Charging station dynamic adaptive planning method based on model deduction

Abstract

The invention belongs to the technical field of charging infrastructure planning, and discloses a charging station dynamic adaptive planning method based on model deduction, which comprises the following steps: dividing a target city into a plurality of planning areas, collecting multi-source heterogeneous data of each planning area, and constructing a data block containing time-space coupling information; establishing a simulation environment of the multi-agent game based on the data blocks; a reinforcement learning model comprising a spatial-temporal feature extraction module and a strategy evaluation module is constructed, an output game equilibrium result is used for training, and the long-term value of the station building action of each planning area is evaluated; deploying the trained reinforcement learning model in a virtual environment which can be dynamically updated, performing multi-step strategy deduction in combination with an optimization algorithm, and selecting an optimal construction scheme of the charging station; the problems that in the prior art, charging infrastructure layout and dynamic demand matching is insufficient, a multi-body game decision mechanism is lacked, and long-term adaptability optimization capacity is lacked are solved.

Description

Technical Field

[0001] This invention belongs to the field of charging infrastructure planning technology, specifically relating to a dynamic adaptive planning method for charging stations based on model deduction. Background Technology

[0002] With the rapid development of the global electric vehicle industry, the contradiction between the continuous growth of electric vehicle ownership and the planning and construction of charging infrastructure is becoming increasingly prominent. Currently, there is a significant structural imbalance in the layout of charging infrastructure: some areas have highly concentrated charging demand but insufficient facilities, leading to excessively long waiting times and a decline in service experience for users; while other areas suffer from resource idleness and operational losses due to over-construction. This supply-demand mismatch severely restricts the healthy development of the electric vehicle industry. Existing research methods are mostly based on static assumptions and cannot effectively address complex factors such as dynamic urban development, multi-stakeholder interest games, and changes in the market environment. Traditional planning models often ignore the interactive influence between the government, operators, and users, lack the ability to capture spatiotemporal dynamic characteristics, and most solutions remain at the single-decision level, failing to establish an effective feedback optimization mechanism. Furthermore, existing algorithms are computationally inefficient when dealing with large-scale urban real-time planning problems, making it difficult to support dynamic adjustment needs. Therefore, existing technologies suffer from problems such as insufficient matching between charging infrastructure layout and dynamic demand, a lack of multi-stakeholder game-based decision-making mechanisms, and a lack of long-term adaptive optimization capabilities. Summary of the Invention

[0003] To address the shortcomings of existing technologies, the present invention aims to provide a model-based dynamic adaptive planning method for charging stations, which solves the problems of insufficient matching between charging infrastructure layout and dynamic demand, lack of multi-agent game decision-making mechanism, and lack of long-term adaptive optimization capability in existing technologies.

[0004] The objective of this invention can be achieved through the following technical solutions: A model-based dynamic adaptive planning method for charging stations includes the following steps: The target city is divided into multiple planning areas, and multi-source heterogeneous data from each planning area are collected to construct a data block containing spatiotemporal coupling information. A simulation environment for multi-agent game based on data blocks is established, including: S21. Construct a user behavior response model that includes perceived error to simulate the probability distribution of user choices for electric vehicles under different planning areas and price conditions; S22. Construct an operator revenue assessment model to calculate the operator's net revenue based on the user selection probability distribution, the charging price in the planned area, and the construction and operation costs. S23. Construct a government decision-making model to set constraints and subsidy standards for planned regional deployments based on operators' net revenue, and guide operators to participate in site construction; S24. Simulate market interaction behavior through a game equilibrium solution algorithm and output the game equilibrium results of the simulation environment; A reinforcement learning model containing a spatiotemporal feature extraction module and a policy evaluation module is constructed. The model is trained using the output game equilibrium results to evaluate the long-term value of station construction actions in each planning area. The trained reinforcement learning model is deployed in a dynamically updated virtual environment, and multi-step strategy deduction is performed in combination with optimization algorithms to select the optimal construction scheme for the charging station. After implementing the optimal construction plan for charging stations, the virtual environment is corrected by collecting actual operating data of the charging stations, and subsequent charging station planning decisions are continuously optimized through a closed-loop iterative mechanism.

[0005] Constructing a data block containing spatiotemporal coupling information specifically includes the following steps: Collect multi-source heterogeneous data, including spatial and temporal data, from various planning areas; Spatial data includes geographic points of interest, road network density, population density, existing charging facility capacity, grid transformer allocable capacity and its rate of change; Temporal data includes time-series data that organizes spatial data by timestamps; Multi-source heterogeneous data is processed into grids according to spatial coordinates and timestamps to form multi-channel data blocks containing spatiotemporal coupling information.

[0006] Constructing a user behavior response model that includes perceived errors specifically involves the following steps: Based on the number of geographical points of interest and the capacity of existing charging facilities in the planning area, calculate the attractiveness of each planning area; By introducing a random component to represent the perception error and comprehensively considering factors such as attractiveness, distance, charging price, queuing time, and driving time, a utility function for users to choose charging stations is constructed. Based on the utility function, a Logit probability model is used to establish a user charging choice probability function to predict the user choice probability distribution, thereby completing the construction of the user behavior model.

[0007] The specific steps involved in constructing an operator revenue assessment model are as follows: Receive the user selection probability distribution output by the user behavior model; Using the planned areas where stations may be built in the future as candidate areas, the potential charging flow of the candidate areas is calculated based on the electric vehicle ownership and user selection probability distribution in the candidate areas. Based on the charging price, electricity procurement cost, and charging power limit of the candidate area, calculate the theoretical revenue that a single charging station can generate after the candidate area is built and put into operation. Based on theoretical benefits, the investment costs of building sites in candidate areas, penalties for over-deployment, and operation and maintenance costs are added to output the net revenue generated by operators after building sites in candidate areas, thus completing the construction of the operator revenue evaluation model.

[0008] The construction of a government decision-making model includes the following steps: Set grid capacity constraints based on the allocable capacity of grid transformers in the planning area, and set the set of planning areas where substations may be built in the future, as well as the maximum number of substations to be built; Set a constraint that you can select at most one planning area for charging station deployment within each planning cycle; The social welfare contribution of each candidate area is calculated based on the operator's net revenue, the user selection probability distribution, and the road network density. Differentiated subsidies are set based on contributions to social welfare, generating government regulation strategies; To complete the construction of a government decision-making model for outputting the planning areas where station construction is permitted, the maximum number of stations, and subsidy standards.

[0009] The game equilibrium solution algorithm simulates market interaction behavior and outputs the game equilibrium results of the simulation environment. The specific steps include: Based on the operator's net profit excluding government subsidies output by the operator's revenue assessment model, the subsidy standard output by the government decision-making model, and the user choice probability distribution output by the user behavior response model, a set of game equilibrium equations is constructed, which includes the operator's total profit and the user choice probability distribution; wherein, the operator's total profit is obtained by adding the operator's net profit and government subsidies. Initialize the queuing time vector and charging price vector for each planning area; The fixed-point iteration method is used to quickly solve the game equilibrium equations, and the charging price vector and queuing time vector are updated alternately until the preset convergence condition is met. The outputs the equilibrium charging price, equilibrium queuing time, and equilibrium user choice probability distribution for each planning area when the Nash equilibrium is reached, serving as the game equilibrium result of the simulation environment.

[0010] A reinforcement learning model is constructed, comprising a spatiotemporal feature extraction module and a policy evaluation module. The model is trained using the output game equilibrium results to evaluate the long-term value of station construction actions in each planned area. The specific steps include: The data block containing spatiotemporal coupling information is input into the spatiotemporal feature extraction module. The spatiotemporal feature extraction module integrates the residual network and the temporal convolutional network to extract spatiotemporal features containing the spatial difference features and temporal evolution trend features of the planning area. The extracted spatiotemporal features and game equilibrium results are input into the strategy evaluation module to calculate the state-action value function of the station construction actions in each planning area. In a simulation environment, the strategy evaluation module is iteratively trained using an experience replay mechanism and a greedy strategy. When the fluctuation range of the state-action value function is lower than a preset threshold or the total number of training rounds reaches a preset upper limit, the model is determined to have converged, and a reinforcement learning model that can evaluate the long-term value of site construction actions in each planning area is obtained.

[0011] The trained reinforcement learning model is deployed in a dynamically updated virtual environment, and multi-step policy deduction is performed using optimization algorithms to select the optimal construction scheme for charging stations. This process includes the following steps: The virtual environment is initialized based on multi-source heterogeneous data collected from various planning areas. The virtual environment includes user behavior response model, operator revenue assessment model, government decision-making model, subsidy standards and power grid capacity constraints. Obtain the current city status parameters and load them into the virtual environment; The trained reinforcement learning model is invoked and combined with the Monte Carlo tree search algorithm to perform multi-step deduction in a virtual environment, executing the selection, expansion, simulation, and feedback processes to simulate the cumulative social welfare generated in multiple planning cycles under different site construction decisions. The game equilibrium solution algorithm is embedded at each decision point in the simulation process to solve the equilibrium charging price, equilibrium queuing time and equilibrium user selection probability distribution corresponding to the decision to build a station in the candidate area. The present value of the cumulative social welfare over multiple planning cycles is used as the evaluation indicator to update the access frequency and value statistics of each projection path; The simulation stops when a predetermined stopping condition is met. Based on the value statistics, the candidate area with the highest cumulative social welfare after discounted calculation is selected for the station construction decision, and this is determined as the optimal construction plan for the charging station at the current stage.

[0012] Stopping conditions include the upper limit of simulation depth or simulation rounds, or the difference in evaluation indicators being less than a preset difference threshold.

[0013] The virtual environment is corrected by collecting actual operating data from charging stations, and subsequent charging station planning decisions are continuously optimized through a closed-loop iterative mechanism. This process includes the following steps: After the charging station is built and put into operation for a predetermined period, actual operational data including attractiveness, grid transformer capacity, and electric vehicle ownership will be collected in each planned area. Based on the collected actual operation data, the recursive Bayesian regression method was used to update the parameters of attractiveness of each planning area in the virtual environment, the allocable capacity of the power grid transformer and the number of electric vehicles, and to obtain the updated mean and covariance of each parameter. Using the updated data-driven virtual environment, repeat the multi-step strategy deduction process to determine the optimal construction plan for the next stage; When the covariance of the updated attractiveness, the allocable capacity of the grid transformer, or the number of electric vehicles exceeds a set threshold, or when the cumulative social welfare calculated after discounting is lower than expected, incremental training of the reinforcement learning model is triggered, thereby realizing a closed-loop optimization mechanism of planning-deployment-feedback-replanning.

[0014] The beneficial effects of this invention are: This invention constructs data blocks containing spatiotemporally coupled information using multi-source heterogeneous data, establishes a multi-agent game simulation environment integrating user behavior response, operator revenue assessment, and government decision-making and regulation, solves the game equilibrium using a fixed-point iterative method, evaluates the long-term value of site construction actions by combining spatiotemporal-aware neural networks and reinforcement learning models, and performs multi-step strategy deduction based on Monte Carlo tree search. Finally, Bayesian correction is applied to the virtual environment using actual operational data to form a closed-loop optimization mechanism. This effectively solves the problems of mismatch between charging infrastructure layout and dynamic demand, and the lack of a multi-agent game decision-making mechanism, significantly improving the adaptability, accuracy, and long-term economic benefits of the planning scheme. Attached Figure Description To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a schematic diagram of the overall process of the dynamic adaptive planning method for charging stations of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] like Figure 1 As shown, a model-based dynamic adaptive planning method for charging stations includes the following steps: The target city is divided into multiple planning areas, and multi-source heterogeneous data from each planning area are collected to construct a data block containing spatiotemporal coupling information. A simulation environment for multi-agent game based on data blocks is established, including: S21. Construct a user behavior response model that includes perceived error to simulate the probability distribution of user choices for electric vehicles under different planning areas and price conditions; S22. Construct an operator revenue assessment model to calculate the operator's net revenue based on the user selection probability distribution, the charging price in the planned area, and the construction and operation costs. S23. Construct a government decision-making model to set constraints and subsidy standards for planned regional deployments based on operators' net revenue, and guide operators to participate in site construction; S24. Simulate market interaction behavior through a game equilibrium solution algorithm and output the game equilibrium results of the simulation environment; A reinforcement learning model containing a spatiotemporal feature extraction module and a policy evaluation module is constructed. The model is trained using the output game equilibrium results to evaluate the long-term value of station construction actions in each planning area. The trained reinforcement learning model is deployed in a dynamically updated virtual environment, and multi-step strategy deduction is performed in combination with optimization algorithms to select the optimal construction scheme for the charging station. After implementing the optimal construction plan for charging stations, the virtual environment is corrected by collecting actual operating data of the charging stations, and subsequent charging station planning decisions are continuously optimized through a closed-loop iterative mechanism. The target city is divided into multiple planning areas, and multi-source heterogeneous data from each planning area are collected to construct data blocks containing spatiotemporal coupling information. The specific steps include: Based on the geographic information system platform, the target city for the construction of charging stations is divided into multiple planning areas; Collect multi-source heterogeneous data, including spatial and temporal data, from various planning areas; Spatial data includes geographic points of interest, road network density, population density, existing charging facility capacity, grid transformer allocable capacity and its rate of change; Temporal data includes time-series data that organizes spatial data by timestamps; Multi-source heterogeneous data is gridded according to spatial coordinates and timestamps to form multi-channel data blocks containing spatiotemporal coupling information; Constructing a user behavior response model that includes perceived errors specifically involves the following steps: Based on the number of geographical points of interest and the capacity of existing charging facilities in the planning area, calculate the attractiveness of each planning area; By introducing a random component to represent the perception error and comprehensively considering factors such as attractiveness, distance, charging price, queuing time, and driving time, a utility function for users to choose charging stations is constructed. Based on the utility function, a Logit probability model is used to establish a user charging choice probability function to predict the user choice probability distribution, thereby completing the construction of the user behavior model. The specific formula for calculating attractiveness is as follows: in, This represents the attractiveness of planning area i at the t-th planning stage; , Let represent the number of geographic points of interest (POIs) and the existing charging capacity of the planning area i at the t-th planning stage, respectively. , These represent the weighting coefficients for the number of geographic points of interest and the capacity of existing charging facilities, respectively. Utility function The specific expression is as follows: in, and Let them represent the deterministic component and the random component, respectively. They are independent and identically distributed and follow an extreme value distribution; , , , , These are the influencing factors: attractiveness, distance, charging price, queuing time, and driving time. The factor influencing road network density; Indicates the first t The planning area during the second planning stage y Attraction; Indicates the charging price; Indicates from the planning area y To the planned area x The shortest distance to the fixed point; Indicates the first t Planning area for the next planning period y The sum of the queuing times for all electric vehicles that require charging. Indicates the first t Planning area for the next planning period y The sum of the driving time of all electric vehicles in the area that require charging; The specific calculation formula is as follows: Where I represents the set of all planned areas; The free-flow delay time represents the charging time; These are formal parameters; , Let represent the charging capacity and potential charging demand of the planned area y per unit time, respectively. An increase will lead to longer waiting times in queues; The specific calculation formula is as follows: Where D represents the distance union, ; and Let be the charging distance and average speed of the k-th electric vehicle within the t-th planning time period, respectively; and Traffic conditions in the planned area y during the t-th planning period The specific relationship is as follows: Let y be the number of roads in the planning area at the t-th planning iteration; These are adjustment parameters related to road type; By By standardizing to the legal speed limit range, the average speed of electric vehicles within the planned area y can be obtained. Then calculate the average speed of electric vehicle k. The specific calculation formula is as follows: in, Indicates the first t Electric vehicles during the next planning period k The number of planned areas traversed; l represents any planned area passed through; L k Let k be the set of planned areas traversed by electric vehicle k. User selection probability distribution The calculation is performed according to the following formula: in, Characterizes the error perceived by the user.

[0018] The specific steps involved in constructing an operator revenue assessment model are as follows: Receive the user selection probability distribution output by the user behavior model; Using the planned areas where stations may be built in the future as candidate areas, the potential charging flow of the candidate areas is calculated based on the electric vehicle ownership and user selection probability distribution in the candidate areas. Based on the charging price, electricity procurement cost, and charging power limit of the candidate area, calculate the theoretical revenue that a single charging station can generate after the candidate area is built and put into operation. Based on theoretical benefits, the investment costs of building stations in candidate areas, penalties for over-deployment, and operation and maintenance costs are added to output the net revenue generated by operators after building stations in candidate areas, so as to complete the construction of the operator revenue evaluation model. Potential charging flow The calculation formula is as follows: in, and These represent the number of electric vehicles in planning area x and their growth rate during the t-th planning period; The theoretical revenue that a single charging station can generate is determined by its own pricing. Other carrier response strategies The combined effect is expressed as follows: in, This represents the theoretical revenue that a single charging station can generate. This represents the retail electricity cost in planning area i during the t-th planning period; This represents the average charging requirement for each electric vehicle. The operator revenue assessment model is a game equilibrium problem involving multiple stakeholders, including operators and users, and users themselves. When price vector When the following conditions are met: This establishes a spatial pricing equilibrium, where no operator can unilaterally increase profits through price adjustments, thus creating stable user charging demand. ; During the overall planning phase, the government needs to consider the overall benefits rather than the interests of individual charging stations. As compensation, subsidies can be provided for charging stations in certain planned areas to attract their deployment; therefore, the overall net revenue of charging station operators... The expression is as follows: in, , , , Converted to the first t The potential market share, investment costs, operation and maintenance costs, and over-deployment penalties exceeding the rated capacity of transformers for urban operators at the time of this planning are calculated using the following formulas: If in the 1st t The planning area during the second planning stage i If the increase in charging capacity is less than the remaining capacity of the transformer, no penalty is imposed. It is the cash discount rate; , , The lifecycle investment cost conversion factor is used to convert lifecycle investment capital into equivalent costs corresponding to the time step. , , These represent the average service life of charging equipment, power distribution cables, and charging station sites, respectively. and These represent installation costs and land costs, respectively. and These are the unit cost of the cable and the length of the distribution wire, respectively. The unit maintenance cost of charging equipment; The construction of a government decision-making model includes the following steps: Set grid capacity constraints based on the allocable capacity of grid transformers in the planning area, and set the set of planning areas where substations may be built in the future, as well as the maximum number of substations to be built; Set a constraint that you can select at most one planning area for charging station deployment within each planning cycle; The social welfare contribution of each candidate area is calculated based on the operator's net revenue, the user selection probability distribution, and the road network density. Differentiated subsidies are set based on contributions to social welfare, generating government regulation strategies; To complete the construction of a government decision-making model for outputting the planning areas where station construction is permitted, the maximum number of stations, and subsidy standards; The government needs to guide the market through government subsidy mechanisms and subsidize prices. The planning area is determined by the t-th planning time. i Net revenue of operators Average net operator revenue of all operators in the city The following was decided jointly: in, , , These represent the investment cost, operation and maintenance cost, and over-deployment penalty for operator i in the planned area at the time of the t-th planning stage; The mathematical expressions for the constraints of the government decision-making model are as follows: When the site is built in planning region i during the t-th planning iteration, then the decision variables are... The value is 1 if it is not 0 otherwise; Limit the maximum number of stations to be built during the entire planning period and control the overall investment scale: in, This represents the maximum number of stations to be built during the entire planning period. Among them, ensuring the area i In the t The total load balancing of charging stations during the next planning phase and base load Not exceeding the capacity of the local transformer. If the capacity exceeds the transformer's rated capacity, an over-deployment penalty will be calculated.

[0019] The game equilibrium solution algorithm simulates market interaction behavior and outputs the game equilibrium results of the simulation environment. The specific steps include: Based on the operator's net profit excluding government subsidies output by the operator's revenue assessment model, the subsidy standard output by the government decision-making model, and the user choice probability distribution output by the user behavior response model, a set of game equilibrium equations is constructed, which includes the operator's total profit and the user choice probability distribution; wherein, the operator's total profit is obtained by adding the operator's net profit and government subsidies. Initialize the queuing time vector and charging price vector for each planning area; The fixed-point iteration method is used to quickly solve the game equilibrium equations, and the charging price vector and queuing time vector are updated alternately until the preset convergence condition is met. The output shows the equilibrium charging price, equilibrium queuing time, and equilibrium user choice probability distribution for each planning area when the Nash equilibrium is reached, serving as the game equilibrium result of the simulation environment. To improve the efficiency of solving game equilibrium, this invention employs the fixed-point iteration (FPI) method for rapid solution. First, the FPI equation for spatial pricing equilibrium is established, and the expressions for the operator's theoretical revenue and its derivative are as follows: Since the factor outside the parentheses in the above equation is greater than zero, the following FPI equation can be derived: Based on the sum of queuing waiting times Calculation and potential charging flow The calculation formula can be used to establish the FPI iterative equation for queuing time, as follows: in, for The iteration step size adjustment parameter; The fixed-point iteration (FPI) method specifically includes the following steps: Step 1: Initialization ,vector Step 2: Update the queue time; Step 3: Update the equilibrium electricity price; Step 4: Determine the convergence condition; if Output solution Otherwise, update. , , Return to step two; This application uses the Nash equilibrium index. NI Assessing the quality of balance: If Nash equilibrium index NI A value approaching 1 indicates that the strategy is gradually approaching the equilibrium solution, while a value away from 1 indicates that there is significant room for improvement.

[0020] A reinforcement learning model is constructed, comprising a spatiotemporal feature extraction module and a policy evaluation module. The model is trained using the output game equilibrium results to evaluate the long-term value of station construction actions in each planned area. The specific steps include: The data block containing spatiotemporal coupling information is input into the spatiotemporal feature extraction module. The spatiotemporal feature extraction module integrates the residual network and the temporal convolutional network to extract spatiotemporal features containing the spatial difference features and temporal evolution trend features of the planning area. The extracted spatiotemporal features and game equilibrium results are input into the strategy evaluation module to calculate the state-action value function of the station construction actions in each planning area. In a simulation environment, the strategy evaluation module is iteratively trained using an experience replay mechanism and a greedy strategy. When the fluctuation range of the state-action value function is lower than a preset threshold or the total number of training rounds reaches a preset upper limit, the model is determined to have converged, and a reinforcement learning model that can evaluate the long-term value of site construction actions in each planning area is obtained.

[0021] The trained reinforcement learning model is deployed in a dynamically updated virtual environment, and multi-step policy deduction is performed using optimization algorithms to select the optimal construction scheme for charging stations. This process includes the following steps: The virtual environment is initialized based on multi-source heterogeneous data collected from various planning areas. The virtual environment includes user behavior response model, operator revenue assessment model, government decision-making model, subsidy standards and power grid capacity constraints. Obtain the current city status parameters and load them into the virtual environment; The trained reinforcement learning model is invoked and combined with the Monte Carlo tree search algorithm to perform multi-step deduction in a virtual environment, executing the selection, expansion, simulation, and feedback processes to simulate the cumulative social welfare generated in multiple planning cycles under different site construction decisions. The game equilibrium solution algorithm is embedded at each decision point in the simulation process to solve the equilibrium charging price, equilibrium queuing time and equilibrium user selection probability distribution corresponding to the decision to build a station in the candidate area. The present value of the cumulative social welfare over multiple planning cycles is used as the evaluation indicator to update the access frequency and value statistics of each projection path; The simulation stops when a predetermined stopping condition is met. Based on the value statistics, the candidate area with the highest cumulative social welfare after discounting is selected for the station construction decision, and this is determined as the optimal construction plan for the charging station at the current stage. Stopping conditions include the upper limit of simulation depth or simulation rounds, or the difference in evaluation indicators being less than a preset difference threshold.

[0022] The virtual environment is corrected by collecting actual operating data from charging stations, and subsequent charging station planning decisions are continuously optimized through a closed-loop iterative mechanism. This process includes the following steps: After the charging station is built and put into operation for a predetermined period, actual operational data including attractiveness, grid transformer capacity, and electric vehicle ownership will be collected in each planned area. Based on the collected actual operation data, the recursive Bayesian regression method was used to update the parameters of attractiveness of each planning area in the virtual environment, the allocable capacity of the power grid transformer and the number of electric vehicles, and to obtain the updated mean and covariance of each parameter. Using the updated data-driven virtual environment, repeat the multi-step strategy deduction process to determine the optimal construction plan for the next stage; When the covariance of the updated attractiveness, the allocable capacity of the grid transformer, or the number of electric vehicles exceeds the set threshold, or when the cumulative social welfare calculated after discounting is lower than the expected level, incremental training of the reinforcement learning model is triggered, thereby realizing a closed-loop optimization mechanism of planning-deployment-feedback-replanning. The specific methods for modifying the virtual environment are as follows: Define state vector , , , These represent the attractive force, the allocable capacity of the power grid transformer, and the number of electric vehicles, respectively; exogenous inputs. It consists of the government's decision on station construction and the result of the game between operators and users (equilibrium electricity price); The expression for the prediction model is as follows: The parameter to be updated (d = input dimension) Indicates noise It follows a normal distribution with a mean of 0 and a variance of . ; Posterior knowledge observations are smoothed using KF to obtain the posterior mean and covariance. After smoothing with a window size of N, the posterior mean and covariance are then... t and t-1 The best estimate of the state at any given time (mean vector) , ; Posterior covariance matrix of state estimation and , described in t and t-1 The magnitude and correlation of uncertainty in time-smoothing estimation; posterior covariance matrix across time steps. , indicating after smoothing, t Time state and t−1 Uncertainty relationships between states at different times; This application will use exogenous input Treating them as deterministic quantities, we construct joint regression independent variables and calculate the expected second moments using the posterior mean and covariance. This leads to the second moments of the independent variables and the cross-second moments of the dependent variable and independent variables, as detailed below: in, It is the second moment of the regression independent variable; The second moments of the dependent and independent variables are used; the posterior moments are used to correct the bias. This application references the forgetting factor. The update formulas for the second moments of the independent variables and the cross second moments of the dependent variable versus the independent variables are as follows: First update order , ; It is an extremely small random number. It is the identity matrix; The formula for updating model parameters is: ; Update process noise covariance The expression is as follows: The expression for the pre-defined mean is as follows: The expression for the prediction covariance matrix is: .

[0023] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0024] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the present invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A model-based deduction method for dynamic adaptive planning of charging stations, characterized in that, The method comprises the following steps: dividing a target city into multiple planning areas, collecting multi-source heterogeneous data of each planning area, and constructing a data block containing spatio-temporal coupling information; based on the data block, a simulation environment of multi-agent game is established, the game is simulated and the game equilibrium result is output, including: S21. Construct a user behavior response model containing perception error, which is used to simulate the user selection probability distribution of electric vehicles in different planning areas and price conditions; S22. Construct an operator revenue evaluation model for calculating the net revenue of the operator according to the user selection probability distribution, the charging price of the planning area and the construction and operation cost; S23. Construct a government decision-making model for setting the constraints and subsidy standards of the planning area deployment according to the net revenue of the operator, and guide the operator to participate in the station construction; S24. Simulate the market interaction behavior by using the game equilibrium solving algorithm, and output the game equilibrium result of the simulation environment; construct a reinforcement learning model containing a spatio-temporal feature extraction module and a strategy evaluation module, train the output game equilibrium result, and evaluate the long-term value of the station construction action in each planning area; deploy the trained reinforcement learning model in a virtual environment that can be dynamically updated, combine the optimization algorithm to perform multi-step strategy deduction, and select the optimal construction scheme of the charging station; after implementing the optimal construction scheme of the charging station, collect the actual operation data of the charging station to correct the virtual environment, and continuously optimize the subsequent charging station planning decision through a closed-loop iteration mechanism.

2. The model-based deduction method for dynamic adaptive planning of charging stations according to claim 1, characterized in that, Construct a data block containing spatio-temporal coupling information, which comprises the following steps: collect multi-source heterogeneous data including spatial data and temporal data in each planning area; wherein the spatial data includes geographic interest points, road network density, population density, existing charging facility capacity, power grid transformer allocatable capacity and its change rate; the time data includes time series data sorted by time stamp of the spatial data; perform grid processing on the multi-source heterogeneous data according to the spatial coordinates and time stamp to form a multi-channel data block containing spatio-temporal coupling information.

3. The model-based deduction method for dynamic adaptive planning of charging stations according to claim 2, characterized in that, The user behavior response model containing perception error comprises the following steps: based on the number of geographic interest points and the existing charging facility capacity of the planning area, calculate the attractiveness of each planning area; introduce a random component representing perception error, and comprehensively consider the attractiveness, distance, charging price, queuing waiting time and travel time to construct a utility function for user selection of charging stations; based on the utility function, a Logit probability model is used to establish a user charging selection probability function for predicting the user selection probability distribution, thereby completing the construction of the user behavior model.

4. The model-based deduction method for dynamic adaptive planning of charging stations according to claim 3, characterized in that, The operator revenue evaluation model comprises the following steps: receive the user selection probability distribution output by the user behavior model; take the planning area where the charging station may be built as a candidate area, calculate the potential charging flow of the candidate area based on the electric vehicle ownership and the user selection probability distribution of the candidate area; calculate the theoretical revenue that a single charging station can generate after the candidate area is built and operated according to the charging price, the electric energy procurement cost and the charging power limit of the candidate area; Based on the theoretical revenue, superimposed candidate area investment cost, excess deployment penalty and operation and maintenance cost, the output of the candidate area station after the operation of the operator net income, in order to complete the construction of the operator income evaluation model.

5. The model-based deduction method for dynamic adaptive planning of charging stations according to claim 4, characterized in that, The construction of government decision model specifically includes the following steps: According to the transformer allocable capacity of the planning area power grid, set the power grid capacity constraint, set the future possible planning area set of station construction and the maximum number of station construction constraints; Set the constraint of selecting at most one planning area for charging station deployment in each planning period; According to the operator net income, user selection probability distribution and road network density, calculate the social welfare contribution degree of each candidate area; Based on the social welfare contribution degree, set the differential subsidy, and generate the government regulation strategy; To complete the construction of the government decision model for outputting the planning area allowed to build station, the maximum number of station construction and the subsidy standard.

6. The model-based deduction method for dynamic adaptive planning of charging stations according to claim 5, characterized in that, Through the game equilibrium solving algorithm to simulate the market interaction behavior, the game equilibrium result of the simulation environment is output, which specifically includes the following steps: Based on the operator net income without government subsidy output by the operator income evaluation model, the subsidy standard output by the government decision model and the user selection probability distribution output by the user behavior response model, the game equilibrium equation set containing the total profit of the operator and the user selection probability distribution is constructed; wherein, the total profit of the operator is obtained by adding the net income of the operator and the government subsidy; Initialize the queuing time vector and charging price vector of each planning area; The game equilibrium equation set is quickly solved by using the fixed point iteration method, and the charging price vector and the queuing time vector are updated alternately until the preset convergence condition is met; The equilibrium charging price, equilibrium queuing time and equilibrium user selection probability distribution of each planning area when reaching the Nash equilibrium state are output as the game equilibrium result of the simulation environment.

7. The model-based deduction method for dynamic adaptive planning of charging stations according to claim 6, characterized in that, The reinforcement learning model including the space-time feature extraction module and the strategy evaluation module is constructed, which is trained by using the output game equilibrium result to evaluate the long-term value of the station construction action of each planning area, which specifically includes the following steps: The data block containing space-time coupling information is input into the space-time feature extraction module, and the space-time feature extraction module fuses the residual network and the time convolution network to extract the space-time features containing the spatial difference features and the time evolution trend features of the planning area; The extracted space-time features and the game equilibrium result are input into the strategy evaluation module together to calculate the state-action value function of the station construction action of each planning area; In the simulation environment, the experience replay mechanism and the greedy strategy are used to iteratively train the strategy evaluation module. When the fluctuation range of the state-action value function is lower than the preset threshold or the total training round number reaches the preset upper limit, it is determined that the model converges, and the reinforcement learning model capable of evaluating the long-term value of the station construction action of each planning area is obtained.

8. The model-based deduction method for dynamic adaptive planning of charging stations according to claim 7, characterized in that, The trained reinforcement learning model is deployed in a virtual environment that can be dynamically updated, and combined with an optimization algorithm to perform multi-step strategy deduction to select the optimal charging station construction scheme, which specifically includes the following steps: Initialize the virtual environment based on the multi-source heterogeneous data collected in each planning area, which includes the user behavior response model, the operator income evaluation model, the government decision model, the subsidy standard and the power grid capacity constraint; Acquire current urban state parameters and load them into the virtual environment; Call the trained reinforcement learning model and combine it with the Monte Carlo tree search algorithm to perform multi-step reasoning in the virtual environment, execute the selection, expansion, simulation, and feedback processes, and simulate the cumulative social welfare generated in the future planning period under different station construction decisions; Embed a game equilibrium solving algorithm at each decision point in the reasoning process to solve the equilibrium charging price, equilibrium queuing time, and equilibrium user selection probability distribution corresponding to the candidate area station construction decision; Use the present value of the discounted cumulative social welfare in the future planning period as the evaluation index, update the access frequency and value statistics of each reasoning path; Pre-set the stopping conditions for reasoning, and stop reasoning when the stopping conditions are met; Select the candidate area station construction decision with the highest discounted cumulative social welfare based on the value statistics results and determine it as the optimal construction scheme for the current stage.

9. The model-based deduction method for dynamic adaptive planning of charging stations according to claim 7, characterized in that, The stopping conditions include the upper limit of reasoning depth or reasoning rounds, or the difference in evaluation index is less than the pre-set difference threshold.

10. The model-based deduction method for dynamic adaptive planning of charging stations according to claim 8, characterized in that, Collect actual operation data of charging stations to correct the virtual environment and continuously optimize subsequent charging station planning decisions through a closed-loop iteration mechanism, including the following steps: After the charging stations are constructed and operated for a pre-set period, collect actual operation data of each planning area, including attractiveness, grid transformer allocable capacity, and electric vehicle ownership; Based on the collected actual operation data, use the recursive Bayesian regression method to update the posterior of the parameters of attractiveness, grid transformer allocable capacity, and electric vehicle ownership in each planning area in the virtual environment, and obtain the updated mean and covariance of each parameter; Use the updated data to drive the virtual environment and repeat the multi-step strategy reasoning process to determine the optimal construction scheme for the next stage; When the covariance of the updated attractiveness, grid transformer allocable capacity, or electric vehicle ownership exceeds the set threshold, or the discounted cumulative social welfare obtained through reasoning is lower than the expected level, trigger the incremental training of the reinforcement learning model, thereby realizing the closed-loop optimization mechanism of planning-deployment-feedback-replanning.

Images