Power distribution network fault self-healing success rate prediction and strategy evaluation method, system, equipment and medium

Abstract

The invention discloses a power distribution network fault self-healing success rate prediction and strategy evaluation method, system and device and a medium, and relates to the technical field of power distribution network self-healing, and the method comprises the steps: collecting power distribution network operation data and historical self-healing case data, analyzing key factors influencing the self-healing strategy success rate after preprocessing, and constructing an influence factor network; a probability graph model is established based on the network to perform success rate prediction, a classification evaluation model is constructed to perform preliminary evaluation, a prediction result is output through multi-model integration optimization fusion, a strategy evaluation index system is constructed to perform comprehensive evaluation on a self-healing strategy, and an adaptive evaluation mechanism is introduced. Dynamically adjusting model parameters according to the network state similarity, and finally outputting an evaluation result. The power distribution network fault self-healing success rate prediction and strategy evaluation method realizes the technical spanning of power distribution network fault self-healing success rate prediction and strategy evaluation from single dimension to multiple dimensions, from static analysis to dynamic optimization, and from simple prediction to comprehensive evaluation, and significantly improves the intelligent level and decision reliability of power distribution network self-healing operation.

Description

Technical Field

[0001] This invention relates to the field of distribution network self-healing technology, specifically to methods, systems, equipment, and media for predicting the success rate of distribution network fault self-healing and evaluating strategies. Background Technology

[0002] In recent years, with the continuous advancement of smart grid construction, distribution network fault self-healing technology, as a key means to improve power supply reliability, has received widespread attention from academia and industry. Traditional distribution network fault self-healing methods mainly rely on preset rules or empirical models, using fixed logic to determine fault types and execute corresponding recovery strategies. However, with the increasing complexity of distribution network structures and the large-scale integration of distributed power sources, traditional methods have gradually shown limitations in terms of dynamic adaptability and prediction accuracy. In recent years, data-driven self-healing strategy evaluation methods have gradually emerged, such as using machine learning models to predict fault recovery success rates or using multi-attribute decision-making methods to comprehensively evaluate strategies. These methods have improved the intelligence level of self-healing systems to some extent, but still suffer from problems such as insufficient model generalization ability, single evaluation indicators, and rigid parameter adjustment mechanisms, making it difficult to cope with the changing operating states and complex fault scenarios of distribution networks.

[0003] Existing technologies still have several shortcomings in predicting the success rate of self-healing in distribution network faults and evaluating strategies. First, most methods rely on a single model for prediction, failing to fully integrate the complex relationships between multi-source heterogeneous data, thus limiting prediction accuracy. Second, strategy evaluation often focuses on single dimensions such as technical feasibility or economic efficiency, lacking comprehensive consideration of multi-dimensional indicators such as reliability and dynamic adaptability, resulting in incomplete evaluation results. Furthermore, existing model parameters are mostly statically set, making it difficult to adaptively adjust according to real-time network conditions, limiting the applicability of the models under different operating conditions. Although some studies have introduced ensemble learning methods to improve prediction performance, they often lack effective design of dynamic weight allocation mechanisms between models and fail to optimize parameters by combining historical cases and real-time conditions. Summary of the Invention

[0004] In view of the above-mentioned existing problems, the present invention provides a method, system, equipment and medium for predicting the success rate of self-healing of distribution network faults and evaluating strategies, in order to solve the problems of limited prediction accuracy, insufficient evaluation results and limited applicability of models under different operating conditions in the prior art.

[0005] To address the aforementioned technical issues, a method for predicting the success rate of self-healing in distribution network faults and evaluating strategies is proposed, including:

[0006] Data collection includes distribution network operation data and historical self-healing case data. Data preprocessing is performed, and key factors influencing the success rate of self-healing strategies are analyzed to construct an influencing factor network. Based on this network, a probabilistic graphical model is established for success rate prediction. A classification evaluation model is constructed to perform a preliminary evaluation of the self-healing strategies. The outputs of the probabilistic graphical model and the classification evaluation model are integrated, and the success rate prediction results are obtained through multi-model ensemble optimization. A strategy evaluation index system is constructed, and the self-healing strategies are comprehensively evaluated based on this system. Evaluation results are output, and an adaptive evaluation mechanism is introduced to adaptively adjust model parameters.

[0007] As a preferred embodiment of the distribution network fault self-healing success rate prediction and strategy evaluation method of the present invention, the data preprocessing includes: performing data integrity processing on the collected raw data, and filling or correcting incomplete data.

[0008] Perform data anomaly identification, filter and exclude abnormal data that deviates from the normal range;

[0009] To unify data scales and eliminate dimensional differences between data from different sources;

[0010] Perform data feature extraction to extract key feature sets from the preprocessed data.

[0011] As a preferred embodiment of the distribution network fault self-healing success rate prediction and strategy evaluation method described in this invention, the construction of the influencing factor network includes: identifying preprocessed data and constructing an influencing factor network reflecting the mutual influence relationship based on the correlation between the data.

[0012] As a preferred embodiment of the distribution network fault self-healing success rate prediction and strategy evaluation method described in this invention, the establishment of the probabilistic graphical model includes constructing a graph structure representing the probabilistic dependency relationship between variables based on the variable relationship defined by the influencing factor network.

[0013] The conditional probability distribution of each node in the graph structure is learned based on historical data;

[0014] By introducing time-dependent reasoning computation, the probability of successfully executing a self-healing strategy under a given state is output.

[0015] As a preferred embodiment of the distribution network fault self-healing success rate prediction and strategy evaluation method described in this invention, the construction of the classification evaluation model includes: using historical case data, using the C4.5 algorithm information gain ratio as the dividing criterion to generate a decision tree model, calculating the classification confidence of each node in the decision tree model based on the number of samples, and applying error complexity-based pruning techniques to the decision tree model to optimize the model structure.

[0016] The C4.5 algorithm formula is expressed as follows:

[0017]

[0018]

[0019] Where GainRatio(S,A) is the information gain ratio, Gain(S,A) is the information gain, SplitInfo(S,A) is the split information, S is the set of all training samples at the current node, A is the feature to be evaluated, and Values(A) is the set of all values ​​of feature A. ν Let |S| be a subset of samples in dataset S with feature A having value v. ν | is a subset S ν Let |S| be the total number of samples in set S, and Entropy(S) be the information entropy of dataset S. ν ) is a subset S ν The information entropy, where v is the index of the value of feature A;

[0020] The pruning technique formula is expressed as follows:

[0021] Error(T) = Error tratn (T)+α·Complexity(T)

[0022] Where Error(T) is the overall estimation error of the decision tree subtree T. tratn (T) represents the error of subtree T on the training data, Complexity(T) is the complexity penalty term, and α is the balance parameter;

[0023] The confidence level calculation formula is expressed as follows:

[0024]

[0025] Where Confidence(x) is the confidence level of the prediction result for the sample reaching leaf node a of the decision tree, a is the leaf node index of the decision tree, and N is the number of leaf nodes. c (a) represents the number of samples belonging to category c among the samples falling on leaf node a, and N(a) represents the total number of samples falling on leaf node a.

[0026] As a preferred embodiment of the distribution network fault self-healing success rate prediction and strategy evaluation method described in this invention, the multi-model integration optimization includes: adopting the Stacking integration method, using the success probability output by the probabilistic graphical model and the classification result and confidence level output by the classification evaluation model as meta-features, and using the logistic regression algorithm as the meta-learner to train and fuse the meta-features. During the training process, an adaptive weight adjustment mechanism based on prediction error is introduced to dynamically optimize the contribution weights of the probabilistic graphical model and the classification evaluation model in the final prediction result.

[0027] The Stacking integration method formula is expressed as follows:

[0028] P final (Success∣X)=Meta(P BN (Success∣X),P DT (Success∣X),X meta )

[0029] Among them, P final (Success|X) represents the final output of the ensemble learning optimization, Meta is the meta-learner, and X... meta As a meta-feature, P BN (Success|X) represents the prediction result of the Bayesian network, P DT (Success|X) represents the prediction result of the decision tree;

[0030] The formula for logistic regression is expressed as:

[0031]

[0032] Among them, P final (Success=1|X) represents the final success probability predicted by the meta-learner, e is the natural constant, β0 is the intercept of the logistic regression model, β1 is the coefficient of the Bayesian network prediction probability, β2 is the coefficient of the decision tree prediction probability, and γ i Let X be the i-th meta-feature. meta,i The coefficient, X meta,i Let m be the i-th meta-feature, and m be the total number of meta-features;

[0033] The adaptive weight adjustment mechanism is expressed as follows:

[0034]

[0035] in, For the weights of the j-th base model in the t-th iteration, Let η be the weight of the j-th base model in the (t+1)-th iteration, and η be the learning rate. Let be the prediction error of the j-th basic model in the t-th iteration, and exp be an exponential function.

[0036] As a preferred embodiment of the distribution network fault self-healing success rate prediction and strategy evaluation method described in this invention, the strategy evaluation index system includes: in the technical feasibility dimension, calculation using a weighted comprehensive approach of three indicators: operational complexity, resource requirements, and implementation time; in the economic dimension, calculation using a weighted comprehensive approach of three indicators: power outage loss reduction rate, operating cost, and equipment loss; and in the reliability dimension, calculation using a weighted comprehensive approach of three indicators: system stability improvement, power supply restoration rate, and risk factors. A comprehensive score for the strategy is obtained by weighted fusion of these three dimensions.

[0037] The adaptive evaluation mechanism includes defining a network state similarity metric, calculating the distance between state vectors and introducing a scaling parameter, adaptively adjusting model parameters based on the network state similarity metric, and designing an adaptive learning rate adjustment mechanism to dynamically update the learning rate according to the prediction error until the minimum learning rate threshold is reached.

[0038] The beneficial effects of this preferred technical solution are as follows: by adopting the Stacking ensemble method, the outputs of the probabilistic graphical model and the classification evaluation model are fused, and logistic regression is used as the meta-learner. An adaptive weight adjustment mechanism based on prediction error is introduced, which realizes the complementarity and dynamic optimization of multiple prediction advantages. The Stacking framework can capture the complex nonlinear relationship between the prediction results of the basic model and the true label. The adaptive weight mechanism can dynamically adjust its importance according to the recent performance of the model, so that the ensemble system always maintains the best performance and improves the overall accuracy and robustness of success rate prediction.

[0039] As a preferred embodiment of the power distribution network fault self-healing success rate prediction and strategy evaluation system of the present invention, it is characterized by including a data acquisition and preprocessing module, an influencing factor analysis module, a Bayesian network prediction module, a decision tree evaluation module, and an integrated learning optimization module.

[0040] The data acquisition and preprocessing module is used to collect operational data and historical fault self-healing case data in real time, clean and organize the raw data, and extract the key feature set that best represents the system state through feature extraction technology.

[0041] The influencing factor analysis module is used to quantitatively identify the intrinsic relationships between fault nature, network topology, real-time operating conditions, and strategy attributes through feature importance assessment and statistical correlation analysis, and to construct an influencing factor network.

[0042] The Bayesian network prediction module is used to construct a probabilistic graphical model by treating influencing factors as network nodes and learning conditional probability distributions. Given the current specific state of the power grid and the proposed strategy, it calculates the probability of the strategy's success.

[0043] The decision tree evaluation module is used to classify new strategies and generate confidence scores by constructing tree-like decision rules.

[0044] The ensemble learning optimization module is used to employ the Stacking ensemble learning framework to fuse the probability prediction results output by the Bayesian network with the classification confidence output by the decision tree, and to learn the optimal combination through a meta-learner to output a comprehensive success probability prediction value.

[0045] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of a method for predicting the success rate of self-healing of distribution network faults and evaluating strategies.

[0046] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of a method for predicting the success rate of self-healing of distribution network faults and evaluating strategies.

[0047] The beneficial effects of this invention are as follows: By collecting and preprocessing distribution network operation data and historical self-healing case data, this invention achieves data integrity processing, anomaly identification, scale unification, and feature extraction, providing a high-quality data foundation for subsequent modeling; the construction of an influencing factor network clarifies the mutual influence relationships among key variables, providing a structured basis for the probabilistic graphical model and enhancing the model's interpretability and relevance; based on this network, a probabilistic graphical model is established, and time-series dependency inference calculation is introduced to achieve dynamic prediction of the success probability of self-healing strategies, improving the timeliness and accuracy of prediction; a classification evaluation model is constructed, and the C4.5 algorithm and pruning techniques are used to optimize the decision tree structure. Combined with confidence calculation, a preliminary classification of self-healing strategies is achieved. The evaluation improved the model's generalization ability and classification accuracy. Through multi-model ensemble optimization, the outputs of the probabilistic graphical model and the classification evaluation model were integrated, and an adaptive weight adjustment mechanism was introduced to further improve the overall performance and stability of the prediction results. A strategy evaluation index system was constructed to comprehensively evaluate the self-healing strategy from three dimensions: technical feasibility, economy, and reliability. The final score was output by combining the prediction success rate, realizing a multi-dimensional quantitative evaluation of the strategy. An adaptive evaluation mechanism was introduced, and the model parameters were dynamically optimized through network state similarity measurement and adaptive learning rate adjustment. This enhanced the model's adaptability and robustness under different operating conditions and significantly improved the overall effect of distribution network fault self-healing success rate prediction and strategy evaluation. Attached Figure Description

[0048] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is a flowchart illustrating the overall process of a distribution network fault self-healing success rate prediction and strategy evaluation method provided in one embodiment of the present invention.

[0050] Figure 2 The flowchart of the system scheme for the distribution network fault self-healing success rate prediction and strategy evaluation system provided in one embodiment of the present invention is shown. Detailed Implementation

[0051] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0052] Example 1, referring to Figure 1 As an embodiment of the present invention, a method for predicting the success rate of self-healing of distribution network faults and evaluating strategies is provided, including:

[0053] S100: Collect power distribution network operation data and historical self-healing case data, perform data preprocessing, analyze the key factors affecting the success rate of self-healing strategies, and construct an influencing factor network.

[0054] S200: Based on the influencing factor network, a probabilistic graphical model is established to predict the success rate, a classification evaluation model is constructed to conduct a preliminary evaluation of the self-healing strategy, and the outputs of the probabilistic graphical model and the classification evaluation model are integrated to obtain the success rate prediction result through multi-model ensemble optimization.

[0055] S300: Construct a strategy evaluation index system, conduct a comprehensive evaluation of the self-healing strategy based on the strategy evaluation index system, output the evaluation results, and introduce an adaptive evaluation mechanism to adaptively adjust the model parameters.

[0056] It should be noted that high-quality input data was provided for model construction through data integrity processing, anomaly identification, scaling, and feature extraction, significantly improving the accuracy of subsequent analysis. By constructing an influencing factor network and a probabilistic graphical model, a structured expression and probabilistic reasoning of the complex relationships among multi-dimensional factors in the distribution network were achieved, enhancing the model's interpretability and its ability to handle uncertainty. The combination of decision tree models and ensemble learning ensured model interpretability while leveraging the complementary advantages and dynamic optimization of different models through stacking ensemble and adaptive weight adjustment mechanisms, significantly improving prediction accuracy and robustness. A multi-dimensional evaluation index system covering technical feasibility, economic efficiency, and reliability was established, providing decision-makers with comprehensive strategic evaluation criteria. Finally, the introduction of an adaptive adjustment mechanism based on state similarity enabled the system to dynamically optimize model parameters according to the real-time state of the power grid, ensuring the applicability and effectiveness of the method in different operating scenarios.

[0057] Example 2, refer to Figure 1 This is a second embodiment of the present invention, which provides a method for predicting the success rate of self-healing of distribution network faults and evaluating strategies, including:

[0058] In this embodiment of the application, the data preprocessing in step S100 includes steps S101 to S104:

[0059] S101: Perform data integrity processing on the collected raw data. Use an improved multiple interpolation method, which comprehensively considers the mean of similar data and the predicted value based on the regression model. Perform weighted calculation using weight coefficients determined by training on historical data to fill in the missing values ​​in the collected data.

[0060] The improved formula for multiple interpolation is expressed as follows:

[0061] X miss =τ·X mean +(1-τ)·X reg

[0062] Among them, X miss X is the estimated value of the missing value. mean X is the mean of similar data, τ is a weighting coefficient between 0 and 1, and X reg These are predicted values ​​based on a regression model.

[0063] S102: Perform data anomaly identification by using the density-based Local Outlier Factor (LOF) method to calculate the relative density of each data point in the neighborhood, compare it with a threshold to identify and exclude abnormal data points caused by momentary equipment failure or communication interference.

[0064] The density-based Local Outlier Factor (LOF) method is expressed as follows:

[0065]

[0066] Among them, LOF o (p) represents the local outlier factor of data point p, where p is the data point to be detected, and Q... o (p) is the set of o nearest neighbors of data point p, that is, the set of o points in the dataset that are closest to data point p. o (p) represents the local reachability density of data point p, lrd o (O) represents the local reachability density of the o nearest neighbors of data point p.

[0067] S103: Data scaling is standardized by using the Robust Z-score method, which uses the median and absolute deviation median to standardize the operating data of line current and node voltage. This method is not sensitive to outliers and can better preserve the overall distribution characteristics of the data.

[0068] The formula for the Robust Z-score method is expressed as follows:

[0069]

[0070] Where V is the standardized value, Y is the original data value, median(Y) is the median of dataset Y, and MAD(Y) is the median absolute deviation of dataset Y.

[0071] S104: Perform data feature extraction by combining principal component analysis (PCA) and wavelet transform. The time-frequency domain features of the data are extracted by wavelet transform, and the high-dimensional features are reduced by PCA to obtain the key feature set that best represents the system state.

[0072] The formula for combining principal component analysis (PCA) with wavelet transform is expressed as follows:

[0073] Y feature =W·PCA(WaveletTransform(Y))

[0074] Among them, Y feature Let W be the extracted feature set, W be the transformation matrix, PCA be the principal component analysis function, and Y be the original data values.

[0075] In an optional implementation, in step S100, the data preprocessing further includes: data integrity processing using the K nearest neighbor (KNN) imputation method, which imputs data based on the mean or weighted value of the K most similar complete samples around the missing data point; data anomaly identification using the 3σ criterion, which considers data points exceeding the mean ± 3 times the standard deviation as outliers and removes or smooths them based on the mean and standard deviation of the data; data scaling using the Min-Max standardization method, which linearly transforms the data to the [0,1] interval; and data feature extraction using linear discriminant analysis (LDA), which attempts to make data of the same category as close as possible and data of different categories as far apart as possible while reducing dimensionality.

[0076] In another optional implementation, in step S100, the data preprocessing may further include: data integrity processing using the expectation-maximization (EM) algorithm for imputation, estimating the maximum likelihood value of missing data through an iterative process; data anomaly identification using the Isolation Forest algorithm, quickly isolating outliers by randomly partitioning the feature space; data scaling using decimal scaling, scaling by moving the decimal point position of the data; and data feature extraction using an autoencoder, compressing data through the encoder and reconstructing data through the decoder, learning an efficient low-dimensional representation of the data.

[0077] Furthermore, in step S100, constructing the influencing factor network includes steps S111 to S112:

[0078] S111: Identify the preprocessed data and key factors affecting the success rate of the self-healing strategy. Feature importance assessment uses a permutation-based method, expressed by the following formula:

[0079]

[0080] Among them, Imp(X) h ) is the feature X h The importance of , G is the total number of samples, L is the loss function, and y is the true label. For the trained prediction model, X(g) is the original feature vector of the g-th sample. For the feature X in the g-th sample h The feature vector after random permutation.

[0081] S112: Based on the relationships between data, construct a network of influencing factors that reflects the mutual influence relationships;

[0082] Feature correlation analysis using the maximum information coefficient (MIC) method is expressed by the following formula:

[0083]

[0084] Where MIC(E,F) is the maximum information coefficient between E and F, I(E,F) is the mutual information between E and F, and n e Let n be the number of grid divisions for E. f Let F be the number of grid divisions, and C(n) be a function that controls the complexity.

[0085] Based on the analysis of influencing factors, a network of factors affecting the success rate of self-healing strategies is constructed, defined as a directed acyclic graph. The main influencing factors include: fault nature (type, location, severity), network characteristics (topology, equipment status, redundancy), operating conditions (load level, distributed power penetration rate, weather conditions), and strategy characteristics (complexity, response time, resource requirements).

[0086] In this embodiment of the application, step S200, the establishment of the probabilistic graphical model includes steps S201 to S203: wherein the probabilistic graphical model is specifically a Bayesian network.

[0087] S201: Based on the variable relationships defined in the influencing factor network, construct a graph structure representing the probabilistic dependencies between variables; specifically, the structure learning adopts a method combining the K2 algorithm and expert knowledge. Given the node order (the importance order determined by the influencing factor analysis module), the optimal directed acyclic graph structure is found by calculating the scores of different network structures (Bayesian Information Criterion BIC), and the prior knowledge of experts in the power distribution network field is introduced to modify the structure.

[0088] The formula is expressed as:

[0089] Score(H,D)=U(H∣D)=U(D∣H)U(H)

[0090] Where Score(H,D) is the posterior probability of network structure H given data D, H is the directed acyclic graph structure of Bayesian network, D is the training dataset, U(H|D) is the likelihood function, and U(H) is the prior probability of structure H.

[0091] S202: The conditional probability distribution of each node in the graph structure is learned based on historical data; the parameter learning adopts maximum a posteriori probability (MAP) estimation, and the conditional probability table (CPT) of each node in the network under different combinations of the state of its parent node is calculated using historical case data, and Dirichlet prior is introduced to avoid the zero probability problem.

[0092] The maximum a posteriori probability (MAP) estimation formula is expressed as:

[0093]

[0094] Where, ∈ qbs M is the conditional probability to be estimated.qbs For counting, θ qbs These are prior parameters.

[0095] S203: By introducing time-dependent inference calculations, the model outputs the success probability of executing a self-healing strategy under a given state. To capture dynamic characteristics, the model also introduces time-dependent terms, using the system state of the previous time step as the parent node of some nodes at the current time step, and balances the influence of historical and current information through time-series fusion weights. It also uses variable elimination to perform precise inference and calculates the posterior probability of successfully executing a specific self-healing strategy under a given observed system state.

[0096] Time-series dependencies are represented as:

[0097]

[0098] in, For variable X d The state at time t, Pa(X) d ) t For variable X d The state of the set of parent nodes at time t. For variable X d The state at time t-1 It is a synchronous dependency relationship. This represents a temporal dependency, where π is the temporal fusion weight.

[0099] Using the variable elimination method for precise reasoning, the posterior probability of successfully executing a specific self-healing strategy under a given observed system state is calculated, expressed by the formula:

[0100]

[0101] Wherein, P(Success=1|J=u) is the conditional probability, which is the posterior probability that the query variable “Success” takes the value “1” (i.e., success) given evidence that the observed variable set J takes the value u. Success is the query variable, J is the evidence variable set, u is the evidence, x is the set of all remaining unobserved variables other than Success and J, and P(Success=1,x,Success,J,J=u) is the joint probability.

[0102] In an optional implementation, in step S200, the establishment of the probabilistic graphical model further includes determining the Markov chain between nodes through a network structure learning algorithm (logistic regression based on L1 regularization) to establish an undirected graphical model. The parameter learning is performed by learning the potential function defined on the maximal clique through maximum likelihood estimation. During inference, an approximate inference algorithm such as cyclic belief propagation is used to calculate the marginal probability of the target variable.

[0103] In another optional implementation, in step S200, the establishment of the probabilistic graphical model may further include: explicitly dividing the time slices into discrete sequences, constructing transition networks between adjacent time slices and static networks within time slices, learning the initial network structure and transition network structure and parameters respectively during the learning process, and using a forward-backward algorithm for probabilistic inference of the state over the entire time series.

[0104] It should be noted that the advantage of using Bayesian networks (combining K2+MAP+temporal terms+variable elimination) lies in their powerful causal expression capabilities and ability to integrate data and knowledge. Compared to Markov networks, the directed edges of Bayesian networks can more intuitively represent the causal relationships between influencing factors, are more consistent with the physical mechanisms of power distribution networks, and have stronger reasoning interpretability. Compared to dynamic Bayesian networks, this invention approximates dynamic effects in a static network by introducing temporal dependencies, significantly reducing computational complexity while ensuring sufficient accuracy, and better meeting the stringent real-time requirements of fault handling.

[0105] Furthermore, in this embodiment of the application, in step S200, the construction of the classification evaluation model includes steps S211 to S213:

[0106] S211: Using historical case data, a decision tree model is generated using the C4.5 algorithm with information gain ratio as the splitting criterion; the C4.5 algorithm formula is expressed as:

[0107]

[0108] Where GainRatio(S,A) is the information gain ratio, Gain(S,A) is the information gain, SplitInfo(S,A) is the split information, S is the set of all training samples at the current node, A is the feature to be evaluated, and Values(A) is the set of all values ​​of feature A. ν Let |S| be a subset of samples in dataset S with feature A having value v. ν | is a subset S ν Let |S| be the total number of samples in set S, and Entropy(S) be the information entropy of dataset S. ν ) is a subset S ν The information entropy is v, where v is the index of the value of feature A.

[0109] S212: Calculate the classification confidence score of each node in the decision tree model based on the number of samples. The confidence score calculation formula is expressed as follows:

[0110]

[0111] Where Confidence(x) is the confidence level of the prediction result for the sample reaching leaf node a of the decision tree, a is the leaf node index of the decision tree, and N is the number of leaf nodes. c (a) represents the number of samples belonging to category c among the samples falling on leaf node a, and N(a) represents the total number of samples falling on leaf node a.

[0112] S213: Apply error complexity-based pruning techniques to the decision tree model to optimize the model structure;

[0113] The pruning technique formula is expressed as follows:

[0114] Error(T) = Error tratn (T)+α·Complexity(T)

[0115] Where Error(T) is the overall estimation error of the decision tree subtree T. tratn (T) represents the error of subtree T on the training data, Complexity(T) is the complexity penalty term, and α is the balancing parameter.

[0116] In an optional implementation, in step S200, the construction of the classification evaluation model further includes mapping historical case data to a high-dimensional feature space, finding an optimal hyperplane that maximizes the margin between successful and unsuccessful cases, and using a kernel function (radial basis function RBF kernel) for nonlinear mapping for nonlinear problems. During decision-making, classification is performed based on the position and distance of the sample points relative to the hyperplane, and the distance value can be converted and used as a measure of classification confidence.

[0117] In another optional implementation, in step S200, the construction of the classification evaluation model may further include generating multiple sub-training sets from the original data through Bootstrap sampling, training a decision tree (using the CART algorithm) for each sub-training set, randomly selecting some features for optimization when splitting each tree, and combining the classification results of all decision trees through a voting method, with the majority vote as the final classification, and the vote ratio can be used as the confidence level.

[0118] Furthermore, in this embodiment of the application, in step S200, the multi-model ensemble optimization includes steps S231 to S233:

[0119] S231: The Stacking ensemble method is used to combine the success probability output by the probabilistic graphical model with the classification result and confidence level output by the classification evaluation model as meta-features.

[0120] The Stacking integration method formula is expressed as follows:

[0121] Pfinal (Success∣X)=Meta(P BN (Success∣X),P DT (Success∣X),X meta )

[0122] Among them, P final (Success|X) represents the final output of the ensemble learning optimization, Meta is the meta-learner, and X... meta As a meta-feature, P BN (Success|X) represents the prediction result of the Bayesian network, P DT (Success|X) represents the prediction result of the decision tree.

[0123] S232: Using logistic regression as a meta-learner, meta-features are trained and fused. The formula for logistic regression is as follows:

[0124]

[0125] Among them, P final (Success=1|X) represents the final success probability predicted by the meta-learner, e is the natural constant, β0 is the intercept of the logistic regression model, β1 is the coefficient of the Bayesian network prediction probability, β2 is the coefficient of the decision tree prediction probability, and γ i Let X be the i-th meta-feature. meta,i The coefficient, X meta,i Let m be the i-th meta-feature, and m be the total number of meta-features.

[0126] S233: During training, an adaptive weight adjustment mechanism based on prediction error is introduced to dynamically optimize the contribution weights of the probabilistic graphical model and the classification evaluation model in the final prediction result.

[0127] The adaptive weight adjustment mechanism is expressed as follows:

[0128]

[0129] in, For the weights of the j-th base model in the t-th iteration, Let η be the weight of the j-th base model in the (t+1)-th iteration, and η be the learning rate. Let be the prediction error of the j-th basic model in the t-th iteration, and exp be an exponential function.

[0130] In an optional implementation, in step S200, the multi-model ensemble optimization further includes directly performing a linear weighted fusion of the success probability output by the Bayesian network and the confidence level output by the decision tree (as an estimate of the success probability).

[0131] In another optional implementation, in step S200, the multi-model ensemble optimization may further include, in the second layer, using a series of ensemble algorithms of Boosting to fit the relationship between meta-features and final labels, by sequentially training multiple weak learners, each new learner focusing on correcting the error of the previous one, to build an ensemble model.

[0132] In step S300, the constructed strategy evaluation index system includes: in the dimension of technical feasibility, calculation using a weighted comprehensive approach of three indicators: operational complexity, resource requirements, and implementation time; in the dimension of economic efficiency, calculation using a weighted comprehensive approach of three indicators: power outage loss reduction rate, operating costs, and equipment losses; and in the dimension of reliability, calculation using a weighted comprehensive approach of three indicators: system stability improvement, power restoration rate, and risk factors. A comprehensive score for the strategy is obtained by weighted fusion of scores from the three dimensions of technical feasibility, economic efficiency, and reliability.

[0133] The technical feasibility indicators are expressed as follows:

[0134] TF = λ1·OC + λ2·RR + λ3·IT

[0135] Where OC is operational complexity, RR is resource requirement, IT is implementation time, λ1, λ2 and λ3 are the weighting coefficients of the three indicators OC, RR and IT, and TF is the technical feasibility indicator.

[0136] The economic evaluation indicators are expressed as follows:

[0137] EC=δ1·PILR-δ2·OC-δ3·ED

[0138] Among them, PILR is the power outage loss reduction rate, OC is the operating cost, ED is the equipment loss, EC is the economic evaluation index, and δ1, δ2 and δ3 are the weighting coefficients of the three indices PILR, OC and ED.

[0139] The reliability assessment metrics are expressed as follows:

[0140] RL=μ1·SIR+μ2·PER-μ3·RF

[0141] Wherein, RL is the reliability assessment index, SIR is the stability improvement, PER is the power supply recovery rate, RF is the risk factor, and μ1, μ2 and μ3 are the weighting coefficients of the three indices SIR, PER and RF.

[0142] The comprehensive scoring formula is expressed as follows:

[0143] Score(B)=ω1·TF(B)+ω2·EC(B)+ω3·RL(B)+ω4·P(Success∣Z,B)

[0144] Wherein, Score(B) is the final comprehensive score of strategy B, ω1, ω2, ω3 and ω4 are the weight coefficients of the four dimensions of technical feasibility, economy, reliability and prediction success rate, TF(B) is the technical feasibility index of strategy B, EC(B) is the economy index of strategy B, RL(B) is the reliability index of strategy B, and P(Success|Z,B) is the prediction success probability of executing strategy B in the current state Z, which is the output of ensemble learning optimization.

[0145] Furthermore, the adaptive evaluation mechanism includes defining a network state similarity metric, calculating the distance between state vectors and introducing a scaling parameter, adaptively adjusting model parameters based on the network state similarity metric, and designing an adaptive learning rate adjustment mechanism to dynamically update the learning rate according to the prediction error until the minimum learning rate threshold is reached.

[0146] The network state similarity metric is represented as:

[0147]

[0148] Where Similarity(Z1,Z2) is the similarity between states Z1 and Z2, d(Z1,Z2) is the distance between states Z1 and Z2, σ is the scaling parameter, and exp is the exponential function;

[0149] The formula for adaptively adjusting model parameters is expressed as:

[0150]

[0151] Where, θ adapt For the adaptively adjusted parameters, θ k Here, K represents the parameters for the k-th historical case, and K is the number of nearest neighbor samples.

[0152] The formula for dynamically updating the learning rate based on the prediction error is expressed as follows:

[0153] η t =η0·exp(-γ·t)+η min

[0154] Where, η t Let η be the learning rate in the t-th iteration, η0 be the initial learning rate, and γ be the decay coefficient. min This is the minimum learning rate.

[0155] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

[0156] Example 3, referring to Figure 2 This is the third embodiment of the present invention. This embodiment provides a distribution network fault self-healing success rate prediction and strategy evaluation system, including a data acquisition and preprocessing module, an influencing factor analysis module, a Bayesian network prediction module, a decision tree evaluation module, and an integrated learning optimization module.

[0157] The data acquisition and preprocessing module is used to collect operational data and historical fault self-healing case data in real time, clean and organize the raw data, and extract the key feature set that best represents the system state through feature extraction technology.

[0158] The influencing factor analysis module is used to quantitatively identify the intrinsic relationships between fault nature, network topology, real-time operating conditions, and strategy attributes through feature importance assessment and statistical correlation analysis, and to construct an influencing factor network.

[0159] The Bayesian network prediction module is used to construct a probabilistic graphical model by treating influencing factors as network nodes and learning conditional probability distributions. Given the current specific state of the power grid and the proposed strategy, it calculates the probability of the strategy's success.

[0160] The decision tree evaluation module is used to classify new strategies and generate confidence scores by constructing tree-like decision rules.

[0161] The ensemble learning optimization module is used to employ the Stacking ensemble learning framework to fuse the probability prediction results output by the Bayesian network with the classification confidence output by the decision tree, and to learn the optimal combination through a meta-learner to output a comprehensive success probability prediction value.

[0162] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

[0163] Example 4, the fourth embodiment of the present invention, differs from the previous three embodiments in that:

[0164] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0165] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0166] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0167] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

Claims

1. A method for predicting the success rate of self-healing of distribution network faults and evaluating strategies, characterized by: include, Collect power distribution network operation data and historical self-healing case data, perform data preprocessing, analyze the key factors affecting the success rate of self-healing strategies, and construct an influencing factor network. Based on the network of influencing factors, a probabilistic graphical model is established to predict the success rate, a classification evaluation model is constructed to conduct a preliminary evaluation of the self-healing strategy, and the outputs of the probabilistic graphical model and the classification evaluation model are integrated and optimized through multi-model ensemble to obtain the success rate prediction result. A strategy evaluation index system is constructed, and the self-healing strategy is comprehensively evaluated based on the strategy evaluation index system. The evaluation results are output, and an adaptive evaluation mechanism is introduced to adaptively adjust the model parameters.

2. The method for predicting the success rate of self-healing of distribution network faults and evaluating strategies as described in claim 1, characterized in that: The data preprocessing includes performing data integrity processing on the collected raw data, filling in or correcting incomplete data; Perform data anomaly identification, filter and exclude abnormal data that deviates from the normal range; To unify data scales and eliminate dimensional differences between data from different sources; Perform data feature extraction to extract key feature sets from the preprocessed data.

3. The method for predicting the success rate of self-healing of distribution network faults and evaluating strategies as described in claim 2, characterized in that: The construction of the influencing factor network includes identifying preprocessed data and constructing an influencing factor network that reflects the mutual influence relationships based on the correlation between the data.

4. The method for predicting the success rate of self-healing of distribution network faults and evaluating strategies as described in claim 3, characterized in that: The establishment of the probabilistic graphical model includes constructing a graph structure representing the probabilistic dependencies between variables based on the variable relationships defined in the influencing factor network. The conditional probability distribution of each node in the graph structure is learned based on historical data; By introducing time-dependent reasoning computation, the probability of successfully executing a self-healing strategy under a given state is output.

5. The method for predicting the success rate of self-healing of distribution network faults and evaluating strategies as described in claim 4, characterized in that: The construction of the classification evaluation model includes: using historical case data, generating a decision tree model with the C4.5 algorithm information gain ratio as the dividing criterion, calculating the classification confidence of each node in the decision tree model based on the number of samples, and applying error complexity-based pruning techniques to the decision tree model to optimize the model structure; The C4.5 algorithm formula is expressed as follows: Where GainRatio(S,A) is the information gain ratio, Gain(S,A) is the information gain, SplitInfo(S,A) is the split information, S is the set of all training samples at the current node, A is the feature to be evaluated, and Values(A) is the set of all values ​​of feature A. ν Let |S| be a subset of samples in dataset S with feature A having value v. ν | is a subset S ν Let |S| be the total number of samples in set S, and Entropy(S) be the information entropy of dataset S. ν ) is a subset S ν The information entropy, where v is the index of the value of feature A; The pruning technique formula is expressed as follows: Error(T) = Error tratn (T) + a - Complexity(T) where Error(T) is the overall estimated error of the decision tree sub-tree T, Error tratn (T) is the error of sub-tree T on the training data, Complexity(T) is the complexity penalty term, and a is a balancing parameter; The confidence level calculation formula is expressed as follows: where Confidence(x) is the confidence of the prediction result for the sample reaching the leaf node a of the decision tree, a is the leaf node index of the decision tree, N c (a) is the number of samples belonging to class c among the samples falling into the leaf node a, N(a) is the total number of samples falling into the leaf node a.

6. The method for predicting the success rate of self-healing of distribution network faults and evaluating strategies as described in claim 5, characterized in that: The multi-model ensemble optimization includes using the Stacking ensemble method, taking the success probability output by the probabilistic graphical model and the classification result and confidence level output by the classification evaluation model as meta-features, and using the logistic regression algorithm as the meta-learner to train and fuse the meta-features. During the training process, an adaptive weight adjustment mechanism based on prediction error is introduced to dynamically optimize the contribution weights of the probabilistic graphical model and the classification evaluation model in the final prediction result. The Stacking integration method formula is expressed as follows: P final (Success∣X)=Meta(P BN (Success∣X),P DT (Success∣X),X meta ) where P final (Success | X) is the final output of ensemble learning optimization, Meta is a meta-learner, X meta is a meta-feature, P BN (Success | X) is the prediction result of the Bayesian network, P DT (Success | X) is the prediction result of the decision tree; The formula for logistic regression is expressed as: where P final (Success = 1 | X) is the final success probability predicted by the meta-learner, e is a natural constant, β0is the intercept of the logistic regression model, β1is the coefficient of the Bayesian network prediction probability, β2is the coefficient of the decision tree prediction probability, γ i is the coefficient of the i-th meta-feature X meta,i , X meta,i is the i-th meta-feature, and m is the total number of meta-features. The adaptive weight adjustment mechanism is expressed as follows: in, For the weights of the j-th base model in the t-th iteration, Let η be the weight of the j-th base model in the (t+1)-th iteration, and η be the learning rate. Let be the prediction error of the j-th basic model in the t-th iteration, and exp be an exponential function.

7. The method for predicting the success rate of self-healing of distribution network faults and evaluating strategies as described in claim 6, characterized in that: The constructed strategy evaluation index system includes the following: in terms of technical feasibility, it is calculated using a weighted comprehensive approach of three indicators: operational complexity, resource requirements, and implementation time; in terms of economic efficiency, it is calculated using a weighted comprehensive approach of three indicators: power outage loss reduction rate, operating costs, and equipment losses; in terms of reliability, it is calculated using a weighted comprehensive approach of three indicators: system stability improvement, power restoration rate, and risk factors; and a comprehensive score for the strategy is obtained by weighted integration of the scores from these three dimensions. The adaptive evaluation mechanism includes defining a network state similarity metric, calculating the distance between state vectors and introducing a scaling parameter, adaptively adjusting model parameters based on the network state similarity metric, and designing an adaptive learning rate adjustment mechanism to dynamically update the learning rate according to the prediction error until the minimum learning rate threshold is reached.

8. A distribution network fault self-healing success rate prediction and strategy evaluation system, employing the distribution network fault self-healing success rate prediction and strategy evaluation method as described in any one of claims 1 to 7, characterized in that, It includes modules for data acquisition and preprocessing, influencing factor analysis, Bayesian network prediction, decision tree evaluation, and ensemble learning optimization. The data acquisition and preprocessing module is used to collect operational data and historical fault self-healing case data in real time, clean and organize the raw data, and extract the key feature set that best represents the system state through feature extraction technology. The influencing factor analysis module is used to quantitatively identify the intrinsic relationships between fault nature, network topology, real-time operating conditions and strategy attributes through feature importance assessment and statistical correlation analysis, and to construct an influencing factor network. The Bayesian network prediction module is used to construct a probabilistic graphical model by treating influencing factors as network nodes and learning conditional probability distributions. Given the current specific state of the power grid and the proposed strategy, it calculates the probability of the strategy's success. The decision tree evaluation module is used to classify new strategies and generate confidence scores by constructing tree-like decision rules. The ensemble learning optimization module is used to employ the Stacking ensemble learning framework to fuse the probability prediction results output by the Bayesian network with the classification confidence output by the decision tree, and to learn the optimal combination through a meta-learner to output a comprehensive success probability prediction value.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the distribution network fault self-healing success rate prediction and strategy evaluation method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the distribution network fault self-healing success rate prediction and strategy evaluation method as described in any one of claims 1 to 7.

Images