Motor vehicle charging pile management optimization method and system based on artificial intelligence
By constructing a spatiotemporal flow characteristic matrix and an obstacle Lyapunov constraint strategy, the scheduling strategy of the charging pile management system is optimized, which solves the adaptability and coordination problem of the charging pile management system in the face of dynamic changes, and realizes the dynamic maintenance and efficient management of power grid safety and service quality.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-25
- Publication Date
- 2026-04-03
AI Technical Summary
Existing charging pile management systems lack the ability to comprehensively model traffic flow, power load, and user behavior when facing dynamic changes in electric vehicle charging demand. This results in insufficient adaptability and real-time performance of scheduling strategies, making it difficult to balance grid security and service quality. Furthermore, cross-regional collaborative scheduling lacks accuracy and global optimization.
By collecting operational data from charging pile controllers, vehicle terminals, and power distribution monitoring systems, a spatiotemporal flow characteristic matrix is constructed. The state space and constraints are initialized using the obstacle Lyapunov constraint strategy, scheduling and control actions are calculated and optimized, and feedback data updates are implemented through the system execution module to achieve online closed-loop optimization.
It improves the safety, stability and coordination of charging pile scheduling, can dynamically maintain electrical safety and service quality in uncertain environments, and achieves efficient management of the charging pile network.
Smart Images
Figure CN121787619A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent management technology for charging infrastructure, and in particular to an artificial intelligence-based method and system for optimizing the management of motor vehicle charging piles. Background Technology
[0002] Currently, with the rapid growth in the number of new energy vehicles, the construction and operation scale of charging piles, as their supporting infrastructure, continues to expand. However, in existing technologies, the management and scheduling of charging piles mainly rely on control strategies based on historical statistical data or simple rules, which are difficult to effectively cope with the dynamic changes in electric vehicle charging demand in time and space. Existing systems generally lack the ability to comprehensively model traffic flow, power load, and user behavior when allocating charging pile resources, resulting in insufficient adaptability and real-time performance of scheduling strategies. On the one hand, when charging demand fluctuates significantly in a short period, traditional scheduling methods often fail to balance grid security and service quality, easily leading to local overload or service congestion. On the other hand, in cross-regional, multi-network collaborative scheduling of charging piles, the lack of accurate spatiotemporal traffic characteristics results in low prediction accuracy, further leading to a lack of coordination and global optimality in scheduling decisions.
[0003] In existing intelligent scheduling research, although some methods have attempted to introduce predictive models such as machine learning and deep learning to improve the accuracy of charging demand forecasting, these methods mostly focus on optimization at a single level and fail to simultaneously consider the dual-layer structure characteristics of transportation and power networks. Furthermore, existing constraint handling mechanisms are mostly based on static or linear safety conditions, lacking dynamic adjustment capabilities. When significant uncertainties or fluctuations occur in the external environment, the strategies are prone to violating electrical safety constraints and service quality constraints, leading to potential operational risks. In addition, for the optimization process of scheduling strategies, traditional methods often remain at the stage of single-time calculation or offline optimization, lacking a closed-loop mechanism for updating the model using real-time execution feedback data, making it impossible for the strategies to continuously adapt to environmental changes.
[0004] Therefore, how to provide an AI-based method and system for optimizing the management of motor vehicle charging stations 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 an artificial intelligence-based method and system for optimizing the management of motor vehicle charging piles. This invention collects operational data and constructs a spatiotemporal flow feature matrix, generates prediction results using a flow prediction module, initializes the state space and constraints using an obstacle Lyapunov constraint strategy, calculates and optimizes scheduling control actions, implements these actions through a system execution module and obtains feedback data, updates flow components using decomposition operations, and achieves online closed-loop optimization, effectively improving the safety, stability, and coordination of charging pile scheduling.
[0006] An artificial intelligence-based method for optimizing the management of motor vehicle charging stations according to an embodiment of the present invention includes the following steps:
[0007] The system collects operational data from the charging pile controller, vehicle terminal, and power distribution monitoring system. The operational data includes the number of vehicles flowing into the station, charging power, voltage and current, transformer load rate, and feeder current. The system then performs timestamp alignment and geographic coordinate unification processing on the operational data to generate a raw traffic dataset with spatiotemporal tags.
[0008] The original traffic dataset is mapped to a two-layer graph structure containing a traffic network layer and a power network layer. The traffic data in the two-layer graph structure is decomposed to obtain gradient flow, curl flow and harmonic flow components. The spatiotemporal traffic feature matrix between stations is constructed based on the component results.
[0009] The spatiotemporal flow feature matrix is input into the flow prediction module to obtain the prediction results of the station inflow rate and flow distribution in future discrete time periods. The prediction results describe the potential load migration trend with gradient flow, describe the local circulating flow intensity with curl flow, and describe the global equilibrium state with harmonic flow.
[0010] The prediction results, along with electrical safety constraint parameters and service quality constraint parameters, are input into the scheduling strategy optimization module to initialize the state space and constraint conditions of the obstacle Lyapunov constraint strategy model.
[0011] Based on the state space and constraints, the predicted gradient flow, curl flow and harmonic flow values are used as environmental state variables to calculate the initial control action that satisfies the stability conditions.
[0012] Substituting the initial control action into the obstacle constraint optimization process, the optimal scheduling action under the constraint conditions is obtained by iterative solution. The scheduling action includes the station-level access power setting value and the queuing release order.
[0013] The optimal scheduling action is input into the system execution module to control the charging pile network in real time, and the traffic and power flow change data after the execution is collected and verified to obtain a feedback dataset.
[0014] The flow and tidal change data are decomposed to update the gradient flow, curl flow, and harmonic flow components, thereby achieving online closed-loop optimization of the obstacle Lyapunov constraint strategy model.
[0015] Optionally, the generation of the original traffic dataset specifically includes:
[0016] The system obtains the operating data of each charging pile in the current time period from the charging pile controller, the vehicle location and charging status data related to the current charging pile from the vehicle terminal, and the power distribution load data corresponding to the site where the current charging pile is located from the power distribution monitoring system. The three types of data are recorded in the raw data table according to the collection time.
[0017] The original data table is sorted according to a unified time base, and the time field of all records is normalized using a unified time base. The time representation is unified to Coordinated Universal Time format and accurate to milliseconds, eliminating clock deviations between different devices and generating an aligned data table with consistent time order and no missing data.
[0018] All location-related fields in the alignment data table are unified to the same coordinate reference, and the location data is merged with the corresponding time labels and operational status data to generate a spatiotemporally labeled raw traffic dataset with time labels and geographic coordinates.
[0019] Optionally, the construction of the spatiotemporal flow feature matrix specifically includes:
[0020] Read the original traffic dataset with spatiotemporal labels, group and sort it according to the unique site code and unified time base, retain the time label, spatial coordinates and operation status data, and form a site-level spatiotemporal sequence arranged in chronological order;
[0021] Based on the site-level spatiotemporal sequence, a two-layer graph structure including a traffic network layer and a power network layer is constructed. In the traffic network layer, site nodes and travel path edges are established, and in the power network layer, site nodes, transformer nodes, feeder nodes and electrical connection edges are established, and cross-layer mapping relationships are established.
[0022] By combining a two-layer graph structure with a station-level spatiotemporal sequence, the flow data of nodes and edges is decomposed and reconstructed. For each time window, each travel path edge is assigned a directed flow value at the traffic network layer according to the convention that the starting station number is less than the ending station number. The directed flow is then decomposed into three components to obtain components reflecting the potential load migration direction, components reflecting the local cycle intensity, and components reflecting the global balance characteristics. The three components are recorded for each station under each time label.
[0023] The three components are combined according to site pairing and time label to generate a spatiotemporal traffic feature matrix between sites that covers all sites and time ranges.
[0024] Optionally, obtaining the measurement result specifically includes:
[0025] Read the spatiotemporal traffic feature matrix between stations, organize it according to the pairing relationship between stations and adjacent stations as the row order and the order of time tags as the column order, and maintain the fixed arrangement of the traffic network layer and power network layer components in each column to complete the construction of the input data block;
[0026] Based on the current decision moment, a fixed-length historical time window is continuously extracted from the input data block to form a historical sequence arranged in ascending order of time. At the same time, several consecutive discrete time periods immediately following the current moment are set as the prediction time domain, ensuring that the site pairing index and component order are completely consistent between the historical sequence and the prediction time domain.
[0027] The historical sequence is input into the traffic prediction module, and prediction result tables are generated one by one according to the site pairing and prediction period. The prediction result table provides four types of indicators for each site in each prediction period: site inflow rate, charging start rate, queue length increment and available access power, and the generated fields are recorded in sequence.
[0028] The system performs integrity checks and boundary checks on the field records, removes records with missing time tags, and prunes negative values to zero or values exceeding the capacity limit to the corresponding limit. The system then summarizes the data in ascending order of site as primary key and time tag to obtain the prediction results for future discrete time periods.
[0029] Optionally, the initialization of the state space and constraints specifically includes:
[0030] The predicted inflow rate, predicted charging start rate, predicted queue length change and predicted available access power of each site in the prediction results of future discrete time periods are combined into a state vector in a fixed order. At the same time, the site-level access power allocation, queue release order and cross-site diversion ratio to be optimized are combined into a control vector to form a state control pair.
[0031] Receive the state control pair, combine the electrical safety constraint parameters and service quality constraint parameters, establish a set of feasible domain functions in the space of the state control pair, and map the values of the feasible domain functions to the state control pair one by one;
[0032] A Lyapunov constraint strategy model is constructed, which receives a set of feasible domain functions and state control pairs. The composite energy evaluation value for each time period is calculated. The time-period difference of composite energy is calculated item by item along the site pairing dimension, and then superimposed in chronological order to form a time-period prefix difference sequence. This sequence is matched one-to-one with convergence coefficients and written into constraint slots to generate the first type of constraint, requiring that the composite energy in the next time period be lower than the composite energy of the current time period scaled by the convergence coefficient. Simultaneously, a forward difference approximation is performed on the feasible domain function for each site pair to obtain the numerical vector of the directional derivative. The weighted sum of the directional change rate of the feasible domain function and the current numerical vector is calculated to generate the second type of constraint, requiring that the current result is not less than zero for all site pairs. The two types of constraints are bound to the state control pairs to form a constraint set, completing the constraint initialization.
[0033] The system receives a set of constraints and a state control pair, sets a reference control quantity for each prediction period, generates a corresponding state space weight matrix, and combines them to form an initial objective function. Using the reference control quantity as the starting value, iterative update calculations are performed, and the updated control quantity is substituted into the constraint set for feasibility verification. Control quantities that do not meet the constraints are corrected by reverting to the previous iteration, and the corrected control quantities continue to participate in the next iteration. When the change in the objective function is less than a set threshold for two consecutive iterations and all constraints are met, the current control quantity is used as the optimal scheduling state space, and the initialization of the state space of the obstacle Lyapunov constraint policy model is completed. The set threshold is based on the 95th quantile of the historical objective function change distribution. During the iteration process of binding the reference control quantity to constraints, a Bregman-Logit type regularization term is used to form the initial iterative optimization cost.
[0034] Optionally, the calculation process for the initial control action specifically includes:
[0035] The state space and constraints are combined with the configured electrical safety constraint parameters and service quality constraint parameters to form a set of environmental state variables for the current prediction period;
[0036] Based on the set of environmental state variables, a mapping relationship from the current state and control variables to the state of the next time period is established, and stability descent constraints are generated by combining the convergence coefficient, while directional safety constraints are generated by combining the feasible region function.
[0037] Stability degradation constraints and safety feasibility constraints are applied sequentially to environmental state variables to determine the operational status of each site during the prediction period. The constraints are then contracted in conjunction with stability conditions to form a set of constraints that can be used for calculation. The stability condition is that the change in the objective function of the obstacle Lyapunov constraint strategy model is less than a set threshold for two consecutive iterations.
[0038] The initial control quantity is gradually adjusted based on the set of constraints available for calculation. In each adjustment, the control components that cause the stability condition to become unbalanced are corrected first, and then the control components that cause the feasibility constraint to be violated are corrected. After each round of adjustment, the stability function value and safety margin are recalculated. The current control quantity obtained by adjustment is used as the initial control action that satisfies the stability condition.
[0039] Optionally, obtaining the optimal scheduling action specifically includes:
[0040] The initial control action is used as the starting solution for iterative optimization. The obstacle Lyapunov constraint policy model is loaded, and the stability descent constraint and safety feasibility constraint are imported into the optimization solver.
[0041] Construct an iterative optimization objective function, where the first part is a quadratic penalty term for the control quantity deviating from the reference control quantity, and the second part is a weighted barrier function penalty term. The barrier function consists of the logarithmic sum of each feasible domain function and is used to increase the penalty value when approaching the constraint boundary. The two parts are combined with a fixed proportional coefficient to form the complete objective function.
[0042] Based on the complete objective function, the control quantity is updated according to the gradient direction of the objective function and the constraints. Then, the control quantity that does not meet the constraints is back-adjusted along the outward normal direction to obtain the corrected control quantity, which is then substituted into the objective function to continue iterating until the control quantity meets the constraints and re-enters the feasible region, thus obtaining the optimal scheduling action.
[0043] Optionally, the collection of the flow and current change data specifically includes:
[0044] Receive the optimal scheduling action, and combine the control instructions in the optimal scheduling action, including site access power allocation, queuing release order and cross-site diversion ratio, according to the execution order, and specify the corresponding execution time and target location for each item to form a sequence of instructions that can be directly issued;
[0045] The instruction sequence is input into the system execution module to complete the power adjustment, queue update and flow direction switching operations in sequence, while continuously recording relevant operating data during the execution process;
[0046] Real-time collection of traffic and power flow change data of the charging pile network. The traffic change data includes changes in the number of vehicles accessing each station per unit time and changes in the charging start rate. The power flow change data includes changes in the station power curve and changes in the feeder current. The traffic and power flow change data are stored in correspondence with the execution time label and the station number.
[0047] The flow and current change data are timestamped and validated for consistency. Missing or abnormal data points are removed, and the flow and current change data after the validation is performed are collected to obtain a feedback dataset.
[0048] Optionally, the implementation of the online closed-loop optimization specifically includes:
[0049] Receive feedback datasets, extract traffic change data and power flow change data related to the operation of each site, and combine the two types of data into a unified spatiotemporal data matrix according to time order and site number;
[0050] The spatiotemporal data matrix is mapped to a graph structure, and Hodge decomposition is performed on the edge flow distribution in the graph. The original flow field is decomposed into three components: gradient flow component, curl flow component, and harmonic flow component. The gradient flow component corresponds to the global situational energy change, the curl flow component corresponds to the local loop flow, and the harmonic flow component corresponds to the inherent pattern of the network topology.
[0051] The state space of the obstacle Lyapunov constraint strategy model is updated using ternary component components. The gradient flow component is used to correct the state prediction of the global stability function, the curl flow component is used to correct the directional changes of local constraints, and the harmonic flow component is used to correct the balance condition of cross-regional coordination constraints. The updated state space is then rewritten into the obstacle Lyapunov constraint strategy model, so that the model can directly use the updated state information in the next round of scheduling optimization to achieve online closed-loop optimization. The updated result is then output to the flow prediction module as input data for the new round of prediction.
[0052] An artificial intelligence-based vehicle charging pile management optimization system according to an embodiment of the present invention includes:
[0053] The data acquisition and spatiotemporal labeling module is used to collect operational data from the charging pile controller, vehicle terminal and power distribution monitoring system, and generate raw flow datasets with spatiotemporal labels;
[0054] The spatiotemporal feature extraction and two-layer graph construction module is used to map the original traffic dataset to a two-layer graph structure composed of the traffic network layer and the power network layer, and to construct the spatiotemporal traffic feature matrix between stations;
[0055] The traffic prediction module receives the spatiotemporal traffic feature matrix between sites and outputs the prediction results.
[0056] The obstacle Lyapunov constraint initialization module is used to generate stability descent constraints and directional safety constraints, and to initialize the obstacle Lyapunov constraint policy model.
[0057] The initial control action generation module is used to generate initial control actions that satisfy stability and feasibility conditions by iteratively adjusting the control quantity based on the state space and constraint conditions in the obstacle Lyapunov constraint policy model.
[0058] The optimal scheduling calculation module is used to substitute the initial control action into the obstacle constraint optimization process, perform iterative updates and back-submission corrections until the optimal scheduling action that satisfies all constraints is output.
[0059] The system execution module is used to receive the optimal scheduling action, convert it into a sequence of execution instructions, and collect traffic and power flow change data during the execution process to generate a feedback dataset.
[0060] The online closed-loop optimization module is used to correct the state space of the obstacle Lyapunov constraint strategy model and feed the updated results back to the traffic prediction module to achieve online closed-loop optimization.
[0061] The beneficial effects of this invention are:
[0062] 1. This invention collects multi-source operational data from charging pile controllers, vehicle terminals, and power distribution monitoring systems, and performs unified time and space processing to comprehensively reflect the dynamic changes in charging demand in different regions and time periods, providing high-quality input data for subsequent prediction.
[0063] 2. This invention utilizes a spatiotemporal feature matrix with a two-layer graph structure and a flow prediction module, which not only improves prediction accuracy but also enhances the coordination of cross-regional and cross-network scheduling, effectively avoiding resource waste or power grid risks caused by prediction deviations.
[0064] 3. This invention introduces the obstacle Lyapunov constraint strategy model into the scheduling strategy optimization, which can dynamically maintain the dual constraints of electrical safety and service quality under uncertain environments. It also achieves online closed-loop optimization through Hodge decomposition, enabling the system to adjust the strategy in real time according to execution feedback, maintain long-term stability and adaptability, and thus significantly improve the safety, stability and overall operating efficiency of charging pile management. Attached Figure Description
[0065] 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:
[0066] Figure 1 This is a flowchart of an artificial intelligence-based optimization method for the management of motor vehicle charging piles proposed in this invention;
[0067] Figure 2 This is a schematic diagram of the algorithm structure of an artificial intelligence-based optimization method for the management of motor vehicle charging piles proposed in this invention. Detailed Implementation
[0068] 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.
[0069] refer to Figure 1-2 A method and system for optimizing the management of motor vehicle charging stations based on artificial intelligence, comprising the following steps:
[0070] The system collects operational data from the charging pile controller, vehicle terminal, and power distribution monitoring system. The operational data includes the number of vehicles flowing into the station, charging power, voltage and current, transformer load rate, and feeder current. The system then performs timestamp alignment and geographic coordinate unification processing on the operational data to generate a raw traffic dataset with spatiotemporal tags.
[0071] The original traffic dataset is mapped to a two-layer graph structure containing a traffic network layer and a power network layer. The traffic data in the two-layer graph structure is decomposed to obtain gradient flow, curl flow and harmonic flow components. The spatiotemporal traffic feature matrix between stations is constructed based on the component results.
[0072] The spatiotemporal flow feature matrix is input into the flow prediction module to obtain the prediction results of the station inflow rate and flow distribution in future discrete time periods. The prediction results describe the potential load migration trend with gradient flow, describe the local circulating flow intensity with curl flow, and describe the global equilibrium state with harmonic flow.
[0073] The prediction results, along with electrical safety constraint parameters and service quality constraint parameters, are input into the scheduling strategy optimization module to initialize the state space and constraint conditions of the obstacle Lyapunov constraint strategy model.
[0074] Based on the state space and constraints, the predicted gradient flow, curl flow and harmonic flow values are used as environmental state variables to calculate the initial control action that satisfies the stability conditions.
[0075] Substituting the initial control action into the obstacle constraint optimization process, the optimal scheduling action under the constraint conditions is obtained by iterative solution. The scheduling action includes the station-level access power setting value and the queuing release order.
[0076] The optimal scheduling action is input into the system execution module to control the charging pile network in real time, and the traffic and power flow change data after the execution is collected and verified to obtain a feedback dataset.
[0077] By decomposing the flow and tidal change data and updating the gradient flow, curl flow, and harmonic flow components, online closed-loop optimization of the obstacle Lyapunov constraint strategy model is achieved.
[0078] This invention achieves accurate modeling of charging pile operation status by acquiring multi-source operational data and using unified spatiotemporal labeling, combined with two-layer graph structure decomposition to obtain gradient flow, curl flow, and harmonic flow components. It utilizes fractionalized features for flow and power flow prediction, and generates optimal scheduling actions that satisfy stability conditions under a barrier Lyapunov constraint strategy model, ensuring grid safety and service quality. Through system execution and feedback closed-loop updates, it achieves dynamic optimization of the scheduling strategy, effectively improving the coordination, stability, and operational efficiency of the charging pile network.
[0079] In this embodiment, the generation of the original traffic dataset specifically includes:
[0080] The system obtains the operating data of each charging pile in the current time period from the charging pile controller, the vehicle location and charging status data related to the current charging pile from the vehicle terminal, and the power distribution load data corresponding to the site where the current charging pile is located from the power distribution monitoring system. The three types of data are recorded in the raw data table according to the collection time.
[0081] The original data table is sorted according to a unified time base, and the time field of all records is normalized using a unified time base. The time representation is unified to Coordinated Universal Time format and accurate to milliseconds, eliminating clock deviations between different devices and generating an aligned data table with consistent time order and no missing data.
[0082] All location-related fields in the alignment data table are unified to the same coordinate reference, and the location data is merged with the corresponding time labels and operational status data to generate a spatiotemporally labeled raw traffic dataset with time labels and geographic coordinates.
[0083] This step acquires multi-source data from the charging pile controller, vehicle terminal, and power distribution monitoring system, and sorts and normalizes it according to a unified time base to eliminate device clock deviations and ensure data consistency and accuracy. Through time alignment and location unification processing, a spatiotemporally labeled raw traffic dataset with time labels and geographic coordinates is generated. This dataset provides reliable input for subsequent traffic feature extraction, prediction, and optimization, ensuring the real-time performance, accuracy, and efficiency of the charging pile management system.
[0084] In this embodiment, the construction of the spatiotemporal flow feature matrix specifically includes:
[0085] Read the original traffic dataset with spatiotemporal labels, group and sort it according to the unique site code and unified time base, retain the time label, spatial coordinates and operation status data, and form a site-level spatiotemporal sequence arranged in chronological order;
[0086] Based on the site-level spatiotemporal sequence, a two-layer graph structure including a traffic network layer and a power network layer is constructed. In the traffic network layer, site nodes and travel path edges are established, and in the power network layer, site nodes, transformer nodes, feeder nodes and electrical connection edges are established, and cross-layer mapping relationships are established.
[0087] By combining a two-layer graph structure with a station-level spatiotemporal sequence, the flow data of nodes and edges is decomposed and reconstructed. For each time window, each travel path edge is assigned a directed flow value at the traffic network layer according to the convention that the starting station number is less than the ending station number. The directed flow is then decomposed into three components to obtain components reflecting the potential load migration direction, components reflecting the local cycle intensity, and components reflecting the global balance characteristics. The three components are recorded for each station under each time label.
[0088] The three components are combined according to site pairing and time label to generate a spatiotemporal traffic feature matrix between sites that covers all sites and time ranges.
[0089] This step involves grouping and sorting the original spatiotemporally labeled traffic data dataset to construct a two-layer graph structure encompassing both the transportation network and power network layers. Cross-layer mapping relationships are then established to achieve a multi-dimensional fusion representation of charging pile operation data. Combining decomposition and reconstruction methods, the traffic data is decomposed into three components reflecting potential load migration, local cycle intensity, and global balance characteristics. This ensures the comprehensiveness and accuracy of spatiotemporal feature extraction. The resulting inter-site spatiotemporal traffic feature matrix provides a high-quality input foundation for subsequent prediction and scheduling optimization.
[0090] In this embodiment, obtaining the measurement result specifically includes:
[0091] Read the spatiotemporal traffic feature matrix between stations, organize it according to the pairing relationship between stations and adjacent stations as the row order and the order of time tags as the column order, and maintain the fixed arrangement of the traffic network layer and power network layer components in each column to complete the construction of the input data block;
[0092] Based on the current decision moment, a fixed-length historical time window is continuously extracted from the input data block to form a historical sequence arranged in ascending order of time. At the same time, several consecutive discrete time periods immediately following the current moment are set as the prediction time domain, ensuring that the site pairing index and component order are completely consistent between the historical sequence and the prediction time domain.
[0093] The historical sequence is input into the traffic prediction module, and prediction result tables are generated one by one according to the site pairing and prediction period. The prediction result table provides four types of indicators for each site in each prediction period: site inflow rate, charging start rate, queue length increment and available access power, and the generated fields are recorded in sequence.
[0094] The system performs integrity checks and boundary checks on the field records, removes records with missing time tags, and prunes negative values to zero or values exceeding the capacity limit to the corresponding limit. The system then summarizes the data in ascending order of site as primary key and time tag to obtain the prediction results for future discrete time periods.
[0095] This step ensures the spatiotemporal consistency of traffic prediction by constructing a fixed arrangement of input data blocks and combining the strict correspondence between historical time windows and the prediction time domain. The traffic prediction module generates multi-dimensional prediction indicators for each site, comprehensively reflecting future operational trends. Before outputting the results, integrity checks and boundary checks are performed to eliminate abnormal or excessive data, thereby effectively improving the accuracy and reliability of the prediction results and providing high-quality input data support for subsequent scheduling optimization.
[0096] In this embodiment, the initialization of the state space and constraints specifically includes:
[0097] The predicted inflow rate, predicted charging start rate, predicted queue length change and predicted available access power of each site in the prediction results of future discrete time periods are combined into a state vector in a fixed order. At the same time, the site-level access power allocation, queue release order and cross-site diversion ratio to be optimized are combined into a control vector to form a state control pair.
[0098] Receive the state control pair, combine the electrical safety constraint parameters and service quality constraint parameters, establish a set of feasible domain functions in the space of the state control pair, and map the values of the feasible domain functions to the state control pair one by one;
[0099] A Lyapunov constraint strategy model is constructed, which receives a set of feasible domain functions and state control pairs. The composite energy evaluation value for each time period is calculated. The time-period difference of composite energy is calculated item by item along the site pairing dimension, and then superimposed in chronological order to form a time-period prefix difference sequence. This sequence is matched one-to-one with convergence coefficients and written into constraint slots to generate the first type of constraint, requiring that the composite energy in the next time period be lower than the composite energy of the current time period scaled by the convergence coefficient. Simultaneously, a forward difference approximation is performed on the feasible domain function for each site pair to obtain the numerical vector of the directional derivative. The weighted sum of the directional change rate of the feasible domain function and the current numerical vector is calculated to generate the second type of constraint, requiring that the current result is not less than zero for all site pairs. The two types of constraints are bound to the state control pairs to form a constraint set, completing the constraint initialization.
[0100] The system receives a set of constraints and a state-control pair, sets a reference control quantity for each prediction period, generates a corresponding state-space weight matrix, and combines them to form an initial objective function. Using the reference control quantity as the starting value, iterative update calculations are performed. Each updated control quantity is substituted into the constraint set for feasibility verification. Control quantities that do not meet the constraints are corrected by reverting to the previous iteration. When the objective function change is below a set threshold for two consecutive iterations and all constraints are met, the current control quantity is used as the optimal scheduling state space, and the state space of the obstacle Lyapunov constraint policy model is initialized. The set threshold is based on the 95th quantile of the historical objective function change distribution. During the iteration process of binding the reference control quantity to constraints, a Bregman-Logit regularization term forms the initial iterative optimization cost.
[0101]
[0102] Among them, u k J is the control vector for predicting time period k. k (u k ) represents the initial iteration optimization cost. R is the reference control vector for the prediction period k. k Let η be the symmetric positive definite weight matrix for the prediction period k, η be the regularization coefficient, and Ω be the adjustment regularization term. k The weights of Ω(u) in the objective function k ) is a Bregman-Logit type regularization term used to penalize the nonlinearity of the control vector. for The transpose of .
[0103] This step constructs the prediction results into state and control vectors, generates a feasible region function by combining electrical safety and quality of service constraints, and introduces a Lyapunov constraint strategy model to form dual constraints of stability and directionality, thus ensuring the safety and controllability of the scheduling optimization process. An initial objective function is constructed using the reference control quantity and the state space weight matrix, and a feedback correction mechanism is incorporated during iterative updates to ensure that the control quantity satisfies all constraints within the convergence threshold. This significantly improves the stability, accuracy, and adaptability of the charging pile scheduling strategy in dynamic environments.
[0104] In this embodiment, the calculation process for the initial control action specifically includes:
[0105] The state space and constraints are combined with the configured electrical safety constraint parameters and service quality constraint parameters to form a set of environmental state variables for the current prediction period;
[0106] Based on the set of environmental state variables, a mapping relationship from the current state and control variables to the state of the next time period is established, and stability descent constraints are generated by combining the convergence coefficient, while directional safety constraints are generated by combining the feasible region function.
[0107] Stability degradation constraints and safety feasibility constraints are applied sequentially to environmental state variables to determine the operational status of each site during the prediction period. The constraints are then contracted in conjunction with stability conditions to form a set of constraints that can be used for calculation. The stability condition is that the change in the objective function of the obstacle Lyapunov constraint strategy model is less than a set threshold for two consecutive iterations.
[0108] The initial control quantity is gradually adjusted based on the set of constraints available for calculation. In each adjustment, the control components that cause the stability condition to become unbalanced are corrected first, and then the control components that cause the feasibility constraint to be violated are corrected. After each round of adjustment, the stability function value and safety margin are recalculated. The current control quantity obtained by adjustment is used as the initial control action that satisfies the stability condition.
[0109] This step generates a set of environmental state variables by fusing state space, constraints, and safety parameters, and establishes a mapping relationship from state to the next time period. It then combines convergence coefficients and feasible region functions to generate stability degradation constraints and directional safety constraints, ensuring the safety and stability of the scheduling strategy in dynamic environments. By gradually adjusting the initial control variables and prioritizing the correction of key constraint violations, it achieves dual guarantees of stability and feasibility, thereby improving the accuracy and robustness of charging pile scheduling.
[0110] In this embodiment, obtaining the optimal scheduling action specifically includes:
[0111] The initial control action is used as the starting solution for iterative optimization. The obstacle Lyapunov constraint policy model is loaded, and the stability descent constraint and safety feasibility constraint are imported into the optimization solver.
[0112] Construct an iterative optimization objective function, where the first part is a quadratic penalty term for the control quantity deviating from the reference control quantity, and the second part is a weighted barrier function penalty term. The barrier function consists of the logarithmic sum of each feasible domain function and is used to increase the penalty value when approaching the constraint boundary. The two parts are combined with a fixed proportional coefficient to form the complete objective function.
[0113] Based on the complete objective function, the control quantity is updated according to the gradient direction of the objective function and the constraints. Then, the control quantity that does not meet the constraints is back-adjusted along the outward normal direction to obtain the corrected control quantity, which is then substituted into the objective function to continue iterating until the control quantity meets the constraints and re-enters the feasible region, thus obtaining the optimal scheduling action.
[0114] This step ensures efficient control of site power allocation, queuing, and cross-site traffic diversion by translating optimal scheduling actions into precise sequences of execution instructions. The execution module records traffic flow and power flow changes in real time during scheduling and ensures the accuracy and completeness of the feedback data through timestamp alignment and consistency checks. This method effectively improves the real-time performance and accuracy of charging pile management, enhances the system's adaptability, and provides reliable data support for subsequent scheduling optimization, thereby achieving continuous closed-loop optimization of charging pile management.
[0115] In this embodiment, the recovery of flow and tidal current change data specifically includes:
[0116] Receive the optimal scheduling action, and combine the control instructions in the optimal scheduling action, including site access power allocation, queuing release order and cross-site diversion ratio, according to the execution order, and specify the corresponding execution time and target location for each item to form a sequence of instructions that can be directly issued;
[0117] The instruction sequence is input into the system execution module to complete the power adjustment, queue update and flow direction switching operations in sequence, while continuously recording relevant operating data during the execution process;
[0118] Real-time collection of traffic and power flow change data of the charging pile network. The traffic change data includes changes in the number of vehicles accessing each station per unit time and changes in the charging start rate. The power flow change data includes changes in the station power curve and changes in the feeder current. The traffic and power flow change data are stored in correspondence with the execution time label and the station number.
[0119] The flow and current change data are timestamped and validated for consistency. Missing or abnormal data points are removed, and the flow and current change data after the validation is performed are collected to obtain a feedback dataset.
[0120] This step transforms optimal scheduling actions into a sequence of instructions with execution times and target locations, precisely controlling site power allocation, queuing, and cross-site diversion to achieve orderly and efficient scheduling execution. During execution, traffic and power flow change data are collected synchronously, and high-quality feedback datasets are generated through timestamp alignment and consistency checks, providing reliable input for subsequent optimization. This method improves the controllability of the execution process and the accuracy of data collection, laying the foundation for real-time closed-loop optimization and continuous improvement of scheduling strategies.
[0121] In this embodiment, the implementation of the online closed-loop optimization specifically includes:
[0122] Receive feedback datasets, extract traffic change data and power flow change data related to the operation of each site, and combine the two types of data into a unified spatiotemporal data matrix according to time order and site number;
[0123] The spatiotemporal data matrix is mapped to a graph structure, and Hodge decomposition is performed on the edge flow distribution in the graph. The original flow field is decomposed into three components: gradient flow component, curl flow component, and harmonic flow component. The gradient flow component corresponds to the global situational energy change, the curl flow component corresponds to the local loop flow, and the harmonic flow component corresponds to the inherent pattern of the network topology.
[0124] The state space of the obstacle Lyapunov constraint strategy model is updated using ternary component components. The gradient flow component is used to correct the state prediction of the global stability function, the curl flow component is used to correct the directional changes of local constraints, and the harmonic flow component is used to correct the balance condition of cross-regional coordination constraints. The updated state space is then rewritten into the obstacle Lyapunov constraint strategy model, so that the model can directly use the updated state information in the next round of scheduling optimization to achieve online closed-loop optimization. The updated result is then output to the flow prediction module as input data for the new round of prediction.
[0125] This step involves performing Hodge decomposition on the feedback dataset, breaking down the flow field into gradient flow, curl flow, and harmonic flow components. The decomposition results are then used to dynamically correct the state space of the obstacle Lyapunov constraint policy model, achieving simultaneous optimization of global stability, local constraint directionality, and cross-regional coordination. Combined with an online closed-loop optimization mechanism, the model can utilize the latest state information for prediction and decision-making in each scheduling round, thereby improving the real-time performance, accuracy, and overall system stability of charging pile scheduling.
[0126] An artificial intelligence-based vehicle charging station management optimization system includes:
[0127] The data acquisition and spatiotemporal labeling module is used to collect operational data from the charging pile controller, vehicle terminal and power distribution monitoring system, and generate raw flow datasets with spatiotemporal labels;
[0128] The spatiotemporal feature extraction and two-layer graph construction module is used to map the original traffic dataset to a two-layer graph structure composed of the traffic network layer and the power network layer, and to construct the spatiotemporal traffic feature matrix between stations;
[0129] The traffic prediction module receives the spatiotemporal traffic feature matrix between sites and outputs the prediction results.
[0130] The obstacle Lyapunov constraint initialization module is used to generate stability descent constraints and directional safety constraints, and to initialize the obstacle Lyapunov constraint policy model.
[0131] The initial control action generation module is used to generate initial control actions that satisfy stability and feasibility conditions by iteratively adjusting the control quantity based on the state space and constraint conditions in the obstacle Lyapunov constraint policy model.
[0132] The optimal scheduling calculation module is used to substitute the initial control action into the obstacle constraint optimization process, perform iterative updates and back-submission corrections until the optimal scheduling action that satisfies all constraints is output.
[0133] The system execution module is used to receive the optimal scheduling action, convert it into a sequence of execution instructions, and collect traffic and power flow change data during the execution process to generate a feedback dataset.
[0134] The online closed-loop optimization module is used to correct the state space of the obstacle Lyapunov constraint strategy model and feed the updated results back to the traffic prediction module to achieve online closed-loop optimization.
[0135] This invention improves scheduling accuracy through dual-layer graph structure modeling and traffic prediction, ensures electrical safety and service quality in dynamic environments by combining obstacle Lyapunov constraint strategies, and achieves real-time feedback and adaptive adjustment through system execution and online closed-loop optimization, thereby improving the stability, coordination and operational efficiency of charging pile scheduling.
[0136] Example 1:
[0137] To verify the feasibility of this invention in practice, it was applied to a pilot project for intelligent scheduling of new energy vehicle charging piles in a city. This pilot project is located at the border of the city's high-tech zone and main urban area, encompassing 8 main charging stations and a total of 112 charging piles, including 72 DC fast charging piles and 40 AC slow charging piles. The service area covers a 5-kilometer radius of business districts, residential areas, and industrial parks. Traffic flow in this area fluctuates significantly, with concentrated vehicle access during morning and evening peak hours. Furthermore, some charging stations are located at the end of feeders where the grid load is nearing saturation. Historically, the project has frequently encountered problems such as excessively long queuing times, delayed cross-station scheduling, and transformer overload during peak hours, severely impacting user experience and grid safety.
[0138] In this scenario, the method of this invention is deployed on the cloud scheduling platform of a charging pile operator. The platform accesses real-time data from the charging pile controller, vehicle terminal, and power distribution monitoring system. Every 30 seconds, it collects the number of vehicles flowing into the station, the charging power of each charging pile, voltage and current parameters, the station transformer load rate, and feeder current values. This operational data is then timestamped and processed to unify geographic coordinates, generating a raw traffic dataset with precise time labels and geographic information. Subsequently, the system maps the raw traffic data to a two-layer graph structure consisting of a traffic network layer and a power network layer. Through Hodge decomposition, the traffic field is decomposed into three components: gradient flow, curl flow, and harmonic flow, to respectively characterize the potential load migration trend, local circulating flow intensity, and global equilibrium state.
[0139] In daily operation, the system constructs a spatiotemporal flow characteristic matrix between stations using the decomposed components. This matrix is then input into the flow prediction module to predict the inflow rate and power flow distribution at three time scales: 15 minutes, 30 minutes, and 60 minutes. The prediction results, along with electrical safety constraint parameters and service quality constraint parameters, are input into the scheduling strategy optimization module. This module generates initial control actions that satisfy stability conditions using a barrier Lyapunov constraint strategy model and outputs the optimal scheduling action after iterative solution. After issuing commands and executing field actions, the system execution module performs time alignment and consistency verification on the flow and power flow change data collected during execution. It then collects the generated feedback dataset and performs Hodge decomposition again to update the gradient flow, curl flow, and harmonic flow components, thereby achieving online closed-loop optimization of the barrier Lyapunov constraint strategy model.
[0140] Table 1 Comparison of Operation Before and After Application in the Pilot Area of the High-tech Zone
[0141]
[0142]
[0143] As shown in Table 1, after implementing this invention, the operating efficiency and safety of charging stations have been significantly improved. The average queuing time has been significantly reduced in all time intervals, decreasing from 12.3 minutes to 6.8 minutes throughout the entire period, a reduction of 44%. During the morning peak, the queuing time at station C3 decreased from 15 minutes to 7 minutes. The grid load control effect is outstanding, with a significant decrease in the peak load rate of transformers, such as at station C3, from 96% to 88%, effectively preventing grid overload. The cross-station dispatch response time has been shortened from 3 minutes to less than 2.6 minutes, greatly improving dispatch efficiency and alleviating pressure during peak hours. The utilization rate of charging piles has increased by 11.6%, with C2 and C5 increasing by 18.0% and 9.5% respectively, resulting in more balanced resource utilization. Overall, this method demonstrates that it can improve service quality and system stability while ensuring electrical safety.
[0144] 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 the management of motor vehicle charging stations based on artificial intelligence, characterized in that, Includes the following steps: The system collects operational data from the charging pile controller, vehicle terminal, and power distribution monitoring system, and performs timestamp alignment and geographic coordinate unification processing on the operational data to generate a raw traffic dataset with spatiotemporal tags. The original traffic dataset is mapped to a two-layer graph structure, and the traffic data in the two-layer graph structure is decomposed and reconstructed to build a spatiotemporal traffic feature matrix between sites. Input the spatiotemporal flow characteristic matrix into the flow prediction module to obtain the prediction results for future discrete time periods; The prediction results, along with electrical safety constraint parameters and service quality constraint parameters, are input into the scheduling strategy optimization module to initialize the state space and constraint conditions of the obstacle Lyapunov constraint strategy model. Based on the state space and constraints, the initial control action that satisfies the stability condition is calculated using environmental state variables. Substitute the initial control action into the obstacle constraint optimization process, and iteratively solve to obtain the optimal scheduling action under the constraint conditions; The optimal scheduling action is input into the system execution module to control the charging pile network in real time, and the traffic and power flow change data after the execution is collected and verified to obtain a feedback dataset. By decomposing the flow and tidal change data and updating the gradient flow, curl flow, and harmonic flow components, online closed-loop optimization of the obstacle Lyapunov constraint strategy model is achieved.
2. The method for optimizing the management of motor vehicle charging piles based on artificial intelligence according to claim 1, characterized in that, The generation of the original traffic dataset specifically includes: Acquire the operation data, vehicle location and charging status data, and corresponding power distribution load data of each charging pile, and record the three types of data in the raw data table according to the collection time; The original data table is sorted according to a unified time base, and the time field of all records is normalized using a unified time base to generate an aligned data table with consistent time order and no missing records. All location-related fields in the alignment data table are unified to the same coordinate reference, and the location data is merged with the corresponding time labels and operational status data to generate a spatiotemporally labeled raw traffic dataset with time labels and geographic coordinates.
3. The method for optimizing the management of motor vehicle charging piles based on artificial intelligence according to claim 1, characterized in that, The construction of the spatiotemporal flow feature matrix specifically includes: Read the original traffic dataset with spatiotemporal labels, group and sort it according to the unique site code and unified time base to form a site-level spatiotemporal sequence arranged in chronological order; A two-layer graph structure containing a transportation network layer and a power network layer is constructed based on the site-level spatiotemporal sequence, and cross-layer mapping relationships are established. By combining a two-layer graph structure with a site-level spatiotemporal sequence, the traffic data of nodes and edges is decomposed and reconstructed to obtain three types of components, and recorded for each site under each time label. The three components are combined according to site pairing and time label to generate a spatiotemporal traffic feature matrix between sites that covers all sites and time ranges.
4. The method for optimizing the management of motor vehicle charging piles based on artificial intelligence according to claim 1, characterized in that, The measurement results are obtained specifically in the following ways: Read the spatiotemporal traffic feature matrix between stations, organize it according to the pairing relationship between stations and adjacent stations as the row order and the order of time tags as the column order to complete the construction of the input data block; Based on the current decision-making moment, a fixed-length historical time window is continuously extracted from the input data block and arranged in ascending order of time to form a historical sequence. Input the historical sequence into the traffic prediction module, generate prediction result tables one by one according to the site pairing and prediction period, and record the generated fields in order; The system performs integrity checks and boundary checks on the field records, removes records with missing time tags, and prunes negative values to zero or values exceeding the capacity limit to the corresponding limit. The system then summarizes the data in ascending order of site as primary key and time tag to obtain the prediction results for future discrete time periods.
5. The method for optimizing the management of motor vehicle charging piles based on artificial intelligence according to claim 1, characterized in that, The initialization of the state space and constraints specifically includes: The prediction results of future discrete time periods are combined into a state vector in a fixed order for each station, and the station-level access power allocation, queuing release order and cross-station diversion ratio to be optimized are combined into a control vector to form a state control pair. Receive the state control pair, combine the electrical safety constraint parameters and service quality constraint parameters, establish a set of feasible domain functions in the space of the state control pair, and map the values of the feasible domain functions to the state control pair one by one; A Lyapunov constraint strategy model for obstacles is constructed. It receives a set of feasible domain functions and state control pairs, calculates the composite energy evaluation quantity for each time period, calculates the time period difference of composite energy item by item in the site pairing dimension, and then superimposes them in time order to form a time period prefix difference sequence. The sequence is matched with the convergence coefficients one by one and written into the constraint slot to generate the first type of constraint conditions. At the same time, a forward difference approximation is performed on the feasible domain function for each site pair to obtain the numerical vector of the directional derivative. The weighted sum of the directional change rate of the feasible domain function and the current numerical vector is calculated to generate the second type of constraint conditions. The two types of constraint conditions are bound to the state control pairs to form a constraint set and complete the constraint condition initialization. The system receives a set of constraints and a state control pair, sets a reference control quantity for each prediction period, generates a corresponding state space weight matrix, and combines them to form an initial objective function. Using the reference control quantity as the starting value, it performs iterative update calculations, and substitutes the updated control quantity into the constraint set for feasibility verification. Control quantities that do not meet the constraints are reverted for correction, and the corrected control quantities continue to participate in the next round of iteration. When the change in the objective function of two consecutive iterations is less than a set threshold and all constraints are met, the current control quantity is used as the optimal scheduling state space, and the initialization of the state space of the obstacle Lyapunov constraint policy model is completed.
6. The method for optimizing the management of motor vehicle charging piles based on artificial intelligence according to claim 1, characterized in that, The calculation process for the initial control action specifically includes: The state space and constraints are combined with the configured electrical safety constraint parameters and service quality constraint parameters to form a set of environmental state variables for the current prediction period; Based on the set of environmental state variables, a mapping relationship from the current state and control variables to the state of the next time period is established, and stability descent constraints are generated by combining the convergence coefficient, while directional safety constraints are generated by combining the feasible region function. Stability degradation constraints and safety feasibility constraints are applied sequentially to environmental state variables to determine the operational status of each site during the prediction period. The constraints are then contracted in conjunction with stability conditions to form a set of constraints that can be used for calculation. The initial control quantity is gradually adjusted based on the set of constraints that can be used for calculation. In each adjustment, the control component that causes the stability condition to become unbalanced is corrected first. The current control quantity obtained by adjustment is used as the initial control action that satisfies the stability condition.
7. The method for optimizing the management of motor vehicle charging piles based on artificial intelligence according to claim 1, characterized in that, The optimal scheduling action is obtained specifically by: The initial control action is used as the starting solution for iterative optimization. The obstacle Lyapunov constraint policy model is loaded, and the stability descent constraint and safety feasibility constraint are imported into the optimization solver. Construct an iterative optimization objective function, where the first part is a quadratic penalty term and the second part is a weighted barrier function penalty term, and combine the two parts with a fixed ratio coefficient to form the complete objective function; Based on the complete objective function, the control quantity is updated according to the gradient direction of the objective function and the constraints. Then, the control quantity that does not meet the constraints is back-adjusted along the outward normal direction to obtain the corrected control quantity, which is then substituted into the objective function to continue iterating until the control quantity meets the constraints and re-enters the feasible region, thus obtaining the optimal scheduling action.
8. The method for optimizing the management of motor vehicle charging piles based on artificial intelligence according to claim 1, characterized in that, The collection of flow and current change data specifically includes: Receive the optimal scheduling action, combine the control instructions within the optimal scheduling action in the execution order, and assign the corresponding execution time and target position to each item to form a sequence of instructions that can be directly issued; The instruction sequence is input into the system execution module to complete the power adjustment, queue update and flow direction switching operations in sequence, while continuously recording relevant operating data during the execution process; Real-time collection of traffic and power flow change data of the charging pile network, and corresponding storage of traffic and power flow change data according to execution time label and station number; The flow and current change data are timestamped and validated for consistency. Missing or abnormal data points are removed, and the flow and current change data after the validation is performed are collected to obtain a feedback dataset.
9. The method for optimizing the management of motor vehicle charging piles based on artificial intelligence according to claim 1, characterized in that, The implementation of the online closed-loop optimization specifically includes: Receive feedback datasets, extract traffic change data and power flow change data related to the operation of each site, and combine the two types of data into a unified spatiotemporal data matrix according to time order and site number; The spatiotemporal data matrix is mapped to a graph structure, and Hodge decomposition is performed on the edge flow distribution in the graph to decompose the original flow field into ternary components of gradient flow, curl flow, and harmonic flow. The state space of the obstacle Lyapunov constraint policy model is updated using the ternary component quantities, and the updated state space is rewritten into the obstacle Lyapunov constraint policy model to achieve online closed-loop optimization. The updated result is then output to the flow prediction module as input data for the next round of prediction.
10. An artificial intelligence-based vehicle charging pile management optimization system, comprising executing the artificial intelligence-based vehicle charging pile management optimization method according to any one of claims 1 to 9, characterized in that, include: The data acquisition and spatiotemporal labeling module is used to collect operational data from the charging pile controller, vehicle terminal and power distribution monitoring system, and generate raw flow datasets with spatiotemporal labels; The spatiotemporal feature extraction and two-layer graph construction module is used to map the original traffic dataset to a two-layer graph structure composed of the traffic network layer and the power network layer, and to construct the spatiotemporal traffic feature matrix between stations; The traffic prediction module receives the spatiotemporal traffic feature matrix between sites and outputs the prediction results. The obstacle Lyapunov constraint initialization module is used to generate stability descent constraints and directional safety constraints, and to initialize the obstacle Lyapunov constraint policy model. The initial control action generation module is used to generate initial control actions that satisfy stability and feasibility conditions by iteratively adjusting the control quantity based on the state space and constraint conditions in the obstacle Lyapunov constraint policy model. The optimal scheduling calculation module is used to substitute the initial control action into the obstacle constraint optimization process, perform iterative updates and back-submission corrections until the optimal scheduling action that satisfies all constraints is output. The system execution module is used to receive the optimal scheduling action, convert it into a sequence of execution instructions, and collect traffic and power flow change data during the execution process to generate a feedback dataset. The online closed-loop optimization module is used to correct the state space of the obstacle Lyapunov constraint strategy model and feed the updated results back to the traffic prediction module to achieve online closed-loop optimization.