Low-voltage transformer area power equipment fault prediction method, device and equipment based on big data

Abstract

The invention provides a low-voltage transformer area power equipment fault prediction method, device and equipment based on big data, and relates to the technical field of electric power. The method comprises the following steps: acquiring a multi-dimensional feature; extracting a first preset feature set from the multi-dimensional features, performing first-level rough prediction according to the first preset feature set, and determining high-risk candidate equipment; extracting a second preset feature set from the multi-dimensional features, performing secondary fine classification according to the second preset feature set, and determining the fault type and confidence of the high-risk candidate equipment; extracting a third preset feature set from the multi-dimensional features, performing three-level risk assessment according to the third preset feature set and the fault type and confidence of the high-risk candidate equipment, and determining the risk level of the high-risk candidate equipment; and extracting verification data from the standardized data, performing cross verification on the fault type, the confidence coefficient and the risk level according to the verification data, and outputting a fault prediction result. According to the invention, the accuracy and adaptability of equipment fault prediction can be improved.

Description

Technical Field

[0001] This invention relates to the field of power technology, and in particular to a method, apparatus and equipment for predicting faults in low-voltage power distribution area equipment based on big data. Background Technology

[0002] As the end-point of the power system facing users, the equipment failures of low-voltage distribution areas directly affect the reliability of power supply. With the widespread adoption of distributed power sources and the diversification of user electricity demands, the causes of failures have expanded from traditional equipment aging and environmental impacts to include fluctuations in user behavior and grid control interactions. Traditional failure prediction methods have many shortcomings, as follows: First, the data dimensions are limited. Traditional methods only focus on the collection of equipment electrical parameters and environmental data, resulting in insufficient support for fault prediction. Second, feature engineering is rigid, resulting in insufficient accuracy in matching features with fault types; Third, the prediction model lacks a refined design. For example, the rough screening and fine judgment of faults are disconnected, which not only results in low accuracy in screening high-risk equipment, but also lacks a multi-source data cross-validation mechanism, leading to insufficient reliability in the determination of fault type and risk level. Fourth, the failure to fully integrate the characteristic data of the transformer area with the results of multi-source verification resulted in a poor match between the fault prediction results and the actual operation and maintenance needs, which could easily lead to a waste of operation and maintenance resources or untimely fault response. Summary of the Invention

[0003] This invention provides a method for predicting faults in low-voltage distribution area power equipment based on big data, in order to improve the accuracy and adaptability of equipment fault prediction.

[0004] In a first aspect, embodiments of the present invention provide a method for predicting faults in low-voltage distribution area power equipment based on big data, including: Acquire multi-source data from low-voltage distribution areas and preprocess it to obtain standardized data; Extract fault-related multidimensional features from the standardized data; A first preset feature set is extracted from the multidimensional features, and a first-level coarse prediction is performed based on the first preset feature set to determine high-risk candidate devices. A second preset feature set is extracted from the multidimensional features, and a secondary fine classification is performed based on the second preset feature set to determine the fault type and confidence level of the high-risk candidate device. A third preset feature set is extracted from the multidimensional features. Based on the third preset feature set, the fault type and confidence level of the high-risk candidate device, a three-level risk assessment is performed to determine the risk level of the high-risk candidate device. Validation data is extracted from the standardized data. Based on the validation data, the fault type and confidence level and the risk level are cross-validated, and the fault prediction result is output.

[0005] In one possible implementation, the step of extracting a first preset feature set from the multidimensional features and performing a first-level coarse prediction based on the first preset feature set to determine high-risk candidate devices includes: Extract the core sub-features of environmental impact from the device operating status features, device physical status features, and third-party data association features, as well as the high-priority sub-features from the user feedback association features from the multi-dimensional features. Then, use the variance filtering method to delete redundant features with variance less than a preset variance threshold to obtain the first preset feature set. The first preset feature set is input into a pre-constructed first-level coarse prediction model, and the failure probability of each device is output. Devices whose probability of failure is greater than a preset probability threshold are identified as high-risk candidate devices.

[0006] In one possible implementation, the step of extracting a second preset feature set from the multidimensional features, performing secondary fine classification based on the second preset feature set, and determining the fault type and confidence level of the high-risk candidate device includes: The equipment operating status features, equipment physical status features, third-party data association features, user feedback association features, power grid control features, power grid control association features, and operation behavior association features are extracted from the multidimensional features, and mutual information entropy is calculated to filter features to obtain the second preset feature set. The second preset feature set is input into a pre-constructed hybrid algorithm model of convolutional neural network and long short-term memory network, and the fault type and confidence level of the high-risk candidate device are output.

[0007] In one possible implementation, the step of extracting a third preset feature set from the multidimensional features, and performing a three-level risk assessment based on the third preset feature set, the fault type and confidence level of the high-risk candidate device, to determine the risk level of the high-risk candidate device includes: Extract user feedback correlation features and station area features from the multidimensional features; Based on the user feedback correlation characteristics and the transformer area characteristics, several core evaluation indicators are determined; Based on the aforementioned core evaluation indicators and the risk assessment system established in advance through the analytic hierarchy process, the comprehensive risk value of the high-risk candidate equipment is determined. The risk level of the high-risk candidate device is determined based on its comprehensive risk value, fault type, and confidence level.

[0008] In one possible implementation, the multiple core evaluation metrics include: fault impact range, fault urgency, equipment importance, user importance, and fault propagation probability.

[0009] In one possible implementation, the step of extracting validation data from the standardized data, and then cross-validating the fault type and confidence level and the risk level based on the validation data, and outputting a fault prediction result, includes: User feedback data, operational behavior data, and third-party data are extracted from the standardized data. Based on the user feedback data, the operation behavior data, and the third-party data, cross-validation is performed on the fault type and confidence level, and the risk level to obtain validation scores for multiple validation dimensions. The verification scores of the multiple verification dimensions are weighted and fused to calculate the comprehensive verification score; If the comprehensive verification score is satisfactory, the fault prediction result will be output. If the comprehensive verification score is unqualified, the high-risk candidate device will be deleted.

[0010] In one possible implementation, after the output fault prediction result, the following is also included: Based on the standardized data and the fault prediction results, the model parameters are dynamically updated through transfer learning algorithms, and the overall performance of the model is evaluated periodically to achieve adaptive optimization of the model.

[0011] In one possible implementation, after the output fault prediction result, the following is also included: Obtain operation and maintenance resource data, and generate a fault solution based on the operation and maintenance resource data and the fault prediction results; Based on the fault solution, an execution report is generated; wherein the execution report includes: implementation steps, technical parameter requirements, resource allocation plan, expected effect evaluation and solution adjustment mechanism.

[0012] In one possible implementation, the preprocessing includes: Data cleaning, outlier removal, missing value imputation, data standardization, cross-dimensional data consistency verification, multi-source cross-validation, semantic parsing and noise reduction of user feedback data.

[0013] Secondly, embodiments of the present invention provide a fault prediction device for low-voltage distribution area power equipment based on big data, comprising: The processing module is used to acquire multi-source data from the low-voltage distribution area and preprocess it to obtain standardized data; and extract fault-related multidimensional features from the standardized data. The first-level coarse prediction module is used to extract a first preset feature set from the multi-dimensional features, perform a first-level coarse prediction based on the first preset feature set, and determine high-risk candidate devices. The secondary fine classification module is used to extract a second preset feature set from the multi-dimensional features, perform secondary fine classification based on the second preset feature set, and determine the fault type and confidence level of the high-risk candidate device. The three-level risk assessment module is used to extract a third preset feature set from the multi-dimensional features, and perform a three-level risk assessment based on the third preset feature set, the fault type and confidence level of the high-risk candidate device, to determine the risk level of the high-risk candidate device. The four-level multi-source verification module is used to extract verification data from the standardized data, and after cross-validating the fault type and confidence level and the risk level based on the verification data, output the fault prediction result.

[0014] Thirdly, embodiments of the present invention provide an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method described in the first aspect or any possible implementation thereof.

[0015] The beneficial effects of the embodiments of the present invention compared with the prior art are as follows: This invention first acquires and preprocesses multi-source data from low-voltage distribution areas to obtain standardized data, thus overcoming the predictive limitations of single-data-dimensional data and providing sufficient basis for accurate prediction. Next, fault-related multi-dimensional features are extracted from the standardized data to strengthen the correlation between features and fault types, thereby improving the input quality of subsequent prediction models.

[0016] This invention extracts a first preset feature set from multi-dimensional features, performs a first-level coarse prediction based on the first preset feature set to identify high-risk candidate devices; extracts a second preset feature set from multi-dimensional features, performs a second-level fine classification based on the second preset feature set to determine the fault type and confidence level of the high-risk candidate devices; extracts a third preset feature set from multi-dimensional features, performs a third-level risk assessment based on the third preset feature set, the fault type and confidence level of the high-risk candidate devices, and determines the risk level of the high-risk candidate devices; and extracts validation data from standardized data, performs cross-validation on the fault type, confidence level, and risk level based on the validation data, and outputs the fault prediction result. Through a hierarchical model of "coarse screening-fine classification-risk assessment-multi-source verification," it first focuses on high-risk candidate devices, then accurately determines the fault type and risk level, and finally, through multi-source data cross-validation, significantly improves the accuracy and reliability of the prediction results. The output fault prediction results are highly targeted and can provide precise guidance for the allocation of operation and maintenance resources and the prioritization of fault handling, reducing operation and maintenance costs, improving fault handling efficiency, and ensuring the stability of power supply in low-voltage distribution areas. Attached Figure Description

[0017] Figure 1 This is the implementation flow of the low-voltage distribution area power equipment fault prediction method based on big data provided in the embodiments of the present invention. Figure 1 ; Figure 2 This is the implementation flow of the low-voltage distribution area power equipment fault prediction method based on big data provided in the embodiments of the present invention. Figure 2 ; Figure 3 This is a schematic diagram of the structure of the low-voltage distribution area power equipment fault prediction device based on big data provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0018] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0019] See Figure 1 The document illustrates a flowchart of the implementation of the low-voltage distribution area power equipment fault prediction method based on big data provided in this embodiment of the invention, detailed below: Step S101: Obtain multi-source data from the low-voltage distribution area and perform preprocessing to obtain standardized data.

[0020] In this embodiment, multi-source data may include, but is not limited to: operating status data of low-voltage distribution area power equipment, environmental perception data, historical fault data, power grid topology data, user-side electricity consumption behavior data, equipment physical status, maintenance data, power grid operation and control data, equipment operation behavior data, third-party monitoring data, user feedback and complaint suggestion data, and distribution area characteristic data.

[0021] After collecting data from multiple sources, preprocessing operations such as data cleaning, outlier removal, missing value imputation, data standardization, cross-dimensional data consistency verification, multi-source cross-validation, semantic parsing of user feedback data, and noise reduction can be performed to obtain standardized data.

[0022] Step S102: Extract fault-related multidimensional features from standardized data.

[0023] This embodiment performs dynamic feature engineering based on a standardized dataset to extract features such as equipment operating status, user behavior correlation features, equipment physical status features, power grid control features, operation behavior correlation features, third-party data correlation features, and user feedback correlation features, and constructs a multi-dimensional feature set.

[0024] Step S103: Extract the first preset feature set from the multidimensional features, perform a first-level coarse prediction based on the first preset feature set, and determine high-risk candidate devices.

[0025] In this embodiment, the first-level coarse prediction model, based on the input multi-dimensional feature set, filters out high-risk candidate devices whose failure probability exceeds a threshold from all devices in the distribution area.

[0026] In one possible implementation, the first-level coarse prediction model extracts equipment operating status features, equipment physical status features, environmental impact core sub-features from third-party data association features, and high-priority sub-features from user feedback association features from multi-dimensional features. Then, it removes redundant features with variance less than a preset variance threshold using a variance filtering method to obtain a first preset feature set. The first preset feature set is input into the pre-constructed first-level coarse prediction model to output the failure probability of each device. Devices with a failure probability greater than a preset probability threshold are identified as high-risk candidate devices.

[0027] Step S104: Extract the second preset feature set from the multidimensional features, perform secondary fine classification based on the second preset feature set, and determine the fault type and confidence level of high-risk candidate equipment.

[0028] In this embodiment, the secondary classification model determines the fault type and confidence level of high-risk candidate devices based on high-risk candidate devices and their feature data.

[0029] In one possible implementation, the two-level fine classification model can extract equipment operating status features, equipment physical status features, third-party data correlation features, user feedback correlation features, power grid control features, power grid control correlation features, and operation behavior correlation features from multi-dimensional features, and calculate mutual information entropy to filter features, thereby obtaining a second preset feature set; the second preset feature set is then input into a pre-constructed hybrid algorithm model of convolutional neural network and long short-term memory network to output the fault type and confidence level of high-risk candidate equipment.

[0030] Step S105: Extract the third preset feature set from the multidimensional features. Based on the third preset feature set, the fault type and confidence level of the high-risk candidate equipment, conduct a three-level risk assessment to determine the risk level of the high-risk candidate equipment.

[0031] In this embodiment, the three-level risk assessment model is based on the output results of the first-level coarse prediction model and the second-level fine classification model, and integrates user feedback correlation features and transformer area characteristic data to calculate the comprehensive risk value of high-risk candidate devices and classify the risk level.

[0032] In one possible implementation, the three-level risk assessment model can extract user feedback correlation features and transformer area features from multi-dimensional features; determine multiple core assessment indicators based on user feedback correlation features and transformer area features; determine the comprehensive risk value of high-risk candidate equipment based on multiple core assessment indicators and a risk assessment system established in advance through the analytic hierarchy process; and determine the risk level of high-risk candidate equipment based on the comprehensive risk value, fault type, and confidence level of the high-risk candidate equipment.

[0033] Step 106: Extract validation data from the standardized data. Based on the validation data, perform cross-validation on the fault type, confidence level, and risk level, and then output the fault prediction results.

[0034] In this embodiment, the four-level multi-source verification model combines user feedback data, operational behavior data, and third-party monitoring data to cross-validate the risk level output by the three-level risk assessment model and the fault type output by the two-level fine classification model, and outputs fault prediction results and verification reports that include the fault type, risk level, and confidence level of high-risk candidate devices.

[0035] In one possible implementation, the four-level multi-source verification model can extract user feedback data, operational behavior data, and third-party data from standardized data; cross-validate the fault type, confidence level, and risk level based on the user feedback data, operational behavior data, and third-party data respectively to obtain verification scores for multiple verification dimensions; weightedly fuse the verification scores of multiple verification dimensions to calculate a comprehensive verification score; if the comprehensive verification score is qualified, the fault prediction result is output; if the comprehensive verification score is unqualified, high-risk candidate devices are deleted.

[0036] This invention first acquires and preprocesses multi-source data from low-voltage distribution areas to obtain standardized data, overcoming the limitations of prediction based on a single data dimension and providing sufficient basis for accurate prediction. Next, it extracts fault-related multi-dimensional features from the standardized data to strengthen the correlation between features and fault types, improving the input quality of subsequent prediction models. Further, this invention extracts a first preset feature set from the multi-dimensional features, performs a first-level coarse prediction based on the first preset feature set to identify high-risk candidate devices; extracts a second preset feature set from the multi-dimensional features, performs a second-level fine classification based on the second preset feature set to determine the fault type and confidence level of the high-risk candidate devices; extracts a third preset feature set from the multi-dimensional features, performs a third-level risk assessment based on the third preset feature set, the fault type and confidence level of the high-risk candidate devices, and determines the risk level of the high-risk candidate devices; finally, it extracts validation data from the standardized data, performs cross-validation on the fault type, confidence level, and risk level based on the validation data, and outputs the fault prediction result. By employing a hierarchical model of "coarse screening, fine classification, risk assessment, and multi-source verification," the system first focuses on high-risk candidate devices, then accurately determines the fault type and risk level, and finally conducts cross-validation using multi-source data. This significantly improves the accuracy and reliability of the prediction results, and the output fault prediction results are highly targeted. They can provide precise guidance for the allocation of operation and maintenance resources and the prioritization of fault handling, thereby reducing operation and maintenance costs, improving fault handling efficiency, and ensuring the stability of power supply in low-voltage distribution areas.

[0037] To facilitate understanding of this solution, a specific embodiment will be used for illustration below.

[0038] See Figure 2 As shown, the steps in this embodiment include: S1. Obtain multi-source data for the low-voltage distribution area.

[0039] Here, a multi-source data acquisition system is constructed, encompassing "equipment-environment-user-grid-operation-third party-feedback," to collect operational status data of low-voltage distribution area power equipment, environmental perception data, historical fault data, grid topology data, user-side electricity consumption behavior data, equipment physical status, maintenance data, grid operation and control data, equipment operation behavior data, third-party monitoring data, user feedback and complaint / suggestion data, and distribution area characteristic data. Distribution area characteristic data includes equipment nameplate parameters and user importance classification data. User feedback and complaint / suggestion data includes fault descriptions, electricity experience complaints, maintenance service suggestions, and equipment anomaly feedback submitted by users through APP / hotline / grid staff. Third-party monitoring data includes meteorological warning data, equipment testing report data issued by third-party testing institutions, and warning data pushed by the user's electricity safety platform.

[0040] For example, the steps for collecting user feedback and complaint / suggestion data include: S11. Establish a multi-channel user feedback collection platform, including a mobile APP, customer service hotline, and community grid worker feedback channel; the mobile APP supports uploading fault descriptions in text / images / videos, selecting complaint categories, and setting satisfaction ratings; the customer service hotline records complaint content through voice-to-text technology; and the community grid worker feedback channel is used to regularly summarize centralized issues in the area. S12. Set up a data collection and synchronization mechanism. APP / hotline feedback adopts a real-time collection method. Grid worker feedback is summarized and uploaded at a fixed time every day (e.g., 16:00 daily). All collected data is synchronized to the user feedback database through an encrypted interface to ensure data transmission security. S13. Establish a mechanism to link user feedback with equipment. Based on the address information contained in the user feedback, match the corresponding transformer substation number and equipment number to achieve precise binding of "user feedback - specific equipment". For example, when a user reports a problem related to "Building 3, XX Community", it is automatically linked to the corresponding branch box D and transformer A of that community. This linking result will be used for the extraction of user feedback linking features in step S3 and the four-level multi-source verification in step S4.

[0041] S2. Preprocess multi-source data to obtain standardized data.

[0042] Preprocessing includes data cleaning, outlier removal, missing value imputation, data standardization, cross-dimensional data consistency verification, multi-source cross-validation, and semantic parsing and noise reduction of user feedback data. Outlier removal employs a 3D model. The missing value filling criterion is to use the moving average method.

[0043] For example, the steps of semantic parsing and noise reduction processing of user feedback data include: S21. Semantic parsing and fault label mapping: The BERT pre-trained model is used to perform semantic analysis on the user feedback text collected in step S1 to extract key fault-related information. For example, the core information of "peak electricity consumption - voltage fluctuation - lighting equipment" is extracted from "lights flickering at 8 pm". A semantic-fault label mapping library is established to uniformly map "lights flickering" to "voltage fluctuation fault". This mapping result will be used for mining the user feedback association features in step S3. S22. Data noise reduction processing: Remove invalid feedback from user feedback collected in step S1 that is repeatedly submitted or lacks specific fault descriptions, such as feedback that only states "the equipment is broken"; use keyword matching rules to filter valid feedback, and feedback containing keywords such as "voltage / power outage / abnormal noise / heating" is considered valid feedback; correct feedback with vague time / location descriptions, determine the accurate power outage period by combining the power grid operation and control data collected in step S1, match the corresponding transformer area through the user's registered address, and the corrected feedback data will be used for standardization processing in step S23; S23. Feedback data standardization: Store the user feedback processed in step S22 in a structured format of "station number-equipment number-fault type-feedback time-severity". The severity is quantified into 1-5 points according to the user description, where "occasional flickering" is defined as level 1, "complete power outage" is defined as level 5, and intermediate levels are divided according to the degree of fault impact. This standardized feedback data will be used as a component of the multi-dimensional feature set in step S3.

[0044] S3. Extract fault-related multidimensional features from standardized data.

[0045] This embodiment uses a standardized data base to perform dynamic feature engineering, extracting features related to equipment operating status, user behavior, equipment physical status, power grid control, operational behavior, third-party data, and user feedback to construct a multi-dimensional feature set.

[0046] The extraction process for each feature includes: obtaining equipment operating status features (including time-series features, trend features, and abrupt change features) based on equipment runtime sequence data in the standardized dataset through sliding window statistics and trend fitting; obtaining user behavior correlation features based on user-side electricity consumption behavior data in the standardized dataset through load fluctuation analysis and time period correlation calculation; obtaining equipment physical status features based on equipment physical status in the standardized dataset through index quantification and health mapping; obtaining power grid control features based on power grid operation and control data in the standardized dataset through control action correlation and response analysis; obtaining operation behavior correlation features based on operation behavior data in the standardized dataset through operation type matching and influencing factor extraction; and obtaining third-party data correlation features based on third-party monitoring data in the standardized dataset through environmental / detection index screening and fault correlation analysis.

[0047] For example, the steps for extracting user feedback-related features include: S31. Extract feedback frequency and concentration features: Based on the standardized feedback data obtained in step S23, calculate the weekly feedback frequency of a specific device or distribution area, such as transformer A receiving 5 voltage-related feedbacks in the past 7 days; calculate the feedback concentration, which is the proportion of users who report the same problem to the total number of users in the distribution area, such as when 15 out of 30 households in a certain community report the same problem, the concentration is 50%; set threshold conditions, such as marking features with a frequency ≥ 3 times / week or a concentration ≥ 30% as high-priority features, and these high-priority features will be input into the first-level coarse prediction module in step S4; S32. Mining Feedback-Fault Correlation Features: The Apriori algorithm is used to analyze the correlation between the standardized feedback data in step S23 and the historical fault data collected in step S1. For example, when the support of "voltage fluctuation feedback" and "voltage regulating device fault" is ≥0.65 and the confidence is ≥0.82, it is determined that there is a strong correlation between the two. The correlation strength is included in the feature set, and the correlation feature will be used for fault type determination in the secondary fine classification module of step S4. S33. Extract feedback time correlation features: Based on the feedback time information in the standardized feedback data in step S23, analyze the time distribution pattern of user feedback, such as feedback concentrated in the 18:00-22:00 period; Combine the user-side electricity consumption behavior data collected in step S1 to calculate the overlap between the feedback period and the equipment load peak. An overlap of ≥80% is marked as a strong correlation; quantify the overlap and fault severity into a fixed-dimensional feature vector, which will be used for the three-level risk assessment in step S4.

[0048] S4. Construct a hierarchical fault prediction model, including a first-level coarse prediction model, a second-level fine classification model, a third-level risk assessment model, and a fourth-level multi-source verification model.

[0049] In this embodiment, the first-level coarse prediction model, based on the input multi-dimensional feature set, filters out high-risk candidate devices whose failure probability exceeds a threshold from all devices in the distribution area. For example, the implementation steps of the first-level coarse prediction model include: S411. Primary Feature Screening and Optimization: Core input features are screened from the multi-dimensional feature set constructed in step S3. These include equipment operating status features (time-series features, trend features, and abrupt change features), key sub-features in equipment physical status features (current deviation, temperature fluctuation amplitude, voltage stability index), core environmental impact sub-features in third-party data association features (meteorological disaster warning level, abnormal signals of third-party detection core indicators, duration of humidity exceeding standards, frequency of extreme temperature occurrences), and high-priority sub-features in user feedback association features (weekly feedback frequency ≥ 3 times, feedback concentration ≥ 30%). Redundant features with variance < 0.01 are eliminated using variance filtering, retaining effective features to form a primary prediction feature subset. Specifically, time-series features provide the dynamic change patterns of equipment parameters over time, trend features reflect the long-term trend of parameter changes, abrupt change features capture abnormal parameter jump signals, physical status features directly reflect the health of equipment hardware operation, third-party data association features incorporate the influence of the external environment and professional detection, and user feedback association features supplement user subjective demand signals. All these features work together to support the preliminary judgment of the probability of fault occurrence.

[0050] Let's take the selection of key sub-features from the physical state characteristics of equipment as an example: For current deviation: calculate the ratio of the difference between the real-time current and the rated current of the equipment to the rated current, i.e., real-time current - rated current | / rated current. ≥10% is marked as an abnormal feature.

[0051] For temperature fluctuation range: the difference between the highest and lowest temperatures of the equipment within 24 hours is statistically analyzed, and values ​​≥20℃ are marked as abnormal characteristics.

[0052] For voltage stability indicators: calculate the deviation rate between the real-time voltage and the rated voltage of the equipment. A deviation rate of ≥±5% for three consecutive acquisition cycles is marked as an abnormal feature.

[0053] S412. Coarse Prediction Model Construction and Training: The LightGBM algorithm is used to construct a first-level coarse prediction model. The "fault occurred (label 1) / not occurred (label 0)" in the historical fault data collected in step S1 is used as the binary classification label. The model is trained on the first-level prediction feature subset. During the training process, 5-fold cross-validation can be used to optimize the hyperparameters. The learning rate is set to 0.05, the tree depth is 5, the number of leaf nodes is 31, and the number of iterations is 100 to ensure the model fitting effect.

[0054] S413. Initial screening and output of high-risk equipment: Extract features from the standardized dataset in step S2, input them into the trained first-level coarse prediction model, and output the failure probability of each device (the probability threshold can be set to ≥0.6); Based on the set probability threshold, select devices with probabilities exceeding the threshold as high-risk candidate devices, and generate a list of first-level coarse prediction results containing device number and failure probability.

[0055] In this embodiment, the two-level precision classification model determines the fault type and confidence level of high-risk candidate devices based on their feature data. For example, the implementation steps of the two-level precision classification model include: S421. Secondary Feature Expansion and Refinement: Using the primary coarse prediction result list output in step S413 as the screening condition, extract the feature data of high-risk candidate equipment; based on the primary prediction feature subset, expand the physical state features, power grid control features, correlation features, and equipment physical state subdivision features, power grid control correlation features, and operation behavior correlation features of high-risk candidate equipment; calculate the correlation degree between the above-collected and extracted features and the fault type through mutual information entropy, and remove weak correlation features with mutual information values ​​<0.1 to form a secondary classification feature set; S422. Multi-classification model construction and training: A CNN-LSTM hybrid algorithm is used to construct a two-level fine classification model to classify fault types. The fault types can be divided into 5 core categories and 1 other category, specifically "voltage fluctuation fault, insulation aging fault, overload fault, poor contact fault, outdoor equipment water ingress fault, and other faults". The fault type labels in the historical fault data collected in step S1 are used as the training target to train the model on the two-level classification feature set. An early stopping mechanism is adopted during the training process (training is stopped if the validation set loss does not decrease for 5 consecutive rounds) to avoid model overfitting. S423. Accurate Fault Type Determination and Output: Input the features extracted and expanded in step S421 into the trained secondary classification model, and output the confidence level of each fault type of high-risk candidate equipment; select the fault type with the highest confidence level as the predicted fault type of the equipment, requiring the highest confidence level to be greater than the preset value (the highest confidence level can be set to ≥0.8), otherwise mark it as "fault type to be verified"; generate a secondary classification result containing the equipment number, predicted fault type, and corresponding confidence level.

[0056] In this embodiment, the three-level risk assessment model is based on the outputs of the first-level coarse prediction model and the second-level fine classification model, and integrates user feedback correlation features and transformer area characteristic data to calculate the comprehensive risk value of high-risk candidate devices and classify their risk levels. For example, the implementation steps of the three-level risk assessment model in step S4 include: S431. Construct a risk assessment indicator system: Determine the core assessment indicators, including the scope of the fault impact (number of affected users), the urgency of the fault (severity level 1-5, equipment importance (core equipment / non-core equipment), user importance (based on user importance classification data in the transformer area characteristic data collected in step S1, such as hospitals and schools as level 1 users, key livelihood / commercial users as level 2 users, and ordinary residents as level 3 users), and the probability of fault propagation (historical spread rate of similar faults). S432. Weighting of indicators: The weights of each evaluation indicator are determined by the analytic hierarchy process. The weights can be set as follows: scope of impact of the fault 0.3, urgency of the fault 0.3, importance of equipment 0.2, importance of user 0.1, and probability of fault propagation 0.1. S433. Quantitative Scoring of Indicators: Core evaluation indicators are quantified, with the following settings: 10 points for fault impact range ≥ 50 households, 6 points for 20-50 households, and 2 points for < 20 households; 2 points for fault urgency level 1, 4 points for level 2, 6 points for level 3, 8 points for level 4, and 10 points for level 5; 10 points for equipment importance (core equipment), and 5 points for non-core equipment; 10 points for user importance (Level 1 users), 7 points for Level 2 users, and 3 points for Level 3 users; and 10 points for fault propagation probability ≥ 80%, 6 points for 50%-79%, and 2 points for < 50%. S434. Comprehensive Risk Value Calculation and Classification: Set risk level classification standards, calculate the comprehensive risk value (range 0-10 points) of high-risk candidate equipment according to "Comprehensive Risk Value = Σ (Indicator Score × Corresponding Weight)", determine the risk level of the equipment based on the risk level classification standards and comprehensive risk value (risk value ≥ 8 points is "extremely high risk", 5-7 points is "high risk", 3-4 points is "medium risk", < 3 points is "low risk"). Based on the comprehensive risk value and integrating the secondary classification results, output the three-level risk assessment results including equipment number, fault type, comprehensive risk value, and risk level.

[0057] In this embodiment, the four-level multi-source verification model combines user feedback data, equipment operation behavior data, third-party monitoring data, and standardized datasets to cross-validate the risk level output by the three-level risk assessment model and the fault type output by the two-level fine classification model. The model outputs fault prediction results and verification reports that include the fault type, risk level, and confidence level of high-risk candidate devices.

[0058] For example, the verification steps of the four-level multi-source verification model combined with operational behavior data include: S441. Operation behavior data extraction: Extract the operation records of the high-risk candidate equipment to be verified within a preset time period from the equipment operation behavior data collected in step S1. These records include operation type (such as maintenance, parameter adjustment, and component replacement), operation time, operator, and post-operation equipment status detection data. S442. Operation-Fault Correlation Verification: Based on the predicted fault types of high-risk candidate devices, analyze the correlation between operation records and fault types, and verify whether there is corresponding operation behavior data within a preset time period before the operation. If corresponding operation behavior data exists and the device's detection data after the operation is qualified, the operation behavior data is determined to support the output results of the secondary classification model, and the verification passes; if corresponding operation behavior data exists but the detection data of the high-risk candidate device after the operation is abnormal, the operation is determined to have residual problems and needs to be reviewed in detail; if no corresponding operation behavior data exists, the confidence level of the predicted fault type is lowered. S443. Operation timeliness verification: If the operation record is less than the current time (which can be set to 7 days), and the operation type conflicts with the fault type output by the secondary classification model in step S4, then a secondary data acquisition is triggered to reacquire the equipment operation data and user feedback data in step S1 to verify the accuracy of the output results of the secondary classification model.

[0059] To make it easier to understand, here is an example: A high-risk candidate device in a certain distribution area is circuit breaker B. The secondary precision classification model predicts its fault type as "poor contact fault." After executing S441, check whether there have been any "component replacement" operations within the past 30 days. If so, and the post-operation test data shows a contact resistance value ≤ 50 ohms... If the operation supports the prediction result, the verification is passed; otherwise, if an operation exists but the detection data is abnormal (contact resistance value > 50), the verification is passed. If the operation is not found to have any outstanding issues, it is determined that there are problems that need to be reviewed. If there are no related operations, the prediction confidence level is lowered by 15%-20%. If the operation record shows that a poor contact was repaired 3 days ago, but the fault is still predicted, S443 is triggered to re-collect equipment operation data and user feedback data to verify the accuracy of the prediction.

[0060] In the example above, since the secondary classification model predicts the fault type as "poor contact fault," the core characteristic of this fault is "excessive contact resistance at the equipment connection (>50 ohms)." This can cause problems such as heat generation and voltage fluctuations when current flows through it. If the test data after operation shows a contact resistance ≤50... This indicates that the equipment contact status is normal after the operation, ruling out the current existence of "poor contact fault" based on physical parameters. However, combining the results of the first-level coarse prediction and the second-level fine classification, the core logic of "verification passed" here is: there is a "component replacement" operation (strongly correlated with "poor contact fault," consistent with historical data patterns); the normal detection data after the operation indicates that the fault may be newly added during use after the operation (such as vibration causing loose contacts), rather than a problem left over from the operation. Therefore, the prediction result is not negated. Instead, the contradiction of "operation correlation + current parameters normal but predicted fault" further focuses the direction of subsequent investigation (such as recent equipment operating environment, load changes). When the operation record is ≤7 days from the present, and the operation type conflicts with the predicted fault type (such as "insulation aging repair" just completed, but "insulation aging fault" is predicted), it indicates that the original fault type prediction contradicts the effect of recent maintenance operations. It is necessary to re-collect new data such as equipment operation data and user feedback to verify whether the "fault type" prediction is accurate, avoiding erroneous conclusions due to data lag or model misjudgment.

[0061] For example, the verification steps in step S4, which combine the four-level multi-source verification model with third-party data, include: S451. Third-party data matching: Based on the transformer area number and equipment number of the high-risk candidate equipment to be verified, match the corresponding meteorological warning data, third-party test report data and power safety platform warning data from the third-party monitoring data collected in step S1. S452. Fault Verification: Analyze the correlation between the predicted fault types of high-risk candidate devices and their matching third-party monitoring data, and verify the fault prediction results and confidence level based on the correlation.

[0062] To facilitate understanding, step S452 will be illustrated with an example: Example of fault verification based on meteorological data: If the secondary classification model in step S4 predicts the fault type as "water ingress fault of outdoor equipment", check the current and the meteorological warning data of the past 24 hours. If there is a rainstorm warning (rainfall ≥ 50mm) and the equipment is installed outdoors, then the meteorological data supports the prediction result and the verification passes; if there is no rainstorm warning, then mark it as "on-site inspection of equipment waterproofing measures is required".

[0063] Example of fault verification based on test report: If the secondary classification model in step S4 predicts the fault type as "insulation aging fault", retrieve the most recent insulating oil test report issued by a third-party testing agency. If the report shows a dielectric loss value > 0.005 (pass standard ≤ 0.005), the verification is successful. If the dielectric loss value is qualified but other indicators are abnormal (such as excessive moisture content), the predicted fault type is adjusted to "insulating oil contamination fault". If all indicators in the test report are qualified, the original prediction confidence level is lowered by 25%-30%.

[0064] For example, the steps for integrating and outputting the verification results of a four-level multi-source verification model include: S461. Verification Dimension Weight Allocation: Set the weight allocation for user feedback verification, operation behavior verification, and third-party data verification to 0.4, 0.3, and 0.3 respectively. Use the standardized dataset generated in step S2 as the unified benchmark for all-dimensional verification. Adapt the units, formats, and quantification standards (such as voltage deviation rate calculation rules and temperature fluctuation amplitude statistical standards) of the verification data from various sources to ensure that the comparison criteria of data from different sources are consistent. The verification results of each dimension are quantified and scored as "pass = 10 points, pending review = 5 points, fail = 0 points". S462. Calculation of comprehensive verification score: Calculate the comprehensive verification score of the high-risk candidate device to be verified based on the verification scores of each dimension and their corresponding weights. Specifically, the comprehensive verification score is calculated as follows: the comprehensive verification score of the device to be verified is calculated as "Comprehensive Verification Score = User Feedback Verification Score × 0.4 + Operation Behavior Verification Score × 0.3 + Third-Party Data Verification Score × 0.3", with a score range of 0-10. The validity of the verification data in each dimension should be determined with reference to the quality standards of the standardized dataset (such as data integrity ≥ 95% and outlier rate < 3%), and invalid verification data that does not meet the standards should be removed.

[0065] S463. Verification result determination: The verification result level is determined based on the comprehensive verification score. If the verification passes, the final fault prediction result is output. If the verification requires supplementary data, new data is collected and re-verified after preprocessing. If the verification fails, high-risk candidate equipment is removed. Specifically, the verification result is determined as follows: a comprehensive verification score of ≥8 is considered "verification passed", 5-7 is considered "requires supplementary data for review", and <5 is considered "verification failed". Devices that "pass verification" directly output the final fault prediction result (including fault type, risk level, and confidence level). Devices that "require supplementary data for review" trigger additional data collection (such as re-acquiring the device operation data and user feedback data from step S1). After collection, the data is processed into standardized data according to the standard of step S2, supplemented into the original standardized dataset, and then re-verified. Devices that "fail verification" are removed from the high-risk list and marked as "low-risk and under observation".

[0066] S464. Verification Report Generation: Generate a verification report that includes verification results for each dimension, comprehensive score, judgment conclusion, review suggestions, and standardized benchmark adaptation instructions. The "Standardized Benchmark Adaptation Instructions" should clearly specify the adaptation status of each source data with the standardized dataset (e.g., "Third-party meteorological data has been converted according to the standardized format, and the temperature unit is uniformly ℃"), and be stored in conjunction with the fault prediction results.

[0067] S5. Based on standardized data and fault prediction results, dynamically update model parameters through transfer learning algorithms and periodically evaluate the overall performance of the model to achieve adaptive optimization of the model.

[0068] In this embodiment, based on the historical fault data, equipment operating status data, power grid operation and control data, third-party monitoring data, user feedback and complaint suggestions data collected in step S1, and the output results of the four-level multi-source verification model in step S4, the hierarchical fault prediction model is adaptively optimized, and the model parameters are updated through transfer learning algorithm to adapt to the operating characteristics of different transformer substations.

[0069] For example, the specific steps of model adaptive optimization include: S51. Construct a training dataset: Divide the historical fault data collected in step S1 into a training set, a validation set, and a test set, with a ratio of 7:2:1. The historical fault data should include user feedback related cases collected in step S1, and incorporate fault cases caused by user triggers and harmonic triggers to enrich the data types of the training dataset. This training dataset will be used for transfer learning adaptation in step S52. S52. Transfer Learning Adaptation: Transfer the model parameters of mature distribution areas to newly put into operation distribution areas. This is especially suitable for new distribution areas with distributed power sources. Transfer learning can reduce the sample requirement by 40% and speed up the adaptation of models in new distribution areas. During the transfer learning process, the physical status data of the equipment in the new distribution area and the power consumption behavior data of the user side collected in step S1 should be referenced.

[0070] S53. Dynamic parameter update: Set a model performance monitoring threshold. When the prediction error exceeds 10% for 5 consecutive times or the false negative rate of new faults exceeds 5%, the parameter update process is automatically triggered. Incremental training is performed based on the user feedback cases added in step S1, the verification feedback data output in step S4, and the fault handling data. The model feature weights are updated, and the updated model parameters will be used to optimize the hierarchical fault prediction model in step S4.

[0071] S54. Periodic performance evaluation: Accuracy, recall, and F1 score are used as model performance evaluation indicators, and each indicator is required to be ≥88%. Periodic evaluation is used to ensure that the overall performance of the model meets the standards. During the evaluation process, the fault prediction results output in step S4 are compared with the actual fault situation. If the standards are not met, the parameter optimization process in step S53 is restarted.

[0072] S6. Obtain operation and maintenance resource data, and generate a fault solution based on the operation and maintenance resource data and the fault prediction results.

[0073] This embodiment generates a customized solution based on operation and maintenance resource data, transformer area characteristic data collected in step S1, user feedback and complaint suggestion data, and fault prediction results output in step S4; wherein, operation and maintenance resource data includes personnel configuration data, tool reserve data, and spare parts inventory data.

[0074] For example, the steps for generating a customized solution include: S61. Matching standard templates: Based on the correspondence between "fault type - fault cause", match the corresponding standard solution template from the dynamic mapping library; for example, the standard template corresponding to "insulation aging - excessive dielectric loss" is "replace insulating oil + dielectric loss value detection". The fault type and cause need to be determined based on the fault prediction results output in step S4. S62. Distribution Area Characteristics Adaptation: Adjust the standard template by combining the equipment nameplate parameters (such as transformer capacity and circuit breaker model) collected in step S1 with the user importance classification data; for example, when the transformer capacity is 500kVA, a maintenance plan of the corresponding specification needs to be matched. The solution for distribution areas where first-level users such as hospitals and large supermarkets are located needs to be configured with an additional backup power supply. This adaptation adjustment needs to refer to the user needs reflected in the user side power consumption behavior data in step S1. S63. Adapting to User Needs: Based on the specific needs in the user feedback and complaint suggestions collected in step S1, adjust the implementation details of the solution; for example, when users report that "power outages during the day affect online classes", adjust the construction period to 22:00-6:00 the next day to avoid affecting users' normal lives. The adjusted implementation details need to be included in the solution execution report in step S7. S64. Resource Constraint Adaptation: Query the inventory status of maintenance resources collected in step S1 in real time. If there is a shortage of spare parts, such as a lack of 500L insulating oil, adjust the solution to a combination of "temporary oil replenishment + allocation of sufficient spare parts within 2 hours" to ensure the feasibility of the solution. This adjustment result needs to be synchronized to the resource allocation plan in step S7.

[0075] S7. Generate an execution report based on the fault solution.

[0076] In this embodiment, a solution execution report is output based on the customized solution, including implementation steps, technical parameter requirements, resource allocation plan, expected effect evaluation, and solution adjustment mechanism. The implementation steps are divided into three stages: preparation, execution, and acceptance. The technical parameter requirements are used to clarify the qualified standards for fault handling. The resource allocation plan is formulated in combination with operation and maintenance resource data.

[0077] For example, the steps for generating a solution execution report include: S71. Implementation steps are broken down into three stages: preparation, execution, and acceptance. The customized solution implementation process generated in step S6 is broken down into three stages: preparation, execution, and acceptance. The time nodes for each stage are clearly defined. For example, the preparation stage is set to 0-2 hours, the execution stage is set to 2-5 hours, and the acceptance stage is set to 5-5.5 hours. The time nodes for each stage need to be determined in conjunction with the operation and maintenance resource scheduling efficiency data collected in step S1. S72. Establish technical parameters: Clarify the qualified standards for fault handling in each stage, such as insulating oil dielectric loss value ≤0.005 and equipment contact resistance ≤50μΩ; For extreme weather scenarios, such as rainstorms and high temperatures, the technical parameter requirements need to be adjusted in conjunction with the meteorological warning data collected in step S1. For example, tools with IP67 protection level need to be used when working in rainstorms. The technical parameter requirements need to refer to the equipment testing standards of the third-party testing agency in step S1. S73. Generate a resource plan: Define the scheduling path and arrival time for maintenance personnel (including electrician's name and skill level), tool requirements (including outbound order number), and spare parts allocation (including delivery method); for example, if two electricians with high-voltage operation qualifications are required to arrive at the work site before 22:00, the personnel and resource allocation must be based on the maintenance resource data collected in step S1 to ensure consistency with the resource constraint adaptation results in step S6; S74. Establish an adjustment mechanism: When there are emergencies such as resource shortages (e.g., delayed delivery of spare parts) or environmental changes (e.g., sudden rainstorms), the backup plan will be automatically triggered. For example, if spare parts cannot arrive on time, insulating oil regenerator will be used as a temporary substitute to ensure the continuous operation and maintenance work. The backup plan should refer to the real-time update of the third-party monitoring data in step S1 (e.g., weather warning upgrades).

[0078] S8. Collect operation and maintenance execution data, actual fault resolution status, and user satisfaction feedback on the solution. Compare and analyze these data with the fault prediction results output in step S4 and the customized solution generated in step S6 to form a closed-loop optimization mechanism. The satisfaction feedback adopts a 1-5 star rating system, and the comparative analysis needs to verify the implementation effect of the customized solution in step S6.

[0079] For example, the implementation steps of the closed-loop optimization mechanism include: S81. Data Collection: Comprehensively collect operation and maintenance execution data, actual fault resolution status, and user satisfaction feedback; operation and maintenance execution data includes implementation time, resource consumption, etc., and user satisfaction feedback adopts a 1-5 star rating system. This data needs to be compared with the expected effect evaluation in the solution execution report output in step S7. S82. Metric Calculation: Calculate core performance metrics, including prediction accuracy (the matching rate between the fault prediction results in step S4 and the actual fault situation), solution execution accuracy (the fit between the customized solution in step S6 and the actual operation and maintenance), feedback-fault matching rate (the correlation between user feedback data in step S1 and the actual fault), and solution satisfaction (user ratings for the solution in step S6); set qualification thresholds for the metrics, such as requiring a feedback-fault matching rate ≥75% and a solution satisfaction rate ≥85%. S83. Deviation Analysis: Analyze the reasons for deviations for indicators that do not meet the standards. If the feedback-fault matching rate is low, optimize the semantic parsing model in step S2 and add semantic-label mapping cases for new fault types. If the satisfaction is low, adjust the demand adaptation logic in step S6 and add a user preference survey step. The reasons for deviation need to be verified in conjunction with the new user feedback data collected in step S1. S84. Model and Template Update: Based on the deviation analysis results, optimize the feature extraction algorithm in step S3 and the parameter configuration of the hierarchical fault prediction model in step S4, and update the "fault-solution" dynamic mapping library; for example, add a standard rule of "arriving on-site to handle emergency faults within 2 hours". The updated model and template will be used to optimize the prediction in subsequent step S4 and the solution generation in step S6.

[0080] The embodiments of the present invention achieve the following effects: (1) Comprehensive coverage of multi-source data: By collecting multi-dimensional data such as equipment operation, power grid topology, user behavior, and third-party monitoring, a complete data support system is built to make up for the prediction limitations of a single data dimension and provide sufficient basis for accurate prediction; (2) Precise and efficient feature extraction: Dynamic feature engineering is carried out based on standardized datasets to extract multiple types of related features, strengthen the correlation between features and fault types, and improve the input quality of subsequent prediction models; (3) Clear hierarchical prediction logic: Through the hierarchical model of "coarse screening-fine classification-risk assessment-multi-source verification", we first focus on high-risk candidate equipment, then accurately determine the fault type and risk level, and finally significantly improve the accuracy and credibility of the prediction results through cross-validation of multi-source data. (4) Adapt to actual operation and maintenance needs: By integrating the characteristic data of the distribution area and the multi-source verification results, the output fault prediction results are highly targeted, which can provide accurate guidance for the allocation of operation and maintenance resources and the priority division of fault handling, reduce operation and maintenance costs, improve fault handling efficiency, and ensure the stability of power supply in low-voltage distribution areas.

[0081] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0082] The following are device embodiments of the present invention. For details not described in detail, please refer to the corresponding method embodiments described above.

[0083] Figure 3 The diagram shows a schematic representation of a low-voltage distribution area power equipment fault prediction device based on big data, as provided in an embodiment of the present invention. For ease of explanation, only the parts relevant to the embodiment of the present invention are shown, and are described in detail below: like Figure 3 As shown, the low-voltage distribution area power equipment fault prediction device 3 based on big data includes: Processing module 31 is used to acquire multi-source data from the low-voltage distribution area and preprocess it to obtain standardized data; and extract fault-related multi-dimensional features from the standardized data. The first-level coarse prediction module 32 is used to extract a first preset feature set from the multi-dimensional features, perform a first-level coarse prediction based on the first preset feature set, and determine high-risk candidate devices. The secondary fine classification module 33 is used to extract a second preset feature set from the multi-dimensional features, perform secondary fine classification based on the second preset feature set, and determine the fault type and confidence level of the high-risk candidate device. The three-level risk assessment module 34 is used to extract a third preset feature set from the multi-dimensional features, and perform a three-level risk assessment based on the third preset feature set, the fault type and confidence level of the high-risk candidate device to determine the risk level of the high-risk candidate device. The four-level multi-source verification module 35 is used to extract verification data from the standardized data, and after cross-verifying the fault type and confidence level and the risk level based on the verification data, output the fault prediction result.

[0084] In one possible implementation, the first-level coarse prediction module 32 is used for: Extract the core sub-features of environmental impact from the device operating status features, device physical status features, and third-party data association features, as well as the high-priority sub-features from the user feedback association features from the multi-dimensional features. Then, use the variance filtering method to delete redundant features with variance less than a preset variance threshold to obtain the first preset feature set. The first preset feature set is input into a pre-constructed first-level coarse prediction model, and the failure probability of each device is output. Devices whose probability of failure is greater than a preset probability threshold are identified as high-risk candidate devices.

[0085] In one possible implementation, the secondary fine classification module 33 is used for: The equipment operating status features, equipment physical status features, third-party data association features, user feedback association features, power grid control features, power grid control association features, and operation behavior association features are extracted from the multidimensional features, and mutual information entropy is calculated to filter features to obtain the second preset feature set. The second preset feature set is input into a pre-constructed hybrid algorithm model of convolutional neural network and long short-term memory network, and the fault type and confidence level of the high-risk candidate device are output.

[0086] In one possible implementation, the three-level risk assessment module 34 is used for: Extract user feedback correlation features and station area features from the multidimensional features; Based on the user feedback correlation characteristics and the transformer area characteristics, several core evaluation indicators are determined; Based on the aforementioned core evaluation indicators and the risk assessment system established in advance through the analytic hierarchy process, the comprehensive risk value of the high-risk candidate equipment is determined. The risk level of the high-risk candidate device is determined based on its comprehensive risk value, fault type, and confidence level.

[0087] In one possible implementation, the multiple core evaluation metrics include: fault impact range, fault urgency, equipment importance, user importance, and fault propagation probability.

[0088] In one possible implementation, the four-level multi-source verification module 35 is used for: User feedback data, operational behavior data, and third-party data are extracted from the standardized data. Based on the user feedback data, the operation behavior data, and the third-party data, cross-validation is performed on the fault type and confidence level, and the risk level to obtain validation scores for multiple validation dimensions. The verification scores of the multiple verification dimensions are weighted and fused to calculate the comprehensive verification score; If the comprehensive verification score is satisfactory, the fault prediction result will be output. If the comprehensive verification score is unqualified, the high-risk candidate device will be deleted.

[0089] In one possible implementation, after the output fault prediction result, the four-level multi-source verification module 35 is further configured to: Based on the standardized data and the fault prediction results, the model parameters are dynamically updated through transfer learning algorithms, and the overall performance of the model is evaluated periodically to achieve adaptive optimization of the model.

[0090] In one possible implementation, after the output fault prediction result, the four-level multi-source verification module 35 is further configured to: Obtain operation and maintenance resource data, and generate a fault solution based on the operation and maintenance resource data and the fault prediction results; Based on the fault solution, an execution report is generated; wherein the execution report includes: implementation steps, technical parameter requirements, resource allocation plan, expected effect evaluation and solution adjustment mechanism.

[0091] In one possible implementation, the preprocessing includes: Data cleaning, outlier removal, missing value imputation, data standardization, cross-dimensional data consistency verification, multi-source cross-validation, semantic parsing and noise reduction of user feedback data.

[0092] This invention first acquires and preprocesses multi-source data from low-voltage distribution areas to obtain standardized data, thus overcoming the predictive limitations of single-data-dimensional data and providing sufficient basis for accurate prediction. Next, fault-related multi-dimensional features are extracted from the standardized data to strengthen the correlation between features and fault types, thereby improving the input quality of subsequent prediction models.

[0093] This invention extracts a first preset feature set from multi-dimensional features, performs a first-level coarse prediction based on the first preset feature set to identify high-risk candidate devices; extracts a second preset feature set from multi-dimensional features, performs a second-level fine classification based on the second preset feature set to determine the fault type and confidence level of the high-risk candidate devices; extracts a third preset feature set from multi-dimensional features, performs a third-level risk assessment based on the third preset feature set, the fault type and confidence level of the high-risk candidate devices, and determines the risk level of the high-risk candidate devices; and extracts validation data from standardized data, performs cross-validation on the fault type, confidence level, and risk level based on the validation data, and outputs the fault prediction result. Through a hierarchical model of "coarse screening-fine classification-risk assessment-multi-source verification," it first focuses on high-risk candidate devices, then accurately determines the fault type and risk level, and finally, through multi-source data cross-validation, significantly improves the accuracy and reliability of the prediction results. The output fault prediction results are highly targeted and can provide precise guidance for the allocation of operation and maintenance resources and the prioritization of fault handling, reducing operation and maintenance costs, improving fault handling efficiency, and ensuring the stability of power supply in low-voltage distribution areas.

[0094] Figure 4 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. For example... Figure 4 As shown, the electronic device 4 in this embodiment includes a processor 40 and a memory 41. The memory 41 stores a computer program 42. When the processor 40 executes the computer program 42, it implements the steps in the various method embodiments described above. Alternatively, when the processor 40 executes the computer program 42, it implements the functions of each module in the various device embodiments described above.

[0095] For example, computer program 42 may be divided into one or more modules / units, which are stored in memory 41 and executed by processor 40 to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of computer program 42 in electronic device 4.

[0096] Electronic device 4 may include, but is not limited to, processor 40 and memory 41. Those skilled in the art will understand that... Figure 4 This is merely an example of electronic device 4 and does not constitute a limitation on electronic device 4. It may include more or fewer components than shown, or combine certain components, or different components. For example, electronic device 4 may also include input / output devices, network access devices, buses, etc.

[0097] For the sake of simplicity and clarity, only the above-described functional modules / units are used as examples. In practical applications, the functions described above can be assigned to different functional modules / units as needed. These modules / units can be implemented in hardware, software, or a combination of both.

[0098] In the above embodiments, the descriptions of each embodiment have their own emphasis. Parts not detailed or described in a particular embodiment can be referred to in the relevant descriptions of other embodiments. Unless otherwise specified or in conflict with logic, the terminology and / or descriptions between different embodiments are consistent and can be referenced interchangeably. Technical features in different embodiments can be combined to form new embodiments based on their inherent logical relationships.

[0099] The above-described 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 the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for predicting faults in low-voltage distribution area power equipment based on big data, characterized in that, include: Acquire multi-source data from low-voltage distribution areas and preprocess it to obtain standardized data; Extract fault-related multidimensional features from the standardized data; A first preset feature set is extracted from the multidimensional features, and a first-level coarse prediction is performed based on the first preset feature set to determine high-risk candidate devices. A second preset feature set is extracted from the multidimensional features, and a secondary fine classification is performed based on the second preset feature set to determine the fault type and confidence level of the high-risk candidate device. A third preset feature set is extracted from the multidimensional features. Based on the third preset feature set, the fault type and confidence level of the high-risk candidate device, a three-level risk assessment is performed to determine the risk level of the high-risk candidate device. Validation data is extracted from the standardized data. Based on the validation data, the fault type and confidence level and the risk level are cross-validated, and the fault prediction result is output.

2. The method for predicting faults in low-voltage distribution area power equipment based on big data as described in claim 1, characterized in that, The step of extracting a first preset feature set from the multidimensional features and performing a first-level coarse prediction based on the first preset feature set to determine high-risk candidate devices includes: Extract the core sub-features of environmental impact from the device operating status features, device physical status features, and third-party data association features, as well as the high-priority sub-features from the user feedback association features from the multi-dimensional features. Then, use the variance filtering method to delete redundant features with variance less than a preset variance threshold to obtain the first preset feature set. The first preset feature set is input into a pre-constructed first-level coarse prediction model, and the failure probability of each device is output. Devices whose probability of failure is greater than a preset probability threshold are identified as high-risk candidate devices.

3. The method for predicting faults in low-voltage distribution area power equipment based on big data as described in claim 1, characterized in that, The step of extracting a second preset feature set from the multidimensional features, performing secondary fine classification based on the second preset feature set, and determining the fault type and confidence level of the high-risk candidate device includes: The equipment operating status features, equipment physical status features, third-party data association features, user feedback association features, power grid control features, power grid control association features, and operation behavior association features are extracted from the multidimensional features, and mutual information entropy is calculated to filter features to obtain the second preset feature set. The second preset feature set is input into a pre-constructed hybrid algorithm model of convolutional neural network and long short-term memory network, and the fault type and confidence level of the high-risk candidate device are output.

4. The method for predicting faults in low-voltage distribution area power equipment based on big data according to claim 1, characterized in that, The step of extracting a third preset feature set from the multidimensional features, and performing a three-level risk assessment based on the third preset feature set, the fault type and confidence level of the high-risk candidate device, to determine the risk level of the high-risk candidate device includes: Extract user feedback correlation features and station area features from the multidimensional features; Based on the user feedback correlation characteristics and the transformer area characteristics, several core evaluation indicators are determined; Based on the aforementioned core evaluation indicators and the risk assessment system established in advance through the analytic hierarchy process, the comprehensive risk value of the high-risk candidate equipment is determined. The risk level of the high-risk candidate device is determined based on its comprehensive risk value, fault type, and confidence level.

5. The method for predicting faults in low-voltage distribution area power equipment based on big data according to claim 4, characterized in that, The core evaluation metrics include: scope of fault impact, urgency of fault, equipment importance, user importance, and probability of fault propagation.

6. The method for predicting faults in low-voltage distribution area power equipment based on big data according to claim 1, characterized in that, The step of extracting validation data from the standardized data, and then cross-validating the fault type and confidence level based on the validation data, and outputting the fault prediction result, includes: User feedback data, operational behavior data, and third-party data are extracted from the standardized data. Based on the user feedback data, the operation behavior data, and the third-party data, cross-validation is performed on the fault type and confidence level, and the risk level to obtain validation scores for multiple validation dimensions. The verification scores of the multiple verification dimensions are weighted and fused to calculate the comprehensive verification score; If the comprehensive verification score is satisfactory, the fault prediction result will be output. If the comprehensive verification score is unqualified, the high-risk candidate device will be deleted.

7. The method for predicting faults in low-voltage distribution area power equipment based on big data according to any one of claims 1 to 6, characterized in that, Following the output of the fault prediction result, the following is also included: Based on the standardized data and the fault prediction results, the model parameters are dynamically updated through transfer learning algorithms, and the overall performance of the model is evaluated periodically to achieve adaptive optimization of the model.

8. The method for predicting faults in low-voltage distribution area power equipment based on big data according to any one of claims 1 to 6, characterized in that, Following the output of the fault prediction result, the following is also included: Obtain operation and maintenance resource data, and generate a fault solution based on the operation and maintenance resource data and the fault prediction results; Based on the fault solution, an execution report is generated; wherein the execution report includes: implementation steps, technical parameter requirements, resource allocation plan, expected effect evaluation and solution adjustment mechanism.

9. A fault prediction device for low-voltage distribution area power equipment based on big data, characterized in that, include: The processing module is used to acquire multi-source data from low-voltage distribution areas and preprocess it to obtain standardized data; Extract fault-related multidimensional features from the standardized data; The first-level coarse prediction module is used to extract a first preset feature set from the multi-dimensional features, perform a first-level coarse prediction based on the first preset feature set, and determine high-risk candidate devices. The secondary fine classification module is used to extract a second preset feature set from the multi-dimensional features, perform secondary fine classification based on the second preset feature set, and determine the fault type and confidence level of the high-risk candidate device. The three-level risk assessment module is used to extract a third preset feature set from the multi-dimensional features, and perform a three-level risk assessment based on the third preset feature set, the fault type and confidence level of the high-risk candidate device, to determine the risk level of the high-risk candidate device. The four-level multi-source verification module is used to extract verification data from the standardized data, and after cross-validating the fault type and confidence level and the risk level based on the verification data, output the fault prediction result.

10. An electronic device, characterized in that, It includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method as described in any one of claims 1 to 8.

Images