time step, and performing difference calculation with the current position state of the vehicle to form a time sequence error signal; according to the time sequence error signal, defining the energy value of each vehicle at each time step as the square weighted value of the time sequence error, and stacking to form an initial energy tensor in the vehicle number and time dimensions; using a bilinear interpolation method to continuously process the initial energy tensor, constructing an energy function that is continuous and differentiable in the vehicle number and time dimensions, and taking the energy function as the tension energy field to reflect the deviation between the vehicle trajectory state and the target trajectory. 9.The method of claim 1, wherein, The equipotential propagation optimization controller takes the current state of the electric vehicle, the optimal intended trajectory, and the tension energy field as input; taking the energy value at each time step as a reference, the equipotential propagation algorithm is used to backtrack from the end time step to the front, and the action with the minimum energy value is selected at each time step; the charging and discharging actions of each electric vehicle in the entire scheduling period are generated in sequence as the charging and discharging action plan of the electric vehicle. 10.A deep reinforcement learning based electric vehicle group charging and discharging optimization control system, performing a deep reinforcement learning based electric vehicle group charging and discharging optimization control method according to any one of claims 1 to 9, characterized in that, The system comprises the following modules: a state perception module for collecting state information and environment information of the electric vehicle group and constructing a state input vector; a causal extraction module configured to perform causal key variable extraction on the state input vector based on a Granger causality test to generate a compressed state vector; a trajectory generation module configured to input the compressed state vector into a conditional diffusion model to generate a plurality of target electric quantity trajectories corresponding to each electric vehicle, thereby forming an intention trajectory candidate set; a trajectory optimization module configured to perform trajectory simulation evaluation on the intention trajectory candidate set based on an improved dream optimization algorithm, to perform belief attribution and sampling correction on intention trajectories with substandard performance through a self-supervised multi-resolution belief error correction stage, and to output an optimal intention trajectory; a tension energy field construction module configured to construct a tension energy field defined in the vehicle and time dimensions based on the optimal intention trajectory; an optimal control module configured to input the current state of the electric vehicle, the optimal intention trajectory, and the tension energy field into an equipotential propagation optimal controller, and to generate a charging and discharging action plan for each electric vehicle according to an energy value minimum value criterion.

Images