A power distribution network data distortion identification and repair method and system

Abstract

The application discloses a power distribution network data distortion identification and repair method and system, relates to the technical field of power distribution network data quality management, and comprises the following steps: firstly, the time series of voltage, current, active power and reactive power are denoised, aligned and interpolated; distortion such as drift, spike and loss is identified by combining statistical screening, machine learning and deep network; the abnormal section is respectively reconstructed by using Bayesian, GDBT and time sequence neural network, and the repair value is obtained by adaptively weighting according to historical error, type adaptation, feature correlation and working condition adaptability; and the feasibility check is carried out by introducing constraints such as power flow consistency, voltage range and line thermal limit. The method can significantly improve the detection rate and positioning accuracy, inhibit false repair, ensure that the result meets the physical law and equipment boundary, is still stable in the scene of sparse observation, noise and working condition switching, reduces manual review and outage risk, and is convenient for online landing.

Description

Technical Field

[0001] This invention relates to the field of power distribution network data quality management technology, and in particular to a method and system for identifying and repairing power distribution network data distortion. Background Technology

[0002] Data distortion identification and repair technology for power distribution networks plays a crucial role in today's information security field. With the rapid development of information technology, users' demands for file security and privacy protection are increasing daily. Traditional encryption methods often rely on fixed keys and algorithms, resulting in complex key management and limited resistance to attacks, failing to meet users' needs for personalized and intelligent encryption. Therefore, how to utilize advanced technologies to improve the intelligence and security of data distortion identification and repair in power distribution networks has become one of the urgent problems to be solved.

[0003] In the field of data distortion identification and repair in power distribution networks, traditional encryption methods mainly rely on symmetric encryption algorithms and public-key encryption algorithms, such as AES, DES, and RSA. These algorithms ensure data confidentiality to a certain extent, but they are difficult to cope with dynamic changes in user behavior and diverse needs. At the same time, due to the fixed nature of encryption algorithms and key management, there are security risks such as key leakage and insufficient encryption strength. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a method for identifying and repairing data distortion in power distribution networks to solve the problem of unstable identification of distortions such as drift, spikes, and missing data in power distribution network monitoring data.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for identifying and repairing data distortion in a power distribution network, comprising: Acquire time-series data of the distribution network and perform preprocessing.

[0007] In the time-series data, a three-stage joint identification architecture is used to identify distorted data.

[0008] The distorted data is repaired through a voting-weighted repair mechanism.

[0009] As a preferred embodiment of the distribution network data distortion identification and repair method of the present invention, the preprocessing includes denoising, deduplication, time alignment, and missing interpolation of the time series of voltage, current, active power, and reactive power to obtain the time series to be measured. .

[0010] As a preferred embodiment of the distribution network data distortion identification and repair method of the present invention, the three-stage joint identification architecture includes a statistical screening stage, a random forest classification stage, and a deep learning fine classification stage.

[0011] The statistical screening stage includes: outlier screening, time-series pattern screening, and dynamic amplitude screening.

[0012] The outlier screening includes removing outliers from the time series under test according to Z-score and IQR rules, and removing samples with abnormal amplitudes.

[0013] The time series pattern screening includes calculating the autocorrelation coefficient of the time series to be tested on a preset lag set and comparing it with a reference band obtained based on historical data; when the autocorrelation coefficient of any lag deviates from the reference band and the deviation exceeds a preset threshold, it is determined to be a time series anomaly and marked.

[0014] The dynamic amplitude screening includes calculating the time series to be tested within a sliding window indexed by time t. The mean and standard deviation are used to define a dynamic threshold: in, Indicates time The dynamic threshold, Indicates time The mean, Indicates time standard deviation Indicates time The dynamic coefficient.

[0015] when If the value is abnormal, it is determined to be an abnormal amplitude and deleted; otherwise, it is retained for subsequent processing. As a preferred embodiment of the distribution network data distortion identification and repair method of the present invention, the random forest classification stage includes: extracting time-domain features, frequency-domain features and relational features from the time series data to be tested, inputting them into a hierarchical random forest model to output normal and distorted data and classifying distorted data.

[0016] The hierarchical random forest model consists of two layers: the first layer distinguishes between normal data and distorted data; and the second layer classifies distorted data by type.

[0017] The deep learning fine-classification stage includes constructing a hybrid neural network architecture: in, This indicates categorized output. Describes the multilayer perceptron function. This represents the attention mechanism function. This represents the function of a convolutional neural network. This represents the Long Short-Term Memory (LSTM) network function. Representing spatial feature data, Represents time-series characteristic data.

[0018] The classification output is computed with attention enhancement based on a hybrid neural network: in, Represents the query matrix. Represents the key matrix. Represents a value matrix, This represents the dimension of the key vector.

[0019] The attention-enhanced feature representation is input into the classifier to obtain the score vector for each distortion category; the category corresponding to the component with the largest value in the score vector is used as the judgment result.

[0020] As a preferred embodiment of the distribution network data distortion identification and repair method of the present invention, the voting weighted repair mechanism includes weighted repair based on Bayesian inference, repair based on gradient boosting decision tree, and repair based on recurrent neural network.

[0021] The Bayesian inference-based repair includes constructing a data probability distribution model based on Bayesian theory: in, Indicates observed data Real data under conditions The posterior probability, Represents the likelihood function. This represents the prior probability.

[0022] When the posterior distribution is obtained When the position is judged to be distorted, the representative value is directly taken from the posterior distribution as the repair value.

[0023] The gradient boosting decision tree-based repair includes constructing a gradient boosting decision tree model: in, Indicates the first The model of the next iteration Indicates the first The model of the next iteration Indicates the learning rate. Indicates the first A decision tree.

[0024] The context sequence of the distorted position is input into the trained gradient boosting decision tree model, and the regression prediction value output by the model is used as the repair value; the context sequence of the distorted position is an ordered sequence formed by arranging adjacent observations within a preset window on both sides of the distorted position in chronological order, combined with the lagged observations of the same period.

[0025] The recurrent neural network-based repair includes constructing a recurrent neural network model; introducing an attention-enhanced sequence-to-sequence model, wherein the decoder of the sequence-to-sequence model outputs the context vector at the i-th time step as follows: in, This represents the context vector at time i in the output sequence; This represents the attention weight of the output sequence at time i to the input sequence at time j; This represents the hidden state at time j of the input sequence; Indicates the length of the input sequence.

[0026] The context sequence of the distorted location is input into the sequence model, and the decoder obtains the reconstructed value of the time series under test as the repair value based on the context sequence.

[0027] As a preferred embodiment of the distribution network data distortion identification and repair method of the present invention, the voting weighted repair mechanism includes fusing three repair results through an adaptive weight formula to obtain the final repair result.

[0028] The adaptive weighting formula: in, Indicates the first The weight of each repair This represents the weight adjustment parameter. Indicates the first Historical errors of each repair module; This represents the index variable, ranging from 1 to 3.

[0029] Calculation fusion results: in, This indicates the final repair result. Indicates the first Repair results of each repair module; Indicates the first The weight of each repair.

[0030] As a preferred embodiment of the distribution network data distortion identification and repair method of the present invention, the voting weighted repair mechanism includes subjecting the repair results to domain knowledge constraints during the repair process. Based on the AC power flow relationship, power balance constraints are applied to the calculated power and target power of each node.

[0031] Violations of power balance, permissible voltage range, and line thermal limits are assessed. Violation of any one of these criteria results in non-compliance with physical constraints and is treated as unqualified.

[0032] Based on the engineering boundaries of the equipment, the candidate results are limited or prohibited.

[0033] Set the sequence, thresholds, and consistency checks based on operating procedures and experience.

[0034] Secondly, the present invention provides a distribution network data distortion identification and repair system, including a preprocessing module for preprocessing distribution network data.

[0035] Identification Module: Construct a three-stage joint identification architecture to identify distorted data in the distribution network.

[0036] Weighted module: Uses a voting weighted repair mechanism to repair the identified distorted data.

[0037] Constraint Module: Ensures the physical rationality of the repair results through the fusion of domain knowledge constraints.

[0038] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the distribution network data distortion identification and repair method as described in the first aspect of the present invention.

[0039] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the distribution network data distortion identification and repair method as described in the first aspect of the present invention.

[0040] The beneficial effects of this invention are as follows: This invention significantly improves the accuracy and stability of identifying distortions such as drift, spikes, and missing values ​​in distribution network monitoring data through statistical screening, machine learning, and cascaded identification and multi-model adaptive fusion repair using deep networks; it dynamically allocates weights based on historical errors, type adaptation, feature correlation, and operating condition adaptability to avoid single-model failure; it incorporates power balance, voltage range, and line thermal limits as physical and engineering soft constraints into the optimization, ensuring that the repair results meet the physical laws of the power grid and the equipment capability boundaries; it remains robust under conditions of sparse observations, strong noise, and operating condition switching, reducing false repairs and jitter; the computational load of the process is controllable, it supports batch and streaming online deployment, and it outputs distortion type, confidence level, and constraint residual information for easy auditing and traceability; it is compatible with multi-source heterogeneous data, significantly reducing manual review costs and improving governance efficiency. Attached Figure Description

[0041] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the 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.

[0042] Figure 1 This is a flowchart of a method for identifying and repairing data distortion in power distribution networks.

[0043] Figure 2 A flowchart of a three-stage joint identification architecture for a data distortion identification and repair method for power distribution networks.

[0044] Figure 3 A schematic diagram of a voting-weighted repair mechanism for data distortion identification and repair methods in power distribution networks.

[0045] Figure 4 A diagram illustrating the knowledge constraint fusion mechanism for methods of identifying and repairing data distortion in power distribution networks.

[0046] Figure 5 This is a performance comparison chart of the data distortion identification and repair method in a real-world distribution network scenario. Detailed Implementation

[0047] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0048] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0049] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0050] Example 1, referring to Figures 1-4 This is one embodiment of the present invention, which provides a method for identifying and repairing data distortion in a power distribution network, comprising the following steps: S1: Preprocess the distribution network data.

[0051] The system collects operational data such as voltage, current, and power from the distribution network monitoring system, along with relevant load characteristics, weather conditions, and system event information. The raw data typically contains data streams with different time resolutions and measurement units, requiring unified processing. For voltage data, the three-phase voltage amplitude and phase angle are recorded; for current data, the three-phase current amplitude and phase angle are recorded; and for power data, active and reactive power are recorded. Each data point includes a timestamp, measurement point ID, data value, and data quality identifier. The time resolution of data acquisition varies depending on the monitoring point type: 1 second for main transformer monitoring points, 5 seconds for distribution line monitoring points, and 15 seconds for end-user monitoring points.

[0052] For obvious errors generated during the data acquisition process, such as null values ​​due to communication interruptions or zero or maximum values ​​due to sensor malfunctions, preliminary cleaning is performed. A sliding window mid-range filter is used to remove instantaneous spike noise. The window size is dynamically set according to the data type: 5 sampling points for voltage data, 7 sampling points for current data, and 9 sampling points for power data. For missing data, different strategies are adopted based on the missing percentage: linear interpolation is used for missing data below 5%; similar day pattern filling is used for missing data between 5% and 20%; and missing data exceeding 20% ​​is marked as invalid and excluded from subsequent analysis.

[0053] To eliminate the influence of different measurement dimensions, the cleaned data is normalized. For voltage data, a linear normalization method is used: in, This represents the normalized data value. Represents the original data value. This represents the minimum value of this type of data. This indicates the maximum value of this type of data.

[0054] For current and power data, considering their distribution characteristics, the Z-score normalization method is adopted: in, This represents the standardized data value. Represents the original data value. This represents the mean of this type of data. This represents the standard deviation of this type of data.

[0055] To standardize the temporal resolution of data from different monitoring points, all data were resampled and converted into time series with 5-minute intervals. For data with sampling rates higher than 5 minutes, downsampling was performed using the averaging method; for data with sampling rates lower than 5 minutes, upsampling was performed using cubic spline interpolation. The resampling process ensured the continuity and smoothness of the time series, providing a unified time benchmark for subsequent analysis.

[0056] Features for distortion identification are extracted from the preprocessed data. Time-domain features include statistics such as mean, variance, kurtosis, skewness, range, and interquartile range; frequency-domain features include dominant frequency components, harmonic content, and spectral energy distribution extracted via Fast Fourier Transform (FFT); and relational features include Pearson correlation coefficient with relevant measurement points, mutual information, and Granger causality. Based on the characteristics of the power system, power-specific features such as three-phase imbalance, power factor, and load curve characteristics are also extracted. All features are combined to form a feature vector, which serves as input to the subsequent model.

[0057] S2: Identify distorted data in the distribution network through a three-stage joint identification architecture.

[0058] like Figure 2 As shown, the three-stage cascaded identification architecture gradually improves the recognition accuracy from simple to complex, while avoiding unnecessary waste of computing resources.

[0059] Statistical screening stage: This stage uses statistical principles to quickly screen the data and identify obvious outliers, serving as the first line of defense against distortion.

[0060] Dynamic adaptive threshold setting. Based on historical data from the same period, combined with current load levels, weather conditions, and system event information, the abnormal threshold for each time period is dynamically calculated: in, Indicates time The dynamic threshold, Indicates time The mean, Indicates time standard deviation Indicates time The dynamic coefficient, This represents the base coefficient (usually set to 3). This represents an adaptive function related to load, weather, and events. The calculation of the adaptive function takes into account the deviation of the current load level from the same period in history, the influence factor of weather temperature on the load, and the influence weight of known events in the system (such as equipment maintenance).

[0061] Different statistical methods are used for different types of data. For voltage data, based on the power system steady-state operation standard, a formula for calculating the voltage qualification rate is set: in, Indicates the voltage qualification rate. This represents the number of samples where the voltage is within the allowable range. This indicates the total number of samples. The permissible voltage deviation is typically ±7% of the rated value.

[0062] For current data, considering load fluctuation characteristics, a segmented statistical boundary based on load intervals is adopted: in, Indicates the lower and upper limits of the current. Indicates load range The mean value of the internal current. Indicates load range The standard deviation of the internal current Indicates load range The coefficient is 3.5 for light load, 3.0 for medium load, and 2.5 for heavy load.

[0063] For the power data, considering power balance constraints, calculate the power deviation index: in, Indicates power deviation. Indicates the measured power value. This represents the theoretical power value based on power flow calculations. Indicates the normalized power deviation. Indicates the reference power value. When Data exceeding a preset threshold (usually 5%) is marked as suspicious.

[0064] To analyze the inherent dependencies of time series data, considering the time-series characteristics of distribution network data, an autocorrelation function is used. in, Indicates lag The autocorrelation coefficient of the period, Indicates time Data values, Represents the mean of a time series. Indicates the sequence length.

[0065] Time series anomalies can be identified by comparing the deviations of actual autocorrelation coefficients with historical patterns. in, Indicates the deviation of the autocorrelation coefficient. This represents the autocorrelation coefficient for the same historical period. Indicates a timing anomaly flag. This represents the autocorrelation bias threshold (usually set to 0.15). Indicates an indicator function.

[0066] Based on the results of the above statistical methods, an anomaly score is calculated for each data point: in, The total score represents the degree of abnormality. Indicates the score of the threshold method. Indicates the score for type-specific methods. This represents the score obtained using the autocorrelation method. , , These represent the weighting coefficients for each method (typically set to 0.5, 0.3, and 0.2). Based on the anomaly score, the data is labeled into three categories: normal data (…). Suspicious data ) and clearly abnormal data ( Abnormal data will proceed directly to the repair phase, suspicious data will proceed to the next stage for further analysis, and normal data will remain unchanged.

[0067] Random Forest Classification Stage: This stage mainly processes data marked as suspicious in the statistical screening stage, and further refines the identification of distortion types through machine learning methods.

[0068] An enhanced feature set is constructed for suspicious data, including: original features (time-domain features, frequency-domain features, and relational features) and derived features (historical deviation features, environmentally relevant features, and equipment status features). Feature importance is calculated using the information gain ratio. in, Representation of features Classification results Information gain Entropy represents the classification result. Representing a given feature Conditional entropy of classification results under given conditions Representation of features Information gain ratio, Representation of features The entropy is calculated. Features are ranked according to their information gain ratio, and the top 80% of important features are selected for model training.

[0069] Hierarchical Random Forest Model Construction: A two-layer random forest classifier is constructed: the first layer distinguishes between normal and distorted data, and the second layer classifies distorted data by type. The first-layer random forest model configuration includes 100 decision trees, a maximum depth of 15 per tree, a feature random selection ratio of 0.7, a sample random selection ratio of 0.8, and uses the Gini coefficient as the splitting criterion. The second-layer random forest model further subdivides different types of distortion (including zero drift, gain error, random noise, communication interruption, sensor failure, equipment malfunction, etc.). The model configuration includes 150 decision trees, a maximum depth of 20 per tree, a feature random selection ratio of 0.6, a sample random selection ratio of 0.7, and uses entropy as the splitting criterion.

[0070] Feature importance adaptive mechanism implementation: Feature weights are dynamically adjusted based on historical classification results. The weight update formula is as follows: in, Representation of features The weight, This represents the weight update coefficient (usually set to 0.8). Representation of features The importance score is obtained through feature importance evaluation using a random forest model. After each classification task, the contribution of each feature is evaluated based on the accuracy of the classification results, the feature weights are updated, and the updated weights are applied in the next round of classification. The initial values ​​of the feature weights are set based on domain knowledge: voltage-related features have an initial weight of 0.8, current-related features have an initial weight of 0.7, power-related features have an initial weight of 0.75, and relationship features have an initial weight of 0.6.

[0071] Imbalanced Sample Handling: To address the imbalanced distribution of different types of distortion in distribution network data, the SMOTE (Synthetic Minority Over-sampling Technique) oversampling technique is used to generate minority class samples. The basic steps of the SMOTE algorithm are: For each sample in the minority class sample set... Calculate its value in the feature space. The nearest neighbors ( (Usually set to 5); randomly select one of these neighbors, denoted as... Generate a random number between 0 and 1. Create new synthetic samples For severely imbalanced categories (such as sensor failures, which typically account for less than 1%), the number of synthesized samples generated is 5 times that of the original samples; for moderately imbalanced categories (such as communication interruptions, which typically account for 1%-5%), the number of synthesized samples generated is 3 times that of the original samples; and for slightly imbalanced categories (such as zero drift, which typically account for 5%-10%), the number of synthesized samples generated is 1 times that of the original samples.

[0072] Weighted loss function design: To further balance classification performance, a weighted loss function is used during the training of the random forest. in, This represents the weighted loss function. Indicates sample The weight of the category Indicates sample The original loss function, Indicates sample The true label, Indicates sample The predicted label. The class weight is inversely proportional to the proportion of that class in the training set: ,in Represents the total number of samples. Indicate category The number of samples, This indicates the total number of categories.

[0073] Classification results output and confidence assessment: For each data point, the random forest model outputs the probability distribution of the distortion type and calculates the classification confidence score. in, Indicates the first Classification confidence of each data point This indicates the number of votes for the category that received the most votes. This represents the total number of decision trees in the random forest. The classification threshold is set based on confidence level: high confidence (…). The classification results of ) are directly accepted, with a medium confidence level ( The classification results are marked as requiring further confirmation, with low confidence ( The classification results will then be used for the next stage of deep learning fine classification.

[0074] Deep learning fine classification stage: This stage deals with complex distorted patterns that cannot be classified with high confidence in the random forest stage, and uses deep learning methods for fine recognition.

[0075] Hybrid neural network architecture design: such as Figure 2 As shown, a hybrid neural network architecture integrating CNN, LSTM, and attention mechanisms is constructed to collaboratively extract spatial and temporal features. The network structure expression is: in, This indicates categorized output. Describes the multilayer perceptron function. This represents the attention mechanism function. This represents the function of a convolutional neural network. This represents the Long Short-Term Memory (LSTM) network function. Representing spatial feature data, Represents time-series characteristic data.

[0076] The CNN spatial feature extraction module employs a three-layer convolutional structure: the first layer uses 32 3×3 convolutional kernels with a stride of 1 and padding of 1, followed by ReLU activation and 2×2 max pooling; the second layer uses 64 3×3 convolutional kernels with a stride of 1 and padding of 1, followed by ReLU activation and 2×2 max pooling; the third layer uses 128 3×3 convolutional kernels with a stride of 1 and padding of 1, followed by ReLU activation. Batch normalization is applied after each convolutional layer, and residual connections are introduced to alleviate the vanishing gradient problem. The CNN module primarily processes the spatial correlation features between multiple measurement points in the power distribution network, capturing the topological relationships and mutual influences between nodes in the power system.

[0077] LSTM Temporal Feature Extraction Module: This module adopts a bidirectional LSTM structure, consisting of three stacked layers: the first layer uses 128 hidden units followed by Dropout (probability 0.3); the second layer uses 96 hidden units followed by Dropout (probability 0.3); and the third layer uses 64 hidden units. The calculation formulas for the forget gate, input gate, and output gate of the LSTM unit are as follows: in, , , These represent the forget gate, input gate, and output gate, respectively. This represents the sigmoid activation function. , , , These represent the corresponding weight matrices. , , , These represent the corresponding bias vectors. This indicates the hidden state at the previous moment. This represents the input at the current moment. Indicates the cell state. Indicates the candidate cell state. This indicates the hidden state at the current moment. This represents the element-wise multiplication operation. The LSTM module primarily processes the time-series dependencies of distribution network data, capturing time-series patterns such as power system load changes and equipment operating state transitions.

[0078] Feature fusion enhanced by attention mechanisms: Self-attention and cross-attention mechanisms are employed to enhance feature extraction capabilities. Self-attention mechanism calculation formula: in, Represents the query matrix. Represents the key matrix. Represents a value matrix, The dimension representing the key. This represents the softmax function. A cross-attention mechanism is used in the fusion of CNN and LSTM features. in, This represents the feature representation after fusion. This represents the features extracted by the CNN. This represents the features extracted by LSTM. This represents the feature dimension. The attention mechanism enables the model to focus on the most relevant features, improving classification accuracy.

[0079] Multilayer Perceptron Classifier: Based on fused feature representations, a multilayer perceptron is constructed for final classification: the first layer contains 256 neurons using the ReLU activation function; the second layer contains 128 neurons using the ReLU activation function; the third layer contains 64 neurons using the ReLU activation function; the output layer has the same number of neurons as the number of distortion types and uses the softmax activation function. The classification loss function is cross-entropy. in, Represents cross-entropy loss, Indicates the number of categories. The first one representing the true label One element (one-hot encoded) The first digit representing the predicted probability Each element.

[0080] Implementation of the adversarial training strategy: To improve the model's generalization ability, an adversarial training method is adopted. First, adversarial examples are generated: in, Indicates adversarial examples, Represents the original sample. This indicates the size of the perturbation (usually set between 0.01 and 0.05). Represents a symbolic function. This represents the gradient of the loss function with respect to the input. The model represents the input The prediction This represents the true label. Then, the model is trained using both the original samples and adversarial samples, with the objective function being: in, Represents the total loss function. This represents the weighting coefficient (usually set to 0.6). Represents the loss of the original sample. This represents the loss from adversarial examples. Adversarial training improves the model's robustness to small changes in the data, effectively preventing overfitting.

[0081] Domain Adaptation Mechanism Introduction: To address the differences in data distribution across different distribution network areas, a domain adaptation mechanism is introduced. in, Indicates the loss of the domain discriminator. Indicates the number of fields. Indicates the real domain tag, Indicates the predicted domain label. Represents the final loss function. The domain adaptation weight coefficients are typically increased from 0 to 0.1. By training the model to simultaneously minimize classification error and maximize domain discrimination error, the extracted features are made domain invariant, improving the model's generalization ability in different power distribution network environments.

[0082] Distortion type and severity output: The model ultimately outputs a distortion type and severity assessment for each data point. The distortion type is determined by the maximum probability class in the softmax layer. in, Indicates the type of distortion in the prediction. This represents the logits value of the output layer. Indicates the first The softmax probability of each category. The severity of distortion is calculated based on the probability values ​​and feature bias: in, Indicates the degree of distortion (between 0 and 1). This represents the probability value of the predicted category. This represents the characteristic bias after normalization. This represents the weighting coefficient (usually set to 0.7). Distortion severity is categorized into three levels: mild (…). ), moderate ( ) and severe ( ).

[0083] S3: Repair the identified distorted data through a voting-weighted repair mechanism.

[0084] like Figure 3 As shown, this mechanism leverages the advantages of different algorithms to repair the identified distorted data and obtains the optimal repair result through adaptive weight fusion.

[0085] The method of repairing the identified distorted data through a voting weighted repair mechanism includes weighted repair based on Bayesian inference, repair based on gradient boosting decision trees, and repair based on recurrent neural networks.

[0086] The Bayesian inference-based repair includes constructing a data probability distribution model based on Bayesian theory: in, Indicates observed data Real data under conditions The posterior probability, Represents the likelihood function. This represents the prior probability.

[0087] When the posterior distribution is obtained When the position is judged to be distorted, the representative value is directly taken from the posterior distribution as the repair value.

[0088] The gradient boosting decision tree-based repair includes constructing a gradient boosting decision tree model: in, Indicates the first The model of the next iteration Indicates the first The model of the next iteration Indicates the learning rate. Indicates the first A decision tree.

[0089] The context sequence of the distorted position is input into the trained gradient boosting decision tree model, and the regression prediction value output by the model is used as the repair value; the context sequence of the distorted position is an ordered sequence formed by arranging adjacent observations within a preset window on both sides of the distorted position in chronological order, combined with the lagged observations of the same period.

[0090] The recurrent neural network-based repair includes constructing a recurrent neural network model and introducing an attention-enhanced sequence-to-sequence model.

[0091] in, Represents the context vector. Indicates attention weights. Indicates the encoder is hidden; Indicates the length of the input sequence.

[0092] The context sequence of the distorted location is input into the sequence model, and the decoder obtains the reconstructed value of the time series under test as the repair value based on the context sequence.

[0093] As a preferred embodiment of the distribution network data distortion identification and repair method of the present invention, the voting weighted repair mechanism includes fusing three repair results through an adaptive weight formula to obtain the final repair result.

[0094] The adaptive weighting formula: in, Indicates the first The weight of each repair This represents the weight adjustment parameter. Indicates the first Historical errors of each repair module; This represents the index variable, ranging from 1 to 3.

[0095] Calculation fusion results: in, This indicates the final repair result. Indicates the first Repair results of each repair module; Indicates the first The weight of each repair.

[0096] S4: Ensure the physical rationality of the repair results through domain knowledge constraint fusion.

[0097] like Figure 4 The constraint fusion shown includes three key components: power system physical model constraints, equipment characteristic constraints, and expert rule constraints.

[0098] The physical model constraints of the power system incorporate fundamental physical laws such as power flow equations, power balance, and voltage constraints. in, Represents a node active power, Represents a node reactive power, Represents a node voltage amplitude, Denotes the real part of the admittance matrix. Represents the imaginary part of the admittance matrix. Represents a node With nodes The phase angle difference between them.

[0099] Considering the computational complexity of traditional power flow calculations in distribution networks, this invention innovatively designs a lightweight distributed power flow model that adapts to the radial topology of distribution networks and takes into account the uncertainties of distributed energy sources and loads. Simultaneously, a physical constraint softening mechanism is introduced. in, Represents the physical constraint loss function. , , This represents the weighting coefficient of each constraint term. This represents the power balance constraint loss. Indicates voltage constraint loss. This represents the thermal limit constraint loss.

[0100] The equipment characteristic constraints take into account the equipment operation laws, such as transformer thermal characteristic model, line load characteristics, and protection device operation characteristics. An innovative equipment aging characteristic model is introduced to consider the impact of equipment lifespan on parameters, and multi-level equipment collaborative constraints are established. Through the cascading relationship diagram between equipment and the equipment state propagation model, it is ensured that the repair results conform to the mutual influence laws between equipment.

[0101] The expert rule constraint system establishes a rule base based on power system expertise and innovatively designs a fuzzy rule reasoning system. It models uncertainty through fuzzy membership functions and triggers rules progressively based on distortion levels. Simultaneously, an automatic rule extraction and optimization mechanism is introduced, mining rules from historical data to evaluate rule effectiveness and dynamically update the rule base.

[0102] This embodiment also provides a distribution network data distortion identification and repair system, including: a preprocessing module for preprocessing distribution network data.

[0103] Identification module: Identifies distorted data in the distribution network through a three-stage joint identification architecture.

[0104] Weighted module: Repairs the identified distorted data through a voting weighted repair mechanism.

[0105] Constraint Module: Ensures the physical rationality of the repair results through the fusion of domain knowledge constraints.

[0106] This embodiment also provides a computer device applicable to the method for identifying and repairing data distortion in power distribution networks, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the method for identifying and repairing data distortion in power distribution networks as proposed in the above embodiment.

[0107] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0108] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the method for identifying and repairing data distortion in a power distribution network as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0109] Example 2, refer to Figure 5 This invention provides a method for identifying and repairing data distortion in power distribution networks, as one embodiment of the present invention. To verify the beneficial effects of the invention, scientific demonstration is conducted through economic benefit calculations and simulation experiments.

[0110] Experimental environment: This embodiment uses actual operational data from a municipal power distribution network for testing and verification. This network comprises 23 substations, 218 lines, and 1567 distribution sites. The experimental dataset covers time-series data of various power parameters, including voltage, current, power, frequency, and phase angle, for the entire year of 2023, with a sampling frequency of 1 minute / time, totaling approximately 286 million data records. Data with various typical weather conditions (sunny, rainy, thunderstorms) and load characteristics (weekdays, holidays, seasonal variations) were specifically selected to verify the method's adaptability under different environmental conditions. The hardware environment utilizes a high-performance computing cluster with a 32-core CPU, 256GB of memory, and four NVIDIA A100 GPUs. The software environment includes Python 3.10, TensorFlow 2.11, PyTorch 1.13, Scikit-learn 1.2.0, and related deep learning and distributed computing frameworks. The datasets used for algorithm training and testing were divided in an 8:2 ratio, and 5-fold cross-validation was employed to ensure the reliability of the results.

[0111] like Figure 5As shown, firstly, a three-stage cascaded identification process is used to determine the distortion location and improve the recognition accuracy. Then, for each distortion location, three repair channels are initiated in parallel: Bayesian inference for probability distribution-type distortion repair, GBDT regression for nonlinear relational distortion repair, and RNN reconstruction for time-dependent distortion repair. Adaptive weights are calculated based on the errors and confidence levels of each channel in historical verification, and the three candidate values ​​are weighted and fused to obtain a preliminary repair result. Consistency verification and amplitude limiting are performed by introducing power grid domain knowledge constraints (such as power balance, rated voltage range, and line thermal limits). If the constraints are not met, a backtracking process is initiated using the highest confidence channel or the median. Finally, the repaired value that satisfies the constraints is output to reduce repair errors while ensuring physical consistency.

[0112] Explanation of Comparison Option 1: A single-model deep learning approach: This approach employs a deep convolutional long short-term memory (ConvLSTM) network for end-to-end identification and restoration of distorted data. It combines the spatial feature capture capabilities of CNNs with the temporal dependency modeling capabilities of LSTMs. The network structure consists of four ConvLSTM layers (each with 64, 128, 256, and 128 convolutional kernels), followed by three fully connected layers (with 512, 256, and 128 units). The loss function uses a weighted combination of mean squared error and cross-entropy. This method performs well in complex data pattern recognition but lacks effective integration of knowledge from the power system domain and has high computational resource requirements.

[0113] Explanation of Comparison Option 2: An integrated learning and physical constraint post-processing scheme is proposed: An integrated model of XGBoost and SVM is used for distortion identification, and an integrated model of ARIMA and Prophet time series models is used for data repair. Post-processing steps enforce compliance with basic power system constraints. In the identification phase, feature engineering is used to extract 173-dimensional features, and a voting mechanism is employed to integrate the results of different base models. In the repair phase, the ARIMA model is used to capture linear time series relationships, and the Prophet model is used to handle periodic variations. Finally, simplified distribution network power flow calculations are used to correct the results. This method combines the advantages of multiple algorithms and considers basic physical constraints; however, the sub-modules are relatively independent and lack deep fusion and collaborative optimization mechanisms.

[0114] Experimental steps: First, a comprehensive data quality assessment and preprocessing were conducted, including outlier detection, missing value handling, timestamp calibration, and data standardization. To ensure the objectivity of the assessment, seven typical distortion modes were manually injected into the raw data: random noise, system drift, plateau effect, spike interference, periodic distortion, intermittent missing values, and data freezing. The distortion rate was set at 15%, and the data was injected in layers according to mild (5%), moderate (7%), and severe (3%) distortion levels to make the distorted data more closely resemble the distribution of real-world scenarios.

[0115] In the implementation of this invention, the parameters of the three-stage cascaded identification architecture are configured as follows: In the statistical screening stage, the base coefficient k_base is set to 2.5 and dynamically adjusted according to real-time load and weather conditions, with an adjustment range of ±30%; in the random forest classification stage, 500 decision trees are used with a maximum depth of 12, and recursive feature elimination is employed for feature selection, combining domain knowledge to select 87 key features; in the deep learning fine classification stage, an Adam optimizer with a learning rate of 0.001, a batch size of 128, and 150 training epochs are used, with an early stopping mechanism applied to prevent overfitting. The weight ratio of temporal attention to spatial attention is set to 3:2 to better capture the temporal characteristics of distribution network data.

[0116] In the voting-weighted repair mechanism, the Bayesian repair module employs variational Bayesian inference, defining the Markov random field neighborhood as directly adjacent nodes in the distribution network topology. The GBDT repair module is implemented using LightGBM, with 800 trees, a learning rate of 0.05, and a minimum sample size of 20 for leaf nodes. The RNN repair module uses a bidirectional GRU structure with 256 hidden layers, 3 layers, and 8 heads for the enhanced attention mechanism. In the adaptive weight fusion mechanism, the initial value of the weight adjustment parameter λi is set to [1.0, 1.2, 0.8], and it is updated every 200 batches of data based on the repair effect.

[0117] In the physical constraint softening process of domain knowledge constraint fusion, the weighting coefficients β for power balance constraints, voltage constraints, and thermal limit constraints are set to 0.5, 0.3, and 0.2, respectively. The lightweight distributed power flow model employs a forward-backward scan method with a maximum of 20 iterations and a convergence threshold of 1e-4. In the equipment characteristic constraints, the equipment aging characteristic model considers parameter drift rates based on runtime, and the line insulation aging rate is set to an average of 0.5% per year. Expert rule constraints utilize a hybrid rule base consisting of 95 rules mined from historical data and 42 rules defined by domain experts.

[0118] Test methods and standards: This embodiment employs a more comprehensive and rigorous performance evaluation system, testing and comparing the three schemes from multiple dimensions. In terms of distortion recognition performance evaluation, in addition to the conventional accuracy, precision, recall, and F1 score, Balanced Accuracy and Matthews Correlation Coefficient (MCC) are introduced to address data imbalance issues. Furthermore, the Area Under the ROC Curve (AUC) and the Precision-Recall Area Under the Curve (PR-AUC) are used to evaluate the model's discriminative ability. The performance calculation formula for distortion recognition is as follows: Wherein, TP represents the number of true positive samples, TN represents the number of true negative samples, FP represents the number of false positive samples, and FN represents the number of false negative samples.

[0119] The repair quality assessment employs a multi-indicator comprehensive evaluation, including mean absolute error (MAE), mean relative error (MRE), root mean square error (RMSE), peak signal-to-noise ratio (PSNR), and structural similarity index (SSIM). Crucially, assessments are conducted under different operating conditions (normal, abnormal, and extreme weather conditions) to verify the robustness of the method. The repair quality assessment formula is as follows: Where yi represents the true value, ŷi represents the predicted value, n represents the number of samples, MAX_I represents the maximum possible value of the data, and MSE represents the mean squared error.

[0120] The physical rationality assessment is more refined than the first embodiment. In addition to basic power flow constraint violation rates, power balance violation rates, and voltage constraint violation rates, it introduces specialized indicators such as three-phase unbalance, deviations between line loss rates and theoretical values, and consistency of protection strategies. These are verified using distribution network simulation software (OpenDSS and GridLAB-D). Strict standards are used for violation rate calculations. In light of the characteristics of power systems, an additional index for the degree of physical constraint deviation is defined: Where Ci represents the calculated constraint value, and Ci,threshold represents the constraint threshold.

[0121] In addition to directly measuring processing time and resource consumption, the computational efficiency evaluation also introduced throughput (number of data points processed per second) and energy efficiency ratio (amount of data processed per watt of energy consumption) indicators, and conducted scalability tests to evaluate the performance trend of the evaluation method when the data size ranges from 1GB to 100GB.

[0122] Experimental results table:

[0123] Interpretation of experimental results: The experimental results of this embodiment show that the proposed solution significantly outperforms the comparative solution in all key performance indicators, especially in handling complex scenes and extreme conditions. Regarding distortion recognition performance, the proposed solution achieves an overall recognition accuracy of 96.7%, an improvement of 9.9 percentage points compared to the best comparative solution. Particularly noteworthy is its ability to recognize mild distortions, achieving an accuracy of 94.3%, 14.8 percentage points higher than comparative solution 2. This significant improvement is mainly attributed to the hybrid neural network architecture in the deep learning fine-classification stage of the three-stage cascaded discrimination architecture, whose attention mechanism effectively captures subtle features of mild distortions. The Matthews correlation coefficient reaches 0.932, far exceeding the 0.723 of the comparative solution, indicating that the proposed solution maintains excellent classification performance even under imbalanced data conditions.

[0124] In terms of repair quality, the average relative error of the proposed solution is only 1.7%, a 45.2% reduction compared to the 3.1% of the comparative solution 2. Particularly under extreme weather conditions, the repair error of this solution is 3.6%, a 53.8% reduction compared to the 7.8% of the comparative solution 2, demonstrating the robustness advantage of the voting-weighted repair mechanism. The peak signal-to-noise ratio (PSNR) reaches 38.5 dB, an improvement of 5.8 dB compared to the comparative solution 2, and the structural similarity index (SSIM) reaches 0.967, indicating that the repair results are not only close to the true values ​​numerically, but also highly consistent in data structure and trends. These improvements are mainly due to the complementarity of the three repair algorithms and the adaptive weight fusion mechanism. Especially under extreme conditions, the system can dynamically adjust the weights of each algorithm, prioritizing the repair strategy most suitable for the current scenario.

[0125] The physical rationality assessment results are particularly outstanding. The overall physical constraint violation rate of the proposed solution is as low as 1.1%, a reduction of 84.7% compared to the 7.2% of the comparative solution 2. The three-phase imbalance violation rate decreased from 8.3% to 1.5%, the consistency of protection strategies improved to 97.8%, and the degree of physical constraint deviation decreased to 4.8%. These indicators fully demonstrate the effectiveness of the domain knowledge constraint fusion mechanism, especially the introduction of the lightweight distributed power flow model and equipment characteristic constraints, which ensures that the repair results are not only statistically reasonable but also conform to the physical laws of the power system.

[0126] In terms of computational efficiency, although the offline training time of this invention is slightly longer (15.6 hours vs. 9.8 hours), the online inference efficiency is significantly improved. The time to process 1000 data points is reduced by 29.4% compared to the comparative solution 2, and the throughput is increased by 41.3%. In particular, in large-scale data processing tests, the time to process 100GB of data is reduced from 134 minutes to 96 minutes, an improvement of 28.4%. The energy efficiency ratio is improved by 36.4%, indicating that this solution not only has a faster processing speed but also higher energy utilization efficiency. This efficiency improvement mainly comes from the early screening mechanism of the three-stage cascaded architecture and the lightweight model design. A large amount of obviously distorted data is identified in the statistical screening stage, without the need for processing by complex deep learning models.

[0127] The results of the adaptability evaluation best demonstrate the technical advantages of this invention. Under complex scenarios such as non-stationary operating conditions, equipment switching events, and topology changes, the accuracy of this scheme remains above 88.9%, an improvement of 15.8% to 18.6% compared to the comparative scheme 2. In particular, the recognition rate for novel distortion patterns reaches 84.5%, 18.8 percentage points higher than the comparative scheme 2, indicating that this scheme has excellent generalization ability. The model transfer performance retention rate reaches 90.2%, demonstrating that this scheme has good adaptability in different distribution network environments. This is mainly due to the introduction of adversarial training strategies and domain knowledge constraints, which make the model focus more on the essential characteristics of distribution network data rather than the surface patterns of specific datasets.

[0128] In summary, this invention achieves high detection rate and low error correction in complex scenarios such as voltage, current, active power, and reactive power measurement data by performing denoising, deduplication, time alignment, and missing data imputation to form the time series to be measured; statistical screening based on dynamic thresholds and hysteresis autocorrelation reference bands is used to achieve initial anomaly detection; hierarchical random forest is used to complete the coarse classification of normal and distorted data driven by artificial features; further, a network that combines CNN-LSTM and attention mechanism is used to refine the distortion type and output confidence scores; for the positions identified as distorted, three repair channels are constructed: Bayesian posterior estimation, gradient boosting decision tree regression, and attention-enhanced sequence-to-sequence reconstruction, and adaptive weighted fusion is performed according to historical error and adaptability to obtain the final repair value; on this basis, power balance, rated voltage range, line thermal limit, and other power physics and engineering boundaries are introduced to implement feasibility verification and amplitude limiting processing; finally, the distortion type, repair value, confidence score, and constraint residual are output and used for subsequent statistics and model updates, thereby achieving high detection rate, low error repair, and ensuring that the results meet the physical laws of the power grid and the operating boundaries of the equipment in complex scenarios such as peaks, missing data, and drift.

[0129] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended 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.

Claims

1. A method for identifying and repairing data distortion in a power distribution network, characterized in that: This includes acquiring and preprocessing time-series data of the distribution network; In the time-series data, a three-stage joint identification architecture is used to identify distorted data; The distorted data is repaired through a voting-weighted repair mechanism.

2. The method for identifying and repairing data distortion in a power distribution network as described in claim 1, characterized in that: The preprocessing includes denoising, deduplication, time alignment, and missing interpolation of the time series of voltage, current, active power, and reactive power to obtain the time series to be measured. .

3. The method for identifying and repairing data distortion in a power distribution network as described in claim 2, characterized in that: The three-stage joint identification architecture includes a statistical screening stage, a random forest classification stage, and a deep learning fine classification stage. The statistical screening stage includes: outlier screening, time-series pattern screening, and dynamic amplitude screening; The outlier screening includes removing outliers from the time series under test according to Z-score and IQR rules, and removing samples with abnormal amplitudes. The time series pattern screening includes calculating the autocorrelation coefficient of the time series to be tested on a preset lag set and comparing it with a reference band obtained based on historical data; when the autocorrelation coefficient of any lag deviates from the reference band and the deviation exceeds a preset threshold, it is determined to be a time series anomaly and marked. The dynamic amplitude screening includes calculating the time series to be tested within a sliding window indexed by time t. The mean and standard deviation are used to define a dynamic threshold: in, Indicates time The dynamic threshold, Indicates time The mean, Indicates time standard deviation Indicates time The dynamic coefficient; when If the amplitude is abnormal, it is determined to be abnormal and deleted; otherwise, it is retained for subsequent processing.

4. The method for identifying and repairing data distortion in a power distribution network as described in claim 3, characterized in that: The random forest classification stage includes: extracting time-domain features, frequency-domain features, and relational features from the time series data to be tested, inputting them into the hierarchical random forest model, outputting a normal and distorted judgment, and classifying the distorted data; The hierarchical random forest model consists of two layers: the first layer distinguishes between normal data and distorted data; and the second layer classifies distorted data by type. The deep learning fine-classification stage includes constructing a hybrid neural network architecture: in, This indicates categorized output. Describes the multilayer perceptron function. This represents the attention mechanism function. This represents the function of a convolutional neural network. This represents the Long Short-Term Memory (LSTM) network function. Representing spatial feature data, Represents time-series characteristic data; The classification output is computed with attention enhancement based on a hybrid neural network: in, Represents the query matrix. Represents the key matrix. Represents a value matrix, Indicates the dimension of the key vector; The attention-enhanced feature representation is input into the classifier to obtain the score vector for each distortion category; the category corresponding to the component with the largest value in the score vector is used as the judgment result.

5. The method for identifying and repairing data distortion in a power distribution network as described in claim 4, characterized in that: The voting weighted repair mechanism includes weighted repair based on Bayesian inference, repair based on gradient boosting decision trees, and repair based on recurrent neural networks. The Bayesian inference-based repair includes constructing a data probability distribution model based on Bayesian theory: in, Indicates observed data Real data under conditions The posterior probability, Represents the likelihood function. Represents prior probability; When the posterior distribution is obtained When the position is judged to be distorted, the representative value is directly taken from the posterior distribution as the repair value; The gradient boosting decision tree-based repair includes constructing a gradient boosting decision tree model: in, Indicates the first The model of the next iteration Indicates the first The model of the next iteration Indicates the learning rate. Indicates the first A decision tree; The context sequence of the distorted position is input into the trained gradient boosting decision tree model, and the regression prediction value output by the model is used as the repair value; the context sequence of the distorted position is an ordered sequence formed by arranging adjacent observations within a preset window on both sides of the distorted position in chronological order, combined with lagged observations of the same period. The recurrent neural network-based repair includes constructing a recurrent neural network model; introducing an attention-enhanced sequence-to-sequence model, wherein the decoder of the sequence-to-sequence model outputs the context vector at the i-th time step as follows: in, This represents the context vector at time i in the output sequence; This represents the attention weight of the output sequence at time i to the input sequence at time j; This represents the hidden state at time j of the input sequence; Indicates the length of the input sequence; The context sequence of the distorted location is input into the sequence model, and the decoder obtains the reconstructed value of the time series under test as the repair value based on the context sequence.

6. The method for identifying and repairing data distortion in a power distribution network as described in claim 5, characterized in that: The voting weighted repair mechanism includes fusing three repair results through an adaptive weight formula to obtain the final repair result; The adaptive weighting formula: in, Indicates the first The weight of each repair This represents the weight adjustment parameter. Indicates the first Historical errors of each repair module; Indicates the index variable, from 1 to 3; Calculation fusion results: in, This indicates the final repair result. Indicates the first Repair results of each repair module; Indicates the first The weight of each repair.

7. The method for identifying and repairing data distortion in a power distribution network as described in claim 6, characterized in that: The voting-weighted repair mechanism includes applying domain knowledge constraints to the repair results during the repair process: Based on the AC power flow relationship, power balance constraints are applied to the calculated power and target power of each node. Violations of power balance, permissible voltage range and line thermal limit are judged. Violation of any one of them is judged as not meeting physical constraints and is treated as unqualified. Based on the engineering boundaries of the equipment, the candidate results are limited or prohibited; Set the sequence, thresholds, and consistency checks based on operating procedures and experience.

8. A distribution network data distortion identification and repair system, based on the distribution network data distortion identification and repair method according to any one of claims 1 to 7, characterized in that: This includes a preprocessing module: which preprocesses the distribution network data; Identification module: Constructs a three-stage joint identification architecture to identify distorted data in the distribution network; Weighted module: Employs a voting-weighted repair mechanism to repair identified distorted data; Constraint Module: Ensures the physical rationality of the repair results through the fusion of domain knowledge constraints.

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 power distribution network data distortion identification and repair method according to 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 data distortion identification and repair method according to any one of claims 1 to 7.

Images