A wide-area charging facility-oriented operation and maintenance resource dynamic scheduling method, medium and system

Abstract

The application provides a kind of wide-area charging facility-oriented operation and maintenance resource dynamic scheduling method, medium and system, belongs to wide-area charging facility operation and maintenance technical field, the present application is by receiving charging facility fault information and carrying out edge preprocessing, based on manifold learning driven multi-scale district adaptive fractal division algorithm dynamically generates operation and maintenance responsibility partition, match fault information to corresponding operation and maintenance responsibility district, start fault cascade propagation prediction model to deduce and calculate operation and maintenance resource pre-deployment location, execute multi-objective optimization operation and maintenance personnel scheduling decision, generate electronic work order and execute spare parts intelligent linkage, implement whole-process dynamic tracking and closed-loop management, through artificial intelligence prediction fault cascade propagation path and combine dynamic partition division and resource pre-deployment mechanism, solve the technical problems that wide-area charging facility operation and maintenance resource scheduling response lag and difficult to adapt to the dynamic change of fault space-time distribution.

Description

Technical Field

[0001] This invention belongs to the field of operation and maintenance technology of wide-area charging facilities. Specifically, it relates to a method, medium and system for dynamic scheduling of operation and maintenance resources for wide-area charging facilities. Background Technology

[0002] In the current operation and maintenance management of electric vehicle charging infrastructure, traditional operation and maintenance resource scheduling methods mainly rely on static partitioning and passive response mechanisms to handle faults. Existing technologies typically divide charging facilities into several operation and maintenance responsibility areas based on administrative divisions or fixed geographical grids, allocate maintenance tasks using the nearest distance principle or a simple work order queuing mechanism, and rely on manual experience to judge spare parts needs and formulate maintenance routes. Because the fault distribution of charging facilities exhibits obvious spatiotemporal dynamics and clustering characteristics, fixed partitioning methods cannot adapt to the time-varying characteristics of fault density. This leads to insufficient operation and maintenance personnel in some high-fault-density areas while resources are idle in low-density areas. Furthermore, traditional methods lack the ability to predict the cascading propagation of faults, and can only respond passively after a fault occurs, unable to deploy operation and maintenance resources in advance to deal with potential cascading fault spread. In other words, existing technologies suffer from technical problems such as lagging response times in the scheduling of operation and maintenance resources for wide-area charging facilities and difficulty in adapting to the dynamic changes in the spatiotemporal distribution of faults. Summary of the Invention

[0003] In view of this, the present invention provides a method, medium and system for dynamic scheduling of operation and maintenance resources for wide-area charging facilities, which can solve the technical problems in the prior art of delayed response of operation and maintenance resource scheduling for wide-area charging facilities and difficulty in adapting to the dynamic changes in the spatiotemporal distribution of faults.

[0004] This invention is implemented as follows: The first aspect of this invention provides a method for dynamic scheduling of operation and maintenance resources for wide-area charging facilities, including receiving charging facility fault information and performing edge preprocessing; dynamically generating operation and maintenance responsibility zones based on a manifold learning-driven multi-scale adaptive fractal partitioning algorithm; mapping the geographical coordinates, historical fault density, road travel time, and equipment health scores of charging facilities to a low-dimensional manifold space; recursively partitioning operation and maintenance responsibility zones according to the self-similarity principle of fractal geometry; calculating the Hausdorff fractal dimension of the operation and maintenance responsibility zone boundaries to evaluate the partitioning quality; and matching fault information to... The system corresponds to the maintenance responsibility area and calculates the geodesic distance between the low-dimensional manifold space coordinates and the centroid of each maintenance responsibility area. It then initiates a fault cascade propagation prediction model to deduce the fault propagation path and outputs a list of potential faulty equipment and their probability distribution within the next 30 minutes. Based on the list of potential faulty equipment and their probability distribution, it calculates the pre-deployment location of maintenance resources, executes multi-objective optimization of maintenance personnel scheduling decisions, and solves the personnel allocation scheme corresponding to the maximum value of the scheduling fitness function through the particle swarm optimization algorithm. It then generates electronic work orders and executes intelligent spare parts linkage, implementing full-process dynamic tracking and closed-loop management.

[0005] Specifically, the edge preprocessing involves deploying a fault preprocessing unit on the charging pile side to perform multi-level priority classification on the original fault data. Emergency faults are pushed to the central scheduling platform in real time via the MQTT protocol's Quality of Service Level 2 mode, while non-emergency faults are stored in a local cache queue and transmitted in batches after data compression.

[0006] The fault preprocessing unit is an embedded computing module deployed locally on the charging pile. It has edge data filtering and preliminary diagnosis functions and prioritizes fault signals such as voltage over-limit, current abnormality, and communication interruption through preset rules.

[0007] Specifically, the manifold learning-driven multi-scale area adaptive fractal partitioning algorithm uses local linear embedding technology to construct a four-dimensional feature vector from the geographical coordinates of charging facilities, historical fault density, road travel time, and equipment health score, and then extracts low-dimensional manifold space coordinates. In the low-dimensional manifold space, the boundaries of the operation and maintenance responsibility areas are recursively partitioned according to the fractal generation rules of the Koch curve.

[0008] The calculation of the Hausdorff fractal dimension uses the box counting method. The boundary curve of the maintenance responsibility area is covered with squares of side length ε, and the minimum number of squares N required for coverage is counted. The value of ε is changed to calculate multiple sets of correspondences between N and ε. The slope of the fitted straight line in the double logarithmic coordinate system is the Hausdorff fractal dimension.

[0009] Among them, when the Hausdorff fractal dimension is in the range of 1.2 to 1.6, it is considered a reasonable division, indicating that the boundary of the operation and maintenance responsibility area is neither too regular nor too fragmented, and meets the operability requirements of actual operation and maintenance.

[0010] The geodesic distance is the shortest path length between two points on the surface of a low-dimensional manifold space, calculated using Dijkstra's algorithm on the adjacency graph formed by the charging facility points, which is different from the straight-line Euclidean distance that ignores terrain and road network constraints.

[0011] If the minimum geodetic distance exceeds 1.5 times the average radius of the maintenance responsibility area, the cross-area collaborative scheduling mechanism will be triggered, allowing the call to maintenance personnel from adjacent maintenance responsibility areas.

[0012] The fault cascade propagation prediction model is a time-series prediction system based on a liquid neural network architecture. It consists of an input layer, a liquid reservoir layer, and a readout layer. The liquid reservoir layer contains eight hundred leakage integral firing type spiking neurons, which transmit pulse signals through synaptic connections.

[0013] The synaptic connection weights possess short-term plasticity characteristics. After each pulse is fired by the presynaptic leak integral firing spiking neuron, the synaptic connection weights decay exponentially. The decay time constant of the synaptic connection weights is dynamically determined according to the fault propagation rate adjustment function.

[0014] The fault propagation rate adjustment function calculates the fault propagation rate index based on three parameters: the number of newly faulty devices in the past ten minutes, the average topological distance between faulty devices, and the grid load rate. The synaptic connection weight decay time constant is set according to the numerical range of the fault propagation rate index.

[0015] Specifically, the calculation of the pre-deployment location of the operation and maintenance resources is obtained by multiplying the geographical coordinates of each device in the list of potential faulty devices by its probability distribution and then calculating the weighted centroid.

[0016] The scheduling fitness function comprehensively considers four dimensions: the shortest road network time from the current location of the maintenance personnel to the fault point, the matching degree between the personnel's skill level and the fault type, the length of the current task queue of the maintenance personnel, and the historical repair success rate. The weight coefficients of the four normalized indicators are 0.4, 0.3, 0.2, and 0.1, respectively.

[0017] The electronic work order includes a unique identifier for the faulty equipment, a fault type code, the equipment's geographical coordinates, the assigned maintenance personnel's employee number, the estimated arrival time, and a recommended repair plan. The system automatically queries the spare parts inventory database to obtain the inventory quantity and location information of the spare parts corresponding to the fault type code.

[0018] When the inventory of spare parts is lower than the safety threshold, the nearest spare parts allocation path is pushed to the maintenance personnel. The nearest spare parts allocation path is the shortest travel route from the current location to the nearest spare parts warehouse.

[0019] Specifically, the closed-loop management involves maintenance personnel reporting their arrival time, start time, completion time, and results via mobile terminals. Once the electronic work order is closed, the maintenance time, replacement spare parts model, and root cause analysis are automatically written into the equipment health record. The information on the charging facility failure is then used as a new sample to input into the fault cascading propagation prediction model for incremental learning.

[0020] A second aspect of the present invention provides a computer-readable storage medium storing program instructions, which, when executed in a computer, are used to perform the above-described method for dynamic scheduling of operation and maintenance resources for wide-area charging facilities.

[0021] A third aspect of the present invention provides a dynamic scheduling system for operation and maintenance resources of wide-area charging facilities, comprising the aforementioned computer-readable storage medium, wherein the system is a computer, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor for executing program instructions stored in the computer-readable storage medium.

[0022] This invention dynamically generates operation and maintenance responsibility zones through a multi-scale adaptive fractal partitioning algorithm driven by manifold learning. It maps the geographical coordinates of charging facilities, historical fault density, road travel time, and equipment health scores to a low-dimensional manifold space and recursively partitions the operation and maintenance responsibility zones according to the self-similarity principle of fractal geometry. The partitioning quality is evaluated using Hausdorff fractal dimension, and geodesic distance is used instead of Euclidean distance to calculate the compactness of the operation and maintenance responsibility zones, solving the problem that traditional fixed partitioning cannot adapt to the dynamic changes in the spatiotemporal distribution of faults. Furthermore, this invention employs a fault cascading propagation prediction model based on a liquid neural network architecture to deduce fault propagation paths, outputting a list of potential faulty devices and their probability distribution within the next 30 minutes. Based on the prediction results, it calculates the pre-deployment locations of operation and maintenance resources and adjusts the standby coordinates of idle operation and maintenance personnel, solving the problem that traditional passive response mechanisms cannot proactively address cascading faults. In summary, this invention solves the technical problems mentioned in the background art, such as the lagging response of wide-area charging facility operation and maintenance resource scheduling and its inability to adapt to the dynamic changes in the spatiotemporal distribution of faults. Attached Figure Description

[0023] Figure 1 This is a flowchart of the method of the present invention.

[0024] Figure 2 A manifold learning-driven multi-scale regional adaptive fractal partitioning boundary map for operation and maintenance responsibility partitioning.

[0025] Figure 3 The probability distribution of charging facility failures over the next 30 minutes, as output by the fault cascading propagation prediction model. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0027] like Figure 1 The diagram shown is a flowchart of a dynamic scheduling method for operation and maintenance resources of wide-area charging facilities provided by the first aspect of the present invention. This method includes the following steps:

[0028] S01. Receive charging facility fault information and perform edge preprocessing. Deploy a fault preprocessing unit on the charging pile side to perform multi-level priority classification on the original fault data. Push emergency faults to the central dispatch platform in real time through the MQTT protocol service quality level 2 mode. Store non-emergency faults in the local cache queue and transmit them in batches after data compression.

[0029] S02. A multi-scale adaptive fractal partitioning algorithm based on manifold learning dynamically generates operation and maintenance responsibility zones. The geographical coordinates of charging facilities, historical fault density, road travel time, and equipment health scores are mapped to a low-dimensional manifold space. In the low-dimensional manifold space, operation and maintenance responsibility zones are recursively partitioned according to the self-similarity principle of fractal geometry. The Hausdorff fractal dimension of the boundary of the operation and maintenance responsibility zone is calculated to evaluate the partitioning quality. When the Hausdorff fractal dimension is in the range of 1.2 to 1.6, it is judged as a reasonable partition.

[0030] S03. Match the received charging facility fault information to the corresponding operation and maintenance responsibility area, extract the low-dimensional manifold space coordinates of the faulty equipment, calculate the geodesic distance between the low-dimensional manifold space coordinates and the centroid of each operation and maintenance responsibility area, and allocate the charging facility fault information to the operation and maintenance responsibility area with the smallest geodesic distance. If the smallest geodesic distance exceeds 1.5 times the average radius of the operation and maintenance responsibility area, the cross-area collaborative scheduling mechanism is triggered.

[0031] S04. Start the fault cascading propagation prediction model to deduce the fault propagation path. Input the electrical topology connection relationship of the current faulty equipment, the historical cascading fault sequence, and the real-time load data into the fault cascading propagation prediction model. The fault cascading propagation prediction model outputs a list of potential faulty equipment and their probability distribution within the next 30 minutes. Calculate the pre-deployment location of operation and maintenance resources based on the list of potential faulty equipment and their probability distribution.

[0032] S05. Perform multi-objective optimization of the scheduling decision of maintenance personnel. The scheduling fitness function is constructed by comprehensively considering four dimensions: the shortest road network time from the current location of maintenance personnel to the fault point, the matching degree between personnel skill level and fault type, the current task queue length of maintenance personnel, and the historical repair success rate. The personnel allocation scheme corresponding to the maximum value of the scheduling fitness function is solved by particle swarm optimization algorithm. At the same time, the standby coordinates of idle maintenance personnel are adjusted according to the pre-deployment location of maintenance resources.

[0033] S06. Generate an electronic work order and execute intelligent spare parts linkage. The electronic work order includes the unique identifier of the faulty equipment, the fault type code, the geographical coordinates of the equipment, the assigned maintenance personnel employee number, the estimated arrival time, and the recommended repair plan. The system automatically queries the spare parts inventory database to obtain the inventory quantity and storage location information of the spare parts corresponding to the fault type code. When the inventory quantity of spare parts is lower than the safety threshold, the system pushes the nearest spare parts allocation path to the maintenance personnel. The electronic work order is pushed to the maintenance personnel via mobile terminal and the push timestamp is recorded.

[0034] S07. Implement full-process dynamic tracking and closed-loop management. Maintenance personnel report arrival time, start time, completion time and results via mobile terminals. The system updates the visualized maintenance resource distribution map in real time, showing the location and task status of each maintenance personnel. After the electronic work order is closed, the maintenance time, replacement spare parts model and root cause analysis are automatically written into the equipment health record. The maintenance personnel performance database is updated synchronously with the number of maintenance operations and the historical maintenance success rate. The information on the current charging facility failure is used as a new sample to input into the fault cascading propagation prediction model for incremental learning.

[0035] The fault preprocessing unit is an embedded computing module deployed locally on the charging pile. It has edge data filtering and preliminary diagnosis functions, and prioritizes fault signals such as voltage over-limit, current abnormality, and communication interruption through preset rules.

[0036] Among them, the MQTT protocol's Quality of Service Level 2 mode is the highest reliability transmission level defined by the Message Queuing Telemetry Transport Protocol. The sender and receiver ensure that the message is received only once and is not lost through a four-way handshake.

[0037] Among them, the manifold learning-driven multi-scale adaptive fractal partitioning algorithm is a dynamic region partitioning method that combines manifold dimensionality reduction and fractal geometry. First, it uses local linear embedding technology to convert the geographical coordinates of charging facilities into two-dimensional planar coordinates. The historical fault density is defined as the number of faults within a unit area over the past 90 days, normalized to the range of 0 to 1. Road travel time is calculated using the road network shortest path algorithm to obtain the minute-level reachability time between each charging facility, and then normalized by dividing by the maximum travel time of the area. The equipment health score is calculated by comprehensively considering equipment runtime, historical fault frequency, and the remaining lifespan of key components, and then normalized. After constructing a four-dimensional feature vector, local linear embedding technology is used to calculate the reachability of each charging facility. The k-nearest neighbor set of each facility point is used to solve for the local linear reconstruction weight matrix. Eigenvalue decomposition is performed on the local linear reconstruction weight matrix to extract the eigenvectors corresponding to the two largest eigenvalues, which are then used as coordinates in the low-dimensional manifold space. In the low-dimensional manifold space, a fractal seed point is initialized at the location with the highest charging facility density. The boundary of the maintenance responsibility area is expanded outward according to the fractal generation rules of the Koch curve. In each iteration, the boundary segment of the maintenance responsibility area is trisected, and an equilateral triangle is formed by bulging outward at the middle third. The iteration depth is adaptively adjusted according to the local charging facility density gradient. For areas with a charging facility density gradient greater than 5 units per square kilometer, 4 to 6 iterations are used; for areas with a charging facility density gradient less than 5 units per square kilometer,... The region is iterated in 2 to 3 layers to calculate the Hausdorff fractal dimension of the generated operation and maintenance responsibility area boundary. The Hausdorff fractal dimension is achieved through box counting. The boundary curve of the operation and maintenance responsibility area is covered with squares of side length ε. The minimum number of squares N required to cover the boundary curve is calculated. The correspondence between N and ε is calculated by changing the value of ε. The slope of the fitted straight line in the double logarithmic coordinate system is the Hausdorff fractal dimension. When the Hausdorff fractal dimension is in the range of 1.2 to 1.6, it indicates that the boundary of the operation and maintenance responsibility area is neither excessively regular nor excessively fragmented, meeting the operability requirements of actual operation and maintenance. If the Hausdorff fractal dimension deviates from the range of 1.2 to 1.6, the fractal iteration parameters are adjusted and the boundary is redrawn. The compactness of the maintenance responsibility area is calculated by using geodesic distance instead of Euclidean distance. The geodesic distance is the shortest path length between two points on the surface of a low-dimensional manifold space. It is calculated on the adjacency graph formed by charging facility points using the Dijkstra algorithm. The compactness of the maintenance responsibility area is defined as the average geodesic distance from all charging facility points in the maintenance responsibility area to the centroid of the maintenance responsibility area divided by the equivalent radius of the maintenance responsibility area. The smaller the compactness of the maintenance responsibility area, the closer the shape of the maintenance responsibility area is to a circle, which is beneficial to shortening the response time. The multi-scale area adaptive fractal partitioning algorithm driven by manifold learning performs a full repartition every 7 days. When the number of faults in a single day exceeds 1.8 times the historical average for the same period, an immediate repartition is triggered.

[0038] The manifold learning-driven multi-scale adaptive fractal partitioning algorithm captures the nonlinear manifold structure of charging facilities in a high-dimensional heterogeneous feature space through local linear embedding technology. It projects geographic coordinates, historical fault density, road travel time, and equipment health scores to a lower-dimensional manifold space, avoiding the sensitivity of traditional Euclidean clustering to distance metric distortion. The self-similarity principle of fractal geometry ensures that the boundaries of maintenance responsibility areas maintain morphological consistency at different scales. The adaptively adjusted iteration depth ensures that the refined management of high-density urban areas is coordinated with the coarse-grained coverage of low-density suburbs. The Hausdorff fractal dimension serves as a quantitative evaluation index to mathematically represent the rationality of the partitioning. The geodesic distance metric conforms to the connectivity characteristics of the actual road network. The multi-scale partitioning system takes into account both the global goal of load balancing between regions and the constraints of local traffic conditions. It significantly improves the adaptability of maintenance responsibility partitioning to the dynamic changes in the spatiotemporal distribution of faults, reduces the idle rate of maintenance resources caused by improper maintenance responsibility partitioning, and enhances the robustness of the scheduling system under fault surge scenarios.

[0039] Among them, Local Linear Embedding (MLE) is a nonlinear dimensionality reduction algorithm that assumes each data point in a high-dimensional space is linearly represented by its nearest neighbors, mapping the data to a low-dimensional space by preserving local linear relationships. Geodesic distance is defined in manifold geometry as the shortest path length along the manifold surface between two points, distinct from the linear Euclidean distance that ignores terrain and road network constraints. Hausdorff fractal dimension is a non-integer dimensionality index describing the complexity and space-filling capacity of a geometric figure, calculated by box counting to statistically determine the scaling relationship of the minimum number of units required to cover the figure at different scales. Koch curves are fractal geometric figures generated by recursively trisecting line segments and constructing equilateral triangle protrusions in the middle, exhibiting self-similarity. Dijkstra's algorithm is a greedy algorithm for calculating the single-source shortest path in a graph, progressively expanding from the starting node to all reachable nodes while maintaining the shortest distance. The centroid of the maintenance responsibility area is the geometric center of the low-dimensional manifold coordinates of all charging facility points within the maintenance responsibility area. The equivalent radius of the maintenance responsibility area is the radius of a circle with equal area obtained by dividing the area of ​​the maintenance responsibility area by the square root of pi. The cross-area collaborative scheduling mechanism allows for resource sharing among maintenance personnel in adjacent maintenance responsibility areas when the fault point is too far from the centroid of its respective maintenance responsibility area. The fault cascading propagation prediction model is a time-series prediction system based on a liquid neural network architecture, used to predict the propagation path and impact range of faults in the electrical topology network.

[0040] The fault cascading propagation prediction model consists of an input layer, a liquid reservoir layer, and a readout layer. The input layer receives the node code of the faulty device, the fault type's unique heat vector, the electrical connection adjacency matrix between devices, the current load rate of each device, and the historical cascading fault sequence. The liquid reservoir layer contains 800 leakage integral firing type spiking neurons. The membrane potential of these neurons follows the leakage integral dynamics equation, decays exponentially over time, and jumps upon receiving an input pulse. When the membrane potential exceeds the firing threshold, the neuron generates an output pulse and resets its membrane potential. In the liquid reservoir, leakage integral-fired spiking neurons transmit pulse signals through synaptic connections. The synaptic connection weights are initialized as random values ​​following a Gaussian distribution with a mean of 0 and a standard deviation of 0.3. The connection probability between leakage integral-fired spiking neurons within the liquid reservoir is 0.15. The synaptic connection weights exhibit short-term plasticity. After each pulse fired by a presynaptic leakage integral-fired spiking neuron, the synaptic connection weights decay exponentially. The decay time constant of the synaptic connection weights is dynamically determined according to the fault propagation rate adjustment function. The readout layer consists of 256 sigmoid-activated fully connected neurons. The readout layer outputs the fault probability vector and fault type distribution for each charging facility within the next 30 minutes.

[0041] The steps for establishing the training dataset for the fault cascading propagation prediction model include: extracting all fault work order records from the operation and maintenance system database for the past three years; each fault work order record includes a fault occurrence timestamp, a unique identifier for the faulty equipment, a fault type label, a list of equipment that failed during the same period, and the electrical topology of the equipment; grouping the fault work order records according to the equipment that initiated the fault, with each group constituting a cascading fault event sample; extracting all equipment that failed within 30 minutes of the initial fault as a label sequence; encoding the 10 kV distribution line connection relationship, transformer capacity ratio, and line impedance parameters between equipment into an electrical connection adjacency matrix between equipment; and collecting the active power, reactive power, voltage amplitude, and current amplitude of each equipment at the time of the fault occurrence to form a... The state feature vector is normalized to the fault occurrence timestamp and converted into a relative time offset. Faults occurring in the same cascaded fault event sample in chronological order are constructed into a directed propagation graph. Nodes in the directed propagation graph are devices, edges are fault propagation paths, and edge weights are propagation time intervals. 80% of the cascaded fault event samples are randomly selected as the training set, 10% as the validation set, and 10% as the test set. Data augmentation is performed on the training set by randomly deleting 5% of the edges in the electrical connection adjacency matrix between devices to simulate line maintenance status, and by adding Gaussian noise with a standard deviation of 0.05 to the current load rate of each device to simulate load fluctuations.

[0042] The training steps of the fault cascade propagation prediction model include: initializing the membrane potential of the leakage integral firing type spiking neurons in the liquid reservoir layer to the resting potential; initializing the synaptic connection weights to random values; encoding the input features of the training set samples as pulse sequences and inputting them into the liquid reservoir layer; encoding the fault occurrence time as a high-frequency pulse; mapping the current load rate of each device to the pulse firing frequency; updating the membrane potential of the leakage integral firing type spiking neurons in the liquid reservoir layer according to the received input pulses and cyclic connection pulses; generating transient pulse response patterns in the leakage integral firing type spiking neurons; and collecting the firing rate time series of the leakage integral firing type spiking neurons in the liquid reservoir layer under the action of the pulse sequence as the reservoir layer's response time. The reservoir state is input to the readout layer. The readout layer weights are trained using the backpropagation algorithm. The loss function is the cross-entropy between the predicted fault probability and the actual fault label. The optimizer uses the Adam algorithm with a learning rate of 0.001. Each batch contains 32 cascaded fault event samples, and the training lasts for 50 epochs. Every 5 epochs, the performance of the fault cascade propagation prediction model is evaluated on the validation set. If the validation set loss does not decrease for 3 consecutive times, the learning rate is reduced to 0.5 times the original learning rate value. After training, the liquid reservoir layer weights are fixed, and only the readout layer weights are retained. After the fault cascade propagation prediction model is deployed, incremental learning is performed every 100 new cascaded fault event samples received. During incremental learning, only the readout layer weights are updated, while the liquid reservoir layer structure remains unchanged.

[0043] The fault cascade propagation prediction model simulates the dynamic characteristics of a liquid state machine in a biological nervous system. The heterogeneous connections of numerous leakage integral firing type spiking neurons in the liquid reservoir layer form a high-dimensional transient dynamic system, mapping the time-varying historical cascade fault sequence input into rich spatiotemporal pulse firing patterns. The leakage integral mechanism of the leakage integral firing type spiking neurons naturally possesses a time-scale memory effect. The short-term plasticity characteristics endow the fault cascade propagation prediction model with adaptive sensitivity to the temporal characteristics of fault propagation. The high-dimensional chaotic dynamics of the liquid reservoir layer provide powerful nonlinear representation capabilities, enabling the fault cascade propagation prediction model to capture the nonlinear propagation patterns and branch paths of fault cascades in the electrical topology network. The readout layer extracts fault propagation patterns from the transient response of the liquid reservoir layer through supervised learning, avoiding gradient vanishing and gradient descent issues in the training process of traditional recurrent neural networks. To address the issue of high-explosion, the design of fixed liquid reservoir layer weights significantly reduces the scale of training parameters. The incremental learning mechanism enables the fault cascading propagation prediction model to continuously absorb newly emerging fault propagation patterns without requiring full retraining. The predicted list of potential faulty devices and their probability distribution provide a forward-looking decision-making basis for the pre-deployment location of maintenance resources. By deploying standby maintenance personnel in predicted high-risk areas in advance, the first response time to large-scale cascading faults is significantly shortened, and the scope of power outages caused by fault propagation is reduced. The pre-deployment strategy optimized by Markov decision process balances the cost of preventive deployment with the risk of potential fault losses under the constraint of a limited number of maintenance personnel. The short-term plasticity characteristics ensure the long-term adaptability of the fault cascading propagation prediction model to changes in charging network topology and new fault modes, thereby improving the overall proactive defense level of the scheduling system against sudden cascading faults.

[0044] Among them, the leaky integral firing spiking neuron is a type of biologically inspired neuron model. The membrane potential of the leaky integral firing spiking neuron leaks towards the equilibrium potential in the resting state. When it receives synaptic input, the membrane potential of the leaky integral firing spiking neuron jumps, and when it reaches the firing threshold, it fires a pulse and resets.

[0045] Among them, short-term plasticity is a phenomenon in which the efficiency of synaptic transmission changes briefly with the history of neural activity. High-frequency firing of presynaptic leaky integral firing spiking neurons leads to a temporary enhancement or inhibition of synaptic connection weights, lasting from hundreds of milliseconds to several seconds.

[0046] The fault propagation rate adjustment function is used to dynamically adjust the synaptic connection weight decay time constant of the synaptic connection weight of the leakage integral-release type pulse neuron in the liquid reservoir according to the current fault propagation speed. The fault propagation rate adjustment function calculates the fault propagation rate index based on three parameters: the number of newly faulty devices in the past 10 minutes, the average topological distance between faulty devices, and the grid load factor. The number of newly faulty devices is obtained by statistically analyzing the total number of fault work orders reported within the previous 10-minute time window. The average topological distance between faulty devices is defined as the arithmetic mean of the shortest path hops from the newly faulty device to the initial faulty device in the electrical topology graph. The grid load factor is the total load power of all online charging facilities in the current time period divided by the rated total capacity. Dividing the number of newly faulty devices by the 10-minute time span yields the fault occurrence frequency in units of times per minute. Dividing the average topological distance between faulty devices by the regional electrical topology... The maximum diameter of the plot is normalized to the interval between 0 and 1. The grid load factor is used directly. The fault propagation rate index is defined as the fault occurrence frequency multiplied by the normalized average topological distance between faulty devices and then multiplied by the grid load factor. When the fault propagation rate index a∈[0, 0.3), it is judged as slow propagation, and the synaptic connection weight decay time constant is set to 500 milliseconds. When a∈[0.3, 0.7), it is judged as medium propagation, and the synaptic connection weight decay time constant is set to 200 milliseconds. When a≥0.7, it is judged as fast propagation, and the synaptic connection weight decay time constant is set to 80 milliseconds. The shorter synaptic connection weight decay time constant makes the leakage integral firing type spiking neurons in the liquid reservoir more sensitive to the response of new inputs, adapting to the rapidly evolving fault propagation situation. The longer synaptic connection weight decay time constant maintains the stability of the synaptic connection weight, which is suitable for the slow and gradual fault diffusion process.

[0047] Among them, the liquid reservoir layer is a recurrent neural network module composed of large-scale leakage integral firing type spiking neurons and their dynamic synaptic connections. The state of the leakage integral firing type spiking neurons continuously evolves with the input to form transient trajectories, providing high-dimensional temporal feature representation.

[0048] The electrical connection adjacency matrix between devices is a square matrix that represents the connection relationships between nodes in the graph structure. An element of the electrical connection adjacency matrix between devices is 1 if there is an edge between the corresponding nodes, and 0 if there is no edge. The elements of the weighted electrical connection adjacency matrix between devices are the weight values ​​of the edges.

[0049] Cross-entropy is an information-theoretic metric that measures the difference between the predicted probability distribution and the true distribution. It is used as a loss function for classification tasks, and the smaller the value, the more accurate the prediction.

[0050] Among them, the Adam algorithm is a gradient descent optimization method with an adaptive learning rate. It combines momentum term and gradient second moment estimation, and uses personalized learning rates for different parameters to accelerate convergence.

[0051] Among them, the Markov decision process is the mathematical framework for sequential decision problems. The system state transition depends only on the current state and action, and the strategy is optimized by maximizing the cumulative reward.

[0052] The pre-deployment location of operation and maintenance resources is the geographical coordinates of the operation and maintenance personnel to be on standby in advance, calculated based on the list of potential faulty equipment and its probability distribution. It is obtained by multiplying the geographical coordinates of each equipment in the list of potential faulty equipment by its probability distribution and then calculating the weighted centroid.

[0053] The scheduling fitness function is used to evaluate the overall suitability of maintenance personnel to perform fault repair tasks. It calculates the shortest road network travel time from the maintenance personnel's current location to the fault point using a path planning algorithm, divides it by the maximum travel time in the area, and normalizes the result. Personnel skill levels are categorized into beginner, intermediate, and advanced, assigned values ​​from 1 to 3 respectively. Fault type difficulty is categorized into easy, moderate, and difficult, assigned values ​​from 1 to 3 respectively. Skill matching is defined as the personnel skill level minus the fault type difficulty, plus 2, then divided by 4 and normalized to the range of 0 to 1. The current task queue length of the maintenance personnel is divided by the maximum task capacity per person and normalized. The historical repair success rate is directly normalized by dividing the percentage value by 100. The scheduling fitness function is a weighted sum of the four normalized indicators, with weight coefficients of 0.4, 0.3, 0.2, and 0.1 respectively. A higher scheduling fitness function value indicates that the maintenance personnel are more suitable to perform the current fault repair task.

[0054] The Particle Swarm Optimization (PSO) algorithm is a swarm intelligence optimization method that simulates bird flock foraging behavior. Particles update their velocity and position in the solution space based on their historical best position and the group's best position, iteratively searching for the extreme value of the objective function. The personnel allocation scheme determines the correspondence between maintenance personnel and fault tasks after the PSO algorithm maximizes the scheduling fitness function. The idle maintenance personnel standby coordinates are the geographical locations of maintenance personnel without current tasks, adjusted according to the pre-deployment locations of maintenance resources. The fault type code is a digital identifier for charging facility fault types, including classifications such as voltage exceeding limits, abnormal current, communication interruption, insulation faults, and contactor faults. The spare parts inventory database stores the model specifications, inventory quantity, storage location information, and safety thresholds of various maintenance spare parts. The safety threshold is the minimum safe level for spare parts inventory; an alert is triggered when the inventory quantity falls below the safety threshold. The nearest spare parts allocation path is the shortest route from the current location to the nearest spare parts storage location, calculated by a path planning algorithm. The equipment health record is a database table that records the operational status and maintenance history of the charging facility throughout its entire lifecycle. It includes fields such as equipment commissioning date, cumulative operating hours, historical fault list, replacement spare parts model record, and equipment health score. The maintenance personnel performance database stores data tables on maintenance personnel employee numbers, number of repairs, historical repair success rate, average response time, and skill level certification.

[0055] The specific implementation methods of the above steps are described in detail below.

[0056] The specific implementation of step S01 involves configuring an embedded computing module locally on the charging pile as a fault preprocessing unit. This module receives raw data such as voltage, current, and communication status collected by the charging pile and performs real-time analysis of the data through a preset rule engine. The rule engine classifies faults into two categories, emergency faults and non-emergency faults, based on indicators such as voltage deviation rate, current fluctuation amplitude, and communication response delay. Faults with voltage deviation exceeding 15% of the rated value, current fluctuation exceeding 100A, or communication interruption lasting more than 30 seconds are identified as emergency faults, while the rest are identified as non-emergency faults. Emergency fault data is sent through the QoS level 2 mode of the message queue telemetry transmission protocol. In this mode, the sender publishes a message, the receiver returns confirmation, the sender publishes release, and the receiver returns to complete a four-way handshake to ensure reliable message delivery. Non-emergency fault data is first written to a local circular buffer queue. When the buffer queue accumulates 50 records or the time interval exceeds 5 minutes, batch transmission is triggered. Before transmission, lossless compression is performed on the data to reduce network bandwidth usage. The purpose of this step is to complete fault screening and classification at the edge side to avoid a large amount of redundant data flooding into the central platform and causing network congestion.

[0057] The specific implementation of step S02 involves extracting the latitude and longitude coordinates of all charging facilities from the historical database, including the number of faults per unit area over the past 90 days, the shortest path time of the road network, the equipment runtime, historical fault frequency, and the remaining lifespan of key components. The number of faults is normalized by dividing the number of faults by the maximum number of faults in the region to obtain the historical fault density feature. The road network time is normalized by dividing the road network time by the maximum travel time in the region to obtain the road travel time feature. The equipment runtime, fault frequency, and component lifespan are weighted and summed using preset weighting coefficients of 0.4, 0.3, and 0.3, and then normalized to obtain the equipment health score. This constructs a four-dimensional data structure including geographical coordinates, historical fault density, road travel time, and equipment health score. The four-dimensional feature vector is reduced using a local linear embedding technique. This technique first calculates the k-nearest neighbor set for each charging facility point, with k set to 12. It then minimizes the reconstruction error to solve for the local linear reconstruction weight matrix, ensuring each point can be linearly represented by its nearest neighbors. Eigenvalue decomposition is performed on the reconstruction weight matrix to extract the eigenvectors corresponding to the two largest eigenvalues, which serve as the coordinates for the two-dimensional low-dimensional manifold space. Within this low-dimensional manifold space, the location with the highest charging facility density is selected as the fractal seed point. Following the Koch curve generation rules, the area boundary is expanded outwards, dividing the initial boundary segment into three equal parts and constructing an equilateral triangle protrusion at the middle third. Based on the local charging... The iteration depth is adaptively determined by the facility density gradient. For areas with a density gradient greater than 5 units per square kilometer, 4 to 6 iterations are performed to finely fit the high-density distribution at the boundary. For areas with a density gradient less than 5 units per square kilometer, 2 to 3 iterations are performed to avoid excessive fragmentation. After iteration, the Hausdorff fractal dimension is calculated using box counting. The minimum number of squares required is counted by covering the boundary curve with squares of lengths 10m, 5m, 2m, and 1m. The slope of the linear relationship between the number of squares and the side length is fitted in a double logarithmic coordinate system, which is the Hausdorff fractal dimension. When the Hausdorff fractal dimension is between 1.2 and 1.6, the boundary complexity is considered moderate and meets the practical operational feasibility. If the Hausdorff fractal dimension is between 1.2 and 1.6, the boundary complexity is considered moderate and meets the practical operational feasibility. If the fractal dimension deviates from the interval, the iteration depth is adjusted and the area is re-divided. The Dijkstra algorithm is used to calculate the geodesic distance on the adjacency graph of charging facilities to replace the Euclidean distance to evaluate the compactness of the area. The compactness is defined as the average geodesic distance from all points in the area to the centroid divided by the equivalent radius. The purpose of this step is to capture the nonlinear distribution structure of charging facilities in the heterogeneous feature space through manifold learning, and combine fractal geometry to realize the multi-scale adaptive division of the operation and maintenance responsibility area, so that the area boundary maintains the consistency of shape at different scales. The full re-division is performed every 7 days. When the number of faults in a single day exceeds 1.8 times the historical average for the same period, an immediate re-division is triggered to adapt to the dynamic changes in the spatiotemporal distribution of faults.

[0058] The specific implementation of step S03 is as follows: after receiving the charging facility fault information, extract the unique identifier of the faulty device, query the low-dimensional manifold space coordinates of the device from the database, calculate the geodesic distance between the coordinates and the centroid of each maintenance responsibility area, where the centroid coordinates are the arithmetic mean of the low-dimensional manifold space coordinates of all charging facility points within the area. The geodesic distance is solved by running Dijkstra's algorithm on the charging facility adjacency graph. The fault information is assigned to the maintenance responsibility area with the smallest geodesic distance. If the minimum geodesic distance exceeds 1.5 times the average radius of the area, the fault point is determined to be located at the edge of the area. At this time, the cross-area collaborative scheduling mechanism is activated, allowing maintenance personnel from adjacent areas to participate in task allocation. The purpose of this step is to use the geodesic distance to accurately reflect the actual road network connectivity and achieve precise matching between fault information and maintenance responsibility areas.

[0059] The specific implementation of step S04 involves inputting the node encoding of the current faulty equipment, the fault type, the unique heat vector, the connection relationship of the 10 kV distribution line between the equipment, the transformer capacity ratio, the line impedance parameter encoding, the active power, reactive power, voltage amplitude, current amplitude, and historical cascaded fault sequence of each equipment into a fault cascade propagation prediction model based on a liquid neural network architecture. This model contains a liquid reservoir layer composed of 800 leakage integral firing-type spiking neurons. The connection probability between neurons is set to 0.15, and the synaptic connection weights are initialized as random Gaussian distributions with a mean of 0 and a standard deviation of 0.3. The neuron membrane potential decays exponentially with time according to the leakage integral dynamics equation and jumps upon receiving an input pulse. When the membrane potential exceeds the firing threshold, the neuron generates an output pulse and resets the membrane potential. The synaptic connection weights possess short-term plasticity characteristics; after each firing pulse from the presynaptic neuron, the synaptic connection weight decays exponentially, and the decay time constant is dynamically determined according to the fault propagation rate adjustment function. This function calculates the fault propagation rate index based on the number of newly faulty devices in the past 10 minutes, the average topological distance between faulty devices, and the grid load rate. When the index is in the range of 0 to 0.3, the decay time constant is set to 500 milliseconds to adapt to slow propagation; when it is in the range of 0.3 to 0.7, it is set to 200 milliseconds to adapt to medium-speed propagation; and when it is greater than or equal to 0.7, it is set to 80 milliseconds to adapt to fast propagation. The readout layer consists of 256 sigmoid-activated fully connected neurons. It outputs the fault probability vector and fault type distribution of each charging facility in the next 30 minutes. Based on the output results, potential faulty devices with a probability greater than 0.3 are extracted to form a list of potential faulty devices. The geographical coordinates of each potential faulty device are multiplied by its occurrence probability, and the weighted centroid is calculated to obtain the pre-deployment location of maintenance resources. The purpose of this step is to use the high-dimensional transient dynamics of the liquid neural network to capture the nonlinear cascading propagation pattern of faults in the electrical topology network, and to provide a forward-looking decision-making basis for the pre-deployment of maintenance resources.

[0060] The specific implementation of step S05 involves constructing a scheduling fitness function to comprehensively evaluate the suitability of maintenance personnel in performing tasks. This function includes four normalized indicators. First, the shortest road network time from the maintenance personnel's current location to the fault point is calculated using a path planning algorithm and normalized by dividing by the maximum travel time in the area. Second, the matching degree is calculated based on the personnel's skill level and the difficulty of the fault type. Personnel skill levels are divided into beginner, intermediate, and advanced, and assigned values ​​of 1 to 3 respectively. Fault type difficulty is divided into easy, moderate, and difficult, and assigned values ​​of 1 to 3 respectively. The matching degree is defined as the personnel skill level minus the fault type difficulty, plus 2, then divided by 4 and normalized to the range of 0 to 1. Third, the current task queue length of the maintenance personnel is divided by the maximum single-person task capacity of 8 for normalization. Fourth, the historical repair success rate percentage is divided by 100. Normalization is performed, and the scheduling fitness function is a weighted sum of four indicators with weight coefficients of 0.4, 0.3, 0.2, and 0.1. Particle swarm optimization is used to solve for the personnel allocation scheme corresponding to the maximum value of this function. In the particle swarm optimization algorithm, the number of particles is set to 50, the maximum number of iterations is set to 100, the initial value of the inertia weight is set to 0.9 and linearly decreased to 0.4, and both the individual learning factor and the group learning factor are set to 2.0. The particle position represents the correspondence between personnel and tasks, and the particle velocity represents the adjustment range of the allocation scheme. The coordinates of idle maintenance personnel without current tasks are adjusted according to the pre-deployment location of maintenance resources. The purpose of this step is to achieve the optimal matching of maintenance personnel and faulty tasks through multi-objective optimization and to pre-deploy maintenance resources to shorten the response time based on the predicted potential fault distribution.

[0061] The specific implementation of step S06 involves generating an electronic work order containing the unique identifier of the faulty equipment, the fault type code, the latitude and longitude coordinates of the equipment, the employee number of the maintenance personnel, the estimated arrival time, and the recommended repair plan. The system queries the spare parts inventory database based on the fault type code to obtain the model, specifications, inventory quantity, storage location information, and safety threshold of the corresponding spare parts. The safety threshold is set to 1.5 times the average monthly consumption of spare parts. When the inventory quantity is lower than the safety threshold, the system calculates the road network distance from the current location of the maintenance personnel to each spare parts storage location, selects the storage location with the shortest distance, and generates the nearest spare parts allocation path through a path planning algorithm, which is then pushed to the maintenance personnel's mobile terminal. The electronic work order is sent through the push interface and the push timestamp is recorded. The purpose of this step is to achieve automated linkage between work order generation and spare parts allocation to ensure that spare parts are ready before the maintenance personnel arrive at the site.

[0062] The specific implementation of step S07 involves maintenance personnel reporting timestamps and maintenance results at three time points: arrival at the site, start of maintenance, and completion of maintenance, via a mobile terminal application. After receiving the reported data, the system updates the visualized maintenance resource distribution map in real time, displaying the location icons of each maintenance personnel and indicating task status by color. After the work order is closed, the system automatically calculates the maintenance time as the completion time minus the arrival time at the site, extracts the replacement spare parts model, and writes the root cause analysis text of the fault into the corresponding field of the equipment health record. Simultaneously, the maintenance personnel performance database is updated by incrementing the maintenance count field by 1, and the historical maintenance success rate is recalculated as the number of successful maintenance counts divided by the total number of maintenance counts. The equipment node code, fault type, electrical topology relationship, load data cascading propagation sequence of the current fault are used to construct a new sample. Only the readout layer weights of the fault cascading propagation prediction model are updated to perform incremental learning, while the liquid reservoir layer structure remains unchanged. The purpose of this step is to achieve closed-loop tracking of the entire maintenance process and continuously feed new fault data back to the prediction model to improve the model's adaptability to new fault modes.

[0063] It should be noted that the key technical approach of this invention includes a multi-scale adaptive fractal partitioning algorithm driven by manifold learning. This algorithm projects geographic coordinates, historical fault density, road traffic time, and equipment health scores onto a low-dimensional manifold space using local linear embedding technology. This captures the nonlinear manifold structure of charging facilities in a high-dimensional heterogeneous feature space, avoiding the sensitivity of traditional Euclidean space clustering to distance metric distortion. Combined with the self-similarity principle of fractal geometry, it ensures that the boundaries of the areas maintain morphological consistency at different scales. The Hausdorff fractal dimension serves as a quantitative evaluation index to mathematically represent the rationality of the partitioning. The geodesic distance metric conforms to the actual road network connectivity characteristics. The multi-scale partitioning system takes into account the constraints of load balancing between regions and the differences in local traffic conditions, significantly improving the adaptability of the operation and maintenance responsibility partitioning to the dynamic changes in the spatiotemporal distribution of faults, reducing the idle rate of operation and maintenance resources caused by improper partitioning, and enhancing the robustness of the scheduling system under fault surge scenarios. Another key technical approach is a fault cascade propagation prediction model based on liquid neural networks. This model forms a high-dimensional transient dynamic system through the heterogeneous connections of a large number of leakage integral firing spiking neurons in the liquid reservoir layer. It maps the time-varying historical cascade fault sequence into rich spatiotemporal pulse firing patterns. The leakage integral mechanism naturally has a time-scale memory effect, and the short-term plasticity characteristics endow the model with adaptive sensitivity to the temporal characteristics of fault propagation. The readout layer extracts the fault propagation law from the transient response through supervised learning, avoiding the gradient vanishing and gradient explosion problems of traditional recurrent neural networks. The fixed liquid reservoir layer weight design reduces the scale of training parameters, and the incremental learning mechanism allows the model to continuously absorb new fault propagation patterns without full retraining. The predicted list of potential faulty equipment provides a forward-looking decision-making basis for the pre-deployment of operation and maintenance resources. By deploying standby operation and maintenance personnel in high-risk areas in advance, the first response time of large-scale cascade faults is shortened, and the power outage range caused by fault propagation is reduced. The synergy of the two technical approaches lies in the fact that the manifold learning-driven area partitioning algorithm provides a spatial partitioning framework that conforms to the actual fault distribution characteristics. Within this framework, the fault cascading propagation prediction model achieves forward-looking risk warning in the time dimension. Together, they construct a spatiotemporally coupled dynamic scheduling decision system. The adaptability of area partitioning ensures that the topological features input to the prediction model accurately reflect regional differences. The potential fault distribution output by the prediction model inversely optimizes the area boundaries and the deployment location of operation and maintenance resources, forming a closed-loop feedback mechanism. This enables the scheduling system to have both the static rationality of spatial optimization and the dynamic foresight of time-series prediction, thereby improving the overall response speed, resource utilization efficiency, and fault prevention and control capabilities of wide-area charging facility operation and maintenance scheduling.

[0064] It should be noted that this invention also solves the following technical problem: existing technologies suffer from the problem that scheduling decisions for maintenance personnel rely on human experience and lack multi-dimensional comprehensive optimization. This invention constructs a scheduling fitness function that comprehensively considers four dimensions: the shortest road network travel time from the current location of maintenance personnel to the fault point, the matching degree between personnel skill level and fault type, the length of the current task queue for maintenance personnel, and the historical repair success rate. It then uses a particle swarm optimization algorithm to solve for the personnel allocation scheme corresponding to the maximum value of the scheduling fitness function. This achieves multi-objective collaborative optimization of maintenance personnel scheduling decisions, avoiding the skill mismatch or unbalanced task load problems caused by traditional methods that only consider distance factors. Furthermore, it adjusts the standby coordinates of idle maintenance personnel based on the pre-deployment location of maintenance resources, improving the global optimality and dynamic adaptability of maintenance resource allocation.

[0065] A second aspect of the present invention provides a computer-readable storage medium storing program instructions, which, when executed in a computer, are used to perform the above-described method for dynamic scheduling of operation and maintenance resources for wide-area charging facilities.

[0066] A third aspect of the present invention provides a dynamic scheduling system for operation and maintenance resources of wide-area charging facilities, comprising the aforementioned computer-readable storage medium. The system is any one of a computer, a server, or a microcontroller. The computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes the program instructions stored in the computer-readable storage medium.

[0067] Specifically, the principle of this invention is as follows: This invention projects the high-dimensional heterogeneous features of charging facilities onto a low-dimensional manifold space using local linear embedding technology, capturing the essential characteristics of fault distribution within a nonlinear manifold structure. This avoids the sensitivity of traditional Euclidean space clustering to distance metric distortion. The self-similarity principle of fractal geometry ensures that the boundaries of maintenance responsibility areas maintain morphological consistency at different scales. The adaptively adjusted iteration depth ensures differentiated management needs between high-density urban areas and low-density suburbs. The Hausdorff fractal dimension, as a quantitative evaluation index, provides a mathematical representation of the rationality of the partitioning. The geodesic distance metric conforms to the connectivity characteristics of actual road networks. The multi-scale partitioning system takes into account both inter-regional load balancing and the differential constraints of local traffic conditions. The heterogeneous connections of numerous leaking integral-firing spiking neurons in the liquid neural network form a high-dimensional transient dynamic system, mapping time-varying historical cascading fault sequences into rich spatiotemporal pulse firing patterns. The short-term plasticity characteristic endows the model with adaptive sensitivity to fault propagation temporal characteristics. By predicting a list of potential faulty equipment, it provides a forward-looking decision-making basis for the pre-deployment of maintenance resources, and by pre-deploying standby maintenance personnel in predicted high-risk areas, it shortens the first response time.

[0068] The following provides a specific embodiment 1 of the present invention. The specific implementation of the calculation steps involved in this embodiment 1 is described in detail below.

[0069] The specific implementation of step S02 is to dynamically generate operation and maintenance responsibility partitions through a multi-scale adaptive fractal partitioning algorithm driven by manifold learning. First, a four-dimensional feature vector is constructed, with geographical coordinates directly using latitude and longitude coordinates, and historical fault density... The calculation formula is expressed as follows:

[0070] ;

[0071] In the formula, Historical fault density, expressed in times per square kilometer per day; The number of faults per unit area in the past 90 days is expressed in times. The area is for statistical purposes, and the unit is square kilometers. The maximum statistical duration is set to 90 days.

[0072] Normalized values ​​of road travel time The calculation formula is expressed as follows:

[0073] ;

[0074] In the formula, This is a normalized value of road travel time, dimensionless. Charging facilities To charging facilities The shortest path time of the road network, in minutes, is calculated using Dijkstra's algorithm, where... The numbering system for charging facilities is dimensionless. This represents the maximum travel time within the area, expressed in minutes.

[0075] Device health rating The calculation formula is expressed as follows:

[0076] ;

[0077] In the formula, The equipment health score is dimensionless. This refers to the equipment's operating time, in hours. Historical failure frequency, in times; Remaining life of critical components, in hours; These are weighting coefficients, with empirical values ​​of 0.4, 0.35, and 0.25, respectively, and are dimensionless. This represents the maximum operating time of equipment within the area, in hours. This represents the maximum frequency of equipment failures within the region, expressed in times. This represents the maximum remaining lifespan of equipment within the region, expressed in hours.

[0078] Dimensionality reduction is performed using locally linear embedding techniques, for the ... The low-dimensional manifold space coordinates of each charging facility point. This is obtained by solving the optimization problem, where The charging facility point numbering is dimensionless, and the weight matrix is ​​reconstructed using local linear reconstruction. elements Satisfy constraints ,in For the first The nearest neighbor number of each charging facility, dimensionless. The number of points in the nearest neighbor set, typically ranging from 5 to 15, is dimensionless. (Hausdorff fractal dimension) The formula for calculating using the box counting method is as follows:

[0079] ;

[0080] In the formula, Let be the Hausdorff fractal dimension, which is dimensionless; For a side length of The minimum number of squares required to cover the boundary curve of the maintenance responsibility area, dimensionless; The unit is the side length of the square, in meters.

[0081] Operation and maintenance responsibility area compactness The calculation formula is expressed as follows:

[0082] ;

[0083] In the formula, For the sake of compactness of the operation and maintenance responsibility area, dimensionless; The average geodetic distance from all charging facility points within the maintenance responsibility area to the centroid of the maintenance responsibility area, in kilometers; The equivalent radius of the maintenance responsibility area, in kilometers, is calculated using the following formula: ,in Pi, with a value of 3.14159, is dimensionless. The area of ​​the maintenance responsibility zone is expressed in square kilometers.

[0084] Average geodetic distance The calculation formula is expressed as follows:

[0085] ;

[0086] In the formula, The total number of charging facilities within the maintenance responsibility area, dimensionless; Numbers of charging facilities within the maintenance responsibility area, dimensionless; For the first The geodetic distance from each charging facility point to the centroid of the maintenance responsibility area, in kilometers.

[0087] The specific implementation of step S03 is to match the received charging facility fault information to the corresponding operation and maintenance responsibility area, and measure the distance. The shortest path length between two points on the low-dimensional manifold surface is calculated using Dijkstra's algorithm on the adjacency graph formed by charging facility points, along with the average radius of the maintenance responsibility area. The calculation formula is expressed as follows:

[0088] ;

[0089] In the formula, The average radius of the maintenance responsibility area, in kilometers; This represents the total number of maintenance responsibility areas, dimensionless. This is a dimensionless number assigned to the maintenance responsibility area. For the first The equivalent radius of each maintenance responsibility area, in kilometers. (If the minimum geodetic distance...) Exceed This triggers the cross-regional collaborative scheduling mechanism, in which This is the minimum geodetic distance from the faulty equipment to the centroid of each maintenance responsibility area, expressed in kilometers.

[0090] The specific implementation of step S04 involves activating a fault cascading propagation prediction model to simulate the fault propagation path, and determining the fault propagation rate index. The calculation formula is expressed as follows:

[0091] ;

[0092] In the formula, This is a dimensionless indicator of fault propagation rate. This represents the number of newly faulty devices in the past 10 minutes, expressed in units. This is a time window with a value of 10 minutes. The average topological distance between faulty devices, in hops; The maximum diameter of the regional electrical topology diagram, in units of jumps; The load factor is dimensionless.

[0093] Grid load factor The calculation formula is expressed as follows:

[0094] ;

[0095] In the formula, Total load power of all online charging facilities, in kilowatts; This refers to the rated total capacity, expressed in kilowatts.

[0096] Average topology distance between faulty devices The calculation formula is expressed as follows:

[0097] ;

[0098] In the formula, This is a dimensionless number for newly added faulty equipment. For the first The shortest path hops from a newly added faulty device to the initially faulty device in the electrical topology diagram, in hops.

[0099] coordinates of pre-deployment location of operation and maintenance resources The calculation formula is expressed as follows:

[0100] , ;

[0101] In the formula, The latitude and longitude coordinates of the pre-deployment location of the operation and maintenance resources, in degrees; The total number of devices in the list of potentially faulty devices, dimensionless; Number the potentially faulty equipment; dimensionless. For the first The probability of occurrence of a potential faulty device is dimensionless and satisfies the following condition. ; For the first The geographical coordinates of the potentially faulty device are given in degrees.

[0102] The specific implementation of step S05 is to perform multi-objective optimization of the operation and maintenance personnel scheduling decision, and the scheduling fitness function. The calculation formula is expressed as follows:

[0103] ;

[0104] In the formula, The value of the scheduling fitness function is dimensionless. The shortest road network travel time from the current location of the maintenance personnel to the fault location, in minutes; This represents the maximum travel time within the area, in minutes. For personnel skill levels, the beginner, intermediate, and advanced levels are assigned values ​​of 1 to 3 respectively, which are dimensionless. The difficulty level of the fault type is assigned values ​​from 1 to 3, with easy, moderate, and difficult being dimensionless. The current task queue length for operations and maintenance personnel, in units of tasks; This represents the maximum number of tasks a single user can perform, typically set to 5. Historical repair success rate, expressed as a percentage.

[0105] The formulas for updating particle velocity and position in the particle swarm optimization algorithm are expressed as follows:

[0106] , ;

[0107] In the formula, The particles are numbered, and are dimensionless. The number of iterations is dimensionless. For the first The particle in the first The velocity vector of the next iteration is dimensionless. For the first The particle in the first The velocity vector of the next iteration is dimensionless. The inertial weight has an empirical value of 0.7 and is dimensionless. The learning factor has an empirical value of 2 and is dimensionless. A random number between 0 and 1, dimensionless; For the first The historical optimal position of each particle is dimensionless. The optimal position for the group is dimensionless. For the first The particle in the first The position vector of the next iteration is dimensionless. For the first The particle in the first The position vector of the next iteration is dimensionless.

[0108] To better understand and implement this invention, the following is a specific application scenario of this invention, Example 2:

[0109] To verify the effectiveness of this invention, technicians built a test environment, simulating a wide-area charging facility operation and maintenance scheduling scenario by deploying a small test network of 20 charging piles. Three maintenance personnel were configured in the test environment, initially located in the northwest, southeast, and center of the test area, respectively, with skill levels of advanced, intermediate, and basic. The charging facilities were distributed over approximately 15... Within the area, the historical fault database contains 846 fault records from the past 90 days, covering five types of faults: voltage over-limit, abnormal current, communication interruption, insulation fault, and contactor fault.

[0110] At the start of the test, technicians simulated a voltage over-limit fault at charging station CP-07. The four-dimensional feature vector of this device, composed of its geographical coordinates, historical fault density, road travel time, and device health score, was (0.42, 0.68, 0.35, 0.71). The system first used local linear embedding technology to reduce the dimensionality of this four-dimensional feature vector. By calculating the k-nearest neighbor set of CP-07 and other charging facilities, with k set to 5, a local linear reconstruction weight matrix was constructed. After eigenvalue decomposition, the feature vectors corresponding to the two largest eigenvalues ​​were extracted, yielding the coordinates of CP-07 in the low-dimensional manifold space as (0.58, 0.73). The system then executed a manifold learning-driven multi-scale adaptive fractal partitioning algorithm in real time, dynamically dividing the test area into three maintenance responsibility areas, such as... Figure 2 As shown, these correspond to the jurisdiction of three maintenance personnel. During the partitioning process, the fractal generation rule of the Koch curve was used to expand the boundary of the maintenance responsibility area outwards. Since the charging facility density gradient in the test area is 6.8 units per square kilometer, the system adaptively selects an iteration depth of 5 layers. The generated maintenance responsibility area boundary was calculated using the box counting method to determine the Hausdorff fractal dimension. Coverage statistics were performed using squares with side lengths ε of 0.1, 0.05, 0.02, and 0.01. Fitting the result in a double logarithmic coordinate system yielded a Hausdorff fractal dimension of 1.43, which falls within a reasonable range of 1.2 to 1.6.

[0111] The system calculates the geodesic distances between the low-dimensional manifold coordinates (0.58, 0.73) of the CP-07 device and the centroids of each maintenance responsibility area. The centroid coordinates of the three maintenance responsibility areas are (0.32, 0.48), (0.76, 0.82), and (0.54, 0.61), respectively. Using Dijkstra's algorithm, the geodesic distances from CP-07 to the centroids of the three maintenance responsibility areas are calculated on the adjacency graph formed by the charging facility points. The geodesic distances are 0.38, 0.29, and 0.16, respectively, with the smallest geodesic distance corresponding to the third maintenance responsibility area. The system assigns fault information to this maintenance responsibility area and initiates the fault cascade propagation prediction model. The liquid reservoir layer of the fault cascade propagation prediction model contains 800 leakage integral-emitting spiking neurons, receiving the node code of the CP-07 device, the one-hot vector of the voltage over-limit fault, the electrical connection adjacency matrix between devices, and the current load rate data of each device. As shown in Table 1, the load rate of the 20 charging piles during the test ranged from 0.23 to 0.89.

[0112] Table 1. Load Rate Distribution of Charging Facilities at Test Time

[0113]

[0114] The fault cascading propagation prediction model calculates the fault propagation rate adjustment function based on the input data. It calculates the number of newly faulty devices in the past 10 minutes as 1, the average topological distance between faulty devices as 0, and the grid load rate as 0.58, resulting in a fault propagation rate index of 0, indicating slow propagation. The synaptic connection weight decay time constant is set to 500 milliseconds. In the liquid reservoir layer, leakage integral firing type spiking neurons update the membrane potential based on the received pulse sequence. The membrane potential evolves from a resting potential of -70 mV. When a high-frequency pulse input at the moment of fault occurrence is received, the membrane potential jumps to -52 mV, exceeding the firing threshold of -55 mV, generating an output pulse and resetting to -70 mV. The readout layer receives the firing rate time series of the liquid reservoir layer and outputs the fault probability vector for each charging facility in the next 30 minutes. The prediction results show that the fault probabilities of devices CP-12 and CP-18 are 0.72 and 0.61, respectively. Figure 3 As shown, the failure probability of the remaining devices is less than 0.15. The system calculates the pre-deployment location of maintenance resources based on the list of potentially faulty devices. The weighted centroids of the products of the geographical coordinates of CP-12 and the failure probability of 0.72, and the products of the geographical coordinates of CP-18 and the failure probability of 0.61, are calculated to obtain the coordinates of the pre-deployment location.

[0115] The system then executes a multi-objective optimization-based scheduling decision for maintenance personnel, calculating the scheduling fitness function values ​​for the three maintenance personnel. For the senior-skilled maintenance personnel, the shortest road network time from their current location to CP-07 is 8 minutes (normalized to 0.53), their skill level is 3, fault type difficulty is 2, skill matching degree is 0.75, current task queue length is 0, historical repair success rate is 0.94, and scheduling fitness function value is 0.74. For the intermediate-skilled maintenance personnel, the shortest road network time is 12 minutes (normalized to 0.80), skill matching degree is 0.50, task queue length is 1, historical repair success rate is 0.87, and scheduling fitness function value is 0.48. For the junior-skilled maintenance personnel, the shortest road network time is 15 minutes (normalized to 1.00), skill matching degree is 0.25, task queue length is 0, historical repair success rate is 0.76, and scheduling fitness function value is 0.36. The particle swarm optimization algorithm is used to find the maximum value of the scheduling fitness function, determine the assignment of the CP-07 fault task to senior skilled maintenance personnel, and adjust the standby coordinates of idle junior skilled maintenance personnel to the pre-deployment location of maintenance resources.

[0116] The system generates an electronic work order containing the unique identifier for CP-07, the voltage over-limit fault type code VE-01, the equipment's geographical coordinates, the assigned senior skilled maintenance personnel employee number RW-003, the estimated arrival time of 8 minutes, and the recommended repair plan of checking the charging module's output voltage feedback loop. The system automatically queries the spare parts inventory database, finding that the voltage sensor spare parts corresponding to fault type code VE-01 are in stock at 3 units, located on shelf 12 in warehouse A, with a safety threshold of 5 units. Since the current inventory is below the safety threshold, the system pushes the nearest spare parts allocation path from the pre-deployment location to warehouse A to the maintenance personnel. After receiving the electronic work order via mobile terminal, the maintenance personnel report an arrival time of 7 minutes, a repair start timestamp of 8 minutes after receiving the work order, a repair completion timestamp of 23 minutes after receiving the work order, and a repair result of replacing the voltage sensor after the equipment returned to normal operation. The system updates the visualized maintenance resource distribution map in real time, showing the senior skilled maintenance personnel's location moving from the initial northwest to the CP-07 equipment location, and the task status changing from idle to executing and then to completed. After the electronic work order is closed, the system automatically writes the 15-minute repair time, the replacement spare part model VS-2024, and the root cause analysis of voltage sensor aging into the equipment health record of CP-07. Simultaneously, it updates the maintenance personnel performance database, increasing the repair count for worker RW-003 by one, and updating the historical repair success rate to 0.95. The system uses this CP-07 voltage over-limit fault information as a new sample to input into the fault cascade propagation prediction model for incremental learning, updating only the readout layer weight parameters.

[0117] In the 18th minute of the test, the predicted CP-12 device did indeed experience a communication interruption. Since junior-skilled maintenance personnel had been deployed to the vicinity of the pre-deployment location in advance, the actual response time was only 4 minutes, which was 11 minutes shorter than the traditional response time from a fixed station, thus improving scheduling efficiency.

[0118] It should be noted that the variables involved in this invention are explained in detail in Tables 2 and 3.

[0119] Table 2. Variable Explanation Table (Part 1)

[0120]

[0121] Table 3. Variable Explanation Table (Part Two)

[0122]

[0123] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for dynamic scheduling of operation and maintenance resources for wide-area charging facilities, characterized in that, The process includes receiving charging facility fault information and performing edge preprocessing; dynamically generating maintenance responsibility zones based on a multi-scale adaptive fractal partitioning algorithm driven by manifold learning; mapping the geographical coordinates, historical fault density, road travel time, and equipment health scores of charging facilities to a low-dimensional manifold space and recursively partitioning maintenance responsibility zones according to the self-similarity principle of fractal geometry; calculating the Hausdorff fractal dimension of the maintenance responsibility zone boundaries to evaluate the partitioning quality; matching fault information to the corresponding maintenance responsibility zones and calculating the geodesic distance between the low-dimensional manifold space coordinates and the centroids of each maintenance responsibility zone; initiating a fault cascade propagation prediction model to deduce the fault propagation path and output a list of potential faulty equipment and their probability distribution within the next 30 minutes; calculating the pre-deployment locations of maintenance resources based on the list of potential faulty equipment and their probability distribution; executing multi-objective optimization of maintenance personnel scheduling decisions and solving the personnel allocation scheme corresponding to the maximum value of the scheduling fitness function through a particle swarm optimization algorithm; generating electronic work orders and executing intelligent spare parts linkage; and implementing full-process dynamic tracking and closed-loop management.

2. The method for dynamic scheduling of operation and maintenance resources for wide-area charging facilities according to claim 1, characterized in that, The edge preprocessing specifically involves deploying a fault preprocessing unit on the charging pile side to perform multi-level priority classification on the original fault data. Emergency faults are pushed to the central scheduling platform in real time via the MQTT protocol's Quality of Service Level 2 mode, while non-emergency faults are stored in a local cache queue and transmitted in batches after data compression.

3. The method for dynamic scheduling of operation and maintenance resources for wide-area charging facilities according to claim 2, characterized in that, The fault preprocessing unit is an embedded computing module deployed locally on the charging pile. It has edge data filtering and preliminary diagnosis functions, and prioritizes fault signals such as voltage over-limit, current abnormality, and communication interruption through preset rules.

4. The method for dynamic scheduling of operation and maintenance resources for wide-area charging facilities according to claim 3, characterized in that, The manifold learning-driven multi-scale adaptive fractal partitioning algorithm specifically uses local linear embedding technology to construct a four-dimensional feature vector from the geographical coordinates of charging facilities, historical fault density, road travel time, and equipment health score, and then extracts low-dimensional manifold space coordinates. In the low-dimensional manifold space, the boundaries of the operation and maintenance responsibility areas are recursively partitioned according to the fractal generation rules of the Koch curve.

5. The method for dynamic scheduling of operation and maintenance resources for wide-area charging facilities according to claim 4, characterized in that, The Hausdorff fractal dimension is calculated using the box counting method. The boundary curve of the maintenance responsibility area is covered with squares of side length ε, and the minimum number of squares N required for coverage is counted. The value of ε is changed to calculate multiple sets of correspondences between N and ε. The slope of the fitted straight line in the double logarithmic coordinate system is the Hausdorff fractal dimension.

6. The method for dynamic scheduling of operation and maintenance resources for wide-area charging facilities according to claim 5, characterized in that, When the Hausdorff fractal dimension is in the range of 1.2 to 1.6, it is considered a reasonable division, indicating that the boundary of the operation and maintenance responsibility area is neither too regular nor too fragmented, and meets the operability requirements of actual operation and maintenance.

7. The method for dynamic scheduling of operation and maintenance resources for wide-area charging facilities according to claim 6, characterized in that, The geodesic distance is the shortest path length between two points on the surface of a low-dimensional manifold space. It is calculated on the adjacency graph formed by charging facility points using the Dijkstra algorithm, which is different from the straight-line Euclidean distance that ignores terrain and road network constraints.

8. The method for dynamic scheduling of operation and maintenance resources for wide-area charging facilities according to claim 7, characterized in that, If the minimum geodetic distance exceeds 1.5 times the average radius of the maintenance responsibility area, the cross-area collaborative scheduling mechanism will be triggered, allowing the call to maintenance personnel from adjacent maintenance responsibility areas.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program instructions, which, when executed in a computer, are used to perform a dynamic scheduling method for operation and maintenance resources of wide-area charging facilities as described in any one of claims 1-8.

10. A dynamic scheduling system for operation and maintenance resources of wide-area charging facilities, characterized in that, The system comprises the computer-readable storage medium of claim 9, wherein the system is a computer, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes program instructions stored in the computer-readable storage medium.

Images