Distribution network dynamic topology identification and self-healing control method and device

Abstract

The invention discloses a distribution network dynamic topology identification and self-healing control method and device, and the method comprises the steps: collecting operation data of a distribution network, carrying out the time synchronization processing, data quality evaluation and data restoration, and obtaining a high-quality network state data set; based on the data set, performing dynamic topology identification by adopting a path improved by a relaxation method and a forest enhancement algorithm to obtain a dynamic network topology model; constructing a backpack network model by adopting a cyclic data generation method, and evaluating distributed photovoltaic output and load conditions; network abnormity is monitored in real time, and a machine learning algorithm is adopted to assist in judging fault types, positions and influence ranges; power supply reliability and economy are comprehensively considered, and a multi-objective optimization algorithm is adopted to generate a switch operation sequence; switching operation is executed, a temporary micro-grid is formed or a power supply path is adjusted, and the execution effect is monitored and evaluated in real time. According to the method, the dynamic topology of the high-proportion distributed photovoltaic power distribution network can be accurately identified, and fault positioning and efficient self-healing control are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power system automation, in particular to a distribution network dynamic topology identification and self-healing control method and device. BACKGROUND

[0002] As an important part of the power system, the distribution network is the key infrastructure directly facing the end users for power supply. With the rapid development of smart grid and energy internet, the distribution network is undergoing a transition from the traditional unidirectional radial power supply mode to a complex grid structure containing distributed energy, which poses new challenges to the operation and management of the distribution network.

[0003] Currently, the management of the distribution network mainly relies on the distribution network automation system and the distribution management system (DMS). These systems collect limited network state information through devices such as remote terminal units (RTUs) and feeder terminal units (FTUs), and perform fault location and isolation based on pre-set topology models. For example, the traditional feeder automation (FA) system mainly uses "three remote" (telemetry, remote signaling, and remote control) technology to perform switch operations through pre-set recovery strategies when a fault occurs.

[0004] However, with the large-scale integration of distributed photovoltaic and other renewable energy sources into the distribution network, the traditional topology identification and self-healing control methods face severe challenges. The existing technology is usually based on static topology models and deterministic algorithms, which cannot adapt to the characteristics of distributed photovoltaic output fluctuation and dynamic changes in network topology after a fault. In addition, the traditional methods have low computational efficiency when dealing with complex network structures, making it difficult to meet the time requirements for rapid fault recovery.

[0005] The common problems in the existing technology include: first, the topology identification accuracy is insufficient, especially in the distribution network containing a high proportion of distributed photovoltaic, due to the bidirectional and fluctuating nature of power flow, the traditional topology inference method often makes a wrong judgment; second, the self-healing control strategy lacks flexibility, which cannot fully utilize the ability of distributed photovoltaic to form a temporary microgrid, resulting in low fault recovery efficiency; third, the algorithm complexity is high, which makes it difficult to achieve real-time response in large-scale distribution networks, prolonging the fault recovery time and reducing the power supply reliability. SUMMARY

[0006] The purpose of the present application is to provide a distribution network dynamic topology identification and self-healing control method and device, which aims to solve the problems of insufficient topology identification accuracy, lack of flexibility in self-healing control strategy, and high algorithm complexity in the existing technology, and to realize accurate topology identification and efficient self-healing control of the distribution network containing a high proportion of distributed photovoltaic.

[0007] To achieve the above purpose, the present application provides a distribution network dynamic topology identification and self-healing control method, comprising the following steps:

[0008] collecting operation data of a power distribution network, performing time synchronization processing, data quality evaluation and data repair on the operation data to obtain a high-quality network state data set;

[0009] based on the high-quality network state data set, performing dynamic topology identification using a path and forest enhanced algorithm improved by a relaxation method to obtain a dynamic network topology model;

[0010] based on the high-quality network state data set and the dynamic network topology model, constructing a piggyback network model using a cyclic data generation method, evaluating distributed photovoltaic output and load conditions based on the piggyback network model, and obtaining photovoltaic-load balance state evaluation results;

[0011] based on the high-quality network state data set and the dynamic network topology model, real-time monitoring of network abnormalities, using a machine learning algorithm to assist in determining fault type, location and impact range, and obtaining fault impact range evaluation results;

[0012] based on the dynamic network topology model, the photovoltaic-load balance state evaluation results and the fault impact range evaluation results, comprehensively considering power supply reliability and economy, using a multi-objective optimization algorithm to generate a switch operation sequence, and obtaining a self-healing control strategy;

[0013] based on the self-healing control strategy, performing switch operation to form a temporary microgrid or adjust the power supply path, real-time monitoring and evaluation of the execution effect, and obtaining self-healing control execution results and performance evaluation report.

[0014] Preferably, the operation data of the power distribution network is collected, the operation data is time-synchronized, data quality is evaluated and data is repaired to obtain a high-quality network state data set, comprising:

[0015] collecting operation data containing voltage amplitude, phase angle, current, active power and reactive power data with a frequency not less than 1Hz and an accuracy level of 0.2, calibrating the operation data with GPS clock or network time protocol timestamp to obtain a time-synchronized sensor data set;

[0016] based on the time-synchronized sensor data set, using 3σ criterion and deep learning model to identify abnormal data deviating from the mean value by more than 3 times the standard deviation, and obtaining a data set with quality labels;

[0017] based on the data set with quality labels, using cubic spline interpolation method and similar day pattern replacement based on historical data to repair the abnormal data, performing maximum and minimum value normalization to obtain the high-quality network state data set.

[0018] Preferably, based on the high-quality network state data set, a dynamic topology identification is performed by using a path and forest enhanced algorithm improved by a relaxation method to obtain a dynamic network topology model, including:

[0019] Based on the physical connection relationship and switch state information of the distribution network, combined with the device coordinate information provided by the GIS system and the real-time switch state data provided by the SCADA system, a basic topology graph model G(V, E) containing a node set V and an edge set E is generated;

[0020] Based on the high-quality network state data set and the basic topology graph model, a spatio-temporal correlation analysis is performed on the node electrical measurement values to obtain a node electrical correlation strength matrix;

[0021] Based on the node electrical correlation strength matrix and the basic topology graph model, the topology constraints are converted into penalty terms by using an improved relaxation method, and a path algorithm and a forest enhanced algorithm are used to minimize the objective function f(T, M) + λ1·p1(g(T)) + λ2·p2(h(T)) to obtain a candidate topology structure set;

[0022] Based on the candidate topology structure set, combined with the root mean square error evaluation of the power flow calculation and the actual measurement values, the forest enhanced algorithm is used to improve the identification accuracy of the complex network structure, and the topology structure with the minimum root mean square error is selected as the dynamic network topology model.

[0023] Preferably, based on the high-quality network state data set and the dynamic network topology model, a knapsack network model is constructed by using a cyclic data generation method, the distributed photovoltaic output and load conditions are evaluated based on the knapsack network model, and a photovoltaic-load balance state evaluation result is obtained, including:

[0024] Based on the key parameters and historical power generation data of the photovoltaic inverter output power, component temperature, and environmental irradiance contained in the high-quality network state data set, time characteristics, environmental characteristics, and electrical characteristics are extracted to obtain a photovoltaic power generation feature set;

[0025] Based on the high-quality network state data set and the user electricity consumption history data, a clustering algorithm is used to identify typical load patterns to obtain a load feature set;

[0026] Based on the photovoltaic power generation feature set and the load feature set, time translation transformation and amplitude scaling transformation are performed on the feature data to generate sample data that satisfies physical constraints, and a knapsack network model N(V, E, P, Z) considering node power attributes P and edge set impedance attributes Z is constructed to obtain a photovoltaic-load correlation network model;

[0027] Based on the photovoltaic-load correlation network model and the dynamic network topology model, distributed photovoltaic output prediction and load prediction are performed using an ensemble learning method to obtain the photovoltaic-load balance state evaluation result.

[0028] Preferably, based on the photovoltaic-load correlation network model, combined with 4-hour short-term weather forecast data, the theoretical power generation of the photovoltaic system is calculated using a temperature correction-based photovoltaic system physical model to obtain a preliminary photovoltaic output prediction result;

[0029] Based on the preliminary photovoltaic output prediction result, statistical analysis is performed using a time series model and a regression model to obtain a statistical photovoltaic output prediction result;

[0030] According to the statistical photovoltaic output prediction result, an ensemble learning algorithm including random forest and XGBoost is used for prediction optimization to obtain a final photovoltaic output prediction result;

[0031] Based on the final photovoltaic output prediction result and the dynamic network topology model, the voltage qualification rate of each node, the line load rate, the power balance margin, and the voltage margin are evaluated, the confidence interval of the prediction result is calculated, and the photovoltaic-load balance state evaluation result is obtained.

[0032] Preferably, based on the high-quality network state data set and the dynamic network topology model, network anomalies are monitored in real time, and machine learning algorithms are used to assist in determining fault types, locations, and impact ranges to obtain fault impact range evaluation results, including:

[0033] Based on the high-quality network state data set, CUSUM control charts and EWMA control charts are used to monitor abnormal changes in voltage, current, and frequency parameters, and a 3σ detection threshold is set to obtain preliminary fault alarm signals;

[0034] Based on the preliminary fault alarm signals, basic probability assignments are calculated using Dempster-Shafer evidence theory, and multi-source information fusion results are obtained by fusing node measurement data and protection device action information;

[0035] According to the multi-source information fusion results, decision tree algorithms are used to analyze fault characteristics to obtain fault type determination results;

[0036] Based on the fault type determination results, improved fault traveling wave analysis and impedance distance measurement methods are used in combination with support vector machine algorithms to obtain fault location information;

[0037] Based on the fault location information and the dynamic network topology model, a system average interruption frequency index (SAIFI), a system average interruption duration index (SAIDI), and an expected energy not supplied (EENS) are calculated, the importance level of power interruption load is evaluated, and the fault influence range evaluation result is obtained.

[0038] Preferably, based on the fault type determination result, a fault transient waveform with a sampling frequency not less than 10 kHz is collected, a wave head feature is extracted by using wavelet transform, a wave head arrival time difference is calculated based on a GPS synchronous clock, and a traveling wave positioning result is obtained in combination with a traveling wave speed of 500 km / s.

[0039] Based on the fault transient waveform, a fault point impedance is calculated, and a fault point distance is estimated according to a line unit length impedance, so as to obtain an impedance ranging result.

[0040] Based on the traveling wave positioning result and the impedance ranging result, a support vector machine algorithm is used to classify a fault feature vector, wherein the feature vector includes a voltage sag amplitude, a fault current amplitude, a wave head time difference, and an impedance measurement value, a radial basis kernel function is used for feature mapping, and a preliminary fault location is obtained.

[0041] Based on the preliminary fault location, a position correction is performed in combination with the dynamic network topology model, a fault current contribution and a voltage support effect of distributed photovoltaic are considered, a weighted least squares method is used to optimize the positioning result, and a positioning accuracy not less than ±100 meters is obtained, so as to obtain the fault location information.

[0042] Preferably, based on the dynamic network topology model, the photovoltaic-load balance state evaluation result, and the fault influence range evaluation result, power supply reliability and economy are comprehensively considered, a multi-objective optimization algorithm is used to generate a switch operation sequence, and a self-healing control strategy is obtained, including:

[0043] Based on the fault influence range evaluation result and the dynamic network topology model, a graph theory algorithm is used to determine a minimum isolation area, a switch operation sequence is generated, and a fault isolation scheme is obtained.

[0044] Based on the fault isolation scheme and the photovoltaic-load balance state evaluation result, a multi-objective genetic algorithm is used to generate a power supply recovery candidate scheme, wherein a chromosome code represents a switch state, a fitness function is f=w1·RR+w2·(1 / NS)+w3·(1 / PL), RR is a recovery rate, NS is a number of operation steps, and PL is a system loss, and a candidate scheme set is obtained.

[0045] Based on the candidate scheme set, the feasibility of forming a temporary microgrid is evaluated for each candidate scheme, the island operation capability of distributed photovoltaic and the load control flexibility are considered, and a temporary microgrid scheme set is obtained.

[0046] Based on the candidate scheme set and the temporary micro-grid scheme set, the weight coefficients of power supply recovery rate, operation step number, and system loss are determined as 0.5, 0.3, and 0.2 respectively by using the analytic hierarchy process, a scheme ranking is performed by using a fuzzy comprehensive evaluation method, and a scheme with the highest comprehensive score is selected as the optimal self-healing control strategy;

[0047] Based on the optimal self-healing control strategy, a detailed switch operation timing is generated, and the self-healing control strategy containing specific execution instructions is obtained.

[0048] Preferably, based on the self-healing control strategy, switch operation is performed to form a temporary micro-grid or adjust a power supply path, and real-time monitoring and evaluation of the execution effect are performed to obtain a self-healing control execution result and a performance evaluation report, including:

[0049] Based on the self-healing control strategy, control instructions are issued to intelligent switch devices through a power distribution automation system master station to perform switch operation, and switch operation execution state information is obtained;

[0050] Based on the switch operation execution state information, a droop control method is used to adjust active power of distributed photovoltaics, a voltage-reactive power characteristic curve is used to control charging and discharging of an energy storage system, and interruptible load management is implemented based on load priority, with a control cycle of 100 ms, and micro-grid operation state data is obtained;

[0051] Based on the switch operation execution state information and the micro-grid operation state data, network operation parameters are continuously monitored, and power supply recovery effect is evaluated, and self-healing control effect evaluation data is obtained;

[0052] Based on the self-healing control effect evaluation data, key performance indicators such as power supply recovery time, recovery rate, and energy utilization efficiency are calculated, and a self-healing control execution result is obtained;

[0053] Based on the self-healing control execution result, a case similarity calculation method is used to compare and analyze historical cases, experience knowledge including fault characteristics, disposal methods, and execution effects is extracted, an improved K-means algorithm is used to cluster the experience knowledge, typical cases and disposal rules in the knowledge base are updated, and the performance evaluation report is obtained.

[0054] Preferably, based on the self-healing control effect evaluation data, fault detection time t d , fault isolation time t i , and power supply recovery time t r are calculated respectively, and whether the overall fault handling time T=t d +t i +t r satisfies the requirement of less than 300 seconds is evaluated, and a time performance indicator is obtained.

[0055] Based on the time performance index, the ratio RR=Pr / Po of the total amount of power supply recovery load Pr and the total amount of power-off load Po is counted, whether the power supply recovery rate reaches the target requirement of more than 95% is evaluated, and a recovery rate index is obtained;

[0056] Based on the recovery rate index, the ratio η=Pa / Pm of the actual output power Pa of the distributed photovoltaic in the temporary micro-grid and the theoretical maximum output power Pm under the current meteorological condition is calculated, whether the energy utilization level reaches the target requirement of more than 90% is evaluated, and an energy efficiency index is obtained;

[0057] Based on the time performance index, the recovery rate index and the energy efficiency index, the analytic hierarchy process is used to determine the weight coefficients of 0.3, 0.4 and 0.3 respectively, the fuzzy comprehensive evaluation method is used to calculate the weighted score S=0.3T+0.4RR+0.3η, and the self-healing control execution result is obtained.

[0058] The application also provides a distribution network dynamic topology identification and self-healing control device, which comprises:

[0059] A data acquisition and preprocessing module is used to acquire the operation data of the power distribution network, and the operation data is subjected to time synchronization processing, data quality evaluation and data repair to obtain a high-quality network state data set;

[0060] A dynamic topology identification module is used to identify the dynamic topology based on the high-quality network state data set by using a path and forest enhanced algorithm improved by a relaxation method to obtain a dynamic network topology model;

[0061] A photovoltaic load evaluation module is used to construct a back-loaded network model by using a cyclic data generation method based on the high-quality network state data set and the dynamic network topology model, to evaluate the distributed photovoltaic output and load condition based on the back-loaded network model, and to obtain a photovoltaic-load balance state evaluation result;

[0062] A fault detection and positioning module is used to monitor network abnormalities in real time based on the high-quality network state data set and the dynamic network topology model, to assist in judging the fault type, position and influence range by using a machine learning algorithm, and to obtain a fault influence range evaluation result;

[0063] A self-healing control strategy module is used to comprehensively consider the power supply reliability and economy based on the dynamic network topology model, the photovoltaic-load balance state evaluation result and the fault influence range evaluation result, to generate a switch operation sequence by using a multi-objective optimization algorithm, and to obtain a self-healing control strategy;

[0064] The strategy execution and verification module adopts the switch operation based on the self-recovery control strategy to form a temporary micro-grid or adjust the power supply path, and performs real-time monitoring and evaluation on the execution effect to obtain a self-recovery control execution result and a performance evaluation report.

[0065] The present application has the following advantages:

[0066] 1. A path and forest enhanced algorithm improved by a relaxation method is proposed, which converts strict topological constraints into penalty terms by introducing relaxation variables, and combines the forest enhanced algorithm to improve the complex network structure identification accuracy, and realizes accurate dynamic topology identification of the power distribution network containing a high proportion of distributed photovoltaic power;

[0067] 2. A cyclic data generation method is innovatively used to enrich the sample space, a back-carrying network model containing space-time relationship is constructed, the network topology structure and node power information are integrated, and the accuracy of distributed photovoltaic output and load estimation is improved;

[0068] 3. A temporary micro-grid construction and evaluation method is proposed, which evaluates the feasibility of forming a temporary micro-grid according to fault conditions and distributed photovoltaic availability, considers photovoltaic island operation capability, energy storage capacity and load control flexibility, calculates micro-grid stability indicators, and realizes maximum power supply recovery after fault area isolation;

[0069] 4. An optimization model containing multiple objectives such as power supply recovery rate, operation step number, system loss and power quality is established, an optimal self-recovery control strategy is generated, and the fault recovery efficiency and reliability of the power distribution network are improved;

[0070] 5. A complete closed loop is formed through continuous data acquisition, dynamic topology identification, state evaluation, fault handling, strategy execution and effect monitoring, and the real-time and adaptability of the power distribution network self-recovery control are realized. BRIEF DESCRIPTION OF DRAWINGS

[0071] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0072] Figure 1 The power grid dynamic topology identification and self-recovery control method flowchart provided for the embodiments of the present application;

[0073] Figure 2 The strategy execution and verification flowchart provided for the embodiments of the present application;

[0074] Figure 3 The structure diagram of the power grid dynamic topology identification and self-recovery control device provided for the embodiments of the present application. DETAILED DESCRIPTION

[0075] The advantages and features of the present application will be more apparent from the following detailed description of preferred embodiments of the present application, taken in conjunction with the accompanying drawings.

[0076] Please refer to Figure 1 , Figure 1 The flow chart of the power distribution network dynamic topology identification and self-healing control method provided by the embodiments of the present application is shown in FIG. 1. As shown in FIG. 1, the power distribution network dynamic topology identification and self-healing control method provided by the embodiments of the present application includes the following steps: Figure 1

[0077] Step S101: Collecting the operation data of the power distribution network, performing time synchronization processing, data quality evaluation and data repair on the operation data, and obtaining a high-quality network state data set.

[0078] Step S102: Based on the high-quality network state data set, using the path and forest enhanced algorithm improved by the relaxation method to perform dynamic topology identification, and obtaining a dynamic network topology model.

[0079] Step S103: Based on the high-quality network state data set and the dynamic network topology model, using a cyclic data generation method to construct a piggyback network model, evaluating the distributed photovoltaic output and load based on the piggyback network model, and obtaining a photovoltaic-load balance state evaluation result.

[0080] Step S104: Based on the high-quality network state data set and the dynamic network topology model, real-time monitoring of network abnormalities, using a machine learning algorithm to assist in judging the fault type, location and influence range, and obtaining a fault influence range evaluation result.

[0081] Step S105: Based on the dynamic network topology model, the photovoltaic-load balance state evaluation result and the fault influence range evaluation result, comprehensively considering the power supply reliability and economy, using a multi-objective optimization algorithm to generate a switch operation sequence, and obtaining a self-healing control strategy.

[0082] Step S106: Based on the self-healing control strategy, performing switch operation to form a temporary microgrid or adjust the power supply path, real-time monitoring and evaluation of the execution effect, and obtaining a self-healing control execution result and performance evaluation report.

[0083] The steps are described in detail as follows, and step S101 includes the following sub-steps:

[0084] Step S1011: Collecting operation data including voltage amplitude, phase angle, current, active power and reactive power data with a frequency not less than 1 Hz and an accuracy level of 0.2 level, and calibrating the operation data with GPS clock or network time protocol timestamp to obtain a time-synchronized sensor data set.​

[0085] Specifically, based on the intelligent sensor network (including PMU, smart meter, FTU, etc.) deployed in the smart power distribution network, a wide-area measurement system is established to collect real-time data such as voltage amplitude, phase angle, current, active power, and reactive power of each node and branch of the power distribution network, and output original multi-source heterogeneous sensor data stream. The collected data has a frequency of no less than 1 Hz and a precision level of 0.2, to ensure the accuracy of subsequent analysis.

[0086] The collected data is time-stamped calibrated by GPS clock or Network Time Protocol (NTP) to solve the problem of different sensor data time synchronization, and output time-synchronized sensor data set. The time synchronization accuracy reaches millisecond level, ensuring the comparability of data in time dimension.

[0087] Step S1012: Based on the time-synchronized sensor data set, use 3σ rule and deep learning model to identify abnormal data deviating from the mean by more than 3 times the standard deviation, and obtain data set with quality label.

[0088] Based on the time-synchronized sensor data set, use 3σ rule and deep learning model to identify abnormal data deviating from the mean by more than 3 times the standard deviation, and obtain data set with quality label. Specifically, the deep learning model is an autoencoder (Autoencoder) structure, which includes an input layer, 3 layers of encoder (128, 64, and 32 neurons respectively in each layer, using ReLU activation function), 3 layers of decoder (symmetric structure with encoder), and an output layer. The model uses mean square error (MSE) as the loss function, uses Adam optimizer for training, sets the learning rate to 0.001, the batch size to 64, and the training rounds to 100. The training data comes from high-quality monitoring data during historical normal operation period, and fixed-length time series samples are generated by sliding window method. The model compares the reconstruction error with the set threshold (usually the 95th percentile of the reconstruction error of the validation set) to determine whether the data point is abnormal.

[0089] Specifically, statistical methods (such as 3σ rule) and deep learning models (such as autoencoder) are used for anomaly detection to identify abnormal conditions such as data missing, mutation, and drift. The 3σ rule is based on the assumption of normal distribution, and data points deviating from the mean by more than 3 times the standard deviation are marked as abnormal. The autoencoder learns the characteristics of normal data and marks data points with large reconstruction error as abnormal. The output is a data set with quality labels, where each data point has a corresponding quality evaluation result.

[0090] Step S1013: Based on the data set with quality labels, the abnormal data is repaired using cubic spline interpolation and similar day pattern replacement based on historical data, maximum and minimum value normalization processing is performed, and the high-quality network state data set is obtained.

[0091] Specifically, the interpolation method, similar day pattern replacement or correction method based on physical model are used for data repair of abnormal data. For short-time missing data, cubic spline interpolation method is used for filling; for long-time missing data, similar day pattern replacement method is used, that is, date data with similar characteristics (such as weather conditions, date type) is found from historical data for replacement. The repaired data is subjected to maximum and minimum value normalization processing, and all kinds of data are unified to the interval [0, 1] for subsequent processing. Finally, a high-quality network state data set is output, providing a reliable data basis for subsequent topology identification and evaluation analysis.

[0092] Step S102 includes the following sub-steps:

[0093] Step S1021: Based on the physical connection relationship of the power distribution network and the switch state information, combined with the device coordinate information provided by the GIS system and the real-time switch state data provided by the SCADA system, a basic topology graph model G(V, E) containing a node set V and an edge set E is generated.

[0094] Specifically, the spatial position information and electrical connection relationship of the power distribution network equipment are extracted from the GIS (Geographic Information System) database, including the position coordinates and connection relationship of the transformer substation, distribution line, tower, switch device, transformer and distributed photovoltaic access point, etc. The state information (closing / opening) of the switch device is obtained in real time from the SCADA (Supervisory Control and Data Acquisition) system.

[0095] In mathematical representation, the initial topology graph model G(V, E) is an undirected graph, where V represents the vertex set, corresponding to the nodes in the power distribution network (such as bus, distribution transformer, load point, distributed photovoltaic access point, etc.), E represents the edge set, corresponding to the connection lines and devices in the power distribution network (such as cable, overhead line, switch, etc.). For each edge e E, the system will associate a state flag s(e), indicating whether the connection is in the power-on state: s(e) = 1 indicates that the corresponding switch device is in the closed state, the connection is valid; s(e) = 0 indicates that the corresponding switch device is in the open state, the connection is invalid.

[0096] Step S1022: Based on the high-quality network state data set and the basic topology graph model, the spatio-temporal correlation analysis of the node electrical measurement value is performed, and the node electrical correlation strength matrix is obtained.

[0097] Specifically, a high-quality network state dataset is obtained from step S101, including time-series data such as voltage amplitude, phase angle, current, active power, and reactive power for each node. For data from adjacent sampling points, the system calculates the Pearson correlation coefficient and partial correlation coefficient between each measurement.

[0098] The formula for calculating the Pearson correlation coefficient is: r(X,Y)=cov(X,Y) / (σ X ·σ Y ), where X and Y are time series measurement data from two nodes, cov(X,Y) is their covariance, and σX and σY are the variances of X and Y. Y These are the standard deviations of X and Y, respectively.

[0099] Algorithms such as improved Conditional Random Fields (CRF) or Graph Convolutional Networks (GCN) are applied to analyze the spatial and temporal dependencies of the measurement data. For each pair of nodes i and j, the system calculates an electrical association strength r. ij It comprehensively considers voltage dependence, power dependence, impedance characteristics, and time delay effects. Finally, it outputs a nodal electrical correlation strength matrix R = [r ij ], where r ij This indicates the degree of electrical connection between nodes i and j.

[0100] Step S1023: Based on the node electrical association strength matrix and the basic topology graph model, the topology constraints are transformed into penalty terms using an improved relaxation method. The objective function f(T,M)+λ1·p1(g(T))+λ2·p2(h(T)) is minimized using the path algorithm and forest enhancement algorithm to obtain a set of candidate topologies.

[0101] Specifically, the traditional topology identification problem can be formulated as finding the topology that best matches the measurement data while satisfying the physical constraints of the power system. This is typically modeled as a combinatorial optimization problem, as follows:

[0102] min T f(T,M);

[0103] stg(T) = 0;

[0104] h(T)≤0;

[0105] Where T represents the topology, M represents the measurement data, f(T,M) is the objective function that measures the degree of matching between the topology T and the measurement data M, and g(T) and h(T) are the equality and inequality constraints that the topology must satisfy (such as radial structure, power balance, etc.).

[0106] This embodiment introduces a relaxation method, which converts strict topological constraints into penalty terms in the objective function. Through relaxation processing, the original problem becomes:

[0107] min T f(T,M)+λ1·p1(g(T))+λ2·p2(h(T))

[0108] where p1 and p2 are penalty functions, and λ1 and λ2 are weight coefficients.

[0109] At the same time, a path algorithm is used to handle the topological identification problem, which regards the distribution network as a connected graph and identifies the topological structure by analyzing the path characteristics in the graph. The improved Dijkstra algorithm or A* algorithm is used to search for possible power supply paths, and the rationality of the paths is evaluated in combination with electrical characteristics.

[0110] A forest enhancement algorithm is also introduced, which is based on the idea of random forest. By constructing multiple decision tree models, different aspects of the topological structure are predicted, and then the prediction results are integrated. Each decision tree uses different feature subsets and data subsets, which improves the generalization ability and robustness of the model.

[0111] The algorithm execution process is as follows:

[0112] (1) Initialize the candidate topology set T = {T0}, where T0 is the initial topology constructed in step S1021;

[0113] (2) For each iteration k:

[0114] a. Construct the objective function, including the data fitting term and the constraint penalty term;

[0115] b. Use the path algorithm to explore possible topological changes;

[0116] c. Apply the forest enhancement algorithm to evaluate each candidate topology;

[0117] d. Update the candidate topology set T.

[0118] (3) Stop when the convergence condition is met or the maximum number of iterations is reached.

[0119] Finally, output the candidate topology structure set T = {T1, T2,..., T}, each candidate topology structure has a corresponding confidence score.

[0120] Step S1024: Based on the candidate topology structure set, combined with the root mean square error evaluation of the actual measured value and the power flow calculation, the forest enhancement algorithm is used to improve the identification accuracy of the complex network structure, and the topology structure with the smallest root mean square error is selected as the dynamic network topology model.

[0121] Specifically, the candidate topology structure set T = {T1, T2,..., T

[0122] } generated in step S1023 is evaluated. The evaluation criteria include measurement data consistency, physical constraint satisfaction degree and structure rationality.

[0123] For each candidate topology T i , the system calculates a comprehensive score S(T i ):

[0124] S(T i ) = w1·S data (T i ) + w2·S constraint (T i ) + w3·S structure (T i )

[0125] Where S data , S constraint and S structure represent the scores of data consistency, constraint satisfaction and structure rationality respectively, and w1, w2 and w3 are the corresponding weight coefficients.

[0126] The forest enhancement algorithm is applied to improve the identification accuracy of the complex network structure:

[0127] (1) A plurality of sub-evaluation models are constructed for the candidate topology set, each model focusing on a different evaluation aspect;

[0128] (2) Each sub-model independently evaluates the candidate topology to generate a score vector;

[0129] (3) Use integrated learning methods such as weighted voting or stacking to integrate the evaluation results of all sub-models;

[0130] (4) According to the fused score, determine the final topology structure ranking.

[0131] In particular, for the region containing distributed photovoltaic, the system will additionally consider the influence of photovoltaic output on network state. Finally, the topology structure with the highest score (the smallest root mean square error) is selected as the optimal solution T*, and the final verification is performed. If the verification is passed, the accurate dynamic network topology model T* is output.

[0132] Step S103 includes the following sub-steps:

[0133] Step S1031: Based on the key parameters and historical power generation data in the high-quality network state data set including photovoltaic inverter output power, component temperature, and environmental irradiance, time characteristics, environmental characteristics, and electrical characteristics are extracted to obtain a photovoltaic power generation feature set.

[0134] Specifically, information related to distributed photovoltaics is collected from multiple data sources, including photovoltaic inverter direct monitoring data (DC side voltage, current, AC side voltage, current, active power, reactive power, power factor, etc.), grid side measurement data (electricity parameters of photovoltaic grid-connected points collected by smart meters, PMUs, etc.), environmental parameter data (solar irradiance, ambient temperature, component temperature, wind speed, etc.), and system static information (configuration information such as photovoltaic system installed capacity, panel type, orientation, inclination, inverter model, control mode, etc.).

[0135] Photovoltaic output feature extraction mainly includes the following types of features:

[0136] (1) Time characteristics: Time stamp is decomposed into year, month, day, hour, minute, and other sunlight time characteristics such as season, sunrise and sunset time, solar elevation angle, and azimuth angle are extracted;

[0137] (2) Environment-related features: Based on environmental parameters such as irradiance and temperature, calculate the theoretical maximum power generation, temperature correction coefficient, etc.;

[0138] (3) Electrical characteristics: Extract power fluctuation characteristics, harmonic content, voltage regulation characteristics, etc. to describe the electrical behavior of photovoltaics;

[0139] (4) Historical output features: Extract historical contemporaneous output patterns, output change rates, daily output curve shapes, etc.;

[0140] (5) Spatial correlation features: Consider the output correlation of photovoltaic systems with similar geographic locations, and capture the impact of regional weather changes on photovoltaic groups.

[0141] For each distributed photovoltaic system, a feature vector is constructed:

[0142] X pv =[x time ,x environment, x electrical ,x history ,x spatial ]

[0143] Where each component corresponds to the above five types of features. Finally, the feature vectors of all distributed photovoltaic systems are combined into a photovoltaic power generation feature set F PV .

[0144] Step S1032: Based on the high-quality network state dataset and user electricity consumption history data, a clustering algorithm is used to identify typical load patterns to obtain a load feature set.

[0145] Specifically, load data from multiple sources is collected, including smart meter data (electricity consumption, power factor, maximum demand, etc. of various users), feeder monitoring data (total load data collected from distribution transformers, feeder terminals, etc.), historical electricity consumption data (daily, weekly, monthly, quarterly, and annual historical electricity consumption data), and auxiliary information (such as user type, geographic location, building area, air conditioning installed capacity, etc.).

[0146] A clustering algorithm is used to identify typical load patterns. Common clustering methods include K-means, hierarchical clustering, density clustering (such as DBSCAN), etc. Cluster analysis is usually based on the following features:

[0147] (1) Daily load curve shape (such as peak-valley difference, electricity consumption time period distribution);

[0148] (2) Weekly / monthly load variation pattern;

[0149] (3) Sensitivity to temperature and other meteorological factors;

[0150] (4) Electricity consumption level and volatility.

[0151] Taking K-means clustering as an example, the execution process is as follows:

[0152] (1) Represent each user's daily load curve as a 24-dimensional vector (corresponding to 24-hour power values);

[0153] (2) Determine the optimal number of clusters K (determined by elbow method, silhouette coefficient, etc.);

[0154] (3) Randomly select K initial cluster centers;

[0155] (4) Iteratively perform assignment and update steps until convergence;

[0156] (5) Analyze the characteristics of each cluster to determine the typical load pattern.

[0157] External factors that affect load are also considered, including seasonal factors, weather factors, date types, and social events. Time series analysis methods such as autoregressive integrated moving average (ARIMA) model, exponential smoothing method, etc. are applied to extract the time trend, periodicity, and seasonality characteristics of the load.

[0158] Finally, the identified typical load patterns, load-external factor correlation models, and time series characteristics are integrated into the load feature set F Load :

[0159] F Load = {Patterns, Weather Sensitivity , Temporal Features , User Attributes}

[0160] Step S1033: Based on the photovoltaic power generation feature set and the load feature set, time shift transformation and amplitude scaling transformation are performed on the feature data to generate sample data that satisfies physical constraints, and a power-laden network model N(V, E, P, Z) considering node power attribute P and edge set impedance attribute Z is constructed to obtain a photovoltaic-load correlation network model.

[0161] Specifically, the innovative cyclic data generation method (Cyclic Data Generation) is implemented to enrich the sample space through the following steps:

[0162] (1) Basic data decomposition: The photovoltaic power generation feature set F_PV of step S1031 and the load feature set F_Load of step S1032 are decomposed into a basic pattern library, including typical daily patterns, weather response patterns, mutation event patterns, etc.

[0163] (2) Parameterized modeling: Each pattern is described parametrically, such as using power series, Fourier series, or wavelet transform coefficients to represent the intra-day curve shape;

[0164] (3) Cyclic transformation generation: New samples are generated by applying time shift, amplitude scaling, superposition, etc. to the basic patterns;

[0165] (4) Physical constraint verification: The generated samples are physically reasonable and meet the actual system operation rules;

[0166] (5) Sample diversity enhancement: Introduce controlled random disturbance to increase the diversity and coverage of samples.

[0167] Next, a power-laden network model (Power-Laden Network Model) is constructed, which innovatively integrates the topological structure and node power information:

[0168] (1) Define the network model N(V, E, P, Z), where V is the node set, E is the edge set, P is the power attribute mapping, and Z is the impedance attribute mapping;

[0169] (2) For each node v∈V, associate a power attribute vector P(v) = [P gen ,P load ,Q gen ,Q loadP, Q, QG, QL, etc.

[0170] (3) For each edge e E, associate an impedance attribute Z(e) and capacity limit C(e);

[0171] (4) Establish the coupling relationship between node power and topology, such as power flow, voltage distribution, etc.

[0172] Using graph neural network (GNN) or edge feature convolution network, learn the mutual influence between node power and network topology. Finally, output the photovoltaic-load correlation network model N(V, E, P, Z), which is an enhanced topology model rich in power information.

[0173] Step S1034: Based on the photovoltaic-load correlation network model and the dynamic network topology model, use the ensemble learning method to perform distributed photovoltaic output prediction and load prediction to obtain the photovoltaic-load balance state evaluation result.

[0174] Specifically, based on the photovoltaic-load correlation network model constructed in step S1033 and the dynamic network topology model output in step S102, the state estimation of the distribution network is performed. The mathematical expression of the distribution network state estimation is:

[0175]

[0176] s.t.g(x)=0;

[0177] Where x is the state variable (such as the voltage amplitude and phase angle of each node), z is the measurement vector, h(x) is the measurement equation, g(x) is the system constraint, and W is the weight matrix.

[0178] The distributed state estimation method is particularly suitable for solving the state estimation problem under the condition of high penetration rate of distributed photovoltaic:

[0179] (1) Divide the network into multiple sub-networks, each containing several adjacent nodes;

[0180] (2) Each sub-network performs local state estimation in parallel;

[0181] (3) Exchange information through boundary nodes to coordinate the estimation results of each sub-network;

[0182] (4) Iterate until overall convergence.

[0183] Next, combined with short-term weather forecast data, distributed photovoltaic output prediction and load prediction are performed. Using the ensemble learning method, the prediction results of multiple models are integrated to improve the prediction accuracy:

[0184] (1) Physical model: Based on the physical parameters of photovoltaic systems and the input of irradiance, temperature, etc., calculate the theoretical power generation;

[0185] (2) Statistical model: such as time series model (ARIMA), regression model, etc.;

[0186] (3) Machine learning model: such as random forest, XGBoost, etc. integrated learning algorithm;

[0187] (4) Deep learning model: such as long short-term memory network (LSTM), spatio-temporal graph convolution network, etc.

[0188] Rolling prediction is performed for a time window of 15 minutes to 4 hours in the future, and the prediction result is continuously updated as new data arrives. For the uncertainty of prediction, the system adopts a probabilistic prediction method to output the confidence interval or probability distribution of the predicted value.

[0189] Finally, based on the state estimation and prediction results, the photovoltaic-load balance state is evaluated, including power balance status, voltage qualification rate, line load rate, system stability and reserve capacity, etc. The system finally outputs the photovoltaic-load balance state evaluation result S, including the above indicators and their time evolution trend.

[0190] In a preferred embodiment, step S1034 further comprises the following sub-steps:

[0191] Step S10341: Based on the photovoltaic-load correlation network model, combined with the short-term weather forecast data for the next 4 hours, the theoretical power generation is calculated using the temperature-corrected photovoltaic system physical model to obtain the preliminary photovoltaic power output prediction result.

[0192] Step S10342: Based on the preliminary photovoltaic power output prediction result, statistical analysis is performed using time series model and regression model to obtain the statistical photovoltaic power output prediction result.

[0193] Step S10343: According to the statistical photovoltaic power output prediction result, integrated learning algorithm containing random forest and XGBoost is used for prediction optimization to obtain the final photovoltaic power output prediction result.

[0194] According to the photovoltaic output statistical prediction result, an ensemble learning algorithm containing random forest and XGBoost is used for prediction optimization to obtain a photovoltaic output final prediction result. Specifically, the random forest model is configured with 100 decision trees, the maximum depth is 8, and the number of features considered when each node is split is the square root of the total number of features; the XGBoost model is configured with 200 trees, the learning rate is 0.05, the maximum depth is 6, the regularization parameter λ = 1, γ = 0.1, and the mean squared log error (MSLE) is used as the loss function. The two models are trained on historical photovoltaic output data (including meteorological data, time features and system parameters), the training set and the test set are divided in a ratio of 8:2, and the hyperparameters are optimized through 5-fold cross-validation. Finally, the prediction results of the two models are combined through weighted averaging (random forest weight 0.4, XGBoost weight 0.6), and the weight values are optimized on the validation set through grid search.

[0195] Step S10344: Based on the photovoltaic output final prediction result and the dynamic network topology model, the voltage qualification rate, the line load rate, the power balance margin and the voltage margin of each node are evaluated, the confidence interval of the prediction result is calculated, and the photovoltaic-load balance state evaluation result is obtained.

[0196] Step S104 includes the following sub-steps:

[0197] Step S1041: Based on the high-quality network state data set, CUSUM control chart and EWMA control chart are used to monitor the abnormal changes of voltage, current and frequency parameters, and a 3σ detection threshold is set to obtain a preliminary fault alarm signal.

[0198] Specifically, based on the high-quality network state data set output by step S101, the key electrical parameters of the distribution network are monitored in real time, including voltage parameters (three-phase voltage amplitude, phase, three-phase unbalance degree), current parameters (three-phase current amplitude, phase, three-phase unbalance degree), power parameters (active power, reactive power, power factor), frequency parameters (system frequency and its change rate), power quality parameters (harmonic content, flicker, transient event, etc.), and protection device state (such as relay action signal, circuit breaker state, etc.).

[0199] Statistical Process Control (SPC) method is applied to establish the normal operation reference value and reasonable fluctuation range of each monitoring parameter, and an alarm is triggered when the parameter deviates from the preset range. The system uses CUSUM (Cumulative Sum) control chart and EWMA (Exponential Weighted Moving Average) control chart, and sets a 3σ detection threshold, that is, data exceeding the range of ±3 times the standard deviation of the mean value is considered abnormal.

[0200] For the distribution network with high penetration of distributed photovoltaic, the following abnormal patterns are of particular concern: voltage out-of-limit, power flow reversal, output sudden change and island operation. When an abnormality is detected, the system conducts preliminary assessment and classification, including disturbance degree assessment, time characteristic judgment, spatial characteristic analysis and preliminary type judgment.

[0201] Finally, a preliminary fault alarm signal A is output, containing the abnormality detection timestamp, the degree of deviation of abnormal parameters, the preliminary judgment of abnormal type, the preliminary estimate of abnormal location and the rating of abnormal severity.

[0202] Step S1042: Based on the preliminary fault alarm signal, the Dempster-Shafer evidence theory is used to calculate the basic probability assignment, and the multi-source information fusion result is obtained by fusing the node measurement data and protection device action information.

[0203] Specifically, based on the preliminary fault alarm signal output in step S1041, the system collects and integrates multi-source information, mainly including measurement data, protection device information, topology information, historical fault data and external environment information.

[0204] The evidence theory (Evidence Theory), also known as Dempster-Shafer theory, is used for multi-source information fusion. The basic idea is:

[0205] (1) Define the recognition framework Θ, which contains all possible fault types;

[0206] (2) Assign a basic probability assignment (BPA) to each information source;

[0207] (3) Use Dempster combination rule to fuse the evidence of multiple information sources.

[0208] The Dempster combination rule is calculated as follows:

[0209] m(A) = (1 / K) ∑_{B∩C=A} m1(B) m2(C)

[0210] Where K is the normalization factor used to handle conflicting evidence.

[0211] Through multi-source information fusion, a more comprehensive understanding of the fault situation can be obtained, and the misjudgment caused by a single information source can be reduced. Finally, the multi-source information fusion result is output, providing a more reliable basis for subsequent fault type judgment.

[0212] Step S1043: According to the multi-source information fusion result, the decision tree algorithm is used to analyze the fault characteristics, and the fault type judgment result is obtained.

[0213] Specifically, a decision tree algorithm is applied for fault type judgment. Decision trees have the advantages of strong interpretability and high computational efficiency, making them suitable for fault diagnosis applications:

[0214] (1) A decision tree model is constructed, with features including electrical parameter change patterns, protection device responses, etc.

[0215] (2) Reasoning along the decision tree path, branch selection is made according to the judgment conditions of each node.

[0216] (3) Reach the leaf node and obtain the fault type judgment result.

[0217] The main fault types that can be identified include:

[0218] (1) Short circuit fault: including single-phase ground, two-phase short circuit, three-phase short circuit, etc.

[0219] (2) Open circuit fault: such as wire breakage, device open circuit, etc.

[0220] (3) Voltage anomaly: voltage upper limit, voltage lower limit, voltage sudden drop, etc.

[0221] (4) Device fault: such as transformer overheating, circuit breaker malfunction, capacitor bank fault, etc.

[0222] (5) Photovoltaic system fault: such as inverter fault, grid connection point access problem, etc.

[0223] For complex or uncertain fault conditions, fuzzy reasoning methods are used to handle various uncertainties through fuzzy rules. The time sequence characteristics of the fault are also considered to determine whether it is a single fault or a cascading fault.

[0224] Finally, the fault type judgment result F_type is output, including the fault type, its confidence, key evidence related to the fault, possible development trend of the fault, and potential impact assessment of the fault on the system.

[0225] Step S1044: Based on the fault type judgment result, improved fault traveling wave analysis and impedance distance measurement method are used, combined with support vector machine algorithm, to obtain fault location information.

[0226] Specifically, based on the fault type judgment result of step S1043 and the dynamic network topology model of step S102, the system selects appropriate positioning algorithms combined with fault types.

[0227] For short circuit faults, the following methods are mainly used:

[0228] (1) Improved fault traveling wave analysis: Collect fault transient waveforms recorded at multiple measurement points, calculate the distance from the fault point to each measurement point based on the time difference of wave front arrival and wave speed, and use improved double-end or multi-end positioning algorithm to reduce the influence of measurement error. Mathematical representation: d = (L + v · (t1-t2)) / 2, where d is the distance from the fault point to measurement point 1, L is the line length, v is the wave speed, and t1 and t2 are the times of wave front arrival at both ends.

[0229] (2) Impedance ranging method: Calculate fault impedance based on fault current and voltage, estimate fault point distance according to line unit length impedance, and correct traditional impedance ranging method considering the influence of distributed photovoltaic injection current. Mathematical representation: d = |Z measured |·L / |Z line |, where Z measured is the measured impedance, Z line is the full line impedance, and L is the line length.

[0230] For non-short-circuit faults (such as open circuit, voltage anomaly, etc.), state estimation method and fault feature mapping method are used.

[0231] To improve positioning accuracy, combine support vector machine (SVM) or random forest machine learning model. In the offline phase, use historical fault data or simulation data to train the model; in the online phase, extract the feature vector of the current fault and input it into the trained model for prediction, output the fault location and its confidence.

[0232] For distribution networks with distributed photovoltaic, especially considering the influence of photovoltaic system on fault features, including fault current contribution, voltage support effect and system impedance change. For multiple fault conditions, the system uses a segmented positioning strategy.

[0233] Finally, output the fault section or device location information L fault , including the accurate position description of the fault point, the confidence of the positioning result, the possible positioning error range and the key evidence of the positioning basis.

[0234] Step S1045: Based on the fault location information and the dynamic network topology model, calculate the system average interruption frequency index SAIFI, the system average interruption duration index SAIDI and the expected energy not supplied EENS, evaluate the importance level of power interruption load, and obtain the fault impact range evaluation result.

[0235] Specifically, based on the fault location information of step S1044, the dynamic network topology model of step S102 and the photovoltaic-load balance state evaluation result of step S103, comprehensive impact assessment is carried out.

[0236] First, the direct outage area is determined by analyzing the network topology and protection configuration to identify the area that loses power immediately after the fault occurs. Second, the affected users and critical loads are evaluated, with particular attention to the number and type of outage users, the status of critical users, the calculation of the amount of power interruption, and the classification of load priorities.

[0237] The impact and availability of distributed photovoltaics are also analyzed, including the distributed photovoltaic capacity within the outage area, the operating status of photovoltaic systems, the identification of photovoltaic systems with islanding operation capability, and the electrical distance and connection relationship between photovoltaic systems and critical loads.

[0238] For possible secondary fault risks, an evaluation and warning are conducted, including line overload due to load transfer, voltage problems due to changes in the topology, power balance problems due to photovoltaic system disconnection, and new challenges for protection systems.

[0239] Key indicators are calculated to quantify the impact of the fault, including the System Average Interruption Frequency Index (SAIFI), the System Average Interruption Duration Index (SAIDI), and the Expected Energy Not Supplied (EENS). Sensitivity analysis is performed for different restoration time scenarios.

[0240] Finally, the fault impact range assessment result R_fault is output, including confirmation information of fault type, location, and occurrence time, detailed description and visualization of the direct outage area, classification statistics of affected users and loads, total and classified statistics of outage loads, available distributed resource information, potential secondary fault risk warnings, estimated values of key indicators, and impact assessments under different restoration time scenarios.

[0241] In a preferred embodiment, step S1044 further comprises the following sub-steps:

[0242] Step S10441: Based on the fault type determination result, collect a fault transient waveform with a sampling frequency not less than 10 kHz, extract the wave head feature using wavelet transform, calculate the wave head arrival time difference based on the GPS synchronous clock, and combine the wave speed of 500 km / s to obtain the traveling wave positioning result.

[0243] Step S10442: Based on the fault transient waveform, calculate the fault point impedance, estimate the fault point distance according to the line unit length impedance, and obtain the impedance ranging result.

[0244] Step S10443: Based on the traveling wave positioning result and the impedance ranging result, use the support vector machine algorithm to classify the fault feature vector, where the feature vector includes the voltage sag amplitude, the fault current amplitude, the wave head time difference, and the impedance measurement value, use the radial basis kernel function for feature mapping, and obtain the preliminary fault location.

[0245] Based on the traveling wave positioning result and the impedance ranging result, a support vector machine algorithm is used to classify the fault feature vector, wherein the feature vector includes voltage sag amplitude, fault current amplitude, wave head time difference and impedance measurement value, radial basis kernel function is used for feature mapping to obtain a preliminary fault location. Specifically, the kernel function parameter γ of the support vector machine algorithm is set to 0.01, the penalty parameter C is set to 10, and grid search and 10-fold cross-validation are used to optimize these hyperparameters. The feature vector is input into the model after normalization, and the training data is derived from actual fault cases and electromagnetic transient simulation data, containing feature data of different types of faults (single-phase grounding, two-phase short circuit, three-phase short circuit, etc.) at different positions, with a total sample size of about 5000. The model training uses the sequential minimal optimization (SMO) algorithm, and the convergence accuracy is set to 0.001 and the maximum iteration number is 1000. To solve the sample imbalance problem, the synthetic minority over-sampling technique (SMOTE) is used for data enhancement.

[0246] Step S10444: Based on the preliminary fault location, the position correction is performed in combination with the dynamic network topology model, the fault current contribution and voltage support effect of distributed photovoltaic are considered, the weighted least squares method is used to optimize the positioning result, and the positioning accuracy is not less than ±100 meters, to obtain the fault location information.

[0247] Step S105 includes the following sub-steps:

[0248] Step S1051: Based on the fault influence range evaluation result and the dynamic network topology model, a graph theory algorithm is used to determine the minimum isolation area, generate a switch operation sequence, and obtain a fault isolation scheme.

[0249] Specifically, based on the accurate fault information report of step S104 and the dynamic network topology model of step S102, the system applies a graph theory algorithm to determine the minimum isolation area.

[0250] A mathematical model for fault isolation is constructed. The distribution network is represented as a graph G(V, E), where V is the node set (such as bus, connection point), and E is the edge set (such as line, switch). The fault isolation problem can be formulated as:

[0251] min |S|

[0252]

[0253] Where f is the fault point, S is the set of switches to be operated, R G-S is the connected component of f after removing S, E S is the set of operable switches. The goal is to find the smallest set of switch operations to isolate the fault point from the power source.

[0254] The system uses an improved minimum cut set algorithm to solve the above problem:

[0255] (1) Determine the fault point f and the power supply point set P;

[0256] (2) In the graph G, find all paths connecting f and any p∈P;

[0257] (3) Identify the minimum edge set (switch set) that can cut off all these paths;

[0258] (4) Adjust the minimum cut set considering the operability constraints of the switches.

[0259] For complex network structures, a hierarchical isolation strategy is adopted, including the backbone network layer, branch network layer and terminal network layer. The actions already performed by the protection devices are also considered to avoid repeated operations. For areas containing distributed photovoltaic, the system specially considers the isolation requirements of photovoltaic grid connection points, the protection settings and responses of photovoltaic systems, the confirmation of anti-islanding protection, and the possible re-grid conditions of photovoltaic systems.

[0260] The generated isolation scheme is checked for safety to verify that the fault point is indeed isolated from the power supply after isolation operation, to check possible transient problems during isolation operation, to evaluate the impact of isolation operation on non-fault areas, and to confirm the executability of the isolation scheme. Finally, the switch operation sequence is optimized to generate detailed operation timing.

[0261] Finally, the fault isolation scheme I_plan is output, including the list of switch devices that need to be operated, the detailed operation sequence and time schedule, the expected state of the network topology after operation, the safety precautions for isolation operation, and the operation confirmation and verification method.

[0262] Step S1052: Based on the fault isolation scheme and the photovoltaic-load balance state evaluation result, a multi-objective genetic algorithm is used to generate a power supply recovery candidate scheme, where the chromosome code represents the switch state, and the fitness function is f=w1·RR+w2·(1 / NS)+w3·(1 / PL), RR is the recovery rate, NS is the number of operation steps, and PL is the system loss, to obtain a candidate scheme set.

[0263] Based on the fault isolation scheme and the photovoltaic-load balance state evaluation result, a multi-objective genetic algorithm is used to generate power supply recovery candidate schemes, wherein a chromosome code represents a switch state, and a fitness function is f = w1-RR + w2-(1 / NS) + w3-(1 / PL), RR is a recovery rate, NS is a number of operation steps, and PL is a system loss, to obtain a candidate scheme set. Specifically, the multi-objective genetic algorithm is configured as follows: a population size is 100, a maximum number of iterations is 200, a chromosome uses binary coding, each bit corresponds to a switch state (0 represents open and 1 represents closed); a double-point crossover is used for a crossover operation, and a crossover probability is 0.8; a bit flip mutation is used for a mutation operation, and a mutation probability is 0.05; a tournament selection is used for a selection operation, and a tournament size is 5; and an elite retention strategy retains 10% of individuals in each generation. The algorithm uses a non-dominated sorting (NSGA-II) framework to process a multi-objective optimization problem, and uses a crowding distance calculation to maintain diversity of solutions. To accelerate convergence, 30% of solutions in an initial population are generated based on expert rules, and the rest are randomly generated. All candidate schemes are verified for feasibility by using a Newton-Raphson method for power flow calculation.

[0264] Specifically, based on the fault isolation scheme of step S1051, the dynamic network topology model of step S102, and the photovoltaic-load balance state evaluation result of step S103, the system applies a multi-objective heuristic algorithm to generate a plurality of power supply recovery candidate schemes.

[0265] Firstly, the areas that can be restored to power supply include the isolated fault areas, the areas that are powered off due to fault isolation but are not faulty, and the areas that remain powered.

[0266] The improved genetic algorithm is applied to generate power supply recovery candidate schemes:

[0267] (1) Coding scheme: binary coding is used to represent a switch state, 1 represents closed, and 0 represents open;

[0268] (2) Initial population: an initial solution set that satisfies basic constraints is generated;

[0269] (3) Fitness function: a recovery load, a network loss, a number of operation steps, and the like are comprehensively considered, and are represented as f = w1-RR + w2-(1 / NS) + w3-(1 / PL), wherein RR is a recovery rate, NS is a number of operation steps, and PL is a system loss;

[0270] (4) Selection operation: a roulette wheel selection or a tournament selection is used;

[0271] (5) Crossover operation: information exchange is performed on parent schemes to generate new schemes;

[0272] (6) Mutation operation: randomly change part of the switch states to increase the diversity of solutions;

[0273] (7) Elitism: keep the best solutions in each generation.

[0274] The particle swarm optimization (PSO) algorithm is used as the complementary algorithm. For each candidate solution, the system performs a power flow calculation to verify its feasibility, checks whether the voltage is within the allowed range, checks whether the line load is over the limit, verifies whether the power balance is met, and evaluates the stability margin of the solution.

[0275] Particular attention is paid to the utilization of distributed photovoltaics, taking into account the current and predicted output of photovoltaic systems, photovoltaic grid connection requirements and limitations, the control capabilities of photovoltaic systems, and the degree of geographical distribution matching between photovoltaic and load.

[0276] Finally, the candidate solution set C_plans is output, each solution containing a sequence of switch operations, a list of loads expected to be restored and the total amount, a configuration of power resources utilized, a network topology structure after restoration, a prediction of key electrical parameters, and an analysis of the advantages and disadvantages of the solution.

[0277] Step S1053: Based on the candidate solution set, evaluate the feasibility of forming a temporary microgrid for each candidate solution, taking into account the islanding operation capability of distributed photovoltaics and the flexibility of load control, to obtain a temporary microgrid solution set.

[0278] Specifically, based on the candidate solution set of step S1052 and the photovoltaic-load balance state evaluation result of step S103, a detailed temporary microgrid construction evaluation is performed.

[0279] First, identify the possible temporary microgrid areas formed in each candidate solution, analyze the topology within the outage area, determine the available distributed photovoltaic resources within the area, identify network sub-areas with independent power supply potential, and evaluate the distribution of key loads within the area.

[0280] Next, evaluate the islanding operation capability of distributed photovoltaics:

[0281] (1) Check the operation mode and control function of photovoltaic inverters, including grid-following and grid-forming modes;

[0282] (2) Evaluate the capacity matching of photovoltaic systems, calculate the P PV / P load ratio, evaluate the power supply and demand balance, consider the volatility of photovoltaic output, and determine the power margin for reliable power supply;

[0283] (3) Check the status of the supporting energy storage system, including the energy storage capacity and power capability, current state of charge (SOC), and charge-discharge control capability;

[0284] (4) Analyze the status of auxiliary equipment, including voltage and frequency control equipment, protection device adaptability, communication and control system availability.

[0285] Analyze the load characteristics to assess the controllability under temporary microgrid conditions:

[0286] (1) Load classification and priority determination, including Class I loads (important loads that must be continuously powered), Class II loads (less important loads that can be interrupted for a short time), and Class III loads (non-critical loads that can be delayed for power supply);

[0287] (2) Load response capability assessment, including controllable load proportion, load adjustment speed and range, and automatic load control equipment status;

[0288] (3) Load-side management strategy development, including load shedding sequence, rotation power supply scheme, and demand response trigger conditions.

[0289] Calculate microgrid stability indicators to assess the reliability of temporary microgrids:

[0290] (1) Power balance stability, including static power balance margin and dynamic response capability;

[0291] (2) Voltage stability, including voltage regulation capability and voltage regulation device response time;

[0292] (3) Frequency stability, including frequency regulation capability, rotational inertia, and equivalent inertia response;

[0293] (4) Protection coordination, including whether short-circuit capacity meets protection device action requirements and temporary adjustment requirements of protection settings.

[0294] For each possible temporary microgrid scheme, the system performs time-domain simulation to verify its dynamic performance, simulates photovoltaic output fluctuation scenarios, load step change response, and small disturbance stability, and evaluates the recovery ability and flexibility of the temporary microgrid.

[0295] Finally, output a set of feasible temporary microgrid schemes M_plans, each containing microgrid boundary definition, power supply configuration, load configuration, control scheme, protection scheme, stability evaluation results, and operation time estimation and limitation conditions.

[0296] Step S1054: Based on the candidate scheme set and the temporary microgrid scheme set, the weight coefficients of power supply recovery rate, operation step number, and system loss are determined as 0.5, 0.3, and 0.2 respectively by using the analytic hierarchy process, a scheme ranking is performed by using a fuzzy comprehensive evaluation method, and a scheme with the highest comprehensive score is selected as the optimal self-healing control strategy.

[0297] Specifically, based on the candidate scheme set of step S1052 and the temporary microgrid scheme set of step S1053, a multi-objective optimization model is established.

[0298] Firstly, key targets to be considered are determined:

[0299] (1) Power supply recovery rate (RR): the proportion of loads restored to power supply to total loads without power supply, RR = ∑P restored / ∑P outage ;

[0300] (2) Operation step number (NS): the number of switch operations required for implementing a scheme, NS = |S operate |;

[0301] (3) System loss (PL): active power loss of a network after power supply is restored, PL = ∑I 2 R;

[0302] (4) Power quality index (PQ): including voltage deviation, three-phase unbalance degree, etc., PQ = w1·VD + w2·UF + w3·THD, wherein VD is voltage deviation, UF is unbalance degree, and THD is harmonic distortion degree;

[0303] (5) Reliability index (RI): evaluating reliability and safety margin of a scheme, RI = w4·SC + w5·VM + w6·LM, wherein SC is short-circuit capacity margin, VM is voltage margin, and LM is load margin.

[0304] The analytic hierarchy process (AHP) is applied to determine the weight of each target, a judgment matrix is constructed to represent the relative importance between targets, the maximum eigenvalue and corresponding eigenvector of the judgment matrix are calculated, the eigenvector is normalized to obtain the weight of each target, and consistency test is performed to ensure the rationality of the judgment. According to the result of the analytic hierarchy process, the weight coefficients of power supply recovery rate, operation step number, and system loss are determined as 0.5, 0.3, and 0.2 respectively.

[0305] A fuzzy comprehensive evaluation method is adopted for evaluation:

[0306] (1) An evaluation factor set (i.e., the above five targets) is established;

[0307] (2) An evaluation set (such as "very good", "good", "general", "poor", and "very poor") is determined;

[0308] (3) For each scheme, a fuzzy relation matrix is established;

[0309] (4) Combined with the weight vector, the fuzzy comprehensive evaluation result is calculated;

[0310] (5) Clear processing is performed to obtain the comprehensive score of the scheme.

[0311] The top-ranked schemes are subjected to sensitivity analysis, the weights of each target are adjusted, the changes in scheme ranking are observed, the influence of photovoltaic output change on scheme evaluation is considered, the sensitivity of load change on scheme optimization is evaluated, and the robustness of the scheme under different boundary conditions is verified.

[0312] Finally, the scheme with the highest comprehensive score is selected as the optimal self-healing control strategy.

[0313] Step S1055: Based on the optimal self-healing control strategy, a detailed switch operation timing is generated to obtain the self-healing control strategy containing specific execution instructions.

[0314] Specifically, based on the selected optimal self-healing control strategy, a detailed switch operation timing is generated, including a fault isolation operation sequence, a power supply path reconstruction operation sequence, a distributed resource control instruction, a load management strategy, a temporary microgrid formation and control scheme (if applicable), a timing arrangement and execution condition of the operation, an expected effect and key indicator, and an emergency plan and alternative scheme.

[0315] This optimization decision-making process comprehensively considers power supply reliability, economy, operation complexity and system safety, etc. and can find the optimal solution balancing various targets under complex fault conditions, providing clear guidance for subsequent self-healing control strategy execution.

[0316] Please refer to Figure 2 , Figure 2 the flowchart of strategy execution and verification provided by the embodiment of the present application. As shown in Figure 2 , step S106 includes the following sub-steps:

[0317] Step S1061: Based on the self-healing control strategy, control instructions are issued to intelligent switch devices through the distribution automation system master station to execute switch operations and obtain switch operation execution state information.

[0318] Specifically, based on the optimal self-healing control strategy output by step S105, the system issues control instructions to related intelligent switch devices through the distribution automation system master station to execute switch operations according to the optimized operation sequence, realizing fault isolation and power supply path reconstruction.

[0319] The switching operation sequence in the self-healing control strategy is first converted into a standard control instruction format, and then the instruction is issued to the corresponding intelligent switch device through the communication network of the power distribution automation system. The system monitors the execution status of each switch operation, including instruction reception confirmation, operation execution process, and operation completion feedback.

[0320] For abnormal situations that may occur during operation, such as communication interruption, device refusal, etc., there are corresponding abnormal handling mechanisms, including retransmission of instructions, backup channel communication, and manual intervention, etc. The timestamp of each operation is also recorded for subsequent performance evaluation.

[0321] Finally, the switch operation execution status information E_status is output, including the execution result, execution time, and related abnormal information (if any) of each switch operation.

[0322] Step S1062: Based on the switch operation execution status information, the active power of the distributed photovoltaic is adjusted using the droop control method, the charging and discharging of the energy storage system is controlled using the voltage-reactive power characteristic curve, the interruptible load management is implemented based on the load priority, the control period is 100ms, and the microgrid operation state data is obtained.

[0323] Specifically, based on the switch operation execution status of step S1061 and the optimal self-healing control strategy of step S105, the temporary microgrid formed is implemented coordinated control, including power regulation of distributed photovoltaic, charging and discharging control of energy storage system, and management of controllable load, to maintain microgrid power balance and voltage stability.

[0324] For distributed photovoltaic, the active power is adjusted using the droop control method:

[0325] P=P0-k p ·(f-f0)

[0326] Where P is the photovoltaic output power, P0 is the rated power, k p is the droop coefficient, f is the system frequency, and f0 is the rated frequency. This control method can make the photovoltaic system automatically adjust the output power according to the system frequency to maintain the stability of the system frequency.

[0327] For the energy storage system, the charging and discharging is controlled using the voltage-reactive power characteristic curve:

[0328] Q=Q0-k q ·(V-V0)

[0329] Where Q is the reactive power output of the energy storage system, Q0 is the rated reactive power, k q is the droop coefficient, V is the system voltage, and V0 is the rated voltage. At the same time, the active charging and discharging of the energy storage system is controlled according to the power balance requirement.

[0330] For load management, interruptible load management is implemented based on load priority, and low-priority loads are cut off in the order of preset priority when power is insufficient, and loads are restored in the opposite order when power is restored.

[0331] The control cycle is 100 ms, which ensures a quick response to power fluctuations and voltage changes in the microgrid. The system continuously monitors the operating state of the microgrid, including node voltage, frequency, power flow, and device status.

[0332] Finally, the microgrid operating state data M_state is output, including electrical parameters of each node in the microgrid, operating state and control parameters of each device, load power supply situation, and system stability indicators, etc.

[0333] Step S1063: Based on the switch operation execution state information and the microgrid operating state data, continuously monitor the network operating parameters, evaluate the power supply recovery effect, and obtain self-healing control effect evaluation data.

[0334] Specifically, based on the switch operation execution state of step S1061 and the microgrid operating state data of step S1062, combined with the data acquisition system of step S101, the network operating parameters are continuously monitored, the power supply recovery effect is evaluated, and potential problems are discovered in time.

[0335] The key parameters monitored include:

[0336] (1) Topology parameter: Verify that the actual network topology after switch operation is consistent with the expected one;

[0337] (2) Electrical parameter: Monitor whether the voltage, current, power and other parameters of each node are within the allowed range;

[0338] (3) Device status: Monitor the load rate and operating state of key devices such as transformers, lines, switches, etc.;

[0339] (4) Protection system: Monitor the setting value and action of the protection device;

[0340] (5) Distributed resources: Monitor the output and control state of photovoltaic systems, energy storage systems and other distributed resources;

[0341] (6) Load power supply: Monitor the power supply condition and power quality of various types of loads.

[0342] Real-time analysis is performed on the monitoring data to evaluate the power supply recovery effect:

[0343] (1) Calculate the power supply recovery rate: the proportion of loads restored to power supply to the total power outage load;

[0344] (2) Assess power quality: including voltage deviation, frequency deviation, three-phase imbalance, etc.

[0345] (3) Analyze system stability: assess the response characteristics and stability margin of the system to small disturbances;

[0346] (4) Monitor abnormal conditions: timely discovery of potential overload, overvoltage or other abnormal conditions.

[0347] For the problems found, generate alarm information and take appropriate handling measures according to the severity of the problem, such as adjusting control parameters, modifying operation sequences or starting emergency plans, etc.

[0348] Finally, output the self-healing control effect evaluation data E_data, including various indicators of power supply recovery effect, detailed information of running state, problems found and handling situation, and time series data of continuous monitoring, etc.

[0349] Step S1064: Based on the self-healing control effect evaluation data, calculate the key performance indicators such as power supply recovery time, recovery rate and energy utilization efficiency, and obtain the self-healing control execution result; based on the self-healing control execution result, use case similarity calculation method and historical cases for comparative analysis, extract experience knowledge including fault characteristics, disposal method and execution effect, use improved K-means algorithm for clustering of experience knowledge, update typical cases and disposal rules in knowledge base, and obtain the performance evaluation report.

[0350] Specifically, based on the self-healing control effect evaluation data of step S1063, calculate the key performance indicators (such as power supply recovery time, recovery rate, energy utilization efficiency, etc.), and compare with historical cases for comparative analysis, extract experience knowledge and store in knowledge base.

[0351] The calculated key performance indicators include:

[0352] (1) Time indicators: including fault detection time, fault isolation time, power supply recovery time and total fault handling time;

[0353] (2) Recovery rate indicators: the proportion of power supply recovery load to total power outage load;

[0354] (3) Energy efficiency indicators: utilization efficiency of distributed resources, such as the ratio of actual output to theoretical maximum output of photovoltaic system;

[0355] (4) Operation efficiency indicators: the ratio of actual operation steps to theoretical optimal steps;

[0356] (5) Economic benefit indicators: including reduced power outage loss, saved operation cost, etc.

[0357] The current case is compared and analyzed with historical cases, and a case similarity calculation method is used to evaluate the similarity between cases:

[0358] sim(C1,C2)=∑w i ·sim i (C1,C2)

[0359] wherein sim(C1,C2) is the overall similarity of cases C1 and C2, sim i is the similarity of the i-th feature, and w i is the corresponding weight.

[0360] Experience knowledge is extracted from the current case, including fault characteristics, treatment methods, and execution effects, etc. An improved K-means algorithm is used to cluster the experience knowledge to identify typical patterns and rules. The extracted experience knowledge is updated to the knowledge base, including updating the typical case library and optimizing the treatment rules.

[0361] Finally, a self-healing control execution result and performance evaluation report R performance is generated, which contains detailed data of key performance indicators, comparison and analysis results with historical cases, extracted experience knowledge and improvement suggestions, and knowledge base update status. This report provides valuable reference for future similar situation processing.

[0362] In a preferred embodiment, step S1064 further comprises the following sub-steps:

[0363] Step S10641: Based on the self-healing control effect evaluation data, respectively calculate the fault detection time t d , the fault isolation time t i and the power supply recovery time t r , evaluate whether the overall fault handling time T=t d +t i +t r meets the requirement of less than 300 seconds, and obtain the time performance indicator.

[0364] Step S10642: Based on the time performance indicator, calculate the ratio of the total amount of load restored to power Pr and the total amount of load disconnected Po RR=Pr / Po, evaluate whether the power supply recovery rate meets the target requirement of 95% or more, and obtain the recovery rate indicator.

[0365] Step S10643: Based on the recovery rate indicator, calculate the ratio of the actual output power Pa of distributed photovoltaic in the temporary micro-grid to the theoretical maximum output power Pm under current weather conditions η=Pa / Pm, evaluate whether the energy utilization level meets the target requirement of 90% or more, and obtain the energy efficiency indicator.

[0366] Step S10644: based on the time performance index, the recovery rate index and the energy efficiency index, the weight coefficients are determined as 0.3, 0.4 and 0.3 respectively by using the analytic hierarchy process, and the weighted score S = 0.3T + 0.4RR + 0.3η is calculated by using the fuzzy comprehensive evaluation method, so as to obtain the self-healing control execution result.

[0367] Please refer to Figure 3 , Figure 3 The structural diagram of the power distribution network dynamic topology identification and self-healing control device provided by the embodiment of the application is shown in the figure. Figure 3 As shown in the figure, the power distribution network dynamic topology identification and self-healing control device provided by the embodiment of the application comprises:

[0368] The data acquisition and preprocessing module 801 acquires the operation data of the power distribution network, and performs time synchronization processing, data quality evaluation and data repair on the operation data to obtain a high-quality network state data set.

[0369] The dynamic topology identification module 802 performs dynamic topology identification based on the high-quality network state data set by using the path and forest enhanced algorithm improved by the relaxation method to obtain a dynamic network topology model.

[0370] The photovoltaic load evaluation module 803 constructs a back-loaded network model by using the cyclic data generation method based on the high-quality network state data set and the dynamic network topology model, evaluates the distributed photovoltaic output and load based on the back-loaded network model, and obtains a photovoltaic-load balance state evaluation result.

[0371] The fault detection and positioning module 804 monitors network abnormalities in real time based on the high-quality network state data set and the dynamic network topology model, and assists in judging the fault type, location and influence range by using a machine learning algorithm to obtain a fault influence range evaluation result.

[0372] The self-healing control strategy module 805 considers power supply reliability and economy comprehensively based on the dynamic network topology model, the photovoltaic-load balance state evaluation result and the fault influence range evaluation result, generates a switch operation sequence by using a multi-objective optimization algorithm, and obtains a self-healing control strategy.

[0373] The strategy execution and verification module 806 executes switch operation based on the self-healing control strategy, forms a temporary microgrid or adjusts the power supply path, monitors and evaluates the execution effect in real time, and obtains a self-healing control execution result and performance evaluation report.

[0374] The above merely illustrates the specific embodiments of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of the changes or replacements within the technical range disclosed by the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for dynamic topology identification and self-healing control of network distribution, characterized in that, The method comprises the following steps: Collecting operation data of a power distribution network, performing time synchronization processing, data quality evaluation and data repair on the operation data to obtain a high-quality network state data set; Based on the high-quality network state data set, a path and forest enhanced algorithm improved by a relaxation method is used for dynamic topology identification to obtain a dynamic network topology model; Based on the high-quality network state data set and the dynamic network topology model, a loop data generation method is used to construct a piggyback network model, and based on the piggyback network model, distributed photovoltaic output and load conditions are evaluated to obtain photovoltaic-load balance state evaluation results; Based on the high-quality network state data set and the dynamic network topology model, network anomalies are monitored in real time, and a machine learning algorithm is used to assist in judging fault types, locations and impact ranges to obtain fault impact range evaluation results; Based on the dynamic network topology model, the photovoltaic-load balance state evaluation results and the fault impact range evaluation results, power supply reliability and economy are comprehensively considered, a multi-objective optimization algorithm is used to generate a switch operation sequence to obtain a self-healing control strategy; Based on the self-healing control strategy, switch operations are performed to form a temporary microgrid or adjust the power supply path, and the execution effect is monitored and evaluated in real time to obtain self-healing control execution results and performance evaluation reports.

2. The method of claim 1, wherein, The collecting operation data of a power distribution network, performing time synchronization processing, data quality evaluation and data repair on the operation data to obtain a high-quality network state data set comprises: Collecting operation data including voltage amplitude, phase angle, current, active power and reactive power data with a frequency not less than 1 Hz and an accuracy level of 0.2 level, calibrating the operation data with GPS clock or network time protocol timestamp to obtain a time-synchronized sensor data set; Based on the time-synchronized sensor data set, using 3σ criterion and deep learning model to identify abnormal data deviating from the mean value by more than 3 times the standard deviation to obtain a data set with quality labels; Based on the data set with quality labels, using cubic spline interpolation method and similar day pattern replacement based on historical data to repair the abnormal data, performing maximum and minimum value normalization to obtain the high-quality network state data set.

3. The method of claim 1, wherein, The based on the high-quality network state data set, a path and forest enhanced algorithm improved by a relaxation method is used for dynamic topology identification to obtain a dynamic network topology model, comprising: Based on the physical connection relationship and switch state information of the power distribution network, combining the device coordinate information provided by the GIS system and the real-time switch state data provided by the SCADA system, a basic topology graph model G(V, E) containing a node set V and an edge set E is generated; Based on the high-quality network state data set and the basic topology graph model, a spatio-temporal correlation analysis is performed on the node electrical measurement values to obtain a node electrical correlation strength matrix; Based on the node electrical correlation strength matrix and the basic topology graph model, a path algorithm and a forest enhancement algorithm are used to minimize the objective function f(T, M)+λ1·p1(g(T))+λ2·p2(h(T)) by converting the topology constraint into a penalty term using an improved relaxation method, to obtain a candidate topology structure set; Based on the candidate topology structure set, the root mean square error evaluation of the power flow calculation and the actual measured value is combined, and the forest enhancement algorithm is used to improve the identification accuracy of the complex network structure, and the topology structure with the minimum root mean square error is selected as the dynamic network topology model.

4. The method of claim 1, wherein, Based on the high-quality network state data set and the dynamic network topology model, a knapsack network model is constructed using a cyclic data generation method, and the distributed photovoltaic output and load conditions are evaluated based on the knapsack network model to obtain photovoltaic-load balance state evaluation results, including: Based on the key parameters and historical power generation data in the high-quality network state data set, including photovoltaic inverter output power, component temperature, and environmental irradiance, time characteristics, environmental characteristics, and electrical characteristics are extracted to obtain a photovoltaic power generation feature set; Based on the high-quality network state data set and user power consumption history data, a clustering algorithm is used to identify typical load patterns to obtain a load feature set; Based on the photovoltaic power generation feature set and the load feature set, time translation transformation and amplitude scaling transformation are performed on the feature data to generate sample data that satisfies physical constraints, and a knapsack network model N(V, E, P, Z) is constructed considering node power attribute P and edge set impedance attribute Z to obtain a photovoltaic-load correlation network model; Based on the photovoltaic-load correlation network model and the dynamic network topology model, an integrated learning method is used for distributed photovoltaic output prediction and load prediction to obtain the photovoltaic-load balance state evaluation results.

5. The method of claim 4, wherein, Based on the photovoltaic-load correlation network model, combined with 4-hour short-term weather forecast data, the theoretical power generation is calculated using a photovoltaic system physical model based on temperature correction to obtain a preliminary photovoltaic output prediction result; Based on the preliminary photovoltaic output prediction result, a time series model and a regression model are used for statistical analysis to obtain a photovoltaic output statistical prediction result; According to the photovoltaic output statistical prediction result, an integrated learning algorithm including random forest and XGBoost is used for prediction optimization to obtain a final photovoltaic output prediction result; Based on the final photovoltaic output prediction result and the dynamic network topology model, the voltage qualification rate of each node, the line load rate, the power balance margin, and the voltage margin are evaluated, the confidence interval of the prediction result is calculated, and the photovoltaic-load balance state evaluation results are obtained.

6. The method of claim 1, wherein, Based on the high-quality network state data set and the dynamic network topology model, real-time network anomalies are monitored, and machine learning algorithms are used to assist in determining fault types, locations, and impact ranges to obtain fault impact range evaluation results, including: Based on the high-quality network state data set, the CUSUM control chart and the EWMA control chart are used to monitor the abnormal changes of voltage, current and frequency parameters, a 3σ detection threshold is set, and a preliminary fault alarm signal is obtained; Based on the preliminary fault alarm signal, the Dempster-Shafer evidence theory is used to calculate the basic probability assignment, and the measurement data and the protection device action information of each node are fused to obtain a multi-source information fusion result; According to the multi-source information fusion result, the decision tree algorithm is used to analyze the fault characteristics to obtain a fault type judgment result; Based on the fault type judgment result, an improved fault traveling wave analysis and impedance distance measurement method are used in combination with a support vector machine algorithm to obtain fault location information; Based on the fault location information and the dynamic network topology model, the system average interruption frequency index SAIFI, the system average interruption duration index SAIDI and the expected energy not supplied EENS are calculated to evaluate the importance level of the power interruption load, and a fault influence range evaluation result is obtained.

7. The method of claim 6, wherein, Based on the fault type judgment result, a fault transient waveform with a sampling frequency not less than 10 kHz is collected, the wave head characteristics are extracted by using wavelet transform, the wave head arrival time difference is calculated based on the GPS synchronous clock, and the traveling wave positioning result is obtained in combination with the traveling wave speed of 500 km / s; Based on the fault transient waveform, the fault point impedance is calculated, and the fault point distance is estimated according to the line unit length impedance to obtain an impedance distance measurement result; Based on the traveling wave positioning result and the impedance distance measurement result, the support vector machine algorithm is used to classify the fault feature vector, wherein the feature vector includes the voltage sag amplitude, the fault current amplitude, the wave head time difference and the impedance measurement value, the radial basis kernel function is used for feature mapping, and a preliminary fault location is obtained; Based on the preliminary fault location, the position correction is performed in combination with the dynamic network topology model, the fault current contribution and the voltage support effect of the distributed photovoltaic are considered, the weighted least squares method is used to optimize the positioning result, and the positioning accuracy is not less than ±100 meters, and the fault location information is obtained.

8. The method of claim 1, wherein, Based on the dynamic network topology model, the photovoltaic-load balance state evaluation result and the fault influence range evaluation result, the power supply reliability and economy are comprehensively considered, a multi-objective optimization algorithm is used to generate a switch operation sequence, and a self-healing control strategy is obtained, including: Based on the fault influence range evaluation result and the dynamic network topology model, a graph theory algorithm is used to determine the minimum isolation area, generate a switch operation sequence, and obtain a fault isolation scheme; Based on the fault isolation scheme and the photovoltaic-load balance state evaluation result, a multi-objective genetic algorithm is used to generate a power supply recovery candidate scheme, wherein the chromosome coding represents the switch state, the fitness function is f=w1·RR+w2·(1 / NS)+w3·(1 / PL), RR is the recovery rate, NS is the number of operation steps, and PL is the system loss, and a candidate scheme set is obtained; Based on the candidate scheme set, the feasibility of forming a temporary microgrid is evaluated for each candidate scheme, the distributed photovoltaic island operation capability and the load control flexibility are considered, and a temporary microgrid scheme set is obtained. Based on the candidate scheme set and the temporary micro-grid scheme set, the weight coefficients of power supply recovery rate, operation step number and system loss are determined as 0.5, 0.3 and 0.2 respectively by using the analytic hierarchy process, the scheme is sorted by using the fuzzy comprehensive evaluation method, and the scheme with the highest comprehensive score is selected as the optimal self-healing control strategy; Based on the optimal self-healing control strategy, a detailed switch operation time sequence is generated, and the self-healing control strategy containing specific execution instructions is obtained.

9. The method of claim 1, wherein, The switch operation is performed based on the self-healing control strategy, a temporary micro-grid or a power supply path is formed, and the execution effect is monitored and evaluated in real time to obtain a self-healing control execution result and a performance evaluation report, including: Based on the self-healing control strategy, control instructions are issued to intelligent switch devices through a power distribution automation system master station to perform switch operation, and switch operation execution state information is obtained; Based on the switch operation execution state information, active power regulation of distributed photovoltaics is performed by using a droop control method, charging and discharging of an energy storage system is controlled by using a voltage-reactive power characteristic curve, and interruptible load management is implemented based on load priority, with a control period of 100 ms, to obtain micro-grid operation state data; Based on the switch operation execution state information and the micro-grid operation state data, network operation parameters are continuously monitored, and power supply recovery effect is evaluated to obtain self-healing control effect evaluation data; Based on the self-healing control effect evaluation data, key performance indicators such as power supply recovery time, recovery rate and energy utilization efficiency are calculated to obtain a self-healing control execution result; Based on the self-healing control execution result, case similarity calculation method is used for comparative analysis with historical cases, experience knowledge including fault characteristics, disposal methods and execution effects is extracted, improved K-means algorithm is used for clustering of experience knowledge, and typical cases and disposal rules in the knowledge base are updated to obtain the performance evaluation report.

10. The method of claim 9, wherein, Based on the self-healing control effect evaluation data, respectively calculate fault detection time t d , fault isolation time t i , and power supply recovery time t r , evaluate the overall fault handling time T=t d +t i +t r whether to meet the requirement of less than 300 seconds, and obtain the time performance index; Based on the time performance indicator, the ratio RR=Pr / Po of the total amount of power supply recovery load Pr to the total amount of power outage load Po is calculated, whether the power supply recovery rate reaches the target requirement of more than 95% is evaluated, and a recovery rate indicator is obtained; Based on the recovery rate indicator, the ratio η=Pa / Pm of the actual output power Pa of the distributed photovoltaics in the temporary micro-grid to the theoretical maximum output power Pm under the current meteorological condition is calculated, whether the energy utilization level reaches the target requirement of more than 90% is evaluated, and an energy efficiency indicator is obtained; Based on the time performance indicator, the recovery rate indicator and the energy efficiency indicator, the weight coefficients are determined as 0.3, 0.4 and 0.3 respectively by using the analytic hierarchy process, the weighted score S=0.3T+0.4RR+0.3η is calculated by using the fuzzy comprehensive evaluation method, and the self-healing control execution result is obtained.

11. A device for dynamic topology identification and self-healing control of network configuration, characterized in that, It includes: A data acquisition and preprocessing module that acquires power distribution network operation data, performs time synchronization processing, data quality evaluation and data repair on the operation data, and obtains a high-quality network state data set; A dynamic topology identification module uses the high-quality network state data set to perform dynamic topology identification using a path and forest enhanced algorithm improved by a relaxation method to obtain a dynamic network topology model; A photovoltaic load evaluation module uses the high-quality network state data set and the dynamic network topology model to construct a back-carrying network model using a cyclic data generation method, evaluates distributed photovoltaic output and load conditions based on the back-carrying network model, and obtains photovoltaic-load balance state evaluation results; A fault detection and positioning module uses the high-quality network state data set and the dynamic network topology model to monitor network abnormalities in real time, uses a machine learning algorithm to assist in determining fault types, locations, and impact ranges, and obtains fault impact range evaluation results; A self-healing control strategy module uses the dynamic network topology model, the photovoltaic-load balance state evaluation results, and the fault impact range evaluation results to comprehensively consider power supply reliability and economy, uses a multi-objective optimization algorithm to generate a switch operation sequence, and obtains a self-healing control strategy; A strategy execution and verification module executes switch operations based on the self-healing control strategy, forms a temporary microgrid or adjusts a power supply path, monitors and evaluates the execution effect in real time, and obtains self-healing control execution results and performance evaluation reports.

Images