A power grid operation implicit fault intelligent research and judgment method and system

Abstract

The present application relates to a kind of power grid operation implicit fault intelligent research and system, belong to power system fault monitoring technical field.The present application collects the multi-source data of SCADA, protection device and online monitoring terminal, after washing normalization, extract 9 steady-state core features and 10 transient fault core features.Construct the double-layer architecture consisting of steady-state operation knowledge base and transient operation knowledge base, the former stores the normal operation threshold interval under multiple operating conditions, the latter stores fault feature weight and disposal strategy.Improved random forest algorithm is used to train intelligent research model, real-time access power grid data, combined with current operating mode and operating condition correction coefficient output fault type, location and occurrence probability.Through calculating feature deviation value and deviation growth rate k, realize fault risk classification and deterioration trend prediction.The present application realizes the unification, precision, initiative research and judgment of power grid primary and secondary equipment global implicit fault, significantly improves the intelligent level of power grid operation.

Description

Technical Field

[0001] This invention relates to a method and system for intelligent assessment of latent faults in power grid operation. Background Technology

[0002] Latent faults in power grids are characterized by their concealment, latency, and gradual degradation. In the early stages, there are no obvious abnormal operating conditions. Traditional manual inspections and data analysis are currently the mainstream approach for identifying latent faults in power grid operation and maintenance. However, existing manual techniques have the following limitations:

[0003] First, manual fault analysis relies heavily on the professional experience of operation and maintenance personnel. Due to limitations in personnel's professional level and energy, it is impossible to comprehensively sort out massive historical power grid operation data and fault ledgers, resulting in insufficient fault feature mining and a high rate of missed and misjudged hidden faults.

[0004] Second, manually checking each power grid operating parameter is time-consuming and has poor real-time performance. Some hidden faults have a short interval from budding to sudden failure. Delayed manual investigation can easily cause line tripping and equipment damage, threatening the safe operation of the power grid.

[0005] Third, the lack of a systematic knowledge base to accumulate historical power grid operation experience and fault characteristics means that the applicability of the original manual judgment experience decreases after the power grid operation mode is adjusted, further exacerbating the judgment bias.

[0006] With the advancement of new power system construction, the power grid structure is becoming increasingly complex, and new energy power plants are being connected to the grid on a large scale. As a result, the power grid operation monitoring data is growing explosively, and the traditional manual judgment mode can no longer meet the needs of accurate and rapid judgment of hidden faults under massive data.

[0007] The existing patent application CN117521804A, "A Knowledge Graph-Based Method for Identifying and Reasoning Latent Faults in Relay Protection," has the following drawbacks: First, the current patent's technical system, feature rules, and knowledge graph are all designed around secondary relay protection equipment. It can only identify secondary-side specific latent faults such as loose connections in secondary circuits, incorrect setting values, and incorrect activation / deactivation of pressure plates, and cannot adapt to latent faults in primary power grid equipment such as busbars, transmission lines, main transformers, and switchgear. Second, the current patent only builds a single fault knowledge graph, focusing only on matching fault anomaly features, and does not construct a benchmark knowledge base for normal steady-state operation of the power grid. It lacks standardized, multi-condition normal operation threshold ranges. It cannot effectively distinguish between normal load fluctuations, operating condition adjustment disturbances, and actual latent fault anomalies, and is highly prone to misjudging normal operating condition fluctuations as faults, exhibiting core defects such as high misjudgment rate, poor anti-interference ability, and insufficient judgment accuracy. Third, the patent can only complete static identification and reasoning for latent faults that have already occurred; it cannot quantify the degree of fault deterioration, calculate the fault deterioration rate, or predict the fault development trend. Summary of the Invention

[0008] The purpose of this invention is to address the problems of existing methods for identifying latent faults, such as reliance on manual labor, limited application scenarios, high misjudgment rates, and lack of trend prediction. It provides an intelligent method and system for identifying latent faults in power grid operation. This involves collecting power grid operation data and historical fault data, completing data regularization and extracting key information and fault features, and building a layered steady-state operation knowledge base and a transient operation knowledge base. It utilizes artificial intelligence algorithms to train a fault analysis model, combining real-time power grid operation data, historical knowledge base experience, and the current actual operation mode of the power grid to achieve intelligent and rapid identification of latent faults. Compared to traditional manual fault investigation and analysis methods, this invention effectively improves the accuracy and real-time performance of latent fault identification, enabling early identification of hidden power grid faults, preventing the risk of fault escalation from the source, and ensuring the overall safe and stable operation of the power grid. This invention can be widely applied to the daily operation and maintenance and fault pre-analysis scenarios of transmission and distribution networks, specifically achieving the following functions:

[0009] (1) Dual knowledge base construction: a steady-state operation knowledge base is built based on the extracted steady-state key information, and a transient operation knowledge base is built based on transient fault characteristics, historical fault cases and fault handling experience;

[0010] (2) Online intelligent analysis: Real-time access to on-site measured data of power grid operation, extraction of key operational information and fault characteristics from real-time data, input into the trained intelligent analysis model, combined with the current actual operation mode of the power grid system and the historical operation experience stored in the knowledge base, automatically completes rapid analysis and judgment of hidden faults in the power grid.

[0011] To achieve the above objectives, the technical solution of the present invention is: an intelligent method for judging latent faults in power grid operation, comprising:

[0012] The power grid steady-state operation data and historical fault data are collected as raw data. After preprocessing the raw data, standardized steady-state datasets and standardized fault datasets are obtained.

[0013] Extract a set of steady-state core features from a standardized steady-state dataset. The set of steady-state core features includes voltage deviation rate, three-phase voltage imbalance, load stability coefficient, active load rate, reactive load rate, power factor, equipment steady-state temperature rise, line impedance steady-state value, and bus frequency deviation.

[0014] Extract the transient fault core feature set from the standardized fault dataset. The transient fault core feature set includes voltage fluctuation coefficient, current fluctuation coefficient, total harmonic distortion rate, leakage current mutation coefficient, joint temperature rise mutation rate, voltage sag depth, load instantaneous fluctuation coefficient, phase offset, insulation resistance attenuation rate, and high-frequency pulse distortion rate.

[0015] A steady-state operation knowledge base is constructed based on the steady-state core feature set, and a transient operation knowledge base is constructed based on the transient fault core feature set and historical fault cases.

[0016] A smart analysis model for hidden faults in the power grid is trained using a steady-state operation knowledge base and a transient operation knowledge base.

[0017] Real-time power grid operation data is collected, real-time steady-state features and real-time transient features are extracted, and input into the trained intelligent analysis model of hidden faults in the power grid. Combined with the current power grid operation mode, the model outputs the results of hidden fault assessment.

[0018] Furthermore, the step of constructing the steady-state operation knowledge base includes:

[0019] The standardized steady-state dataset is divided into multiple operating condition categories according to the power grid operating conditions;

[0020] For each operating condition category, calculate the mean of each feature in the steady-state core feature set. and standard deviation ;

[0021] According to the formula Calculate the normal operating threshold range for each steady-state characteristic under various operating conditions. , ];

[0022] The operating condition category, feature name, mean, standard deviation, and corresponding threshold interval are associated and stored to form a steady-state operation knowledge base.

[0023] Furthermore, the step of constructing the transient operation knowledge base includes:

[0024] Calculate the Pearson correlation coefficients between each feature in the core feature set of transient faults and the preset fault types. ;

[0025] According to the formula Calculate the weighting coefficients of each transient fault characteristic. ;

[0026] Establish the correlation between fault types and corresponding transient fault feature combinations, feature weight coefficients, anomaly thresholds, and risk levels;

[0027] The relationships and historical fault handling strategies are stored in a structured manner to form a transient operation knowledge base.

[0028] Furthermore, the steps for training the intelligent analysis model for latent faults in the power grid include:

[0029] By integrating the benchmark data in the steady-state operation knowledge base and the fault feature data in the transient operation knowledge base, a training sample set containing negative samples and positive samples is constructed, where the negative samples are fault-free steady-state operating condition data and the positive samples are latent fault operating condition data.

[0030] An improved random forest algorithm was used to build an intelligent analysis model for hidden faults in power grids. The input layer of the intelligent analysis model for hidden faults in power grids was set to 19-dimensional nodes, which correspond to 9 steady-state core features and 10 transient fault core features, respectively.

[0031] The intelligent analysis model for hidden faults in the power grid is iteratively trained using a training sample set, and the cross-entropy loss function is applied. By optimizing the model parameters, a fixed intelligent analysis model for hidden faults in the power grid is obtained.

[0032] Furthermore, after the step of outputting the latent fault assessment result, the method further includes:

[0033] According to the formula Calculate real-time feature deviation value ,in For real-time feature parameters, The mean of steady-state characteristics, For feature weights;

[0034] According to the formula Initial assessment results of the model output Perform operating condition corrections, including This is the operating condition correction factor determined based on the current power grid load fluctuations and grid adjustment status;

[0035] According to the formula Calculate the failure risk score ,in The probability of failure occurring. This is the maximum permissible deviation value. , These are the weighting coefficients;

[0036] Based on risk scoring Classify the faults and match them with corresponding handling strategies.

[0037] Furthermore, it also includes fault location steps:

[0038] According to the formula Calculate the fault distance coefficient ,in To monitor the characteristics of the points in real time, This serves as a baseline value for fault characteristics.

[0039] Fault distance coefficient The smallest location is determined as the hidden fault location.

[0040] Furthermore, it also includes a failure trend prediction step:

[0041] Calculation of fault deviation growth rate based on least squares method ;

[0042] According to the formula Predict the fault characteristic deviation value at time t, where This is the current deviation value;

[0043] According to the growth rate The numerical range defines the level of fault degradation.

[0044] This invention also provides an intelligent system for assessing latent faults in power grid operation, which executes the method described above, including:

[0045] The data acquisition module is used to collect steady-state operation data, historical fault data, and real-time operation data of the power grid;

[0046] The data processing module is used to clean, denoise, and normalize the collected data, and extract the steady-state core feature set and the transient fault core feature set.

[0047] The knowledge base construction module is used to construct a steady-state operation knowledge base based on the steady-state core feature set, and to construct a transient operation knowledge base based on the transient fault core feature set and historical fault cases.

[0048] The model training module is used to train an intelligent analysis model for hidden faults in the power grid using the steady-state operation knowledge base and the transient operation knowledge base;

[0049] The online analysis module is used to receive real-time extracted feature data and current power grid operation parameters, and call the trained intelligent analysis model to output the latent fault analysis results.

[0050] The results output module is used to output the fault location, risk level, and early warning information.

[0051] Furthermore, the knowledge base construction module includes:

[0052] The operating condition classification unit is used to divide steady-state data into peak load operating conditions, low load operating conditions, line loop closing operating conditions, and line loop breaking operating conditions.

[0053] The threshold calculation unit is used to calculate the mean, standard deviation, and threshold interval based on the [μs−2σs,μs+2σs] rule for steady-state characteristics under various operating conditions.

[0054] The weight calculation unit is used to calculate the weight coefficients of transient fault characteristics based on the Pearson correlation coefficient.

[0055] The storage unit is used to store the data of the steady-state operating knowledge base and the transient operating knowledge base in a multi-dimensional tag index, and is configured with an incremental update interface.

[0056] Furthermore, the online analysis module is also configured as follows:

[0057] Calling formula Calculate the characteristic deviation value. This represents the real-time feature deviation value. For real-time feature parameters, The mean of steady-state characteristics, For feature weights;

[0058] Calling formula Introducing operating condition correction factor Correct the model output. This is the initial assessment result;

[0059] Calling formula Locate the fault location. This is the fault distance coefficient. Features of real-time monitoring points; This serves as a baseline value for fault characteristics.

[0060] Calling formula Perform fault trend prediction. This is the current deviation value. This represents the rate of increase in fault deviation.

[0061] Compared with the prior art, the present invention has the following beneficial effects:

[0062] Technical effects:

[0063] (1) Significantly improve the accuracy and real-time performance of hidden fault assessment, and reduce the rate of missed and false judgments.

[0064] (2) Achieve full-scenario adaptation of the power grid and break through the limitations of existing technology, equipment and scenarios.

[0065] Socioeconomic benefits:

[0066] (1) Reduce the economic losses caused by power outages due to sudden power grid failures.

[0067] (2) Improve the reliability of regional power supply and ensure the stability of basic electricity use for people's livelihood. Attached Figure Description

[0068] Figure 1 This is a flowchart of the intelligent assessment process for latent faults in the power grid according to the present invention. Detailed Implementation

[0069] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings.

[0070] This invention provides an intelligent method for identifying latent faults in power grid operation, comprising:

[0071] The power grid steady-state operation data and historical fault data are collected as raw data. After preprocessing the raw data, standardized steady-state datasets and standardized fault datasets are obtained.

[0072] Extract a set of steady-state core features from a standardized steady-state dataset. The set of steady-state core features includes voltage deviation rate, three-phase voltage imbalance, load stability coefficient, active load rate, reactive load rate, power factor, equipment steady-state temperature rise, line impedance steady-state value, and bus frequency deviation.

[0073] Extract the transient fault core feature set from the standardized fault dataset. The transient fault core feature set includes voltage fluctuation coefficient, current fluctuation coefficient, total harmonic distortion rate, leakage current mutation coefficient, joint temperature rise mutation rate, voltage sag depth, load instantaneous fluctuation coefficient, phase offset, insulation resistance attenuation rate, and high-frequency pulse distortion rate.

[0074] A steady-state operation knowledge base is constructed based on the steady-state core feature set, and a transient operation knowledge base is constructed based on the transient fault core feature set and historical fault cases.

[0075] A smart analysis model for hidden faults in the power grid is trained using a steady-state operation knowledge base and a transient operation knowledge base.

[0076] Real-time power grid operation data is collected, real-time steady-state features and real-time transient features are extracted, and input into the trained intelligent analysis model of hidden faults in the power grid. Combined with the current power grid operation mode, the model outputs the results of hidden fault assessment.

[0077] This invention also provides an intelligent system for assessing latent faults in power grid operation, which executes the method described above, including:

[0078] The data acquisition module is used to collect steady-state operation data, historical fault data, and real-time operation data of the power grid;

[0079] The data processing module is used to clean, denoise, and normalize the collected data, and extract the steady-state core feature set and the transient fault core feature set.

[0080] The knowledge base construction module is used to construct a steady-state operation knowledge base based on the steady-state core feature set, and to construct a transient operation knowledge base based on the transient fault core feature set and historical fault cases.

[0081] The model training module is used to train an intelligent analysis model for hidden faults in the power grid using the steady-state operation knowledge base and the transient operation knowledge base;

[0082] The online analysis module is used to receive real-time extracted feature data and current power grid operation parameters, and call the trained intelligent analysis model to output the latent fault analysis results.

[0083] The results output module is used to output the fault location, risk level, and early warning information.

[0084] The following is a detailed implementation process of the present invention.

[0085] This invention discloses an intelligent method for identifying latent faults in power grid operation. The overall process is as follows: Figure 1 As shown, the main implementation process includes the following steps:

[0086] Step 1 (Multi-source data acquisition and processing): Specific implementation techniques and methods:

[0087] (1) Multi-source data full-domain acquisition: Collect steady-state operation data and historical fault data of the entire power grid. Data sources cover SCADA system, relay protection device, equipment online monitoring terminal, and dispatch ledger system. Among them, steady-state data includes continuous monitoring data of normal operating conditions such as bus voltage, line current, active power, reactive power, equipment temperature, and load rate; historical fault data includes hidden fault ledger, equipment defect record, fault alarm data, and fault handling review data.

[0088] (2) Data classification and archiving: The collected data is divided into steady-state raw datasets. Original dataset of faults The data sampling frequency was standardized at 50ms / point, and the data timestamp format and device number coding rules were standardized to complete the initial classification and organization.

[0089] (3) Data cleaning and noise reduction: Preprocessing is performed on missing values, abnormal jump values, and redundant duplicate values ​​in the original data to remove invalid data and retain valid operating condition data and fault data.

[0090] (4) Standardization processing: unify all data units (voltage: kV, current: A, power: MW), complete data normalization processing, generate standardized original datasets, and provide compliant input data for subsequent feature extraction.

[0091] (5) Data output: Output standardized steady-state dataset Standardized fault dataset This provides unique original input data for feature extraction in step S2 (step two).

[0092] Step Two (Intelligent Inspection and Plan Comparison at Intervals) Specific Implementation Techniques:

[0093] (1) Steady-state feature extraction: Based on the standardized steady-state dataset output by S1 (step one) It extracts the core steady-state characteristics of the power grid's normal operation from all dimensions, without omissions or ambiguities. Specifically, it includes nine core steady-state parameters: voltage deviation rate, three-phase voltage imbalance, load stability coefficient, active load rate, reactive load rate, power factor, equipment steady-state temperature rise, line impedance steady-state value, and bus frequency deviation, forming a complete and standardized set of steady-state characteristics.

[0094] (2) Transient fault feature extraction: based on the standardized fault dataset output by S1 It comprehensively uncovers the unique transient characteristics of hidden faults in the power grid, covering various hidden fault characteristics such as insulation degradation, loose connections, line aging, harmonic anomalies, and load fluctuations. Specifically, it includes ten core transient fault parameters: voltage fluctuation coefficient, current fluctuation coefficient, total harmonic distortion rate, leakage current mutation coefficient, joint temperature rise mutation rate, voltage sag depth, load instantaneous fluctuation coefficient, phase offset, insulation resistance attenuation rate, and high-frequency pulse distortion rate, forming a comprehensive set of transient fault characteristics.

[0095] (3) Feature selection optimization: Remove redundant features with extremely low correlation and no fault correlation, retain high-sensitivity core features, distinguish between steady-state feature parameter groups and transient fault feature parameter groups, and complete feature quantification and annotation.

[0096] Steady-state core feature set The complete definitions of the 9 parameters are as follows: 1. Voltage offset rate; 2. 3. Three-phase voltage imbalance; : Load stability coefficient; 4. Active load factor; 5. : Reactive load rate; 6. Power factor; 7. 8. Steady-state temperature rise of the equipment; : Steady-state value of line impedance; 9. Bus frequency deviation.

[0097] Transient Fault Core Feature Set The complete definitions of the 10 parameters are as follows: 1. Voltage fluctuation coefficient; 2. : Current fluctuation coefficient; 3. : Total harmonic distortion; 4. Leakage current mutation coefficient; 5. : Joint temperature rise abrupt change rate; 6. : Depth of voltage sag; 7. : Instantaneous load fluctuation coefficient; 8. Phase offset; 9. Insulation resistance attenuation rate; 10. High-frequency pulse distortion rate.

[0098] Feature correlation screening model (Pearson correlation coefficient): Eliminating redundant features

[0099]

[0100] In the formula: The correlation coefficient between the feature and the fault is |r|>0.3. If the feature is considered valid, it will be retained; otherwise, it will be considered redundant and will be removed.

[0101] (4) Data output: Output steady-state core feature set Transient Fault Core Feature Set This provides core feature data for the construction of the S3 knowledge base and the training of the S4 model.

[0102] Step 3 (Standardized Comparison of Power Outage and Restoration Signal Sections) Specific Implementation Techniques:

[0103] (1) Construction of Steady-State Operation Knowledge Base

[0104] 1. Multi-condition data classification and aggregation. Retrieve the S1 standardized steady-state dataset. The S2 steady-state nine core characteristic parameters (voltage deviation rate, three-phase voltage imbalance, load stability coefficient, active load rate, reactive load rate, power factor, equipment steady-state temperature rise, line impedance steady-state value, and bus frequency deviation) are divided into four scenarios according to typical power grid operating conditions: peak load condition, off-peak load condition, line loop closing condition, and line loop breaking condition. The characteristic data is classified and archived according to the scenarios to prevent the deviation of the benchmark threshold caused by mixed operating conditions.

[0105] 2. Solving for Steady-State Characteristic Baseline Parameters. For each type of operating condition, all steady-state characteristic sampling data are traversed, and the mean and standard deviation are calculated using the formulas for calculating the mean and standard deviation to determine the mean of the nine steady-state characteristics under each operating condition. Standard deviation Then, relying on the steady-state parameter threshold interval model:

[0106]

[0107] In the formula: , These are the upper and lower threshold values ​​for steady-state parameters; , These are the mean and standard deviation of the steady-state characteristic parameters, respectively. The range represents the normal steady-state operating condition, and the range represents the abnormal operating condition. The upper and lower limits of the normal operating threshold for each steady-state characteristic under different operating conditions are calculated to form a multi-operating-condition-multi-feature benchmark parameter matrix.

[0108] 3. Steady-state knowledge is structured and stored in the database. A steady-state knowledge base data table is constructed, with fields including: operating condition type, characteristic parameter name, characteristic symbol, benchmark mean, upper limit of standard threshold, lower limit of standard threshold, compliance judgment rule, and equipment compatibility model. All steady-state benchmark parameters and threshold ranges obtained in step 2 are sequentially stored in the database, forming a standard knowledge base for normal power grid operation covering all operating conditions and all steady-state characteristics. This knowledge base is used for subsequent real-time compliance comparison and anomaly identification.

[0109] (2) Construction of Transient Operation Knowledge Base

[0110] 1. Fault Feature Weighting. Retrieve the 10 core transient fault features output from S2 and their Pearson correlation coefficients. Through feature matching weight model:

[0111]

[0112] In the formula: The weight coefficient for the i-th fault feature; For the correlation coefficient between the corresponding feature and the fault (S2 output), all weights satisfy... The weights are stored in the transient knowledge base for subsequent analysis and correction; the fault contribution weight corresponding to each transient fault feature is calculated, the feature weight is quantified and assigned, the weight data is permanently bound to the corresponding fault feature and stored in the transient knowledge base, providing a basis for subsequent weighted correction of fault analysis.

[0113] 2. Historical Fault Case Association Matching. Collect historical hidden fault ledgers, equipment defect records, fault review data, and operation and maintenance handling records of the power grid. Identify various typical hidden faults such as insulation degradation, loose connections, line aging, harmonic anomalies, and partial discharge. Bind each type of fault to 10 corresponding transient abnormal characteristics to establish a fixed association relationship of "fault type - abnormal characteristic combination".

[0114] 3. Encapsulation of Fault Risk and Handling Knowledge. For each type of latent fault, based on the historical fault severity, fault deterioration rate, and impact scope, it is divided into three risk levels: high, medium, and low. At the same time, it matches the corresponding fault causes, fault development trends, standardized operation and maintenance handling plans, and troubleshooting processes, forming a six-in-one transient fault knowledge system of "fault type - feature weight - anomaly threshold - risk level - cause analysis - handling strategy".

[0115] 4. Standardize and store transient knowledge. Build a transient knowledge base data table. The data table fields include: fault type, corresponding transient feature set, feature weight, fault anomaly threshold, risk level, fault cause, development trend, handling plan, and number of historical fault samples. Batch and structure all quantitative fault knowledge, case knowledge, and operation and maintenance knowledge are stored in the database to form a fully covered implicit fault knowledge base.

[0116] (3) Global index construction and knowledge base iterative optimization

[0117] 1. Multi-dimensional tag index construction: All data in the steady-state knowledge base and transient knowledge base are tagged, and operating condition tags, equipment type tags, feature parameter tags, fault type tags, and risk level tags are set to build a multi-dimensional retrieval index, which supports millisecond-level accurate retrieval of benchmark data, fault knowledge, and feature weights during subsequent model training and online judgment.

[0118] 2. Dynamic update mechanism configuration: Configure incremental update interface for dual knowledge bases and set periodic update rules. It can access new steady-state operation data and new hidden fault cases of the power grid in real time, and automatically complete feature calculation, threshold update, weight correction and knowledge iteration to realize continuous optimization of knowledge base and adapt to changes in operating conditions caused by power grid structure adjustment, load change and equipment aging.

[0119] (4) Data output

[0120] Output steady-state knowledge base dataset Transient knowledge base dataset (Including fault characteristics, weights, fault types, and risk level benchmarks), providing sample datasets for S4 model training and knowledge support for S5 online assessment.

[0121] Step Four (Artificial Intelligence Model Training): Specific Implementation Techniques

[0122] The steady-state knowledge base dataset output by S3 (step three) Transient knowledge base dataset Using S2 as the sole core input, and relying on 9 steady-state features and 10 transient fault features selected through S2, combined with feature weights and fault label data, the model completes sample set construction, model building, iterative training, accuracy verification, and parameter solidification. This establishes a nonlinear correlation mapping between power grid operation characteristics and latent faults, ultimately yielding an intelligent model that can be directly used for online assessment.

[0123] (1) Construction of standardized training sample set and data partitioning

[0124] 1. Comprehensive sample data collection: Batch retrieval of S3 steady-state knowledge base Internal multi-condition steady-state benchmark characteristic data, threshold range data, and transient knowledge base. Internal fault characteristic parameters and characteristic weights Fault type label Fault risk level label The original sample dataset for the model is integrated and formed. The sample data covers all operating conditions of the power grid, including peak, off-peak, loop closing, and loop opening, and covers all typical hidden fault types.

[0125] 2. Positive and negative sample balance construction: The fault-free steady-state operating condition data is used as negative samples, and the abnormal transient feature data corresponding to various latent faults are used as positive samples. For a small number of scarce fault samples, interpolation is used to complete the sample balance processing to avoid model training bias and ensure the uniformity of model recognition of various faults.

[0126] 3. Dataset Standardization: The balanced total sample set is randomly divided into training, validation, and test sets in a 7:2:1 ratio. The training set is used for iterative learning of model parameters, the validation set is used for hyperparameter tuning during training, and the test set is used for final evaluation of the model's generalization ability. The three datasets are independent and have no data overlap, ensuring the authenticity of the model evaluation. The BayState (S2 output) is associated with the protection device adaptation rules (e.g., reclosing should be disabled during cold standby).

[0127] (2) Customization and construction of intelligent judgment model architecture

[0128] 1. Adaptability Model Architecture Construction: An intelligent assessment model for hidden faults in the power grid is constructed using an improved random forest algorithm.

[0129]

[0130] In the formula: Output the model (latent fault type, fault probability); The model uses a nonlinear mapping function; the inputs are steady-state features, transient features, and baseline data from dual knowledge bases. The model structure is optimized to address the multi-feature, nonlinear, and strongly coupled operational characteristics of the power grid. The model input layer has 19 input nodes, each corresponding to one of the 9 core steady-state features and 10 core transient fault features; the hidden layer uses a multi-branch decision tree structure to adapt to differentiated mapping relationships for different feature weights; the output layer has multi-classification output nodes, corresponding to different types of latent faults and their probabilities of occurrence.

[0131] 2. Model Hyperparameter Initialization: Based on the requirements of power grid fault assessment scenarios, initialize hyperparameters such as the number of decision trees, tree depth, branch sampling ratio, and feature sampling ratio, while importing the feature weights fixed in the S3 knowledge base. Prioritize the allocation of decision weights to highly correlated fault features to enhance the model's ability to perceive key fault features.

[0132] (3) Iterative training and parameter optimization

[0133] Layered iterative training and optimization: Input training set samples into the model for iterative training, relying on the cross-entropy loss function.

[0134]

[0135] In the formula: The number of samples; Number of fault types; Labels for actual faults; The smaller the loss value, the higher the model's accuracy in predicting fault probabilities. The error between the model's predicted values ​​and the actual fault labels is calculated in real time, and the model's internal decision parameters and feature weights are iteratively corrected in reverse. During training, model accuracy is monitored in real time using a validation set, hyperparameters are dynamically optimized, the loss function value is continuously reduced, and the accuracy of the association mapping between model features and faults is improved.

[0136] (4) Data output

[0137] The system outputs a solidified intelligent assessment model M for latent faults, the optimal feature weight parameters of the model, and a fault mapping matrix, providing a core algorithm model for online real-time assessment in S5.

[0138] Standard random forests have inherent flaws when applied to the assessment of latent faults in power grids: Standard random forests use equal-probability random sampling for all features, failing to differentiate the contribution of fault features to latent faults. Weak latent fault features have low weights and are easily overlooked, leading to a high risk of missed detection. The use of simple random sampling to construct subsets results in a scarcity of fault samples and a massive amount of fault-free steady-state samples, leading to model bias towards "no fault" detection and poor accuracy in identifying niche latent faults. Decision tree splitting relies solely on the Gini coefficient / information entropy, without incorporating the power grid's steady-state operating benchmark threshold, making it impossible to distinguish between "normal load fluctuations" and "abnormal disturbances caused by latent faults." The model only outputs fault classification labels, failing to simultaneously output fault occurrence probabilities, and cannot be directly integrated with subsequent operating condition correction models and risk scoring models. The model integration output lacks industry knowledge constraints, deviating from the power grid knowledge base benchmark, and is prone to unfounded misjudgments under changing operating conditions.

[0139] Therefore, combining the 9 steady-state features, 10 transient fault features, and S3 knowledge base feature weights of this application... steady-state threshold range The power grid operating condition correction logic improves the random forest algorithm.

[0140] Improvement 1: Weighted random subspace sampling based on fault feature weights

[0141] Features are randomly and equally extracted from the original random forest; feature weights are calculated using the S3 transient knowledge base. As the sampling probability, the sampling probability of the 19-dimensional feature (9-dimensional steady state + 10-dimensional transient state) is related to... The sampling frequency of features with high fault sensitivity (insulation resistance attenuation rate, leakage current mutation coefficient, etc.) is increased, while the sampling frequency of steady-state features with low sensitivity is decreased, thereby enhancing the ability to identify weak and latent fault features.

[0142] Improvement 2: Layered balanced auto-sampling to solve the problem of positive and negative sample imbalance.

[0143] The original random forest randomly sampled samples; this application uses a stratified sample set:

[0144] Negative samples: Samples under normal operating conditions from the steady-state knowledge base;

[0145] Positive samples: various hidden fault abnormal samples; stratified independent self-sampling, linear interpolation is used to expand the sample for scarce fault types with small sample sizes, ensuring that the positive and negative sample ratios of each decision tree are balanced, and avoiding model bias towards fault-free judgments.

[0146] Improvement 3: Splitting criterion for incorporating steady-state threshold deviation penalty

[0147] Standard decision trees only use the Gini coefficient as a splitting metric; this application reconstructs the splitting loss function by adding a deviation penalty term based on the Gini coefficient, which is applied to features that deviate from the steady-state knowledge base threshold. In the formula: Gini coefficient, γ represents the deviation of the feature from the steady-state baseline threshold, and γ is the penalty coefficient. When the feature exceeds the normal steady-state range, the loss value increases significantly. The decision tree prioritizes splitting based on abnormal fault features, naturally distinguishing between normal operating condition fluctuations and true latent faults.

[0148] Improvement 4: Dual output layer structure (fault type + fault occurrence probability)

[0149] Standard random forests only output a single classification label; the ensemble layer in this application outputs results for both classes simultaneously.

[0150] Discrete output: Latent fault categories (poor connection, insulation aging, harmonic faults, etc.);

[0151] Continuous output: Fault occurrence probability P; the probability value can be directly input into the S6 risk scoring model. It also supports adaptive correction using the S5 operating condition correction formula Y∗=Y×(1+λ).

[0152] Improvement 5: Correction of post-constraints for knowledge base rules

[0153] The original random forest has no prior knowledge constraints; this application embeds dual knowledge base verification logic after integration output: retrieve the steady-state knowledge base threshold and the transient fault feature benchmark interval. If all the features corresponding to the fault type output by the model fall within the steady-state normal interval, it is judged as a misjudgment, and the output is corrected to "no hidden fault", thereby reducing the misjudgment rate caused by operating condition disturbance from the algorithm output end.

[0154] The complete implementation process of the improved random forest algorithm of this invention is as follows:

[0155] Step 1: Algorithm initialization (corresponding to S4 stage (2) Step 2 Model hyperparameter initialization)

[0156] Input data source: S3 steady-state knowledge base Transient knowledge base Input feature set (9-dimensional steady state) (10-dimensional transient), total input dimension D=19;

[0157] Load the weights of each feature (Depend on (Calculated), construct a weighted sampling probability distribution;

[0158] Set the hyperparameters: total number of decision trees T, maximum depth of a single tree, number of samples per tree, penalty coefficient γ, and sample augmentation interpolation step size;

[0159] Divide the samples into stratified sets: negative sample set (normal steady-state operating conditions) and positive sample set for multiple types of faults.

[0160] Step 2: Hierarchical balanced self-service sample subset sampling (corresponding to S4 stage (1) step 2 positive and negative sample balancing)

[0161] Iteratively generate T independent sample subsets corresponding to each decision tree:

[0162] Negative samples and positive samples for each type of fault were extracted in stratified manner.

[0163] For fault categories with insufficient sample size, supplementary samples are generated based on linear interpolation of temporal features;

[0164] Merge stratified sampling data to ensure that the ratio of positive to negative samples within a single tree is uniformly 1:1.

[0165] Step 3: Weighted random feature subset extraction (Improvement 1)

[0166] Based on feature weights Based on the probability distribution, some features are randomly selected as candidate features for splitting the current decision tree, with high-weight fault features having a higher probability of being selected.

[0167] Step 4: Single decision tree training (Improvement point 3)

[0168] Train a single decision tree using the sample subset and feature subset obtained from sampling:

[0169] Traverse candidate features and split threshold;

[0170] The splitting loss is calculated using the loss function Loss=Gini+γ⋅ΔF, which is a penalty for fusion steady-state deviation.

[0171] Select the feature and threshold corresponding to the minimum loss to complete node splitting;

[0172] The splitting process stops when the maximum depth is reached or when the number of node samples is less than the threshold, and a single decision tree is generated.

[0173] Step 5: Integration layer fusion, dual result output (Improvement point 4)

[0174] Classification output: Statistically analyze the fault categories output by T trees, and use a voting method to determine the basic fault type Y;

[0175] Probability output: The percentage of decision trees that are statistically determined to be this type of fault is used as the probability of fault occurrence P; the original output combination (Y,P) of the model is obtained.

[0176] Step 6: Post-correction of knowledge base rules (Improvement point 5)

[0177] Retrieve steady-state threshold range Calculate the current input feature deviation ΔF:

[0178] If all features fall within the steady-state normal range, the forced correction output is "no hidden faults", P=0;

[0179] If the output fault type does not match the fault feature benchmark in the transient knowledge base, the fault probability P is lowered to complete the initial output correction of the model.

[0180] Step 7: Model Iteration Optimization and Accuracy Verification (corresponding to S4 stages (3) and (4))

[0181] Iterative training is performed on the input training set, and the penalty coefficient γ and feature sampling weight distribution are finely adjusted using the cross-entropy loss function.

[0182] Calculate model accuracy using an independent test set. ;

[0183] When the accuracy is ≥95% and there is no serious overfitting, the iteration stops, and all decision tree structures, hyperparameters, and feature weight mapping relationships are solidified to obtain the final intelligent judgment model M.

[0184] Step 5 (One-click inspection and tagging detection of all anomalies) Specific implementation techniques:

[0185] (1) Real-time data access

[0186] By accessing real-time measured data from the power grid, and repeating the preprocessing and feature extraction logic in steps S1 and S2, real-time steady-state features are generated. Real-time transient characteristics .

[0187] (2) Multi-dimensional data fusion

[0188] Retrieve current power grid operating parameters (grid structure, load distribution, equipment operating status) and S3 solidified dual knowledge base. , S4 (Step 4) The intelligent model M is trained.

[0189] (3) Intelligent judgment and calculation

[0190] 1. Initial Fault Assessment and Deviation Calculation: Based on model M, nonlinear mapping operations are performed to output the initial fault type and fault occurrence probability; simultaneously, a real-time feature deviation assessment model is invoked.

[0191]

[0192] In the formula: The characteristic deviation value; These are real-time feature parameters; The mean of the steady-state features of the knowledge base; Weights are assigned to fixed features in the knowledge base; the larger the deviation, the higher the risk of failure. This is combined with the mean of the steady-state features of the knowledge base. With feature weights Calculate the feature deviation values ​​for each dimension. Quantify the degree of anomaly.

[0193] 2. Operating condition correction and precise fault location: Match operating condition correction coefficients based on current grid load fluctuations and grid adjustment status. The initial judgment results of the model are adaptively corrected, and the model is modified according to the power grid operation mode.

[0194]

[0195] In the formula: This is the original output of the model; The operating condition correction factor (dynamically taken as ±0.05~±0.2 under load fluctuation and grid adjustment conditions) is used to obtain accurate fault assessment results. Further, the fault location distance model is invoked:

[0196]

[0197] In the formula: Features of real-time monitoring points; The fault characteristic reference value ( Typical abnormal feature values ​​corresponding to various latent faults are taken from the transient knowledge base; the normal steady-state benchmark values ​​stored in the steady-state knowledge base are only used for preliminary anomaly identification and are not used as distance matching benchmarks. The fault distance coefficient is calculated by comparing the features of real-time monitoring points with the fault benchmark features. The position with the minimum coefficient is the fault location, which helps to pinpoint the precise location of hidden faults and quantify the characteristic deviation dimension corresponding to the fault cause.

[0198] (4) Data output

[0199] The corrected fault assessment data is output as follows: fault location coordinates, fault cause characteristic deviation value, fault occurrence probability, and operating condition adaptation coefficient, providing accurate data support for the graded output and early warning of S6 (step six).

[0200] Step Six (Output of Analysis Results): Specific Implementation Techniques

[0201] (1) Fault risk level classification: Based on the fault probability and feature deviation value output by S5 (step five), and combined with the risk threshold standard of the knowledge base, the latent faults are classified and judged according to the fault risk level scoring model:

[0202]

[0203] In the formula: Score the risk of failure (0~100 points); This represents the probability of a failure occurring. The characteristic deviation value; This is the maximum permissible deviation value; , These are the weighting coefficients. + =1. Risk classification criteria: High risk (S≥80), Medium risk (50≤S≤80), Low risk (S<50). Table 1 shows... , Several possible values ​​for .

[0204] Table 1 , Value examples in different scenarios

[0205]

[0206] (2) Results integration and output: Summarize the fault location, core causes, risk level, and fault development trend to form a standardized analysis report;

[0207] Fault trend prediction model: Predict the development trend of faults

[0208]

[0209] is the fault feature deviation value at time t; is the current deviation value; is the deviation growth rate, used to predict the fault deterioration speed;

[0210] Low-risk latent fault: 0<k≤0.02, the fault deteriorates slowly and poses a long-term existing risk;

[0211] Medium-risk latent fault: 0.02<k≤0.05, the fault continues to deteriorate, and planned maintenance is required;

[0212] High-risk latent fault: k>0.05, the fault deteriorates rapidly, with risks of short-time tripping and equipment damage, and urgent defect elimination is required;

[0213] (3) Operation and maintenance warning generation: Push warning information of corresponding levels according to different risk levels, match the built-in fault handling schemes in the knowledge base, and output operation and maintenance guidance suggestions.

[0214] The above are preferred embodiments of the present invention. Any modifications made according to the technical solution of the present invention, whose generated functions and effects do not exceed the scope of the technical solution of the present invention, shall fall within the protection scope of the present invention.

Claims

1. A method for intelligent assessment of latent faults in power grid operation, characterized in that, include: The power grid steady-state operation data and historical fault data are collected as raw data. After preprocessing the raw data, standardized steady-state datasets and standardized fault datasets are obtained. Extract a set of steady-state core features from a standardized steady-state dataset. The set of steady-state core features includes voltage deviation rate, three-phase voltage imbalance, load stability coefficient, active load rate, reactive load rate, power factor, equipment steady-state temperature rise, line impedance steady-state value, and bus frequency deviation. Extract the transient fault core feature set from the standardized fault dataset. The transient fault core feature set includes voltage fluctuation coefficient, current fluctuation coefficient, total harmonic distortion rate, leakage current mutation coefficient, joint temperature rise mutation rate, voltage sag depth, load instantaneous fluctuation coefficient, phase offset, insulation resistance attenuation rate, and high-frequency pulse distortion rate. A steady-state operation knowledge base is constructed based on the steady-state core feature set, and a transient operation knowledge base is constructed based on the transient fault core feature set and historical fault cases. A smart analysis model for hidden faults in the power grid is trained using a steady-state operation knowledge base and a transient operation knowledge base. Real-time power grid operation data is collected, real-time steady-state features and real-time transient features are extracted, and input into the trained intelligent analysis model of hidden faults in the power grid. Combined with the current power grid operation mode, the model outputs the results of hidden fault assessment.

2. The intelligent judgment method for latent faults in power grid operation according to claim 1, characterized in that, The steps for constructing the steady-state operation knowledge base include: The standardized steady-state dataset is divided into multiple operating condition categories according to the power grid operating conditions; For each operating condition category, calculate the mean of each feature in the steady-state core feature set. and standard deviation ; According to the formula Calculate the normal operating threshold range for each steady-state characteristic under various operating conditions. , ]; The operating condition category, feature name, mean, standard deviation, and corresponding threshold interval are associated and stored to form a steady-state operation knowledge base.

3. The intelligent assessment method for latent faults in power grid operation according to claim 1, characterized in that, The steps for constructing the transient operation knowledge base include: Calculate the Pearson correlation coefficients between each feature in the core feature set of transient faults and the preset fault types. ; According to the formula Calculate the weighting coefficients of each transient fault characteristic. ; Establish the correlation between fault types and corresponding transient fault feature combinations, feature weight coefficients, anomaly thresholds, and risk levels; The relationships and historical fault handling strategies are stored in a structured manner to form a transient operation knowledge base.

4. The intelligent judgment method for latent faults in power grid operation according to claim 1, characterized in that, The steps for training the intelligent analysis model for latent faults in the power grid include: By integrating the benchmark data in the steady-state operation knowledge base and the fault feature data in the transient operation knowledge base, a training sample set containing negative samples and positive samples is constructed, where the negative samples are fault-free steady-state operating condition data and the positive samples are latent fault operating condition data. An improved random forest algorithm was used to build an intelligent analysis model for hidden faults in power grids. The input layer of the intelligent analysis model for hidden faults in power grids was set to 19-dimensional nodes, which correspond to 9 steady-state core features and 10 transient fault core features, respectively. The intelligent analysis model for hidden faults in the power grid is iteratively trained using a training sample set, and the cross-entropy loss function is applied. By optimizing the model parameters, a fixed intelligent analysis model for hidden faults in the power grid is obtained.

5. The intelligent judgment method for latent faults in power grid operation according to claim 1, characterized in that, Following the step of outputting the latent fault assessment results, the method further includes: According to the formula Calculate real-time feature deviation value ,in For real-time feature parameters, The mean of steady-state characteristics, For feature weights; According to the formula Initial assessment results of the model output Perform operating condition corrections, including This is the operating condition correction factor determined based on the current power grid load fluctuations and grid adjustment status; According to the formula Calculate the failure risk score ,in The probability of failure occurring. This is the maximum permissible deviation value. , These are the weighting coefficients; Based on risk scoring Classify the faults and match them with corresponding handling strategies.

6. The intelligent judgment method for latent faults in power grid operation according to claim 5, characterized in that, It also includes fault location steps: According to the formula Calculate the fault distance coefficient ,in To monitor the characteristics of the points in real time, This serves as a baseline value for fault characteristics. Fault distance coefficient The smallest location is determined as the hidden fault location.

7. The intelligent judgment method for latent faults in power grid operation according to claim 5, characterized in that, It also includes a fault trend prediction step: Calculation of fault deviation growth rate based on least squares method ; According to the formula Predict the fault characteristic deviation value at time t, where This is the current deviation value; According to the growth rate The numerical range defines the level of fault degradation.

8. A smart system for assessing latent faults in power grid operation, comprising executing the method described in any one of claims 1-7, characterized in that, include: The data acquisition module is used to collect steady-state operation data, historical fault data, and real-time operation data of the power grid; The data processing module is used to clean, denoise, and normalize the collected data, and extract the steady-state core feature set and the transient fault core feature set. The knowledge base construction module is used to construct a steady-state operation knowledge base based on the steady-state core feature set, and to construct a transient operation knowledge base based on the transient fault core feature set and historical fault cases. The model training module is used to train an intelligent analysis model for hidden faults in the power grid using the steady-state operation knowledge base and the transient operation knowledge base; The online analysis module is used to receive real-time extracted feature data and current power grid operation parameters, and call the trained intelligent analysis model to output the latent fault analysis results. The results output module is used to output the fault location, risk level, and early warning information.

9. The intelligent judgment system for hidden faults in power grid operation according to claim 8, characterized in that, The knowledge base construction module includes: The operating condition classification unit is used to divide steady-state data into peak load operating conditions, low load operating conditions, line loop closing operating conditions, and line loop breaking operating conditions. The threshold calculation unit is used to calculate the mean, standard deviation, and threshold interval based on the [μs−2σs,μs+2σs] rule for steady-state characteristics under various operating conditions. The weight calculation unit is used to calculate the weight coefficients of transient fault characteristics based on the Pearson correlation coefficient. The storage unit is used to store the data of the steady-state operating knowledge base and the transient operating knowledge base in a multi-dimensional tag index, and is configured with an incremental update interface.

10. The intelligent judgment system for latent faults in power grid operation according to claim 8, characterized in that, The online analysis module is also configured to: Calling formula Calculate the characteristic deviation value. This represents the real-time feature deviation value. For real-time feature parameters, The mean of steady-state characteristics, For feature weights; Calling formula Introducing operating condition correction factor Correct the model output. This is the initial assessment result; Calling formula Locate the fault location. This is the fault distance coefficient. Features of real-time monitoring points; This serves as a baseline value for fault characteristics. Calling formula Perform fault trend prediction. This is the current deviation value. This represents the rate of increase in fault deviation.

Images