Electric arc prevention method and device for high-voltage energy storage connector of charging pile

Abstract

The invention relates to an arc prevention method and device for a high-voltage energy storage connector of a charging pile. The method comprises the following steps: constructing a first arc characteristic track based on a current and voltage change curve in a charging process of the high-voltage energy storage connector; performing parameter adjustment on the current and voltage change curve according to the first arc characteristic track to obtain a second arc characteristic track; processing the second arc characteristic track to obtain time sequence data; inputting the time sequence data into a CNN-MMHA model for arc occurrence probability prediction to obtain an arc risk assessment result; performing calculation according to the arc risk assessment result to obtain arc prevention strategy parameters; and according to the arc prevention strategy parameters and based on the three-state Markov chain, realizing a random trigger transmission protocol, and generating a charging control instruction. According to the intelligent arc prevention method, multi-source information can be comprehensively utilized, the dynamic feature extraction capacity is achieved, self-adaptive adjustment can be achieved, and the safety and reliability of the high-voltage energy storage connector are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of energy storage connector detection, and in particular to an arc protection method and device for a high-voltage energy storage connector used in a charging pile. Background Art

[0002] With the rapid development of the electric vehicle industry, high-voltage energy storage charging piles have become an important part of the electric vehicle charging infrastructure. However, during the charging process of the high-voltage energy storage connector, arcing is easily generated due to factors such as current mutation and poor contact, which not only affects the charging efficiency, but may also cause damage to the connector and even cause safety accidents. Traditional arc protection methods mainly rely on hardware protection devices and simple threshold detection, which are difficult to cope with complex and changeable charging environments and different types of arc faults.

[0003] With the development of artificial intelligence and big data technology, data-driven intelligent arc protection methods have gradually attracted attention. However, most of the existing intelligent arc protection algorithms are limited to static feature extraction and single model prediction, which makes it difficult to effectively capture the dynamic process and multi-scale spatiotemporal characteristics of arc formation. In addition, how to ensure the arc protection effect while taking into account the charging efficiency, and how to design a robust control strategy to cope with the uncertainty in the actual charging process are all urgent problems to be solved. Summary of the invention

[0004] The main purpose of the present invention is to provide an arc protection method and device for a high-voltage energy storage connector for a charging pile, which can comprehensively utilize multi-source information, has dynamic feature extraction capabilities, and is an intelligent arc protection method that can be adaptively adjusted, thereby improving the safety and reliability of the high-voltage energy storage connector.

[0005] To achieve the above object, the present invention provides an arc protection method for a high-voltage energy storage connector for a charging pile, comprising the following steps: Based on the current and voltage change curve of the high-voltage energy storage connector during the charging process, a first arc characteristic trajectory is constructed; Performing safety constraint verification according to the first arc characteristic trajectory and adjusting parameters of the current and voltage variation curve to obtain a second arc characteristic trajectory; Processing the second arc characteristic trajectory through a dynamic spatiotemporal graph neural network to obtain time series data; The time series data is input into the CNN-MMHA model to predict the arc occurrence probability, and obtain an arc risk assessment result; Calculate arc protection strategy parameters according to the arc risk assessment result; According to the arc protection strategy parameters and based on a three-state Markov chain, a random trigger transmission protocol is implemented, and a charging control instruction is generated.

[0006] The present invention also provides an arc protection device for a high-voltage energy storage connector of a charging pile, comprising: A construction module, used to construct a first arc characteristic trajectory based on a current and voltage change curve of the high-voltage energy storage connector during charging; an adjustment module, configured to perform safety constraint verification according to the first arc characteristic trajectory and perform parameter adjustment on the current and voltage variation curve to obtain a second arc characteristic trajectory; A processing module, used for processing the second arc characteristic trajectory through a dynamic spatiotemporal graph neural network to obtain time series data; A prediction module, used for inputting the time series data into the CNN-MMHA model to predict the probability of arc occurrence and obtain an arc risk assessment result; A calculation module, used for calculating arc protection strategy parameters according to the arc risk assessment result; A generation module is used to implement a random trigger transmission protocol according to the arc protection strategy parameters and based on a three-state Markov chain, and generate a charging control instruction.

[0007] The present invention also provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any one of the above methods when executing the computer program.

[0008] The present invention also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of any of the above-mentioned methods are implemented.

[0009] In summary, the technical solution provided by the present invention can capture the dynamic process of arc formation more accurately by constructing arc feature trajectories and performing safety constraint verification, and effectively improve the accuracy and reliability of arc detection. The dynamic spatiotemporal graph neural network is used to process arc feature trajectories, which can simultaneously extract spatial and temporal features in the arc formation process, and enhance the model's ability to recognize complex arc patterns. The CNN-MMHA model is introduced to predict the probability of arc occurrence. Combined with the convolutional neural network and the multi-head self-attention mechanism, it can better capture the local and global dependencies of arc features and improve the accuracy of prediction. The learning strategy of the intelligent agent is designed based on the implicit dual gradient tracking algorithm, which realizes the dynamic adjustment of the anti-arc strategy parameters and enhances the adaptability of the method to different charging environments. The Monte Carlo tree search is used to optimize the preliminary anti-arc strategy, which improves the long-term benefits of decision-making and makes the anti-arc strategy more robust and efficient. The random trigger transmission protocol is implemented based on the three-state Markov chain, which effectively reduces the arc risk while ensuring communication efficiency, and achieves a balance between arc protection effect and charging efficiency. The present invention constructs an end-to-end intelligent arc protection system, which greatly improves the safety and reliability of the high-voltage energy storage connector. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 It is a schematic diagram of the steps of an arc protection method for a high-voltage energy storage connector for a charging pile in one embodiment of the present invention; Figure 2 It is a structural block diagram of an arc protection device for a high-voltage energy storage connector used in a charging pile in one embodiment of the present invention; Figure 3 It is a schematic block diagram of the structure of a computer device according to an embodiment of the present invention.

[0011] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0012] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0013] Reference Figure 1 This embodiment provides an arc protection method for a high-voltage energy storage connector for a charging pile, comprising the following steps: S1, constructing a first arc characteristic trajectory based on a current and voltage change curve of the high-voltage energy storage connector during charging; Among them, the current and voltage change curves of the high-voltage energy storage connector during the charging process are sampled, and the continuous current and voltage change curves are converted into discrete time series data. Based on the discrete time series data, the current change amount and voltage change amount between adjacent sampling points are calculated to obtain the initial change rate sequence. By processing the initial change rate sequence, the average change rate in each time window is calculated to obtain a smoothed change rate sequence. Based on a preset threshold, the smoothed change rate sequence is segmented and divided into multiple current and voltage curve segments. Each current and voltage curve segment is modeled using a polynomial fitting method. Through polynomial fitting, a segmented polynomial function is generated. Based on the segmented polynomial function, the instantaneous current change rate and instantaneous voltage change rate at each time point are calculated to obtain a continuous change rate function. The continuous change rate function is discretized and sampled to obtain a discrete change rate point set. These discrete change rate points can represent the current and voltage change characteristics at specific time points during the charging process. By mapping the change rate points into a two-dimensional plane coordinate system, an initial arc feature point set is obtained. The coordinate system uses the current change rate and the voltage change rate as coordinates to form a current and voltage feature space. In order to optimize the arc feature point set, the initial arc feature points are identified for high-density areas. By identifying the areas where the feature points are concentrated, the key feature points of the arc are determined. These target arc feature points are important indicators of arc behavior. Based on the identified target arc feature points, the first arc feature trajectory is constructed to describe the characteristics of arc formation and evolution.

[0014] S2, performing safety constraint verification according to the first arc characteristic trajectory and adjusting parameters of the current and voltage variation curve to obtain a second arc characteristic trajectory; Specifically, the first arc characteristic trajectory is discretely sampled to obtain a trajectory sampling point set. Based on this sampling point set, combined with the preset safety threshold, each sampling point is evaluated for safety and a corresponding safety score sequence is obtained. The safety score sequence can reflect whether each trajectory point meets the safety standard, thereby making a preliminary assessment of the overall arc risk. According to the safety score sequence, the average safety score in each time window is calculated to generate a smoothed safety score sequence. The purpose of the smoothing process is to eliminate the noise in the data so that the score sequence is more stable and coherent. Based on the smoothed safety score sequence, those intervals that are continuously below the preset safety threshold are identified to find potential dangerous intervals. These intervals represent time periods with a greater risk of arc occurrence and need to be analyzed and processed in detail. Density peaks are clustered for each potential dangerous interval to obtain a potential dangerous point set. Through the clustering method, the key points in the dangerous area are found, which are high-risk locations of the arc characteristics. Based on the dangerous point set, a dangerous distribution function is constructed to describe the distribution of dangerous points in time and space. The dangerous distribution function is time-aligned with the original current and voltage change curve to obtain a dangerous time mapping relationship. Based on the dangerous time mapping relationship, the current and voltage change curves are segmented to identify those curve segments that need to be adjusted. Parameters of each curve segment that needs to be adjusted are optimized, including adjusting the slope, curvature or other key parameters of the curve to ensure that the adjusted current and voltage change curves are more in line with safety requirements. Based on the optimized current and voltage curve segments, the current and voltage change curves are reconstructed to obtain the adjusted current and voltage change curves, which reflect the optimized current and voltage changes and ensure that the potential arc risks are effectively reduced or eliminated. Based on the adjusted current and voltage change curves, the arc characteristic trajectory is reconstructed to generate the final second arc characteristic trajectory.

[0015] S3, processing the characteristic trajectory of the second arc through a dynamic spatiotemporal graph neural network to obtain time series data; It should be noted that an initial undirected graph is constructed based on the physical topological structure of the high-voltage energy storage connector. In the undirected graph, each component of the high-voltage energy storage connector is set as a node, and the physical connection between these components is defined as an edge to form the topological structure of the entire system. Based on the electrical characteristics between the components, the weights of the edges in the topological graph are initialized, the initial edge weights are calculated, and a weighted topological graph is generated to reflect the electrical relationship and the strength of the interaction between the components. The second arc characteristic trajectory is discretized into a series of time step sequences, and the eigenvalues ​​in each time step are extracted to obtain a sequence of eigenvectors. Based on the eigenvectors, attributes are assigned to each node in the weighted topological graph to form a dynamic node attribute graph. In this process, the eigenvector is not only used as a node attribute input, but also combined with the topological structure of the system, so that the attributes of each node can reflect its electrical state at a specific time point. Based on the graph convolution layer in the dynamic spatiotemporal graph neural network, spatial feature extraction is performed on the dynamic node attribute graph. Through the graph convolution operation, the spatial relationship between the nodes in the topological graph is captured to generate a spatial feature matrix. Based on the time convolution layer, the spatial feature matrix is ​​extracted in the time dimension to obtain a spatiotemporal feature tensor. The spatiotemporal feature tensor integrates time and space information, and can reflect the dynamic evolution of arc features in the system over time and changes in the connections between components. In order to improve the analysis of arc features, the dynamic spatiotemporal graph neural network introduces an attention mechanism. Through the attention mechanism, the importance weights of each feature in the spatiotemporal feature tensor are calculated to generate a weighted feature tensor, so that the network can automatically focus on the most critical arc features and weaken unimportant noise information. The weighted feature tensor is pooled to obtain a graph-level feature vector, which can condense the key features of the entire system at each time point. Based on the fully connected layer, the network performs a nonlinear transformation on the graph-level feature vector to generate hidden layer features. The hidden layer features can enhance the representation of the complex dynamic behavior of the system. Through the output layer of the dynamic spatiotemporal graph neural network, the hidden layer features are mapped to the target time series space to generate initial time series data, which preliminarily describes the time series evolution of current and voltage in the system. In order to improve the accuracy and practicality of the time series, the local time dependency of the initial time series data is extracted to generate an enhanced time series. The enhanced time series is standardized to obtain time series data that meets the input requirements to support subsequent arc risk assessment and the formulation of protection strategies.

[0016] S4, inputting the time series data into the CNN-MMHA model to predict the arc occurrence probability and obtain the arc risk assessment result; Specifically, the time series data is segmented to obtain multiple data segments. Each data segment represents the current and voltage change characteristics within a specific time range. The data segments are input into the one-dimensional convolutional neural network in the CNN-MMHA model for local time feature extraction. In the convolutional neural network, three convolutional layers are used, and the sizes of the convolution kernels are set to 5, 3, and 3 respectively. Each convolutional layer is equipped with a ReLU activation function to ensure the extraction of nonlinear features, and batch normalization technology is used to stabilize the training process and accelerate convergence. After the convolution operation, the model generates a feature map sequence containing local time features extracted from the original data segments. The generated feature map sequence is input into the multi-head self-attention module in the CNN-MMHA model. The multi-head self-attention module is used to calculate the correlation between different time steps in order to capture the dependency between each moment in the time series. The module contains 8 independent attention heads, each of which calculates the linear projection of the query, key, and value. Through the scaled dot product attention mechanism, the model calculates the attention weights, which can represent the importance correlation between time steps. After attention weighting, the feature map sequence is transformed into a weighted feature with more temporal correlation. The attention-weighted features are transformed nonlinearly to obtain enhanced feature representation. The nonlinear transformation enables the model to mine complex features in arc risk. In this process, the model captures long-term temporal dependencies in the time series and generates context-aware features. These features can not only describe the changes in local time steps, but also capture trends and patterns in a longer time range, providing deeper support for the prediction of arc risk. Global average pooling and global maximum pooling operations are performed on the context-aware features. Through global pooling, the model can compress the information in the time series and obtain a global feature vector. The global feature vector condenses the key information in the entire time series and is used to calculate the probability of arc occurrence. By processing the global feature vector, the model can accurately estimate the probability of arc occurrence. Based on the calculated probability of arc occurrence, a risk score is calculated to quantify the severity of the arc risk. The risk score is compared with the preset safety threshold to obtain a binary risk label, which is used to indicate whether the system is currently in a high risk state of arc occurrence. Combining the risk score and the binary risk label, an arc risk assessment result containing risk degree and risk category is generated. Through this assessment, the severity of the arc risk in the system can be intuitively reflected, helping related equipment to take corresponding protective measures in a timely manner.

[0017] Furthermore, the context-aware features are processed. Global average pooling is performed on the context-aware features along the time dimension. An average pooling feature vector is obtained by calculating the average value of the feature value of each time step on the time series. The role of average pooling is to compress the time series information and retain the global trend of the feature, so that the feature vector can summarize the key information of the entire time series data. At the same time, global maximum pooling is performed on the context-aware features along the time dimension. The maximum pooling feature vector is obtained by selecting the maximum value of the feature value in each time step. Maximum pooling focuses on capturing the most significant feature changes in the time series, retaining the extreme feature values ​​at key moments, and ensuring that important features are not diluted. The average pooling feature vector is concatenated with the maximum pooling feature vector to form a fused feature vector. The fused feature vector combines the global features of the average and maximum values, and contains both the trend information of the entire time series and the prominent local features. According to the fused feature vector, the importance score of each feature is calculated. By weighting the features, the model generates a feature importance vector. This vector is used to evaluate the different roles played by each feature in the arcing process and weight the fused feature vector based on the importance of the feature. The weighted feature vector highlights the features that play a key role in the occurrence of arcs, weakens the influence of minor features, and improves the accuracy of prediction. The weighted feature vector is transformed nonlinearly to obtain the potential feature representation. Through nonlinear transformation, the model can explore the complex relationship between features and enhance its representation ability. Based on the potential feature representation, the correlation between each feature is calculated. By capturing the dependency and interaction information between features, the prediction accuracy of arc risk is improved, and the target weighted feature is finally generated. The target weighted feature is operated by the Sigmoid activation function. The Sigmoid function maps the weighted feature to a probability value between 0 and 1 to obtain the probability of arc occurrence. The probability value represents the possibility of arc occurrence in the system within a specific time period, which can provide effective data support for risk assessment and the formulation of protection strategies.

[0018] S5, calculating arc protection strategy parameters according to arc risk assessment results; Among them, the physical characteristics and operating status of each charging unit are parameterized. The state space model is constructed by analyzing the key physical characteristics of the charging unit, such as voltage, current, and temperature, and the state representation of the intelligent agent is generated. The intelligent agent can perceive the state of the charging unit in real time through the state representation. At the same time, combined with the arc risk assessment results, a reward function is designed to evaluate the effect of each action and obtain the objective function of the intelligent agent. The objective function is Taylor expanded and its first-order derivative and second-order derivative are calculated. The change trend of the objective function is approximated by Taylor expansion, and a Hessian matrix is ​​constructed to better understand the local quadratic approximation of the objective function. Through the local quadratic approximation, the model can capture the direction and rate of change of the objective function under different states. Based on the results of the Hessian matrix, conjugate gradient analysis is performed to calculate the search direction and step size. The conjugate gradient method is an effective optimization method, which can be used to obtain the optimal parameter update vector. This update vector describes how to adjust the parameters of the intelligent agent to better implement the arc protection strategy. Based on the parameter update vector, the update rule of the implicit dual gradient tracking algorithm is designed to ensure that the intelligent agent maintains stability and convergence during the parameter update process. Through this rule, the intelligent agent gradually learns and improves its arc prevention strategy. Offline training is performed according to the learning strategy to obtain the initial strategy parameters. A sensitivity analysis is performed on the initial strategy parameters to identify the most influential target parameters and rank them based on their importance. Based on the importance ranking of the parameters, an adaptive learning rate scheduling scheme is designed to fine-tune the learning strategy of the intelligent agent online. Online fine-tuning allows the intelligent agent to continuously optimize its strategy in the actual operating environment and dynamically adjust the strategy parameters to adapt to environmental changes. Constraint projection is performed on the dynamically adjusted strategy parameters to ensure that these parameters meet the preset physical and safety constraints and generate candidate strategy parameters. The candidate strategy parameters are input into the strategy network to generate a series of arc prevention action sequences. These action sequences are specific operations taken by the charging unit in different states, which constitute the preliminary arc prevention strategy. The preliminary strategy is explored and optimized using Monte Carlo tree search. Through Monte Carlo tree search, various possible strategy paths are evaluated to finally obtain the optimal arc prevention strategy parameters.

[0019] Furthermore, the candidate policy parameters are normalized, and the parameters of different magnitudes are standardized to ensure that the data input into the policy network are within the same range, so as to avoid affecting the training effect of the model due to large numerical differences. The standardized policy parameters are input into the multi-layer policy network. The policy network contains three fully connected layers, each of which uses the Tanh activation function to capture the complex relationship between the parameters and generate a policy distribution. Based on the policy distribution, discrete action samples are generated. The discrete action samples represent the arc protection operations that the system may take at different time points during the charging process. In order to ensure the effectiveness of the action sequence, the generated candidate arc protection action sequence is filtered, and the effective arc protection action sequence that best meets the safety and performance requirements is selected through rules or constraints. The effective arc protection action sequence is input into the value network for evaluation. The value network estimates the potential effect of each action by calculating its long-term benefit and generates an action value estimate. The action value estimate is used to evaluate the contribution of each action to the arc protection effect during long-term operation. Based on the action value estimate, the probability of each action being selected is calculated to form an action selection probability distribution. The action selection probability is used to initialize the root node of the Monte Carlo tree search, and the tree structure is expanded to generate the initial search tree. During the Monte Carlo tree search process, each leaf node is quickly simulated, its corresponding return is evaluated, and the value evaluation of the node is obtained. The return of each leaf node represents the overall benefit that the system may obtain in the future under the condition of the current action sequence. Based on the node value, back propagation is performed to update the parent node in the search tree. By updating the number of visits and cumulative returns of the parent node, the search tree can more accurately represent the potential value of different action paths. After multiple iterations, the search tree will gradually converge to the optimal solution. According to the updated search tree, the optimal action path is selected to form an optimized arc prevention strategy. The optimized arc prevention strategy is weightedly fused with the original strategy parameters. The effective part of the original strategy is retained, and the optimization results of the Monte Carlo tree search are combined to ensure that the final arc prevention strategy not only has the ability of global optimization, but also can adapt to the actual charging scenario, minimize the risk of arc occurrence, and finally obtain the arc prevention strategy parameters, so that the charging pile system can effectively prevent the occurrence of arcs under different operating conditions, and improve the safety and stability of the system.

[0020] S6, implements random triggering transmission protocol according to arc protection strategy parameters and based on a three-state Markov chain, and generates charging control instructions.

[0021] Among them, the communication state of each charging unit is modeled. The constructed three-state Markov chain includes three states: communication success, communication failure and no communication. Based on the historical communication data of the charging unit, the state transition probability between each state is calculated to obtain the initial Markov model, which reflects the probability of each state transition during the communication process. The anti-arc strategy parameters are converted into state transition probability adjustment factors. Through these adjustment factors, the transition probability in the initial Markov model is dynamically adjusted to form a dynamic Markov model. The dynamic model can adjust the transition probability between states in real time according to different anti-arc strategies to improve the adaptability to the actual communication environment. Based on the dynamic Markov model, a random trigger mechanism is generated, which allows the system to randomly select appropriate communication behaviors under different communication states to ensure that the system can run smoothly and reduce the arc risk under different states. The random trigger mechanism is integrated with the update rule of the implicit dual gradient tracking algorithm to design an adaptive update frequency strategy. This strategy can dynamically adjust the update frequency according to the changes in the communication state and the anti-arc strategy, so that the system can maintain high communication efficiency and trigger the anti-arc action at the necessary time to achieve a balance between the two. Based on the adaptive update frequency strategy, the communication efficiency and arc protection effect are taken as optimization goals, and the optimal configuration parameters are obtained by continuously adjusting the model parameters and communication strategies. The optimal configuration parameters can ensure that the system can achieve efficient charging and communication during actual operation, while effectively reducing the arc risk. According to the optimal configuration parameters, a preliminary charging adjustment plan is generated with charging efficiency and arc risk as reward signals. The preliminary charging adjustment plan is combined with the arc risk assessment results to form a more accurate charging control instruction.

[0022] In one example, based on the current and voltage change curve of the high-voltage energy storage connector during charging, a first arc characteristic trajectory is constructed, including: The current and voltage change curves of the high-voltage energy storage connector during the charging process are sampled to obtain discrete time series data, and the current change and voltage change between adjacent sampling points are calculated based on the discrete time series data to obtain an initial change rate sequence; Calculating the average change rate in each window according to the initial change rate sequence to obtain a smoothed change rate sequence, and segmenting the smoothed change rate sequence based on a preset threshold to obtain a plurality of current and voltage curve segments; Performing polynomial fitting on each current-voltage curve segment to obtain a piecewise polynomial function, and calculating the instantaneous current change rate and the instantaneous voltage change rate at each time point based on the piecewise polynomial function to obtain a continuous change rate function; Discretize and sample the continuous rate of change function to obtain a discrete rate of change point set, and construct a two-dimensional plane coordinate system based on the discrete rate of change point set, map each rate of change point to the coordinate system, and obtain an initial arc feature point set; High-density area recognition is performed on the initial arc feature point set to obtain target arc feature points, and a first arc feature trajectory is constructed based on the target arc feature points.

[0023] In this example, the current and voltage curves of the high-voltage energy storage connector during charging are sampled to obtain discrete time series data. The discrete time series records the current and voltage values ​​at each time point. The sampling frequency is set to , at time The current and voltage values ​​at are and , we get a time series dataset where each data point is a binary . According to the discrete time series data, calculate the current change and voltage change between adjacent sampling points. Assume that the current sampling point is , the previous sampling point is , the change in current It can be expressed as: ; Similarly, the change in voltage for: ; By calculation, the change of current and voltage between adjacent sampling points is obtained, and the initial change rate sequence is constructed. This sequence reflects the change rate of current and voltage at each time step. According to the initial change rate sequence, the average change rate in each window is calculated to obtain a smoothed change rate sequence to eliminate noise interference. The initial change rate sequence is processed based on the sliding window to calculate the average change rate in each window. Set a window size of , at every point in time , the smoothed current change rate and voltage change rate can be expressed as: ; ; Calculate in window The average value of the current and voltage change rates within a certain period of time is used to generate a smoothed change rate sequence. The smoothed change rate sequence effectively reduces the random fluctuations in the signal, making the subsequent analysis more stable. According to the smoothed change rate sequence and the preset threshold, the current and voltage curves are segmented. The set threshold can be a certain limit of the current or voltage change rate. The area exceeding the threshold can be regarded as a segment with significant changes, while the area below the threshold represents a relatively stable part of the current or voltage. In this way, the curve is divided into multiple current and voltage curve segments. For each current and voltage curve segment, a polynomial fitting is performed. Polynomial fitting is a commonly used method to fit a smooth polynomial curve through a set of data points to approximate the change trend of these points. Assuming that a cubic polynomial fitting is performed on a certain curve segment, the fitting function can be expressed as: ; in, are the coefficients of the polynomial, which are obtained from the data points by the least squares method. Similarly, the voltage can also be fitted by a similar polynomial to obtain its fitting function: ; The instantaneous current rate of change and the instantaneous voltage rate of change at each time point are calculated based on the piecewise polynomial function. The instantaneous rate of change can be obtained by taking the derivative of the polynomial function. For example, the instantaneous rate of change of current is: ; Similarly, the instantaneous rate of change of voltage is: ; By performing derivative operations on polynomial functions, continuous current and voltage rate of change functions are obtained to generate continuous rate of change functions. The continuous rate of change functions are discretized and sampled again to obtain discrete rate of change point sets. The continuous rate of change functions are converted back to discrete form. By taking each instantaneous rate of change value as a discrete point, a discrete rate of change point set is formed. Based on the discrete rate of change point set, a two-dimensional plane coordinate system is constructed. In this coordinate system, the current rate of change and the voltage rate of change are used as the horizontal axis and the vertical axis respectively, and each rate of change point is mapped to the coordinate system. By graphically processing the combination of the current and voltage rate of change, the initial arc feature point set is obtained, which represents the key features of the arc that may occur during the charging process. The initial arc feature point set is identified for high-density areas. The identification of high-density areas can be achieved through density clustering algorithms (such as DBSCAN), which can effectively identify areas where feature points are concentrated. Through clustering analysis of feature points, those densely distributed areas are found, and target arc feature points with higher arc risks are identified. Based on the target arc feature points, a first arc feature trajectory is constructed to reflect the changing trends of current and voltage during the charging process, as well as the time point and state where the arc may occur.

[0024] In one example, performing safety constraint verification according to the first arc characteristic trajectory and adjusting parameters of the current and voltage variation curve to obtain the second arc characteristic trajectory includes: Discretely sampling the first arc characteristic trajectory to obtain a trajectory sampling point set, and performing a safety assessment on each point in the trajectory sampling point set based on a preset safety threshold to obtain a safety score sequence; According to the safety score sequence, the average safety score in each window is calculated to obtain a smoothed safety score sequence, and based on the smoothed safety score sequence, intervals that are continuously below the safety threshold are identified to obtain a potential danger interval; Perform density peak clustering on each potential danger interval to obtain a potential danger point set, and construct a danger distribution function based on the potential danger point set; The hazard distribution function is time-aligned with the current-voltage variation curve to obtain a hazard-time mapping relationship, and the current-voltage variation curve is segmented based on the hazard-time mapping relationship to obtain the curve segment that needs to be adjusted; Adjusting the curve parameters of each curve segment to be adjusted to obtain an optimized current-voltage curve segment, and reconstructing the current-voltage change curve based on the optimized current-voltage curve segment to obtain an adjusted current-voltage change curve; The arc characteristic trajectory is constructed for the adjusted current and voltage variation curve to obtain a second arc characteristic trajectory.

[0025] In this example, the first arc characteristic trajectory is discretized, and the continuous arc characteristic trajectory is sampled at a fixed time interval or step size to form a trajectory sampling point set. Assume that the time length of the trajectory is The sampling time step is , at each time point At , an arc feature point is obtained, and these points are combined into a discrete trajectory sampling point set .in, Indicates at time The arc feature point at the time includes the current change rate and voltage change rate at that time point. Based on the preset safety threshold, a safety assessment is performed on each point in the trajectory sampling point set to assess whether there is a risk of arc occurrence at each point. Assume that the current change rate of a feature point is and the voltage change rate is Its safety can be evaluated by the following formula: ; in, is the safety score of the feature point, function It is a predefined evaluation function based on the arc occurrence law. It combines the current and voltage change rates to generate a score value. If the change rate exceeds a certain preset threshold, it is considered that the arc is likely to occur, and the safety score is The lower the value, the higher the arc risk. By performing such an evaluation on each sampling point, a set of safety score sequences is generated. Based on the safety score sequence, the average safety score in each window is calculated, and the data is smoothed to remove noise and mutation signals. Assume that the window size is , at each time point The average safety score on can be expressed as: ; This formula represents the nearby The average value of the safety score within the time step. The smoothed safety score sequence generated by the sliding window can more stably reflect the risk trend of arc occurrence. Based on the smoothed safety score sequence, the intervals that are continuously below the preset safety threshold are identified. These intervals represent potential dangerous intervals with higher arc risk. Set a safety score threshold , when continuous scoring Lower than When the arc feature points are in each potential danger zone, the potential danger point sets are determined by the density peak clustering algorithm. These points are concentrated in the arc high risk areas. The clustering algorithm calculates the local density of the points. The distance to the nearest high-density point , identify the gathering center. The potential danger point set reflects the area where arc is most likely to occur and is the most important focus in the arc protection process. Based on the identified potential danger point set, a danger distribution function is constructed. This distribution function It is used to describe the distribution of arc risk during the entire charging process, and reflect the time points and areas with high risks in the current and voltage change curves. Through the distribution function, the arc risk is quantified and used in the subsequent adjustment and optimization process. In order to optimize the current and voltage change curve, the hazard distribution function is time-aligned with the original current and voltage change curve. The alignment process analyzes the time interval where the hazard point set is located, identifies the relationship between the hazard interval and the current and voltage changes, and establishes a hazard time mapping relationship. Based on the time mapping relationship, the current and voltage change curve is divided into several curve segments, some of which are aligned with areas with higher arc risks and marked as curve segments that need to be adjusted. Adjust the curve parameters for each curve segment that needs to be adjusted. By adjusting the slope, curvature and other parameters of the current and voltage curves, the changes of these segments are made smoother, avoiding excessive current and voltage fluctuations and reducing the possibility of arc occurrence. Parameter adjustment can be achieved through polynomial regression or other curve fitting techniques. Set the fitted curve to and , these optimized current and voltage curve segments can better control arc risks. After obtaining the optimized current and voltage curve segments, these segments are reintegrated into the overall current and voltage change curve to generate an adjusted current and voltage change curve. Based on the adjusted current and voltage change curve, the arc characteristic trajectory is reconstructed. This trajectory combines the optimized current and voltage change information and can more accurately reflect the risk points and timing of arc occurrence under new charging conditions. By analyzing the adjusted arc characteristic trajectory, the second arc characteristic trajectory is obtained.

[0026] In one example, the second arc characteristic trajectory is processed by a dynamic spatiotemporal graph neural network to obtain time series data, including: An initial undirected graph is constructed based on the physical topological structure of the high-voltage energy storage connector, each component of the connector is set as a node, and the physical connection between the components is set as an edge to obtain a topological graph, and the edge weight of the topological graph is initialized. The initial edge weight is calculated based on the electrical characteristics between the components to obtain a weighted topological graph; Discretize the second arc characteristic trajectory into a time step sequence, extract the characteristic value of each time step to obtain a characteristic vector sequence, and assign attributes to the nodes of the weighted topological graph based on the characteristic vector sequence to obtain a dynamic node attribute graph; Based on the graph convolution layer in the dynamic spatiotemporal graph neural network, the spatial features of the dynamic node attribute graph are extracted to obtain the spatial feature matrix. Based on the temporal convolution layer in the dynamic spatiotemporal graph neural network, the features of the spatial feature matrix in the time dimension are extracted to obtain the spatiotemporal feature tensor. Based on the attention mechanism in the dynamic spatiotemporal graph neural network, the importance weights of each feature in the spatiotemporal feature tensor are calculated to obtain a weighted feature tensor, and the weighted feature tensor is pooled to obtain a graph-level feature vector. Based on the fully connected layer in the dynamic spatiotemporal graph neural network, the graph-level feature vector is nonlinearly transformed to obtain the hidden layer features, and based on the output layer in the dynamic spatiotemporal graph neural network, the hidden layer features are mapped to the target time series space to obtain the initial time series data; The local time dependency is extracted from the initial time series data to obtain an enhanced time series, and the enhanced time series is standardized to obtain time series data.

[0027] In this example, an initial undirected graph is constructed based on the physical topology of a high-voltage energy storage connector. The high-voltage energy storage connector consists of multiple components, such as battery modules, charging interfaces, control units, and electrical switches. Each component can be considered a node, and the physical connections between these components constitute the edges in the graph. By associating these nodes and edges, an initial undirected graph is constructed. Assume that the set of nodes in the graph is , where each node Corresponding to a connector component, the edge set is ,in Representative components and Build a graph , which is the physical topology of the connector. Initialize the weight of each edge in the topology. The weight of the edge is calculated based on the electrical characteristics between components. These electrical characteristics may include resistance, conductance, power transfer capability, etc. The initial weight of the edge It can be calculated by the following formula: ; in, It is a component and In this way, a weighted topological graph is obtained, which not only represents the physical connection of the components, but also combines the electrical properties between the components, thereby reflecting the dynamic behavior of the system. For the time step sequence obtained by discretizing the second arc characteristic trajectory, the characteristic value is extracted for each time step. The arc characteristic trajectory is divided into time steps. Discretize and get a series of feature vectors , where each eigenvector Contains current change rate, voltage change rate and other electrical characteristics. The feature vector is mapped to the nodes of the weighted topology graph to form a dynamic node attribute graph. In the dynamic node attribute graph, each node It not only contains its electrical characteristics, but also contains the time-related arc characteristics, so that the graph can dynamically reflect the state changes of the system. In order to conduct a deeper analysis of the dynamic characteristics, the graph convolution layer in the dynamic spatiotemporal graph neural network (ST-GCN) is used to extract spatial features. The graph convolution layer integrates the information between nodes through the adjacency relationship in the graph. The characteristics of each node not only depend on its own attributes, but also are affected by its neighboring nodes. For each node , its updated features The propagation rule through the graph convolutional layer can be expressed as: ; in, Is a node The set of neighbor nodes of and The nodes are and The degree of is the weight matrix of the graph convolutional layer, is the activation function. Through this formula, the network can extract the spatial feature matrix of the entire graph, representing the spatial relationship between different nodes. The spatial features are extracted in the time dimension through the temporal convolution layer in the dynamic spatiotemporal graph neural network. The role of the temporal convolution layer is to capture the dynamic pattern of node features changing over time and generate a spatiotemporal feature tensor. This tensor not only contains the node features at each time step, but also reflects the evolution of each node feature over time. Based on the attention mechanism in the dynamic spatiotemporal graph neural network, the importance weights of each feature in the spatiotemporal feature tensor are calculated to obtain a weighted feature tensor. Through the attention mechanism, the model can focus on the part of the spatiotemporal features that has the greatest impact on the occurrence of arcs. The weighted feature tensor is pooled to generate a graph-level feature vector. The pooling operation includes global average pooling and maximum pooling, which can condense the spatiotemporal features of the entire graph into a global feature representation. The dynamic spatiotemporal graph neural network performs nonlinear transformation on the graph-level feature vector through the fully connected layer to generate hidden layer features. The nonlinear transformation of the fully connected layer enables the model to extract the complex relationship between features and obtain a more representative hidden feature representation. Based on the output layer, the network maps the hidden layer features to the target time series space to obtain the initial time series data, that is, the model's preliminary prediction of the system state. The local time dependency is extracted from the initial time series data. By analyzing the dependency between different time steps, the time features that are not captured by the initial model are extracted to form an enhanced time series. The enhanced time series is standardized to unify the numerical ranges of different time steps and features, thereby eliminating the differences in data distribution, improving the stability and prediction performance of the model, and obtaining the final time series data.

[0028] In one example, time series data is input into the CNN-MMHA model to predict the probability of arc occurrence, and the arc risk assessment results are obtained, including: The time series data is segmented to obtain multiple data segments, and each data segment is input into the one-dimensional convolutional neural network of the CNN-MMHA model to extract local time features and obtain a feature map sequence; the one-dimensional convolutional neural network contains 3 convolutional layers, each layer uses the ReLU activation function and batch normalization, and the convolution kernel sizes are 5, 3, and 3 respectively; The feature map sequence is input into the multi-head self-attention module of the CNN-MMHA model to calculate the correlation between different time steps to obtain the attention weighted features; the multi-head self-attention module contains 8 attention heads, each of which independently calculates the linear projection of the query, key and value, and then obtains the attention weight through the scaled dot product attention mechanism; Perform nonlinear transformation on the attention-weighted features to obtain enhanced feature representations, and capture long-term temporal dependencies on the enhanced feature representations to obtain context-aware features. Perform global average pooling and global maximum pooling on the context-aware features to obtain a global feature vector, and calculate the arc occurrence probability of the global feature vector; The risk score is calculated based on the probability of arc occurrence, and the risk score is compared with the preset threshold to obtain a binary risk label. The risk score and the binary risk label are combined to generate an arc risk assessment result including risk degree and risk category.

[0029] In this example, the time series data is segmented into multiple time segments, each of which represents a characteristic behavior within a continuous period of time. Set the time window size to , the entire time series data Split into multiple segments, each segment Length is , get multiple time segment sets ,in . Each data segment is input into the one-dimensional convolutional neural network (CNN) in the CNN-MMHA model for local temporal feature extraction. CNN contains 3 convolutional layers, each layer uses ReLU activation function and batch normalization, and the sizes of convolution kernels are set to 5, 3, and 3 respectively. The convolution kernel is the core of the convolutional network and is used to capture the local pattern of the signal. By using the convolution kernel on each time segment, the network can extract local temporal features, such as the changing trend of current and voltage in a short time. Assume that the data segment input to the convolution layer is , the convolution operation can be expressed as: ; in, is the weight matrix of the convolution kernel, is the bias, and * represents the convolution operation. After three layers of convolution, the network extracts a sequence of feature maps, which represents the local features of each time point in the time segment. The extracted feature map sequence is input into the multi-head self-attention module in the CNN-MMHA model. The function of the multi-head self-attention module is to calculate the correlation between different time steps so that the network can identify the relationship between distant time points in the time series. The multi-head self-attention mechanism contains 8 independent attention heads, each of which calculates the linear projection of the query (Query), key (Key) and value (Value). For a given feature vector , whose query, key, and value are: ; in, , , are the weight matrices for query, key, and value, respectively. The scaled dot product attention mechanism is used to calculate the attention weights between time steps: ; in, is the dimension of the key vector, which is used to scale the dot product value to prevent the dot product value from being too large or too small. In this way, the correlation between the features of each time step and the features of other time steps is calculated to obtain the attention-weighted features. The attention-weighted features are nonlinearly transformed to obtain enhanced feature representations. The complex relationship between features is enhanced by nonlinear activation functions, so that the model can better identify potential temporal patterns. The enhanced feature representation captures long-term temporal dependencies, and the model considers long-term dependencies in the time series. For example, the occurrence of an arc may not be caused by the change of a single feature in a short period of time, but the interaction and accumulation of trends between multiple time steps. Capturing long-term temporal dependencies can better identify the potential risks of arcs and obtain context-aware features. Global average pooling and global maximum pooling are performed on the context-aware features. Global average pooling calculates the average value of the feature over all time steps, while global maximum pooling takes the maximum value of the feature. Through these two operations, two global feature vectors are obtained, which reflect the global trend and the maximum feature change in the time series respectively. The model calculates the probability of an arc based on the global feature vector. Assume that the global feature vector is , the probability of arc occurrence It can be calculated by a simple fully connected layer plus a Sigmoid activation function: ; in, is the weight matrix of the fully connected layer, is the bias term, is a Sigmoid activation function that limits the output value between 0 and 1 and is used to represent probability. Based on the probability of arc occurrence, the model calculates a risk score. The risk score indicates the severity of the arc occurrence and can be calculated based on a nonlinear function of the probability of arc occurrence. The risk score is compared with the preset threshold to generate a binary risk label. This label is used to indicate whether the current arc risk exceeds the safety limit. Combining the risk score and the binary risk label, the model ultimately generates an arc risk assessment result that includes the risk level and risk category. The risk level is used to quantify the possibility and severity of the arc occurrence, while the risk category indicates whether the system is in a high-risk state. Based on this assessment result, corresponding protective measures are taken according to different risk levels to improve the safety of the charging process.

[0030] In one example, global average pooling and global maximum pooling are performed on the context-aware features to obtain a global feature vector, and the arc occurrence probability of the global feature vector is calculated, including: Perform global average pooling on the context-aware features along the time dimension to obtain an average pooled feature vector; Perform global maximum pooling on the context-aware features along the time dimension to obtain the maximum pooling feature vector; Concatenate the average pooling feature vector and the maximum pooling feature vector to obtain a fused feature vector; According to the fused feature vector, the importance score of each feature is calculated to obtain the feature importance vector, and the fused feature vector is weighted based on the feature importance vector to obtain a weighted feature vector; Perform nonlinear transformation on the weighted feature vector to obtain potential feature representation, and calculate the correlation between features based on the potential feature representation to obtain the target weighted feature; The Sigmoid activation function is operated on the target weighted features to obtain the arc occurrence probability.

[0031] In this example, the context-aware features are globally averaged pooled along the time dimension to compress the features in the time dimension. Assume that the context-aware features are the matrix , where each Represents the time step The global average pooling calculates the average value of these features at all time steps and compresses them into an average pooling feature vector The formula is as follows: ; in, is the total number of time steps, is the feature vector after global average pooling, which represents the average behavior of all time step features. The pooling method can effectively extract the global trend of time series data without being affected by sudden changes in a short period of time. At the same time, global maximum pooling is performed to find the most important or most significant feature changes in the time series data. For each feature dimension, the maximum value over all time steps is taken to obtain the maximum pooled feature vector The formula is as follows: ; The operation is to take the maximum value of each feature dimension and generate a maximum pooling feature vector , which can reflect the strongest feature changes in the time series. Compared with average pooling, maximum pooling pays more attention to extreme events or signal peaks, which helps to identify sudden behaviors in arc risks. The average pooling feature vector and the maximum pooling feature vector are concatenated to obtain the fused feature vector The concatenation operation can combine global features and local extreme features to generate a more comprehensive feature representation. The fused feature vector can be expressed as: ; and are directly concatenated into a long vector , which contains both the global trend and the most significant local features of the time series. , calculate the importance score of each feature and obtain the feature importance vector. The feature importance vector reflects the relative importance of each feature in arc risk assessment. The calculation of feature importance can be achieved through a fully connected layer, which performs a linear transformation on the fused features and outputs the importance score of each feature through an activation function: ; in, is the weight matrix, is the bias term, and the softmax function normalizes the result to a probability distribution so that the importance score of each feature is between 0 and 1. Get the feature importance vector After that, the fused feature vector is weighted to generate a weighted feature vector

[0032] ; Here, ⊙ represents element-by-element multiplication. In this way, the contribution of each feature is adjusted by its importance weight, resulting in a more targeted feature representation. Nonlinear transformation of the weighted feature vector captures the complex relationship between features and generates a latent feature representation. . Through activation functions such as ReLU or Tanh, the formula is as follows: ; in, is the weight matrix, is the bias term, and ReLU is the activation function. The role of nonlinear transformation is to increase the performance of the model and extract the high-level features hidden in the fused feature vector. The model calculates the correlation between features based on the potential feature representation to obtain the target weighted features. The correlation between features reflects the mutual influence of each feature in arc risk assessment, which can help the model better capture the complex dynamic behavior within the system. The Sigmoid activation function is performed on the target weighted features to calculate the probability of arc occurrence. The Sigmoid activation function maps the weighted feature value to between 0 and 1, indicating the possibility of arc occurrence. The calculation formula is: ; in, is the Sigmoid function, and are the weight and bias of the output layer respectively. Through this formula, the final output Indicates the probability of arc occurrence.

[0033] In one example, arc protection strategy parameters are calculated based on arc risk assessment results, including: The physical characteristics and operating status of each charging unit are parameterized, and a state space model is constructed to obtain the state representation of the intelligent agent. A reward function is designed based on the state representation and arc risk assessment results to obtain the objective function of the intelligent agent. Perform Taylor expansion on the objective function to obtain the first-order derivative and the second-order derivative of the objective function, and construct the Hessian matrix based on the first-order derivative and the second-order derivative to obtain the local quadratic approximation of the objective function; Perform conjugate gradient analysis on the local quadratic approximation, calculate the search direction and step size, obtain the parameter update vector, and design the update rule of the implicit dual gradient tracking algorithm based on the parameter update vector to obtain the learning strategy of the intelligent agent; Offline training is performed according to the learning strategy to obtain initial strategy parameters, and sensitivity analysis is performed on the initial strategy parameters to identify target parameters and obtain parameter importance ranking; Design an adaptive learning rate scheduling scheme based on the parameter importance ranking, perform online fine-tuning on the learning strategy to obtain dynamically adjusted strategy parameters, and perform constraint projection on the dynamically adjusted strategy parameters to obtain candidate strategy parameters; The candidate strategy parameters are input into the strategy network to generate an arc protection action sequence to obtain a preliminary arc protection strategy, and a Monte Carlo tree search is performed on the preliminary arc protection strategy to obtain the arc protection strategy parameters.

[0034] In this example, the physical characteristics and operating state of each charging unit are parameterized. The physical characteristics of the charging unit include parameters such as current, voltage, and temperature, while its operating state reflects the changes of these variables at different time points. By modeling these physical parameters, the state of the charging unit at each time point is represented as a vector ,in It's time The current at the moment, is the voltage, is the temperature, etc. These state variables constitute a state space model, which represents the dynamic behavior of the system at different times. Based on the state representation of the state space model and the arc risk assessment results, the reward function of the intelligent agent is designed. The learning process of the intelligent agent is guided so that it can choose the best action to reduce the arc risk. Assume that the current state The corresponding arc risk assessment result is , then the reward function of the intelligent agent can be designed as: ; in, Is the intelligent agent in state The action to be performed is: and is a weight coefficient used to balance the trade-off between arc risk and action selection. This reward function indicates that the intelligent agent needs to minimize the arc risk , while considering the execution action Based on the reward function, the objective function of the intelligent agent It can be expressed as its expected cumulative reward during the entire charging process: ; in, are the strategy parameters of the intelligent agent, Is the discount factor, which indicates the decay weight of future rewards. Taylor expansion is performed to better understand the local variation of the objective function in the parameter space. Taylor expansion can obtain the local variation of the objective function in the parameter space by expanding the value of the objective function in the parameter space. The first and second derivatives at . The first derivative of the objective function represents the gradient of the objective function under the current policy parameters, and the second-order derivative is obtained through the Hessian matrix To express: ; By constructing the Hessian matrix from the first and second derivatives, a local quadratic approximation of the objective function can be performed. This approximation provides key information about the trend of the objective function in the parameter space for the optimization process, allowing the intelligent agent to adjust parameters more effectively. Conjugate gradient analysis is performed on the local quadratic approximation of the objective function. The conjugate gradient method is an iterative algorithm for solving quadratic optimization problems. It accelerates the convergence process of gradient descent by finding conjugate directions in the parameter space. Through conjugate gradient analysis, the search direction is calculated. and step length

[0035] ; in, It is The gradient of the step, is the conjugate coefficient, which represents the linear combination coefficient of the current direction and the previous search direction. By calculating the search direction and step size, we can get the parameter update vector , which indicates how to adjust the policy parameters in each iteration step Based on the parameter update vector, the update rule of the implicit dual gradient tracking algorithm is designed to update the strategy of the intelligent agent. The implicit dual gradient tracking algorithm is an efficient method for optimizing nonlinear functions. By simultaneously tracking the gradient of the objective function and the gradient of the constraint conditions, it ensures that the intelligent agent can optimize the strategy under the premise of satisfying the constraint conditions. The update rule can be expressed as: ; in, is the learning rate, which indicates the step size of parameter update in each iteration. After designing the update rule, the intelligent agent is trained offline according to the learning strategy. Through offline training, the intelligent agent can repeatedly try different strategies in a simulated environment to obtain the initial strategy parameters. The initial strategy parameters represent the optimal strategy learned by the intelligent agent during the training process. A sensitivity analysis is performed on the initial strategy parameters to identify the parameters that have the greatest impact on the strategy results, namely the target parameters. The target parameters are sorted to obtain the parameter importance ranking. Based on the importance ranking of the parameters, an adaptive learning rate scheduling scheme is designed, that is, the learning rate is dynamically adjusted according to the importance of the parameters. For parameters with higher importance, the learning rate should be lower to ensure that the adjustment of these parameters is more refined; for parameters with lower importance, the learning rate can be higher to speed up the convergence. After completing the online fine-tuning, the dynamically adjusted strategy parameters are obtained. Constraint projection is performed on the dynamically adjusted strategy parameters. Constraint projection is a commonly used optimization method. By projecting the optimization results into the parameter space that satisfies the constraints, it is ensured that the final strategy parameters can achieve the optimization goals and meet the actual operating conditions of the system. The candidate strategy parameters are input into the strategy network to generate an arc protection action sequence. Based on the input policy parameters, the policy network analyzes the current state and outputs an arc prevention action to guide the intelligent agent to choose appropriate actions under different charging states to reduce the possibility of arcing. In order to optimize the preliminary arc prevention strategy, a Monte Carlo tree search is performed on it. Monte Carlo tree search simulates different policy paths, evaluates the reward value on each path, and finally selects the path with the highest reward as the optimal strategy. The key steps of Monte Carlo tree search include: performing random simulation at each node of the tree, evaluating the reward of the simulation results, and then updating the cumulative reward value of each node through back propagation, and finally selecting the optimal action sequence.

[0036] In one example, candidate strategy parameters are input into a strategy network to generate an arc protection action sequence to obtain a preliminary arc protection strategy, and a Monte Carlo tree search is performed on the preliminary arc protection strategy to obtain arc protection strategy parameters, including: Normalize the candidate policy parameters to obtain standardized policy parameters, and input the standardized policy parameters into a multi-layer policy network. The multi-layer policy network contains three fully connected layers, each of which uses the Tanh activation function to obtain policy distribution. Generate discrete action samples based on strategy distribution to obtain candidate arc protection action sequences, and perform action filtering on the candidate arc protection action sequences to obtain effective arc protection action sequences; Input the effective arc protection action sequence into the value network, evaluate the long-term benefits of each action, obtain the action value estimate, and calculate the action selection probability based on the action value estimate; Use the action selection probability to initialize the root node of the Monte Carlo tree, expand the tree structure to obtain the search tree, and quickly simulate the search tree to evaluate the return of each leaf node to obtain the node value; Back propagation is performed based on the node value to update the number of visits and cumulative returns of the parent node to obtain an updated search tree. Based on the updated search tree, the optimal action path is selected to obtain the optimized arc protection strategy. The optimized arc protection strategy is weightedly fused with the original strategy parameters to obtain the arc protection strategy parameters.

[0037] In this example, the candidate policy parameters are normalized and mapped to a relatively uniform numerical range to obtain standardized policy parameters. The normalized standardized policy parameters are input into the multi-layer policy network. The network contains 3 fully connected layers, and each layer uses the Tanh activation function. The expression of the Tanh activation function is: ; Its output range is between [-1, 1], which can effectively capture nonlinear relationships. The output of each layer of the neural network can be expressed as: ; in, It is The input features of the layer, is the weight matrix, is the bias term. After three layers of calculation, the strategy distribution is obtained. , which represents the various actions under given policy parameters The probability of. Based on the strategy distribution , from which a series of discrete action samples are sampled. Each action sample corresponds to an arc protection operation strategy during charging. Assume that the candidate action sequence is sampled from the strategy distribution , these action sequences represent the arc prevention measures that the charging system may take at different time steps. In order to ensure the effectiveness of these actions, the candidate arc prevention action sequences are filtered, and actions that meet the safety and performance requirements are screened out according to the operating rules or physical constraints of the charging system to generate effective arc prevention action sequences. Through action filtering, those actions that are impractical or high-risk in the actual charging environment are eliminated to obtain a set of effective candidate actions. The effective arc prevention action sequence is input into the value network to evaluate the long-term benefits of each action. The value network calculates the value estimate of each action by analyzing the impact of each action on the long-term behavior of the system. This process can be implemented through a deep value network, where the input is the action sequence and the output is the value of each action. The value function can be expressed as: ; in, is the discount factor, which represents the decay weight of future returns, It's in time Execute action when The immediate reward obtained. Through this formula, the value network can provide an estimate of the long-term benefits for each action. Based on the value estimate of each action , calculate the action selection probability. The action selection probability can be expressed as follows by normalizing the value: ; The formula uses the softmax function to convert the value of all actions into a probability distribution, ensuring that the action selection probability is between 0 and 1 and that the sum of all probabilities is 1. After the action selection probabilities are calculated, the root node of the Monte Carlo tree is initialized with these probabilities. Monte Carlo tree search is an algorithm that finds the optimal strategy by simulating multiple possible paths. Using the action selection probability as the initialization basis, the tree structure is expanded to obtain a search tree containing all possible action sequences. At each node of the tree, there is an action, and the weight of the node is the selection probability of the action. Through fast simulation, each leaf node is simulated and the reward value of each leaf node is evaluated to obtain the value estimate of the node. Through the simulation of the charging system, the reward of each action sequence is evaluated according to the actual arc risk and charging efficiency during the simulation process. After evaluating the rewards of the leaf nodes, back propagation is performed to update the number of visits and cumulative rewards of the parent node. The purpose of back propagation is to pass the reward information of the leaf node back to the parent node, so as to update its reward value and number of visits for each node. Through back propagation, each node in the Monte Carlo tree is gradually adjusted to optimize its corresponding policy path. As the search tree is gradually updated, the optimal action path is selected based on the updated search tree. This optimal path represents the protection strategy that is most likely to reduce the arc risk under the current state, and finally the optimized arc protection strategy is obtained. The optimized arc protection strategy is weightedly fused with the original strategy parameters. On the premise of maintaining the stability of the original strategy, combined with the optimization results of the Monte Carlo tree search, a more robust and effective arc protection strategy parameter is generated. The fusion strategy can be expressed as: ; in, are the original strategy parameters, are the optimized strategy parameters, is a weighting coefficient used to balance the impact between the original strategy and the optimized strategy.

[0038] In one example, a random trigger transmission protocol is implemented based on a three-state Markov chain according to arc protection strategy parameters, and a charging control instruction is generated, including: The communication state of each charging unit is modeled and a three-state Markov chain is constructed. The three states include communication success, communication failure, and no communication. The state transition probability matrix is ​​calculated based on the historical communication data to obtain the initial Markov model. The arc protection strategy parameters are converted into state transition probability adjustment factors, the transition probability of the initial Markov model is adjusted to obtain a dynamic Markov model, and a random trigger mechanism is generated based on the dynamic Markov model; The random trigger mechanism is integrated with the update rules of the implicit dual gradient tracking algorithm to design an adaptive update frequency strategy; Based on the adaptive update frequency strategy, the communication efficiency and arc protection effect are taken as optimization targets to obtain the optimal configuration parameters; According to the optimal configuration parameters, the charging efficiency and arc risk are used as reward signals to obtain a preliminary charging adjustment plan; The preliminary charging adjustment plan is combined with the arc risk assessment results to generate charging control instructions.

[0039] In this example, three communication states of the charging unit are defined: successful communication, failed communication, and no communication. These three states represent the communication status between the charging unit and the control system, corresponding to normal data transmission, interrupted or failed communication, and temporary no communication. Based on these three states, a Markov chain model is constructed to describe the transition process of the system between different communication states. The core of the Markov chain is that the transition between states is based on the state transition probability. Assume that the communication state of the charging unit is at time In a state , then at the next time The probability of transitioning to another state at a certain moment depends only on the current state . Assume that the three communication states are (communication successful), (Communication failed), (no communication), then use the state transition probability matrix To represent the probability of the system transitioning from one state to another. The form is ; in, Indicates that the system is in state Transfer to state These transition probabilities can be estimated by analyzing historical communication data. For example, by recording the number of successful, failed, and non-communication times of the system in different time periods, calculating the transition frequency between each state, and obtaining the state transition probability matrix , generate an initial Markov model, which can preliminarily describe the communication behavior of the charging unit. After obtaining the initial Markov model, the arc protection strategy parameters are converted into adjustment factors for the state transition probability. The arc protection strategy parameters are designed to reduce the risk of arc occurrence. Specific measures may include adjusting voltage, current or other charging control parameters. In order to introduce the strategy parameters into the Markov model, they are converted into adjustment factors for the state transition probability. For example, if the arc risk is high, the system may tend to take more corrective measures when communication fails to prevent further risks caused by communication interruptions. At this time, the transition probability The probability of transferring from successful communication to failed communication will be adjusted according to the parameters of the arc protection strategy. The specific adjustment method can be expressed by a linear function: ; in, is the adjusted transition probability, is the arc protection strategy parameter, function An adjustment factor used to map strategy parameters to state transition probabilities. By adjusting the transition probability in the initial Markov model, a dynamic Markov model is obtained to reflect the transition law of the communication state and also consider the impact of the arc protection strategy on the system state. Based on the dynamic Markov model, a random trigger mechanism is generated. The role of the random trigger mechanism is to randomly trigger the arc protection control strategy at an appropriate time point according to the current communication state of the system and the parameters of the arc protection strategy. Since the Markov model can describe the dynamic evolution of the system state, the random trigger mechanism can trigger different arc protection measures at different time points based on the probability of state transition, thereby ensuring that the system takes corresponding protective actions when the risk is high. The random trigger mechanism is integrated with the update rule of the implicit dual gradient tracking algorithm to design an adaptive update frequency strategy. The implicit dual gradient tracking algorithm is an optimization algorithm that can quickly find the optimal solution by tracking the gradient of the objective function and the constraint conditions. In this embodiment, the goal is to optimize the communication efficiency and arc protection effect of the charging unit, so the design of the adaptive update frequency strategy needs to consider both aspects at the same time. According to the current communication state of the system, the adaptive update frequency strategy dynamically adjusts the update frequency of the parameters. For example, when the communication success rate is high, the update frequency can be low to keep the system stable; when the communication failure rate is high, the update frequency needs to be faster to quickly adjust the system strategy to avoid further communication failures or arc risks. Through the adaptive update frequency strategy, a set of optimal configuration parameters is found that can minimize the arc risk while ensuring communication efficiency. The optimization goal can be set to maximize communication efficiency and minimize arc risk at the same time. The objective function can be expressed as: Communication efficiency Arc risk ; in, is the arc protection strategy parameter, and is the weight coefficient used to balance communication efficiency and arc risk. By solving the optimization problem, a set of optimal configuration parameters is obtained so that the system achieves the best balance between communication efficiency and arc risk. According to the optimal configuration parameters, the charging efficiency and arc risk are used as reward signals to obtain a preliminary charging adjustment plan. For example, when the charging efficiency is high and the arc risk is low, the system will maintain the current charging parameters; when the arc risk increases, the system will automatically adjust the charging parameters to reduce the possibility of arc occurrence. In order to improve the accuracy of the charging adjustment plan, the preliminary charging adjustment plan also needs to be combined with the arc risk assessment results. The arc risk assessment results are based on real-time monitoring data and can reflect the arc risk level of the current system. By combining the charging adjustment plan with the real-time arc risk assessment results, it is ensured that the system dynamically adjusts the charging parameters according to the actual risk during operation and generates the final charging control instructions.

[0040] Reference Figure 2 This embodiment provides an arc protection device for a high-voltage energy storage connector of a charging pile, comprising: A construction module 1 is used to construct a first arc characteristic trajectory based on a current and voltage change curve of the high-voltage energy storage connector during charging; An adjustment module 2 is used to perform safety constraint verification according to the first arc characteristic trajectory and to adjust parameters of the current and voltage variation curve to obtain a second arc characteristic trajectory; Processing module 3, used for processing the second arc characteristic trajectory through a dynamic spatiotemporal graph neural network to obtain time series data; Prediction module 4, used for inputting time series data into CNN-MMHA model to predict arc occurrence probability and obtain arc risk assessment result; A calculation module 5 is used to calculate arc protection strategy parameters according to arc risk assessment results; The generating module 6 is used to implement the random triggering transmission protocol according to the arc protection strategy parameters and based on the three-state Markov chain, and generate the charging control instruction.

[0041] In this embodiment, for the specific implementation of each unit in the above device embodiment, please refer to the above method embodiment, which will not be repeated here.

[0042] Reference Figure 3 In an embodiment of the present invention, a computer device is also provided. The computer device may be a server, and its internal structure may be as follows: Figure 3As shown. The computer device includes a processor, a memory, a display screen, an input device, a network interface and a database connected through a system bus. Among them, the processor designed by the computer is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the above method is implemented.

[0043] Those skilled in the art will understand that Figure 3 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied.

[0044] An embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the above method is implemented. It can be understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0045] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided by the present invention and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double-speed data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM.

[0046] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "includes a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.

[0047] The above description is only a preferred embodiment of the present invention, and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for arc prevention of a high-voltage energy storage connector for a charging pile, characterized in that: The following steps are involved: Based on the current and voltage change curve of the high-voltage energy storage connector during charging, a first arc characteristic trajectory is constructed; Performing safety constraint verification according to the first arc characteristic trajectory and adjusting parameters of the current and voltage variation curve to obtain a second arc characteristic trajectory; Processing the second arc characteristic trajectory by a dynamic spatiotemporal graph neural network to obtain time series data; The time series data is input into the CNN-MMHA model to predict the arc occurrence probability, and obtain an arc risk assessment result; Calculate arc protection strategy parameters according to the arc risk assessment result; According to the arc protection strategy parameters and based on a three-state Markov chain, a random trigger transmission protocol is implemented, and a charging control instruction is generated.

2. The arc protection method for a high-voltage energy storage connector for a charging pile according to claim 1, characterized in that: The method of constructing a first arc characteristic trajectory based on a current and voltage variation curve during the charging process of the high-voltage energy storage connector includes: Sampling the current and voltage change curves of the high-voltage energy storage connector during charging to obtain discrete time series data, and calculating the current change and voltage change between adjacent sampling points based on the discrete time series data to obtain an initial change rate sequence; Calculating an average change rate in each window according to the initial change rate sequence to obtain a smoothed change rate sequence, and segmenting the smoothed change rate sequence based on a preset threshold to obtain a plurality of current and voltage curve segments; Performing polynomial fitting on each of the current-voltage curve segments to obtain a piecewise polynomial function, and calculating the instantaneous current change rate and the instantaneous voltage change rate at each time point based on the piecewise polynomial function to obtain a continuous change rate function; Discretize and sample the continuous rate of change function to obtain a discrete rate of change point set, construct a two-dimensional plane coordinate system based on the discrete rate of change point set, map each rate of change point to the coordinate system, and obtain an initial arc feature point set; High-density area recognition is performed on the initial arc feature point set to obtain target arc feature points, and a first arc feature trajectory is constructed based on the target arc feature points.

3. The arc protection method for a high-voltage energy storage connector for a charging pile according to claim 1, characterized in that: The performing safety constraint verification according to the first arc characteristic trajectory and adjusting parameters of the current and voltage change curve to obtain a second arc characteristic trajectory includes: Discretely sampling the first arc characteristic trajectory to obtain a trajectory sampling point set, and performing a safety assessment on each point in the trajectory sampling point set based on a preset safety threshold to obtain a safety score sequence; According to the safety score sequence, an average safety score in each window is calculated to obtain a smoothed safety score sequence, and based on the smoothed safety score sequence, intervals that are continuously below a safety threshold are identified to obtain a potential danger interval; Performing density peak clustering on each of the potential danger intervals to obtain a potential danger point set, and constructing a danger distribution function based on the potential danger point set; Time-aligning the hazard distribution function with the current-voltage variation curve to obtain a hazard-time mapping relationship, and segmenting the current-voltage variation curve based on the hazard-time mapping relationship to obtain curve segments that need to be adjusted; Adjusting the curve parameters of each curve segment to be adjusted to obtain an optimized current-voltage curve segment, and reconstructing the current-voltage change curve based on the optimized current-voltage curve segment to obtain an adjusted current-voltage change curve; An arc characteristic trajectory is constructed for the adjusted current and voltage variation curve to obtain a second arc characteristic trajectory.

4. The arc protection method for a high-voltage energy storage connector for a charging pile according to claim 1, characterized in that: The processing of the second arc characteristic trajectory by a dynamic spatiotemporal graph neural network to obtain time series data includes: An initial undirected graph is constructed based on the physical topological structure of the high-voltage energy storage connector, each component of the connector is set as a node, and the physical connection between the components is set as an edge to obtain a topological graph, and the edge weight of the topological graph is initialized, and the initial edge weight is calculated based on the electrical characteristics between the components to obtain a weighted topological graph; Discretizing the second arc characteristic trajectory into a time step sequence, extracting the characteristic value of each time step to obtain a characteristic vector sequence, and assigning attributes to the nodes of the weighted topological graph based on the characteristic vector sequence to obtain a dynamic node attribute graph; Based on the graph convolution layer in the dynamic spatiotemporal graph neural network, spatial feature extraction is performed on the dynamic node attribute graph to obtain a spatial feature matrix, and based on the time convolution layer in the dynamic spatiotemporal graph neural network, feature extraction is performed on the spatial feature matrix in the time dimension to obtain a spatiotemporal feature tensor; Based on the attention mechanism in the dynamic spatiotemporal graph neural network, the importance weight of each feature in the spatiotemporal feature tensor is calculated to obtain a weighted feature tensor, and a pooling operation is performed on the weighted feature tensor to obtain a graph-level feature vector; Based on the fully connected layer in the dynamic spatiotemporal graph neural network, the graph-level feature vector is nonlinearly transformed to obtain hidden layer features, and based on the output layer in the dynamic spatiotemporal graph neural network, the hidden layer features are mapped to the target time series space to obtain initial time series data; The local time dependency relationship is extracted from the initial time series data to obtain an enhanced time series, and the enhanced time series is standardized to obtain time series data.

5. The arc prevention method for a high-voltage energy storage connector for a charging pile according to claim 1, characterized in that: The time series data is input into the CNN-MMHA model to predict the arc occurrence probability and obtain the arc risk assessment result, including: The time series data is segmented to obtain multiple data segments, and each of the data segments is input into a one-dimensional convolutional neural network of a CNN-MMHA model to extract local time features to obtain a feature map sequence; the one-dimensional convolutional neural network includes three convolutional layers, each layer uses a ReLU activation function and batch normalization, and the convolution kernel sizes are 5, 3, and 3 respectively; Inputting the feature map sequence into the multi-head self-attention module of the CNN-MMHA model to calculate the correlation between different time steps to obtain attention weighted features; the multi-head self-attention module includes 8 attention heads, each of which independently calculates the linear projection of the query, key and value, and then obtains the attention weight through the scaled dot product attention mechanism; Performing a nonlinear transformation on the attention weighted features to obtain an enhanced feature representation, and capturing long-term temporal dependencies on the enhanced feature representation to obtain a context-aware feature; Performing global average pooling and global maximum pooling on the context-aware features to obtain a global feature vector, and calculating an arc occurrence probability of the global feature vector; A risk score is calculated based on the arc occurrence probability, the risk score is compared with a preset threshold to obtain a binary risk label, and the risk score and the binary risk label are combined to generate an arc risk assessment result including a risk degree and a risk category.

6. The arc prevention method for a high-voltage energy storage connector for a charging pile according to claim 5, characterized in that: The performing global average pooling and global maximum pooling on the context-aware features to obtain a global feature vector, and calculating the arc occurrence probability of the global feature vector, includes: Performing global average pooling on the context-aware features along the time dimension to obtain an average pooled feature vector; Performing global maximum pooling on the context-aware features along the time dimension to obtain a maximum pooling feature vector; Concatenate the average pooling feature vector and the maximum pooling feature vector to obtain a fused feature vector; According to the fused feature vector, an importance score of each feature is calculated to obtain a feature importance vector, and the fused feature vector is weighted based on the feature importance vector to obtain a weighted feature vector; Performing a nonlinear transformation on the weighted feature vector to obtain a potential feature representation, and calculating the correlation between features based on the potential feature representation to obtain a target weighted feature; A Sigmoid activation function operation is performed on the target weighted features to obtain the arc occurrence probability.

7. The arc prevention method for a high-voltage energy storage connector for a charging pile according to claim 1, characterized in that: The arc prevention strategy parameters are calculated according to the arc risk assessment result, including: Parameterizing the physical characteristics and operating state of each charging unit, constructing a state space model, obtaining a state representation of the intelligent agent, and designing a reward function based on the state representation and the arc risk assessment result to obtain an objective function of the intelligent agent; Performing Taylor expansion on the objective function to obtain a first-order derivative and a second-order derivative of the objective function, and constructing a Hessian matrix based on the first-order derivative and the second-order derivative to obtain a local quadratic approximation of the objective function; Performing conjugate gradient analysis on the local quadratic approximation, calculating the search direction and step size, obtaining a parameter update vector, and designing an update rule of an implicit dual gradient tracking algorithm based on the parameter update vector to obtain a learning strategy of the intelligent agent; Performing offline training according to the learning strategy to obtain initial strategy parameters, and performing sensitivity analysis on the initial strategy parameters to identify target parameters and obtain parameter importance ranking; Designing an adaptive learning rate scheduling scheme based on the parameter importance ranking, performing online fine-tuning on the learning strategy to obtain dynamically adjusted strategy parameters, and performing constraint projection on the dynamically adjusted strategy parameters to obtain candidate strategy parameters; The candidate strategy parameters are input into a strategy network to generate an arc prevention action sequence to obtain a preliminary arc prevention strategy, and a Monte Carlo tree search is performed on the preliminary arc prevention strategy to obtain arc prevention strategy parameters.

8. The arc protection method for a high-voltage energy storage connector for a charging pile according to claim 7, characterized in that: The step of inputting the candidate strategy parameters into a strategy network, generating an arc prevention action sequence, obtaining a preliminary arc prevention strategy, and performing a Monte Carlo tree search on the preliminary arc prevention strategy to obtain arc prevention strategy parameters includes: Normalizing the candidate policy parameters to obtain standardized policy parameters, and inputting the standardized policy parameters into a multi-layer policy network, wherein the multi-layer policy network includes three fully connected layers, each of which uses a Tanh activation function to obtain a policy distribution; Generate discrete action samples based on the strategy distribution to obtain candidate arc prevention action sequences, and perform action filtering on the candidate arc prevention action sequences to obtain effective arc prevention action sequences; Input the effective arc protection action sequence into the value network, evaluate the long-term benefit of each action, obtain the action value estimate, and calculate the action selection probability based on the action value estimate; Initializing the root node of the Monte Carlo tree using the action selection probability, expanding the tree structure to obtain a search tree, and quickly simulating the search tree to evaluate the return of each leaf node to obtain the node value; Based on the node value, back propagation is performed to update the number of visits and cumulative returns of the parent node to obtain an updated search tree, and according to the updated search tree, an optimal action path is selected to obtain an optimized arc prevention strategy; The optimized arc protection strategy is weightedly fused with the original strategy parameters to obtain arc protection strategy parameters.

9. The arc prevention method for a high-voltage energy storage connector for a charging pile according to claim 8, characterized in that: The method of implementing a random trigger transmission protocol based on the arc prevention strategy parameters and a three-state Markov chain and generating a charging control instruction includes: The communication state of each charging unit is modeled and a three-state Markov chain is constructed. The three states include communication success, communication failure, and no communication. The state transition probability matrix is ​​calculated based on the historical communication data to obtain the initial Markov model. Converting the arc protection strategy parameters into state transition probability adjustment factors, adjusting the transition probability of the initial Markov model to obtain a dynamic Markov model, and generating a random trigger mechanism based on the dynamic Markov model; The random trigger mechanism is integrated with the update rule of the implicit dual gradient tracking algorithm to design an adaptive update frequency strategy; Based on the adaptive update frequency strategy, the communication efficiency and arc protection effect are taken as optimization targets to obtain the optimal configuration parameters; According to the optimal configuration parameters, a preliminary charging adjustment plan is obtained with charging efficiency and arc risk as reward signals; The preliminary charging adjustment plan is combined with the arc risk assessment result to generate a charging control instruction.

10. An arc protection device for a high-voltage energy storage connector of a charging pile, characterized in that: For implementing the steps of the method according to any one of claims 1 to 9, the device comprises: A construction module, used to construct a first arc characteristic trajectory based on a current and voltage change curve of the high-voltage energy storage connector during charging; an adjustment module, configured to perform safety constraint verification according to the first arc characteristic trajectory and perform parameter adjustment on the current and voltage variation curve to obtain a second arc characteristic trajectory; A processing module, used for processing the second arc characteristic trajectory through a dynamic spatiotemporal graph neural network to obtain time series data; A prediction module, used for inputting the time series data into the CNN-MMHA model to predict the probability of arc occurrence and obtain an arc risk assessment result; A calculation module, used for calculating arc protection strategy parameters according to the arc risk assessment result; A generation module is used to implement a random trigger transmission protocol according to the arc protection strategy parameters and based on a three-state Markov chain, and generate a charging control instruction.

Images