A charging station intelligent operation and maintenance path planning method based on a large language model and a multi-start strategy reinforcement learning

Abstract

This invention provides an intelligent operation and maintenance scheduling method for charging stations based on a large language model and multi-starting-point policy reinforcement learning. The main steps include: collecting alarm and latitude / longitude data of target charging stations, and constructing a symmetrical global time-distance matrix by combining macro and local road networks; initializing the large language model decision center to generate macro control parameters (including weight parameters and penalty coefficients for each region, and the number of clusters) containing the number of operation and maintenance personnel; using a clustering algorithm based on these parameters to divide all stations into several controlled operation and maintenance areas; extracting real two-dimensional coordinates, using a multi-starting-point policy reinforcement learning network to perform path planning for stations within the operation and maintenance areas, and implementing real-time posterior arbitration; calculating the total maintenance time and workload variance of the planned routes; if the target is not met, generating a natural language diagnostic report and feeding it back to the large language model to readjust the weight coefficients, penalty coefficients, and number of clusters, and iterating until the solution meets the workload balance satisfaction threshold; and outputting the final optimal path planning routes and total time consumption for all target operation and maintenance charging stations.

Description

Technical Field

[0001] A method for intelligent operation and maintenance path planning of charging stations based on large language models and multi-starting-point policy reinforcement learning is proposed, belonging to the field of path planning technology. Technical Background

[0002] With the rapid expansion of charging station scale, the massive cross-regional operation and maintenance scheduling faces severe technical challenges: First, traditional operations research algorithms are prone to combinatorial explosion when faced with a massive number of nodes, losing their timeliness. Second, pure reinforcement learning models often produce spatial illusions, planning routes that deviate from reality and failing to balance the workload of personnel. Third, theoretical mathematical models ignore the asymmetry and island effect of complex real-world traffic, causing the model to fail in practical applications. Fourth, most existing systems are one-way "black boxes" lacking dynamic feedback and self-diagnosis capabilities. When encountering extreme allocation anomalies, manual fine-tuning is still required, making it difficult to achieve truly fully automated closed-loop scheduling.

[0003] To address the pain points encountered by the aforementioned industries, this paper proposes an intelligent operation and maintenance path planning scheme for charging stations. This scheme eliminates the need for overall considerations and path planning, using a large-scale language model as the command center and a multi-starting-point strategy reinforcement learning network for local search. Based on natural language diagnostic reports, a closed loop is formed, resulting in a fully intelligent judgment. This saves maintenance time and solves the problems of resource idleness and uneven workload distribution. This algorithm completes the intelligent development process from physical world perception to complex decision-making, reducing the reaction time of maintenance personnel, enhancing the speed and economy of fault handling, and preventing resource waste, unnecessary duplicate dispatches, and uneven workload distribution.

[0004] The technical solution of this patent is not only applicable to the operation and maintenance of charging piles, but can also be extended to other similar distributed facility operation and maintenance fields, with broad application prospects and significant commercial value. Summary of the Invention

[0005] The technical problem this invention aims to solve is to address the shortcomings of current charging station maintenance route selection, which relies excessively on manual experience and suffers from computational exhaustion in traditional heuristic algorithms. This is achieved through a closed-loop iterative architecture combining macro-level decision-making using a large language model, clustering algorithms, and reinforcement learning for local optimization using a multi-starting-point strategy. This architecture divides a massive number of discrete maintenance sites into several controlled maintenance grid regions based on physical resistance and overload penalties. Reinforcement learning algorithms iteratively find the shortest path in each maintenance region, achieving efficient optimization of maintenance efficiency across all sites and balancing of global manpower load. This invention not only solves the persistent problems of low optimization efficiency and tendency to get stuck in local optima when dealing with a large number of path points in traditional path optimization algorithms, but also eliminates the route infeasibility problem caused by the asymmetry of the physical road network at the underlying logic level. It provides a scientific, interpretable, and quantifiable basis for maintenance departments to dynamically increase or decrease staffing in different seasons and periods of high failure incidence, and to accurately assess the workload balance of multiple agents.

[0006] The technical solution adopted by this invention to solve its technical problem is: an intelligent operation and maintenance scheduling method for charging stations based on large language models and multi-starting-point policy reinforcement learning, comprising the following steps:

[0007] Step 1: Collect alarm and latitude / longitude data of the target charging station, and construct a symmetrical global time-distance matrix by combining macro and local road network data;

[0008] Step 2: Initialize the large language model decision center to generate macro-control parameters (including weight parameters and penalty coefficients for each region, and the number of clusters). Based on these parameters, use a clustering algorithm to divide all sites into several operation and maintenance regions.

[0009] Step 3: Extract the real two-dimensional coordinates and combine them with a multi-starting-point strategy reinforcement learning network (POMO) to solve for the optimal operation and maintenance route for each region;

[0010] Step 4: Calculate the total maintenance time and workload variance of the optimal route in each region. If the target is not met, generate a natural language diagnostic report and feed it back to Step 2. The large language model then readjusts the weight coefficients and penalty coefficients for clustering.

[0011] Step 5: Output the final optimal route planning for all target charging stations and the total time for each route.

[0012] Preferably, the specific implementation process of constructing the symmetric global temporal distance feature matrix in step 1 is as follows:

[0013] Step 1-1: Clean the underlying IoT raw alarm data, remove duplicate alarms to obtain the actual number of faults, and calculate and lock the estimated maintenance workload for a single station;

[0014] Steps 1-2: Obtain traffic resistance time of the backbone road network through the Gaode Map API. For isolated and remote stations, estimate the three-dimensional spherical physical distance using the semi-sine formula and connect it to the backbone network node; the semi-sine formula is as follows:

[0015]

[0016] in, This is the baseline value for the Earth's average radius of curvature; These represent the geographical latitude and longitude coordinates of the initial fault site and the target legitimate road network node to be connected, respectively. These represent the geographical longitude coordinates of the starting fault site and the target legitimate road network node to be connected, respectively.

[0017] Steps 1-3: Introduce a local KNN micro-loop detection mechanism to establish forced direct connections between stations that are physically very close, and deduce and generate the original directed time feature matrix;

[0018] Steps 1-4: To eliminate the asymmetry caused by one-way congestion in the real road network, the original feature matrix and its transpose matrix are globally symmetric and smoothed to obtain the final symmetric feature matrix. The purpose of this approach is to solve the problem of charging station location offset.

[0019] Preferably, in step 2, the initialization of the large language model decision center to generate macro-control parameters (including weight parameters and penalty coefficients for each region, and the number of clusters) is used. Based on these parameters, a clustering algorithm is employed to divide all isolated sites into several controlled operation and maintenance grid regions. The specific implementation process is as follows:

[0020] Step 2-1: Define the ideal workload for a single operations and maintenance personnel; the Large Language Model (LLM) inference outputs a key focus area, injects high-order weights into sites within that area, and uses a double-squared weighted optimization K-means++ strategy to select initial cluster centers; candidate cluster centers... The formula for the expected initial probability of being selected by the system is as follows:

[0021]

[0022] in, and Representing candidate physical sites and The minimum of the lowest-level topological time distance to the set of all legal cluster centers determined in the current round (i.e., finding its nearest selected center); This represents the soft weight coefficient of the focus area assigned to a specific site by the large language model; This represents the total number of valid charging station nodes that are actually involved in clustering and segmentation in the current global physical network.

[0023] Step 2-2: When deploying discrete sites, dynamically calculate the expected overload fatigue percentage (the proportion exceeding a set threshold) of personnel in candidate grids; multiply the physical travel time, the estimated overload situation, and the LLM output penalty to calculate the total scheduling cost, avoiding cross-regional dispatching; and integrate the scheduling cost factor. The calculation formula is as follows:

[0024]

[0025] in, The lower the cost, the higher the allocation priority; Representing physical sites Drive to the grid The shortest bottom-level road network time distance of the current cluster gravity center; This is a dynamically adjustable distance multiplier penalty coefficient; This represents the risk ratio that the estimated total workload exceeds the system's set limit for the ideal workload per person.

[0026] Steps 2-3: When updating the grid center, access all valid sites within the grid and select the site that minimizes the sum of actual distances between nodes within the cluster as the new center; if an empty grid is found, enable the forced reset function and place the center at the most heavily affected disaster point globally; its physical optimization objective function is as follows:

[0027]

[0028] in, This represents the controlled mesh after internal optimization, update, and convergence. The internal physical geometric center; Represents the first segment dynamically partitioned by the system. A locally controlled subset of the operation and maintenance grid; These represent elements that are strictly limited to the current grid subset. Any two nodes of the charging station to be repaired within the system can be traversed. Representative node With nodes The values ​​of the symmetric travel time distance matrix are called between them.

[0029] Steps 2-4: Establish a global elite solution retention pool. Find the clustering partitioning scheme with the minimum total time cost within the maximum number of iterations. This scheme is then passed on as the final grid output of this round of macro scheduling to the next stage of micro optimization. The purpose is to guide the large model towards a more reasonable direction and avoid blindly updating the large model.

[0030] Preferably, in step 3, the extraction of the true two-dimensional geometric coordinates of each grid and the combination of a multi-point policy reinforcement learning network (POMO) to solve for the optimal operation and maintenance route for each region are implemented as follows:

[0031] Step 3-1: Extract the real two-dimensional latitude and longitude coordinate tensors of each station as the bottom-level input state features of the deep neural network, preserving the physical spatial attributes;

[0032] Step 3-2: Perform a global isometric normalization stretching operation on the geospatial coordinates to prevent proportional distortion, and add a zero overflow correction term to prevent network paralysis caused by extremely overlapping sites; the processing operator formula is as follows:

[0033]

[0034] in, This represents a two-dimensional normalized spatial sensing coordinate tensor that has been scaled down to the 0-1 feature space range after eliminating the influence of absolute values ​​of spatial scale. This represents the obtained array of original two-dimensional true latitude and longitude values; and These represent the global objective minimum and maximum values ​​of the obtained coordinate tensor in the orthogonal directions of latitude and longitude, respectively; To prevent zero overflow, a minimum positive real number compensation term is used.

[0035] Step 3-3: Utilizing the multi-instance view enhancement mechanism of the POMO framework, affine transformations such as rotation and flipping are applied to the normalized features to form a multi-view enhancement tensor that is fed into the deep network. This acquires geometrically invariant features to prevent overfitting of the model. The rotation geometric matrix transformation formula is as follows:

[0036]

[0037] in, Represents an enhanced coordinate-aware tensor set containing multiple independent view sequences of physical topology; This represents the spatial rotation angle offset of the discretized space.

[0038] Steps 3-4: Invoke the pre-trained tensor network based on the attention mechanism to generate multiple sets of candidate paths without self-intersection in parallel, and simultaneously output the network's internal native features to evaluate the cost.

[0039] Steps 3-5: Substitute the candidate path set back into the actual traffic time matrix for physical time accumulation and filtering; finally, output the optimal execution route. Absolutely optimal executable route. The final stringent judgment is as follows:

[0040]

[0041] in, After filtering, the final output is the optimal deterministic path execution sequence that has the shortest travel time in the objective physical world; This represents the complete set of candidate path policies generated by a multi-starting-point policy reinforcement learning network. Refers to the set of strategies Any single specific candidate exploration path in the; This represents the total number of charging stations awaiting repair within the currently controlled local maintenance grid that is undergoing neural routing scheduling. This indicates that during the execution of the candidate strategy path, the task node... Drive to the next mission node By querying the underlying physical base to obtain the actual time-mapping cost, based on the actual time matrix, the problem of generated paths being unusable in reality can be avoided.

[0042] Preferably, in step 4, the variance of the total physical maintenance time of the optimal route in each region and the workload of the multi-agent system is calculated. When the evaluation does not meet the expected standard, a natural language diagnostic report is generated and sent to the large language model to change its macroscopic parameters (weight coefficient, penalty coefficient, number of clusters). The specific implementation steps are as follows:

[0043] Step 4-1: Calculate the final expected total time per person within each grid, which includes physical travel time and estimated repair time upon arrival at the site; the operator formula is as follows:

[0044]

[0045] in, Representing the The total physical time taken for a single professional operations engineer within a controlled operations and maintenance grid to complete all dispatched tasks within the daily shift schedule; Represents the target grid The total number of sites within the area that have been pre-arranged for repair; This refers to the station currently under maintenance and the next station to be visited. This represents the cost of personnel who remain on duty for maintenance.

[0046] Step 4-2: Obtain the arithmetic mean of the total time spent in the global grid, calculate the standard deviation of the overall manpower workload balance, and use it as the main indicator to measure the manpower load balance of the entire system; calculate the variance. formula:

[0047]

[0048] in, This represents the total pool of personnel for the macro-level operations and maintenance grid, automatically generated by the Large Language Model (LLM) based on the system's previous natural language feedback in the current round. Represents the whole The lower bound of the average total physical time required for each maintenance grid to be used together is the baseline.

[0049] Step 4-3: Activate the anti-oscillation mechanism. If the variance is less than the threshold and there is no illegal passenger loading, the early stop strategy is activated to directly output the optimal solution. If there is a large amount of waiting or idle time, the lowest-level data is translated into a natural language diagnostic report and fed together with the latest simulated historical data into the large language model. This allows the model to reset its global parameters (such as weight coefficients, penalty coefficients, number of clusters, etc.) and iterate repeatedly until convergence or the maximum number of iterations is reached. This continuous modification is intended to replace manual intervention in real life, reducing time and manpower costs.

[0050] Preferably, step 5, which outputs the final optimal route planning for all target charging stations and the total time for each route, is specifically implemented as follows:

[0051] Step 5-1: The system compiles and integrates the shortest time-consuming route generated by multi-starting-point strategy reinforcement learning with the expected total time consumption data per person; combined with the anti-vibration diagnostic report containing working time and variance distribution, it generates standardized scheduling data and feeds it back to the large language model decision module, outputting a fully automated, closed-loop operation and maintenance dispatch list for charging stations in different geographical locations.

[0052] A charging station operation and maintenance path planning system combining large language model decision-making and multi-starting-point policy reinforcement learning, wherein the method is based on large language model decision-making and multi-starting-point policy reinforcement learning, and the system includes:

[0053] The data preprocessing and topology construction module is used to collect and clean the alarm and latitude and longitude data of the target charging station, combine the macro backbone network and local micro-circulation to construct the underlying network, and deduce the symmetric global time distance matrix.

[0054] The large language model decision module is used to receive environmental summaries and historical feedback using a structured output protocol, and adaptively output macro-control parameters including the number of maintenance personnel (including weight parameters and penalty coefficients for each region, and the number of clusters).

[0055] The site clustering module is used to perform flexible memory clustering and classification of sites in different geographical locations by combining the distance multiplier penalty method and the airdrop activation mechanism, as a candidate pool of sites for operation and maintenance routes.

[0056] The reinforcement learning module has built-in real coordinate extraction, multi-start point multi-view augmentation (POMO) and posterior arbitration mechanism, which are used to plan the shortest real traffic time path selected by the operation and maintenance personnel from the candidate pool without human intervention.

[0057] The assessment and diagnosis module generates anti-vibration diagnosis reports in natural language based on working hours and variance distribution, which determines whether to increase or decrease the number of maintenance personnel, modify penalties or weights, etc.

[0058] Compared with the prior art, the beneficial effects of this invention are:

[0059] (1) Breaking the limitations of traditional algorithms in terms of computing power and improving the scientific nature of macro-decision making: By extracting macro-vectors through a large language model, the problem of "combinatorial explosion" and computing power exhaustion caused by the large number of nodes in traditional heuristic algorithms is well solved. Combined with the double-weighted improved iterative algorithm and the distance multiplier penalty method, effective grid management of geographical nodes is realized, avoiding cross-city and cross-level dispatching and reducing the cost of operation and maintenance personnel.

[0060] (2) Resolving the conflict between spatial illusion and physical path to ensure path feasibility: The POMO and arbitration mechanism were used to correct the "spatial illusion" phenomenon generated by the simple neural network model. The mechanism design takes into account the complexity of actual physical road conditions. While ensuring the optimization effect, it ensures that the planned path conforms to the actual situation, thus solving the problem of path infeasibility or excessive deviation from reality.

[0061] (3) Improve the rationality of resource allocation and form a closed-loop management of operation and maintenance system: take the integrity of the entire physical transfer and parking maintenance time as the assessment indicator, and use the natural reporting and feedback closed loop method. This method effectively solves the problems of uneven workload distribution and uneven human resources in traditional operation and maintenance, greatly improves the operability of path planning, and achieves the effect of high efficiency and economy in the overall operation and maintenance process. Attached Figure Description

[0062] Figure 1 This is a flowchart of a charging station operation and maintenance path planning method that combines AI-agent and multi-starting-point policy reinforcement learning.

[0063] Figure 2 This is a flowchart of the multi-starting-point policy reinforcement learning method (POMO) used in this application.

[0064] Figure 3 It is a joint graph of spatial network partitioning and POMO shortest path.

[0065] Figure 4 It is a comparative analysis of the workload balancing of various operation and maintenance network multi-agents. Detailed Implementation

[0066] The present invention will now be described in further detail with reference to the accompanying drawings and implementation process.

[0067] Example: Figure 1 The flowchart shown is a process for an intelligent operation and maintenance scheduling method for charging stations based on a large language model and multi-starting-point policy reinforcement learning. The method includes the following steps:

[0068] Step 1: Collect alarm and latitude / longitude data of the target charging station, and construct a symmetrical global time-distance matrix by combining macro and local road network data. This includes the following steps:

[0069] Step 1-1: Extract the original alarm data, use the compilation tool to remove alarm storms with the same cause that are repeatedly reported by the same site, and obtain the actual number of independent faults at each site; and combine the standard repair time for a single fault with the maximum time limit set by the system to calculate the estimated maintenance workload for a single site.

[0070] Steps 1-2: Use the Amap API to obtain the backbone network time. For isolated sites and networks caused by coordinate deviations, calculate the spherical physical distance based on the Haversine formula and then integrate them into the network. The distance calculation formula is as follows:

[0071]

[0072] in, This is the baseline value for the Earth's average radius of curvature; These represent the geographical latitude and longitude coordinates of the initial fault site and the target legitimate road network node to be connected, respectively. These represent the geographical longitude coordinates of the starting fault site and the target legitimate road network node to be connected, respectively.

[0073] Steps 1-3: Introduce a local KNN micro-loop to establish forced connections for stations with a physical straight-line distance less than a threshold (set to 20,000 meters in this embodiment), and use the Gaode API to obtain the initial time feature matrix. .

[0074] Steps 1-4: To eliminate the directional asymmetry caused by one-way congestion and highway restrictions in the real road network, the original feature matrix and its transpose are globally symmetricized to obtain the final computation matrix. :

[0075]

[0076] in, Represents the original time matrix. This represents the transpose of the original time matrix.

[0077] Step 2: Initialize the large language model decision center to generate macro-control parameters (including weight parameters and penalty coefficients for each region, and the number of clusters). Based on these parameters, a clustering algorithm is used to divide all sites into several operation and maintenance regions. This includes the following steps:

[0078] Step 2-1: Pre-load the global average commute time into the load baseline to calculate the ideal workload threshold for a single person; the Large Language Model (LLM) inference output focuses on key areas, injecting high-order weights into the stations in those areas. The K-means++ strategy, which employs double-squared weighted optimization, is used to select initial cluster centers to break spatial mutual exclusion; candidate center points The formula for the expected initial probability of being selected by the system is as follows:

[0079]

[0080] in, and Representing candidate physical sites and The minimum value of the lowest bottom-level topological time distance from the set of all legal cluster centers determined in the current round; This represents the soft weight coefficient of the focus area assigned to a specific site by the large language model; This represents the total number of valid charging station nodes awaiting repair that are currently participating in the global clustering and segmentation process.

[0081] Step 2-2: When deploying discrete sites, dynamically calculate the expected percentage of overload fatigue for candidate grid personnel (the percentage exceeding a set threshold). Furthermore, a distance multiplier penalty method is employed, which combines the physical travel time with the dynamic penalty coefficient output by the LLM. This involves integrating and calculating the overall scheduling cost to prevent cross-city / regional order dispatch. The calculation formula is as follows:

[0082]

[0083] in, The lower the cost, the higher the allocation priority; Representing physical sites Drive to the grid The shortest road network time distance for the current cluster gravity center.

[0084] Steps 2-3: When updating the grid center, traverse the grid. For all members within the cluster, find the site that minimizes the total sum of real-time distances within the cluster to be the new center. The objective function for this search is as follows:

[0085]

[0086] If the grid is found to have relatively idle time, the airdrop mechanism is activated, forcing the center point to return to the most heavily loaded disaster area site globally.

[0087] Steps 2-4: Establish an elite retention mechanism to save the clustering partitioning scheme with the minimum global total time consumption within the maximum number of iterations specified internally, as the final network output of this round of macro scheduling.

[0088] Step 3: Utilize real 2D coordinates combined with a multi-point policy reinforcement learning network (POMO). The flowchart of the POMO method is shown below. Figure 2 As shown, the optimal operation and maintenance route for each region is determined through the following steps:

[0089] Step 3-1: Extract the true two-dimensional latitude and longitude coordinate tensors of stations within each grid submap As input features, to eliminate the illusion of route intersections and topological collapse generated by traditional dimensionality reduction.

[0090] Step 3-2: Perform global isometric normalization on the real spatial coordinates to prevent scaling distortion, and add a zero overflow compensation term. The operator formula is as follows:

[0091]

[0092] in, This represents a two-dimensional normalized tensor that has been proportionally compressed into the 0-1 space; and These represent the minimum and maximum values ​​in the longitude or latitude directions, respectively.

[0093] Step 3-3: Utilize the multi-instance enhancement mechanism of the POMO framework to rotate the normalized features by different angles. The affine transformation is used to generate a multi-view enhanced tensor input network. The rotation matrix transformation formula is as follows:

[0094]

[0095] in, Represents an enhanced coordinate-aware tensor set containing multiple independent view sequences of physical topology; This represents the spatial rotation angle offset of the discretized space.

[0096] Steps 3-4: Use a pre-trained tensor network based on the attention mechanism to perform greedy autoregressive decoding, and generate multiple sets of candidate paths without self-crossing in parallel. It also outputs the network's internal native evaluation cost simultaneously. .

[0097] Steps 3-5: Perform a posteriori arbitration to arbitrate the candidate set. Substitute the actual traffic time matrix Filtering the route with the shortest physical time; the joint graph of spatial network partitioning and DRL shortest path is shown below. Figure 3 As shown, the optimal executable route The judgment is as follows:

[0098]

[0099] in, After filtering, the final output is the optimal deterministic path execution sequence that has the shortest travel time in the objective physical world; This represents the complete set of candidate path policies generated by a multi-starting-point policy reinforcement learning network. Refers to the set of strategies Any single specific candidate exploration path in the; This represents the total number of charging stations awaiting repair within the currently controlled local maintenance grid that is undergoing neural routing scheduling. This indicates that during the execution of the candidate strategy path, the task node... Drive to the next mission node By querying the underlying physical base to obtain the actual time-mapping cost, based on the actual time matrix, the problem of generated paths being unusable in reality can be avoided.

[0100] Step 4: Calculate the total maintenance time and workload variance of the optimal route for each region. If the target is not met, generate a natural language diagnostic report and feed it back to the large language model to readjust the weight coefficients, penalty coefficients, and number of clusters. This includes the following steps:

[0101] Step 4-1: Calculate each grid set The final expected total time spent by a single person This time includes the physical transfer time along the route and the estimated maintenance operation time to reach the site, as shown in the following formula:

[0102]

[0103] in, This represents the downtime and repair costs for the corresponding site based on the number of faults.

[0104] Step 4-2: Extract the expected arithmetic mean of the total time taken by the global grid. And calculate the variance of the overall manpower workload balance. :

[0105]

[0106] in, This represents the total number of macro-level operation and maintenance grids dynamically assigned by the large language model in this round.

[0107] Step 4-3: Activate the anti-oscillation mechanism. If variance... If the timeout is less than the given limit and there is no personnel overload, the early stop mechanism will be activated to directly provide the best solution; if the timeout is severe, the underlying variance and time consumption information will be converted into natural language diagnostic information, and then combined with the most recent inference information to feed back to the large language model, allowing it to re-determine the parameter selection until the upper limit or the maximum number of times is reached.

[0108] Step 5: Output the final optimal route planning for all target charging stations and the total time for each route, which includes the following steps:

[0109] Step 5-1: The system retains the solutions that reach the satisfaction threshold in the above closed-loop iteration, and organizes the unmanned shortest routes generated by multi-starting-point strategy reinforcement learning, combined with the diagnostic report containing working time and variance distribution, and finally outputs a fully automated operation and maintenance dispatch list for charging stations in each region. The optimal path of a batch of stations is extracted as an example as follows:

[0110]

[0111] It should be noted that the above embodiments are not intended to limit the scope of protection of the present invention. Equivalent transformations or substitutions made based on the above technical solutions all fall within the scope of protection of the claims of the present invention.

Claims

1. A method for intelligent operation and maintenance scheduling of charging stations based on large language models and multi-starting-point policy reinforcement learning, characterized in that, Includes the following steps: Step 1: Collect alarm and latitude / longitude data of the target charging station, and construct a symmetrical global time-distance matrix by combining macro and local road network data; Step 2: Initialize the large language model decision center to generate macro-control parameters, including the weight parameters and penalty coefficients of each region, the number of clusters, and use a clustering algorithm based on these parameters to divide all sites into several operation and maintenance regions. Step 3: Extract real two-dimensional coordinates and combine them with a multi-starting point strategy reinforcement learning network to solve for the optimal operation and maintenance route for each region; Step 4: Calculate the total maintenance time and workload variance of the optimal route in each region. If the target is not met, generate a natural language diagnostic report and feed it back to Step 2. The large language model readjusts the weight coefficients, penalty coefficients and number of clusters for clustering. If the target is met, proceed to Step 5. Step 5: Output the final optimal route planning for all target charging stations and the total time for each route.

2. The intelligent operation and maintenance scheduling method for charging stations based on large language models and multi-starting-point policy reinforcement learning according to claim 1, characterized in that, Step 1 involves collecting alarm and latitude / longitude data of the target charging station, and then constructing a symmetrical global time-distance matrix by combining macroscopic and local road network data. The specific implementation process is as follows: Step 1: Extract the original alarm data, remove alarms with the same cause that are repeatedly reported by the same site, and obtain the actual number of independent faults at each site; and combine the standard repair time for a single fault with the maximum time limit set by the system to calculate the estimated maintenance workload for a single site. Step 2: Obtain the backbone road network time via the map API. For isolated stations and island networks, estimate the spherical physical distance using the Haversine formula and connect them to the network. The distance calculation formula is as follows: in, This is the baseline value for the Earth's average radius of curvature; These represent the absolute geographical latitude and longitude coordinates of the starting fault site and the target legitimate road network node to be connected, respectively. These represent the absolute geographical longitude coordinates of the initial fault site and the target legitimate road network node to be connected to, respectively. Step 3: Introduce a local KNN micro-loop to establish forced connections between stations whose straight-line distance is less than a threshold, and apply for and generate a time matrix through the Gaode API. To eliminate the directional asymmetry of the road network, it is symmetricized to obtain the final computational base. : in, Represents the original time matrix. This represents the transpose of the original time matrix.

3. The intelligent operation and maintenance scheduling method for charging stations based on large language models and multi-starting-point policy reinforcement learning according to claim 1, characterized in that, Step 2 initializes the large language model decision center to generate macro-control parameters, including weight parameters and penalty coefficients for each region, and the number of clusters. Based on these parameters, a clustering algorithm is used to divide all sites into several operation and maintenance regions. The specific implementation process is as follows: Step 1: Embed the global average commute time into the load baseline to calculate the ideal workload for each person; add weights to sites based on the list of key areas. The double-squared weighted K-means++ strategy is used to select initial points to break spatial mutual exclusion, and the center point The probability formula for being selected is as follows: in, and Representing candidate physical sites and The minimum bottom-level topological time distance from the set of all legal cluster centers determined in the current round is used to find the nearest selected center; This represents the soft weight coefficient of the focus area assigned to a specific site by the large language model; This represents the total number of valid charging station nodes awaiting repair that participate in clustering and segmentation in the current global physical network. Step 2: When allocating sites, calculate the overload ratio of the current grid. It employs a distance multiplier penalty method, using physical time as the base, and combines the overload ratio with the penalty coefficient output by the large language model. As a multiplier, strict measures are taken to prevent cross-city order dispatch, considering the overall cost. The calculation formula is as follows: in, The lower the cost, the higher the allocation priority; Representing physical sites Drive to the grid The shortest bottom-level road network time distance of the current cluster gravity center; This is a dynamically adjustable distance multiplier penalty coefficient; This represents the risk ratio that the estimated total workload exceeds the system's set threshold for the ideal workload per person. Step 3: Traverse the mesh when updating the medoid. For all members within the cluster, find the site that minimizes the total sum of real-time distances within the cluster to be the new center. The objective function for this search is as follows: in, This represents the controlled mesh after internal optimization, update, and convergence. The true physical geometric center inside; Represents the first segment dynamically partitioned by the system. A locally controlled subset of the operation and maintenance grid; These represent elements that are strictly limited to the current grid subset. Any two nodes of the charging station to be repaired within the system can be traversed. Representative node With nodes The underlying symmetrical travel time distance to the base is called between them; If an idle grid is detected, the airdrop mechanism is triggered, forcibly resetting the center to the site with the largest global workload; Step 4: Establish an elite retention mechanism. Within the maximum number of iterations, record the clustering partitioning scheme with the minimum global allocation cost as the final output.

4. The intelligent operation and maintenance scheduling method for charging stations based on large language models and multi-starting-point policy reinforcement learning according to claim 1, characterized in that, Step 3 extracts the real two-dimensional coordinates and combines them with a multi-point policy reinforcement learning network (POMO) to solve for the optimal operation and maintenance route for each region. The specific implementation process is as follows: Step 1: Extract the actual 2D latitude and longitude coordinates of each station within the grid submap As input features, to eliminate the illusion of route intersections caused by topological collapse; Step 2: Perform global isometric normalization on the real-space coordinates to prevent scaling distortion. The processing formula is as follows: in, This represents a two-dimensional normalized tensor that has been proportionally compressed into the 0-1 feature space; and These refer to the minimum and maximum values ​​of coordinates in orthogonal directions of longitude and latitude, respectively. To prevent zero overflow from the minimum value; Step 3: Utilize the POMO multi-instance enhancement mechanism to generate different rotation angles. The rotation matrix transformation formula for the enhanced tensor input network is as follows: in, It represents a set of enhanced coordinate-aware tensors containing multiple independent view sequences of physical topology; The spatial rotation angle offset scalar represents the discretized space. Step 4: Use the attention mechanism (Attention) tensor pre-trained network to perform greedy autoregressive decoding and generate multiple sets of parallel candidate paths. It also simultaneously outputs the internal native evaluation cost of multi-starting-point policy reinforcement learning networks. ; Step 5: Perform simulated real-world posterior arbitration to categorize the candidate set. Substitute into the real time matrix Filter the route with the shortest physical time, the optimal route The judgment is as follows: in, After filtering, the final output is the optimal deterministic path execution sequence that has the shortest travel time in the objective physical world; This represents the complete set of candidate path policies generated by a multi-starting-point policy reinforcement learning network. Refers to the set of strategies Any single specific candidate exploration path in the; This represents the total number of charging stations awaiting repair within the currently controlled local maintenance grid that is undergoing neural routing scheduling. This indicates that during the execution of the candidate strategy path, the task node... Drive to the next mission node Between these, query the actual time-mapping cost obtained from the underlying physical base.

5. The intelligent operation and maintenance scheduling method for charging stations based on large language models and multi-starting-point policy reinforcement learning according to claim 1, characterized in that, In step 4, the total maintenance time and workload variance of the optimal route for each region are calculated. If the target is not met, a natural language diagnostic report is generated and fed back to the large language model to readjust the weight parameters, penalty parameters, and number of clusters. The specific implementation process is as follows: Step 1: Calculate each grid set The final expected total time spent by a single person It includes transit time and stopover / repair time, and the formula is as follows: in, After a comprehensive simulation, the representative... Within a controlled operations and maintenance grid, the final expected total physical time for a single professional operations and maintenance engineer to complete all scheduled tasks within the daily shift. Represents the target grid The total number of sites with pending repair tasks that have been centrally allocated within the organization; Refers to the current station under repair and the next station to be visited; This represents the cost of time spent on maintenance and repairs. Step 2: Extract the average total time spent on the global grid And calculate the variance of the workload of the scheme. : in, This represents the pool of total macro-level operation and maintenance grid numbers dynamically and autonomously set by the Large Language Model (LLM) based on the natural language feedback from the previous round of the system in the current round. Represents the whole The arithmetic mean of the expected total physical time cost that each maintenance grid must bear is the baseline. Step 3: Activate the anti-oscillation locking mechanism, if the variance If the value is less than the minimum, a fixed number of grid parameters is sent to the large language model, and instructions can only be used to adjust the penalty factor. If there is a serious timeout or idle period, the analysis results are converted into text and sent to the large language model to reset the parameters until the minimum value or the maximum number of iterations is reached.

6. The intelligent operation and maintenance scheduling method for charging stations based on large language models and multi-starting-point policy reinforcement learning according to claim 1, characterized in that, If the solution reaches the satisfaction threshold in step 5 (the repair and driving time for each cluster is within 10 hours and the time is balanced), the final optimal route planning route and the total time of each route required for all target charging stations will be directly output. The specific implementation process is as follows: The system compiles and integrates the real shortest time route without human intervention generated by multi-starting point strategy reinforcement learning with the expected total time data of a single person; combined with the anti-vibration diagnosis report containing working time and variance distribution, standardized scheduling data is generated and fed back to the large language model decision module, and a fully automated, closed-loop operation and maintenance dispatch list for charging stations in different geographical locations is output.

7. A charging station operation and maintenance path planning system combining large language model decision-making and multi-starting-point policy reinforcement learning, characterized in that, The method is based on large language model decision and reinforcement learning, and the system includes: Data preprocessing and topology construction module: Collects and cleans alarm and latitude and longitude data of target charging stations, combines macro backbone network and local micro-circulation to construct underlying network, and derives symmetric global time distance matrix; Large Language Model Decision Module: Utilizes a structured output protocol to receive environmental summaries and historical feedback, and adaptively outputs macro-control parameters including the number of maintenance personnel (including weight parameters and penalty coefficients for each region, and the number of clusters). Site clustering module: Used to perform flexible memory clustering and classification of sites in different geographical locations by combining distance multiplier penalty method and airdrop activation mechanism, as a candidate pool of sites for operation and maintenance routes; Reinforcement learning module: Built-in real coordinate extraction, multi-starting point and multi-view enhancement and post-arbitration mechanism, used to plan the shortest real traffic time path selected by operation and maintenance personnel from the candidate pool without human intervention; Assessment and Diagnosis Module: Used to generate anti-vibration diagnostic reports with natural language features based on working time and variance distribution, to provide feedback on whether it is necessary to increase or decrease the number of maintenance personnel and adjust penalties or weights.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the intelligent operation and maintenance scheduling method for charging stations based on large language models and multi-starting-point policy reinforcement learning as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing computer instructions thereon, characterized in that, When the computer instruction is executed by the processor, it implements the intelligent operation and maintenance scheduling method for charging stations based on large language models and multi-starting-point policy reinforcement learning as described in any one of claims 1-6.

Images