Power distribution network fault calibration early warning method and system based on data analysis

Abstract

The invention discloses a power distribution network fault calibration early warning method and system based on data analysis. The method comprises the steps of S1, collecting a power distribution network operation data set; s2, obtaining a feature subset highly related to the power distribution network fault; s3, constructing an improved naive Bayes classifier; s4, obtaining an optimized naive Bayes classifier; s5, inputting the operation data set of the power distribution network into the optimized naive Bayes classifier, determining the fault level of the power distribution network according to the output result of the optimized naive Bayes classifier, and calibrating the fault type; and S6, according to the calibrated fault grade and the fault type, when the fault grade reaches a preset threshold value, generating early warning information. According to the method, the problems of insufficient early recognition, limited calibration precision and poor real-time response in the prior art are effectively solved, and innovative technical support is provided for ensuring safe operation of the power distribution network.

Description

Technical Field

[0001] The present invention relates to the field of power distribution technology, and in particular to a distribution network fault calibration early warning method and system based on data analysis. Background Art

[0002] With the rapid development of smart grid technology, the distribution network, as an important part of the power system, has a vital impact on the reliability of power supply due to its safety and stability. However, in the actual operation of the distribution network, the frequency of faults gradually increases due to complex working conditions and environmental conditions. Faults may cause power outages, equipment damage, and even large-scale power outages. Therefore, how to timely detect and accurately calibrate fault levels and types has become an important research topic in the operation and maintenance of distribution networks.

[0003] In recent years, with the development of data analysis and machine learning technologies, some algorithm-based distribution network fault diagnosis methods have gradually been applied to practical scenarios. For example, some methods use models such as support vector machines and random forests to classify fault types, but these models usually require a large amount of labeled data and complex parameter tuning, and cannot adapt well to the processing needs of real-time data in distribution networks. Another part of the research is based on neural networks or deep learning for fault detection. Although it has improved the classification accuracy to a certain extent, the above methods have complex models and large computational workloads, making them difficult to efficiently deploy in edge computing nodes with limited resources. In addition, they have weak adaptability to dynamic data characteristics. As the operating environment of the distribution network changes, their detection effect may gradually decline.

[0004] The existing technology has the following defects in distribution network fault detection:

[0005] Insufficient recognition of early fault signals: Traditional detection methods lack sensitivity to the early characteristics of faults and have difficulty capturing complex dynamically changing signals, which easily leads to delays in fault calibration.

[0006] Limited calibration accuracy and adaptability: Existing models have difficulty achieving high calibration accuracy when faced with multi-dimensional and multi-type data, and have poor adaptability to changes in the operating environment.

[0007] Insufficient real-time and responsiveness: Many existing methods have high computational complexity and perform poorly in scenarios with high real-time and rapid response requirements, and are unable to provide early warning information in a timely manner.

[0008] In summary, the existing technology has significant deficiencies in the accuracy, real-time and dynamic adaptability of distribution network fault detection. A more efficient, intelligent and accurate fault calibration and early warning method is urgently needed to make up for the shortcomings of the existing technology. Summary of the invention

[0009] One object of the present invention is to propose a distribution network fault calibration early warning method and system based on data analysis, which effectively solves the problems of insufficient early identification, limited calibration accuracy and poor real-time response existing in the prior art, and provides innovative technical support for ensuring the safe operation of the distribution network.

[0010] A distribution network fault calibration and early warning method based on data analysis according to an embodiment of the present invention comprises the following steps:

[0011] S1. Collecting a distribution network operation data set, preprocessing the distribution network operation data set, and performing segmentation processing according to time series characteristics to form a preprocessed distribution network operation data set;

[0012] S2. Based on the preprocessed distribution network operation data set, key features related to distribution network faults are extracted, and the extracted key features are screened using a feature selection algorithm to obtain a feature subset that is highly correlated with distribution network faults;

[0013] S3. Construct an improved naive Bayes classifier, and train the improved naive Bayes classifier based on the feature subset and the historical distribution network operation data set;

[0014] S4. The improved naive Bayes classifier is optimized based on the Kepler optimization algorithm. The fitness of each solution is calculated by randomly initializing multiple solutions of the optimization parameters as a population, and the position and trajectory of the solution are iteratively updated until the fitness of the solution reaches the optimal value, thereby obtaining the optimized naive Bayes classifier.

[0015] S5. Input the distribution network operation data set into the optimized naive Bayes classifier, determine the fault level of the distribution network and calibrate the fault type according to the output result of the optimized naive Bayes classifier;

[0016] S6. According to the calibrated fault level and fault type, when the fault level reaches the preset threshold, an early warning information is generated. The early warning information includes the fault level, fault type, fault location and recommended response strategy. The early warning information is sent to the maintenance personnel through the sound and light alarm equipment, mobile terminal or distribution network dispatching center system.

[0017] Optionally, the S1 includes:

[0018] S11. Collecting a distribution network operation data set through multi-point sensors deployed in the distribution network, wherein the distribution network operation data set includes current I(t), voltage V(t), load fluctuation L(t), harmonic distortion rate H(t) and ambient temperature T(t);

[0019] S12. Normalizing the collected distribution network operation data set;

[0020] S13. Fill missing values ​​in the normalized distribution network operation data set and calculate the missing values ​​using linear interpolation;

[0021] S14. Remove noise from the filled distribution network operation data set and smooth the noise interference using a sliding average method;

[0022] S15. According to the time series characteristics, the denoised distribution network operation data set is segmented and divided according to the fixed time window Δt to generate the preprocessed distribution network operation data set D pre :

[0023] D pre ={X smooth (t)|t∈[t0,t0+Δt]};

[0024] Among them, D pre is the preprocessed distribution network operation data set, t0 represents the starting time point of the segment, Δt represents the time window size, X smooth (t) represents the data value at time point t.

[0025] Optionally, S2 includes:

[0026] S21. From the preprocessed distribution network operation dataset D pre Features related to distribution network faults are extracted from the

[0027] The harmonic distortion rate H(t) is defined as the total ratio of each harmonic component current to the fundamental current;

[0028] Voltage fluctuation rate V rate (t), defined as the rate of change of voltage value per unit time;

[0029] Load fluctuation rate L rate (t), defined as the rate of change of load fluctuation;

[0030] The overcurrent ratio C(t) is defined as the ratio of the real-time current value to the rated current value;

[0031] Ambient temperature differential characteristic T diff (t), defined as the change in ambient temperature per unit time;

[0032] S22. Perform feature selection on the extracted features, use a ranking algorithm based on feature importance to evaluate the relevance of each feature with the distribution network fault, and calculate the relevance score R s :

[0033]

[0034] Among them, X featureis a set of eigenvalues, including H(t), V rate (t),L rate (t),C(t),T diff (t), Y is the corresponding fault label set, Cov(X feature ,Y) represents the covariance between the feature value and the fault label, σ(X feature ) and σ(Y) are the standard deviations of the feature value and fault label, respectively;

[0035] S23. According to the relevance score R s The evaluation results screen the correlation scores higher than the set threshold R threshold The features of the distribution network are formed into a feature subset F that is highly correlated with distribution network faults. selected :

[0036] F selected ={X feature |R s ≥R threshold};

[0037] Among them, F selected is the final selected feature subset, R threshold The relevance score threshold set for the data from an experiment or run.

[0038] Optionally, the S3 includes:

[0039] S31. According to the feature subset F selected and historical distribution network operation dataset D history Construct a multi-scale feature training dataset D multi , the feature subset F selected According to the time series, it is divided into short-term feature set F short and the long-term feature set F long , the short-term feature set is used to capture short-term fluctuation information, and the long-term feature set is used to capture operation trend information:

[0040] D multi ={(X short,i ,X long,i ,Y i )}, i = 1, 2, ..., N;

[0041] Among them, X short,i represents the short-term feature vector of the i-th sample, X long,i represents the long-term eigenvector, Y i Indicates the corresponding fault level label;

[0042] S32. Aiming at the multi-scale characteristics of distribution network operation data, the naive Bayes classifier model is improved and the multi-dimensional joint posterior probability calculation formula is designed:

[0043]

[0044] Where C represents the fault level label, including minor fault, medium fault and severe fault, j represents the index of the jth feature in the short-term feature set, k represents the index of the kth feature in the long-term feature set, X = {X short ,X long}, P(C) is the prior probability of the fault level, P(x short,j |C) and P(x long,k |C) represent the conditional probabilities of short-term and long-term characteristics under fault level C, M short and M long are the number of short-term and long-term features, respectively;

[0045] S33. Aiming at the sparsity and feature interaction of distribution network data, an improved conditional probability smoothing strategy is used to calculate the conditional probability P(x|C):

[0046]

[0047] Among them, λ is the dynamic smoothing coefficient, K is the total number of eigenvalues, which is determined by the feature interaction strength, and f(x) represents the interaction strength of eigenvalue x, which is defined as:

[0048]

[0049] Among them, Cov(x,k) is the frequency of feature x under fault level C, and Var(x) represents the variance of feature x;

[0050] S34. Based on the conditional probability calculation results combined with multi-scale feature training data set D multi , train the improved naive Bayes classifier so that the improved naive Bayes classifier takes into account the impact of short-term fluctuations and long-term trends on fault level calibration, and optimizes the improved naive Bayes classifier's ability to distinguish fault modes.

[0051] Optionally, the S4 includes:

[0052] S41. Based on the improved naive Bayes classifier model, the prior probability P(C) of the fault level and the dynamic smoothing factor λ of the conditional probability are extracted, and the optimization parameter vector S = [P(C), λ] is constructed;

[0053] S42. Initialize the population of Kepler optimization algorithm. Each individual S in the population i is an optimization parameter vector:

[0054] S i =[P i (C),λ i ];

[0055] Among them, P i (C) is the prior probability distribution of the i-th individual in the population, satisfying ∑ C P i (C) = 1, λ i is a dynamic smoothing factor used to adjust the conditional probability calculation of sparse data, randomly generate the initial value of each solution in the population, and form an initial population S = {S1, S2, ..., S N};

[0056] S43. For each individual in the population, calculate its fitness value F(S i ):

[0057]

[0058] Among them, E(S i ) represents the classification error, based on the actual fault level C actual and predicted failure level C predicted The difference calculation is:

[0059]

[0060] Among them, N test is the number of samples in the test data set;

[0061] S44. Simulate Keplerian motion to optimize the parameter vector S i The position and velocity are updated:

[0062]

[0063] in, represents the speed of the i-th individual at the t-th iteration, G is the gravitational constant, which controls the convergence speed of the optimization process, It is the individual solution with the highest fitness value in the current iteration;

[0064] S45. Check the boundary of the population solution after each iteration. If the solution exceeds the parameter range, adjust it to the nearest valid value and repeat steps S43 and S44 until one of the following conditions is met:

[0065] The change in the fitness value of the population is less than the set threshold ΔF threshold ;

[0066] The maximum number of iterations has been reached;

[0067] S46. Output the optimized parameter vector and generate the final improved naive Bayes classifier:

[0068] S optimal =[P optimal(C),λ optimal ].

[0069] Optionally, the S5 includes:

[0070] S51. Input the preprocessed distribution network operation data set into the optimized improved naive Bayes classifier S optimal , extract the running data feature X;

[0071] S52. Calculating the posterior probability of the fault level according to the optimized improved naive Bayes classifier;

[0072] S53. Compare the posterior probabilities of all fault levels and determine the most likely fault level C best ;

[0073] S54. According to the fault level C output by the optimized improved naive Bayes classifier best , combined with the characteristic value X, determine the fault type T, which includes:

[0074] When x C ≥I threshold It is judged as an overcurrent fault, where I threshold is the threshold for overcurrent determination;

[0075] When x V ≤V threshold It is judged as undervoltage fault, where V threshold It is the threshold value for undervoltage determination;

[0076] When x H ≥H threshold It is judged to be a harmonic distortion fault, where H threshold is the threshold for harmonic distortion determination;

[0077] S55. Fault level C of output distribution network best and fault type T, and calibrate the specific category v of the fault.

[0078] A distribution network fault calibration early warning system based on data analysis, used to execute a distribution network fault calibration early warning method based on data analysis, the system comprising:

[0079] A data acquisition module is used to collect distribution network operation data in real time through multi-point sensors arranged in the distribution network, wherein the distribution network operation data includes current, voltage, load fluctuation, harmonic distortion rate and ambient temperature; and to preliminarily store the collected data;

[0080] The data preprocessing module is used to normalize, fill missing values ​​and remove noise on the collected distribution network operation data, segment the data according to the time series characteristics, and generate a preprocessed distribution network operation data set;

[0081] The feature extraction and selection module is used to extract features related to distribution network faults from the preprocessed distribution network operation data set, including harmonic distortion rate, fluctuation rate, overcurrent ratio and differential features, and screen the extracted features based on the feature importance sorting algorithm to generate a feature subset that is highly related to distribution network faults;

[0082] The naive Bayes classifier module is used to construct an improved naive Bayes classifier based on multi-scale features, combine short-term features and long-term features for classification, complete the posterior probability calculation of the fault level based on the optimized parameters, and output the optimal judgment result of the distribution network fault level;

[0083] Kepler optimization module, used to optimize the prior probability and conditional probability smoothing factor of the naive Bayes classifier, uses the Kepler optimization algorithm to iteratively optimize the classifier parameters, and outputs the optimized parameters through population initialization, fitness calculation and position update;

[0084] The fault calibration module is used to input the pre-processed distribution network operation data into the optimized naive Bayes classifier, determine the fault level of the distribution network according to the posterior probability calculation result, and calibrate the fault type, including overcurrent fault, undervoltage fault and harmonic distortion fault, by combining the classifier output and the characteristic value range;

[0085] An early warning information generation module is used to generate early warning information according to the fault calibration results. The early warning information includes fault level, fault type, fault location and recommended response strategy, and sends the early warning information to maintenance personnel through sound and light alarm equipment, mobile terminals or distribution network dispatching center system;

[0086] The model update module is used to regularly update the training data set of the naive Bayes classifier based on the distribution network operation history data and new fault data, retrain the classifier model, and optimize the parameters through the Kepler optimization algorithm.

[0087] The beneficial effects of the present invention are:

[0088] (1) The present invention designs a naive Bayes classifier with a multi-scale feature structure, which divides the features into short-term features and long-term features for joint modeling. The short-term features are used to capture transient abnormal signals in the operation of the distribution network, and the long-term features are used to identify operating trends and chronic fault characteristics. The accurate calibration of the fault level is achieved through multi-dimensional joint posterior probability calculation. The improved naive Bayes classifier can more comprehensively utilize the characteristics of distribution network data and significantly improve the ability to distinguish complex fault scenarios.

[0089] (2) The present invention optimizes the core parameters of the naive Bayes classifier, including the prior probability of fault level and the conditional probability smoothing factor, through the Kepler optimization algorithm. The optimization algorithm simulates the motion law of Kepler celestial bodies and realizes efficient parameter optimization through a dynamic gravitational model, thus overcoming the limitation that traditional optimization methods are prone to fall into local optimality. Experimental data show that the Kepler optimization algorithm improves the convergence efficiency by 25% compared with the particle swarm optimization algorithm. At the same time, the fault level calibration error of the optimized classifier under different data distributions is reduced by more than 15%, further improving the robustness and real-time performance of the system.

[0090] (3) The present invention introduces a dynamic update module to regularly update the training data set of the naive Bayes classifier based on the newly added fault data and historical data during operation, and readjusts the model parameters in combination with the Kepler optimization algorithm so that the system can continue to maintain efficient fault calibration capabilities as the distribution network operating environment changes. The dynamic update mechanism avoids the shortcoming of the static model in the traditional method that is prone to aging. BRIEF DESCRIPTION OF THE DRAWINGS

[0091] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0092] Figure 1 The present invention provides a flow chart of a distribution network fault calibration and early warning method and system based on data analysis. DETAILED DESCRIPTION

[0093] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0094] refer to Figure 1 , a distribution network fault calibration and early warning method based on data analysis, comprising the following steps:

[0095] S1. Collect distribution network operation data sets, preprocess the distribution network operation data sets, and perform segmentation processing according to time series characteristics to form a preprocessed distribution network operation data set;

[0096] S2. Based on the preprocessed distribution network operation data set, key features related to distribution network faults are extracted, and the extracted key features are screened using a feature selection algorithm to obtain a feature subset that is highly correlated with distribution network faults;

[0097] S3. Construct an improved naive Bayes classifier, and train the improved naive Bayes classifier based on the feature subset and the historical distribution network operation data set;

[0098] S4. The improved naive Bayes classifier is optimized based on the Kepler optimization algorithm. The fitness of each solution is calculated by randomly initializing multiple solutions of the optimization parameters as a population, and the position and trajectory of the solution are iteratively updated until the fitness of the solution reaches the optimal value, thereby obtaining the optimized naive Bayes classifier.

[0099] S5. Input the distribution network operation data set into the optimized naive Bayes classifier, determine the fault level of the distribution network and calibrate the fault type according to the output result of the optimized naive Bayes classifier;

[0100] S6. According to the calibrated fault level and fault type, when the fault level reaches the preset threshold, an early warning information is generated. The early warning information includes the fault level, fault type, fault location and recommended response strategy. The early warning information is sent to the maintenance personnel through the sound and light alarm equipment, mobile terminal or distribution network dispatching center system.

[0101] In this implementation, S1 includes:

[0102] S11. Collect distribution network operation data sets through multi-point sensors deployed in the distribution network, where the distribution network operation data sets include current I(t), voltage V(t), load fluctuation L(t), harmonic distortion rate H(t) and ambient temperature T(t);

[0103] S12. Normalizing the collected distribution network operation data set;

[0104] S13. Fill missing values ​​in the normalized distribution network operation data set and calculate the missing values ​​using linear interpolation;

[0105] S14. Remove noise from the filled distribution network operation data set and smooth the noise interference using a sliding average method;

[0106] S15. According to the time series characteristics, the denoised distribution network operation data set is segmented and divided according to the fixed time window Δt to generate the preprocessed distribution network operation data set D pre :

[0107] D pre ={X smooth (t)|t∈[t0,t0+Δt]};

[0108] Among them, D pre is the preprocessed distribution network operation data set, t0 represents the starting time point of the segment, Δt represents the time window size, X smooth (t) represents the data value at time point t.

[0109] In this implementation, S2 includes:

[0110] S21. From the preprocessed distribution network operation dataset D pre Features related to distribution network faults are extracted from the

[0111] The harmonic distortion rate H(t) is defined as the total ratio of each harmonic component current to the fundamental current;

[0112] Voltage fluctuation rate V rate (t), defined as the rate of change of voltage value per unit time;

[0113] Load fluctuation rate L rate (t), defined as the rate of change of load fluctuation;

[0114] The overcurrent ratio C(t) is defined as the ratio of the real-time current value to the rated current value;

[0115] Ambient temperature differential characteristic T diff (t), defined as the change in ambient temperature per unit time;

[0116] S22. Perform feature selection on the extracted features, use a ranking algorithm based on feature importance to evaluate the relevance of each feature with the distribution network fault, and calculate the relevance score R s :

[0117]

[0118] Among them, X feature is a set of eigenvalues, including H(t), V rate (t),L rate (t),C(t),T diff (t), Y is the corresponding fault label set, Cov(X feature ,Y) represents the covariance between the feature value and the fault label, σ(X feature ) and σ(Y) are the standard deviations of the feature value and fault label, respectively;

[0119] S23. According to the relevance score R s The evaluation results screen the correlation scores higher than the set threshold R threshold The features of the distribution network are used to form a feature subset F that is highly correlated with distribution network faults. selected :

[0120] F selected ={X feature |R s ≥R threshold};

[0121] Among them, F selected is the final selected feature subset, R thresholdThe relevance score threshold set for the data from an experiment or run.

[0122] In this implementation, S3 includes:

[0123] S31. According to the feature subset F selected and historical distribution network operation dataset D history Construct a multi-scale feature training dataset D multi , the feature subset F selected According to the time series, it is divided into short-term feature set F short and the long-term feature set F long , the short-term feature set is used to capture short-term fluctuation information, and the long-term feature set is used to capture operation trend information:

[0124] D multi ={(X short,i ,X long,i ,Y i )}, i = 1, 2, ..., N;

[0125] Among them, X short,i represents the short-term feature vector of the i-th sample, X long,i represents the long-term eigenvector, Y i Indicates the corresponding fault level label;

[0126] S32. Aiming at the multi-scale characteristics of distribution network operation data, the naive Bayes classifier model is improved and the multi-dimensional joint posterior probability calculation formula is designed:

[0127]

[0128] Where C represents the fault level label, including minor fault, medium fault and severe fault, j represents the index of the jth feature in the short-term feature set, k represents the index of the kth feature in the long-term feature set, X = {X short ,X long}, P(C) is the prior probability of the fault level, P(x short,j |C) and P(x long,k |C) represent the conditional probabilities of short-term and long-term characteristics under fault level C, M short and M long are the number of short-term and long-term features, respectively;

[0129] S33. Aiming at the sparsity and feature interaction of distribution network data, an improved conditional probability smoothing strategy is used to calculate the conditional probability P(x|C):

[0130]

[0131] Among them, λ is the dynamic smoothing coefficient, K is the total number of eigenvalues, which is determined by the feature interaction strength, and f(x) represents the interaction strength of eigenvalue x, which is defined as:

[0132]

[0133] Among them, Cov(x,k) is the frequency of feature x under fault level C, and Var(x) represents the variance of feature x;

[0134] S34. Based on the conditional probability calculation results combined with multi-scale feature training data set D multi , train the improved naive Bayes classifier so that the improved naive Bayes classifier takes into account the impact of short-term fluctuations and long-term trends on fault level calibration, and optimizes the improved naive Bayes classifier's ability to distinguish fault modes.

[0135] In this implementation, S4 includes:

[0136] S41. Based on the improved naive Bayes classifier model, the prior probability P(C) of the fault level and the dynamic smoothing factor λ of the conditional probability are extracted, and the optimization parameter vector S = [P(C), λ] is constructed;

[0137] S42. Initialize the population of Kepler optimization algorithm. Each individual S in the population i is an optimization parameter vector:

[0138] S i =[P i (C),λ i ];

[0139] Among them, P i (C) is the prior probability distribution of the i-th individual in the population, satisfying ∑ C P i (C) = 1, λ i is a dynamic smoothing factor used to adjust the conditional probability calculation of sparse data, randomly generate the initial value of each solution in the population, and form an initial population S = {S1, S2, ..., S N};

[0140] S43. For each individual in the population, calculate its fitness value F(S i ):

[0141]

[0142] Among them, E(S i ) represents the classification error, based on the actual fault level C actual and predicted failure level C predicted The difference calculation is:

[0143]

[0144] Among them, N test is the number of samples in the test data set;

[0145] S44. Simulate Keplerian motion to optimize the parameter vector S i The position and velocity are updated:

[0146]

[0147] in, represents the speed of the i-th individual at the t-th iteration, G is the gravitational constant, which controls the convergence speed of the optimization process, It is the individual solution with the highest fitness value in the current iteration;

[0148] S45. Check the boundary of the population solution after each iteration. If the solution exceeds the parameter range, adjust it to the nearest valid value and repeat steps S43 and S44 until one of the following conditions is met:

[0149] The change in the fitness value of the population is less than the set threshold ΔF threshold ;

[0150] The maximum number of iterations has been reached;

[0151] S46. Output the optimized parameter vector and generate the final improved naive Bayes classifier:

[0152] S optimal =[P optimal (C),λ optimal ].

[0153] In this implementation, S5 includes:

[0154] S51. Input the preprocessed distribution network operation data set into the optimized improved naive Bayes classifier S optimal , extract the running data feature X;

[0155] S52. Calculating the posterior probability of the fault level according to the optimized improved naive Bayes classifier;

[0156] S53. Compare the posterior probabilities of all fault levels and determine the most likely fault level C best ;

[0157] S54. According to the fault level C output by the optimized improved naive Bayes classifier best , combined with the characteristic value X, determine the fault type T, which includes:

[0158] When x C ≥Ithreshold It is judged as an overcurrent fault, where I threshold is the threshold for overcurrent determination;

[0159] When x V ≤V threshold It is judged as undervoltage fault, where V threshold It is the threshold value for undervoltage determination;

[0160] When x H ≥H threshold It is judged to be a harmonic distortion fault, where H threshold is the threshold for harmonic distortion determination;

[0161] S55. Fault level C of output distribution network best and fault type T, and calibrate the specific category v of the fault.

[0162] A distribution network fault calibration early warning system based on data analysis is used to execute a distribution network fault calibration early warning method based on data analysis. The system includes:

[0163] The data acquisition module is used to collect the distribution network operation data in real time through the multi-point sensors arranged in the distribution network. The distribution network operation data includes current, voltage, load fluctuation, harmonic distortion rate and ambient temperature; and to preliminarily store the collected data;

[0164] The data preprocessing module is used to normalize, fill missing values ​​and remove noise on the collected distribution network operation data, segment the data according to the time series characteristics, and generate a preprocessed distribution network operation data set;

[0165] The feature extraction and selection module is used to extract features related to distribution network faults from the preprocessed distribution network operation data set, including harmonic distortion rate, fluctuation rate, overcurrent ratio and differential features, and screen the extracted features based on the feature importance sorting algorithm to generate a feature subset that is highly related to distribution network faults;

[0166] The naive Bayes classifier module is used to construct an improved naive Bayes classifier based on multi-scale features, combine short-term features and long-term features for classification, complete the posterior probability calculation of the fault level based on the optimized parameters, and output the optimal judgment result of the distribution network fault level;

[0167] Kepler optimization module, used to optimize the prior probability and conditional probability smoothing factor of the naive Bayes classifier, uses the Kepler optimization algorithm to iteratively optimize the classifier parameters, and outputs the optimized parameters through population initialization, fitness calculation and position update;

[0168] The fault calibration module is used to input the pre-processed distribution network operation data into the optimized naive Bayes classifier, determine the fault level of the distribution network according to the posterior probability calculation result, and calibrate the fault type, including overcurrent fault, undervoltage fault and harmonic distortion fault, by combining the classifier output and the characteristic value range;

[0169] The warning information generation module is used to generate warning information according to the fault calibration results. The warning information includes fault level, fault type, fault location and recommended response strategy, and sends the warning information to maintenance personnel through sound and light alarm equipment, mobile terminals or distribution network dispatching center system;

[0170] The model update module is used to regularly update the training data set of the naive Bayes classifier based on the distribution network operation history data and new fault data, retrain the classifier model, and optimize the parameters through the Kepler optimization algorithm.

[0171] Embodiment 1:

[0172] The power supply network in Area A, which is under the responsibility of a certain city's power distribution company, covers about 50 square kilometers. There are multiple substations and distribution lines in the area, with a total length of about 120 kilometers. The power supply users exceed 10,000 households. In a monthly maintenance report in May 2024, technicians found that the operating data of some line equipment had frequent abnormal fluctuations, but did not cause obvious faults. Abnormal fluctuations are often difficult to identify in time through traditional fault detection methods, resulting in several small-scale power outages.

[0173] In order to verify the actual application effect of the method of the present invention, the technical team deployed a distribution network fault calibration early warning system based on the method of the present invention on a high-load distribution line in area A, and conducted detailed monitoring of the real-time operation data. The following is a description of the specific application scenario:

[0174] At 10:15 am on June 12, 2024, the distribution network operation monitoring system began to capture abnormal data signals on line "L-1203". After recording by the real-time data acquisition module, it was found that the harmonic distortion rate of the line rose rapidly from the normal value of 0.02 to 0.15, and the load fluctuation rate increased by about 40%. At the same time, the ambient temperature sensor reported that the ambient temperature abnormally rose to 35 degrees Celsius.

[0175] The preprocessing module of the system normalizes and removes noise from the collected data during this period, and generates a feature vector including key features such as harmonic distortion rate, load fluctuation rate and overcurrent ratio. Combined with the naive Bayes classifier of the present invention, real-time analysis is performed. The fault posterior probability calculation results show that the fault level of the line is judged to be a medium fault, and the fault type is marked as a harmonic distortion fault.

[0176] After the classification results are generated, the system automatically generates a detailed early warning report, which records the following information:

[0177] Failure time: 10:18 am, June 12, 2024;

[0178] Fault location: main section of line "L-1203";

[0179] Fault characteristics: The harmonic distortion rate significantly increases to 0.15, the load fluctuation rate exceeds the normal range, and the ambient temperature fluctuates abnormally;

[0180] Fault type: harmonic distortion fault;

[0181] Recommended response measures: Check the operating status of the harmonic filtering equipment and detect whether the related equipment is aging or overloaded.

[0182] Subsequently, the system notified the maintenance team through the mobile terminal, and the maintenance personnel arrived at the site for inspection within 20 minutes. It was confirmed that the source of the fault was the aging of the harmonic filter components due to long-term high-load operation, which significantly reduced the filtering efficiency and generated harmonic distortion. The technicians replaced the filtering equipment as recommended and optimized the line load distribution, avoiding a larger-scale power outage.

[0183] Specific case data include:

[0184] At 12:45 noon on June 13, 2024, the harmonic distortion rate rose to 0.12, and the system successfully warned of a medium fault and located the branch line "L-1203-A1". The maintenance team then discovered and repaired a loose joint problem;

[0185] At 3:10 pm on June 14, 2024, the load fluctuation rate increased from the normal value of 2% to 8%. The system determined it to be a minor fault and calibrated it as an overcurrent problem. Afterwards, the technicians found that a high-power electrical equipment was newly added in a residential area and adjusted the load distribution in time.

[0186] At 4:30 a.m. on June 16, 2024, the system captured abnormal ambient temperature and voltage fluctuations, calibrated them as undervoltage faults, and the maintenance team confirmed that a cooling device in the substation had failed and quickly repaired it.

[0187] The examples show that the method of the present invention can quickly and accurately calibrate the distribution network fault type, significantly improve the detection efficiency, reduce missed alarms and false alarms, optimize the fault response speed, and provide important support for the safe and stable operation of the distribution network.

[0188] The present invention designs a naive Bayes classifier with a multi-scale feature structure, which divides the features into short-term features and long-term features for joint modeling. The short-term features are used to capture transient abnormal signals in the operation of the distribution network, and the long-term features are used to identify operating trends and chronic fault characteristics. Accurate calibration of fault levels is achieved through multi-dimensional joint posterior probability calculation. The improved naive Bayes classifier can more comprehensively utilize the characteristics of distribution network data and significantly improve the ability to distinguish complex fault scenarios.

[0189] The present invention optimizes the core parameters of the naive Bayes classifier through the Kepler optimization algorithm, including the prior probability of the fault level and the conditional probability smoothing factor. The optimization algorithm simulates the motion law of Kepler celestial bodies and realizes efficient parameter optimization through a dynamic gravitational model, thereby overcoming the limitation that traditional optimization methods are prone to fall into local optimality. Experimental data show that the Kepler optimization algorithm improves the convergence efficiency by 25% compared with the particle swarm optimization algorithm. At the same time, the fault level calibration error of the optimized classifier under different data distributions is reduced by more than 15%, further improving the robustness and real-time performance of the system.

[0190] The present invention introduces a dynamic update module, which regularly updates the training data set of the naive Bayes classifier based on the newly added fault data and historical data during operation, and readjusts the model parameters in combination with the Kepler optimization algorithm so that the system can continue to maintain efficient fault calibration capabilities as the distribution network operating environment changes. The dynamic update mechanism avoids the shortcoming of the static model in the traditional method that is prone to aging.

[0191] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A distribution network fault calibration and early warning method based on data analysis, characterized in that: The steps include: S1. Collecting a distribution network operation data set, preprocessing the distribution network operation data set, and performing segmentation processing according to time series characteristics to form a preprocessed distribution network operation data set; S2. Based on the preprocessed distribution network operation data set, key features related to distribution network faults are extracted, and the extracted key features are screened using a feature selection algorithm to obtain a feature subset that is highly correlated with distribution network faults; S3. Construct an improved naive Bayes classifier, and train the improved naive Bayes classifier based on the feature subset and the historical distribution network operation data set; S4. The improved naive Bayes classifier is optimized based on the Kepler optimization algorithm. The fitness of each solution is calculated by randomly initializing multiple solutions of the optimization parameters as a population, and the position and trajectory of the solution are iteratively updated until the fitness of the solution reaches the optimal value, thereby obtaining the optimized naive Bayes classifier. S5. Input the distribution network operation data set into the optimized naive Bayes classifier, determine the fault level of the distribution network and calibrate the fault type according to the output result of the optimized naive Bayes classifier; S6. According to the calibrated fault level and fault type, when the fault level reaches the preset threshold, an early warning information is generated. The early warning information includes the fault level, fault type, fault location and recommended response strategy. The early warning information is sent to the maintenance personnel through the sound and light alarm equipment, mobile terminal or distribution network dispatching center system.

2. A distribution network fault calibration and early warning method based on data analysis according to claim 1, characterized in that: The S1 includes: S11. Collecting a distribution network operation data set through multi-point sensors deployed in the distribution network, wherein the distribution network operation data set includes current I(t), voltage V(t), load fluctuation L(t), harmonic distortion rate H(t) and ambient temperature T(t); S12. Normalizing the collected distribution network operation data set; S13. Fill missing values ​​in the normalized distribution network operation data set and calculate the missing values ​​using linear interpolation; S14. Remove noise from the filled distribution network operation data set and smooth the noise interference using a sliding average method; S15. According to the time series characteristics, the denoised distribution network operation data set is segmented and divided according to the fixed time window Δt to generate the preprocessed distribution network operation data set D pre : D pre ={X smooth (t)|t∈[t0,t0+Δt]}; Among them, D pre is the preprocessed distribution network operation data set, t0 represents the starting time point of the segment, Δt represents the time window size, X smooth (t) represents the data value at time point t.

3. A distribution network fault calibration and early warning method based on data analysis according to claim 1, characterized in that: The S2 includes: S21. From the preprocessed distribution network operation dataset D pre Features related to distribution network faults are extracted from the The harmonic distortion rate H(t) is defined as the total ratio of each harmonic component current to the fundamental current; Voltage fluctuation rate V rate (t), defined as the rate of change of voltage value per unit time; Load fluctuation rate L rate (t), defined as the rate of change of load fluctuation; The overcurrent ratio C(t) is defined as the ratio of the real-time current value to the rated current value; Ambient temperature differential characteristic T diff (t), defined as the change in ambient temperature per unit time; S22. Perform feature selection on the extracted features, use a ranking algorithm based on feature importance to evaluate the relevance of each feature with the distribution network fault, and calculate the relevance score R s : Among them, X feature is a set of eigenvalues, including H(t), V rate (t),L rate (t),C(t),T diff (t), Y is the corresponding fault label set, Cov(X feature ,Y) represents the covariance between the feature value and the fault label, σ(X feature ) and σ(Y) are the standard deviations of the feature value and fault label, respectively; S23. According to the relevance score R s The evaluation results screen the correlation scores higher than the set threshold R threshold The features of the distribution network are used to form a feature subset F that is highly correlated with distribution network faults. selected : F selected ={X feature |R s ≥R threshold }; Among them, F selected is the final selected feature subset, R threshold The relevance score threshold set for the data from an experiment or run.

4. A distribution network fault calibration and early warning method based on data analysis according to claim 1, characterized in that: The S3 includes: S31. According to the feature subset F selected and historical distribution network operation dataset D history Construct a multi-scale feature training dataset D multi , the feature subset F selected According to the time series, it is divided into short-term feature set F short and the long-term feature set F long , the short-term feature set is used to capture short-term fluctuation information, and the long-term feature set is used to capture operation trend information: D multi ={(X short,i ,X long,i ,Y i )},i=1,2,...,N; Among them, X short,i represents the short-term feature vector of the i-th sample, X long,i represents the long-term eigenvector, Y i Indicates the corresponding fault level label; S32. Aiming at the multi-scale characteristics of distribution network operation data, the naive Bayes classifier model is improved and the multi-dimensional joint posterior probability calculation formula is designed: Where C represents the fault level label, including minor fault, medium fault and severe fault, j represents the index of the jth feature in the short-term feature set, k represents the index of the kth feature in the long-term feature set, X = {X short ,X long }, P(C) is the prior probability of the fault level, P(x short,j |C) and P(x long,k |C) represent the conditional probabilities of short-term and long-term characteristics under fault level C, M short and M long are the number of short-term and long-term features, respectively; S33. Aiming at the sparsity and feature interaction of distribution network data, an improved conditional probability smoothing strategy is used to calculate the conditional probability P(x|C): Among them, λ is the dynamic smoothing coefficient, K is the total number of eigenvalues, which is determined by the feature interaction strength, and f(x) represents the interaction strength of eigenvalue x, which is defined as: Among them, Cov(x,k) is the frequency of feature x under fault level C, and Var(x) represents the variance of feature x; S34. Based on the conditional probability calculation results combined with multi-scale feature training data set D multi , train the improved naive Bayes classifier so that the improved naive Bayes classifier takes into account the impact of short-term fluctuations and long-term trends on fault level calibration, and optimizes the improved naive Bayes classifier's ability to distinguish fault modes.

5. A distribution network fault calibration and early warning method based on data analysis according to claim 1, characterized in that: The S4 includes: S41. Based on the improved naive Bayes classifier model, the prior probability P(C) of the fault level and the dynamic smoothing factor λ of the conditional probability are extracted, and the optimization parameter vector S = [P(C), λ] is constructed; S42. Initialize the population of Kepler optimization algorithm. Each individual S in the population i is an optimization parameter vector: S i =[P i (C),λ i ]; Among them, P i (C) is the prior probability distribution of the i-th individual in the population, satisfying ∑ C P i (C) = 1, λ i is a dynamic smoothing factor used to adjust the conditional probability calculation of sparse data, randomly generate the initial value of each solution in the population, and form an initial population S = {S1, S2, ..., S N }; S43. For each individual in the population, calculate its fitness value F(S i ): Among them, E(S i ) represents the classification error, based on the actual fault level C actual and predicted failure level C predicted The difference calculation is: Among them, N test is the number of samples in the test data set; S44. Simulate Keplerian motion to optimize the parameter vector S i The position and velocity are updated: in, represents the speed of the i-th individual at the t-th iteration, G is the gravitational constant, which controls the convergence speed of the optimization process, It is the individual solution with the highest fitness value in the current iteration; S45. Check the boundary of the population solution after each iteration. If the solution exceeds the parameter range, adjust it to the nearest valid value and repeat steps S43 and S44 until one of the following conditions is met: The change in the fitness value of the population is less than the set threshold ΔF threshold ; The maximum number of iterations has been reached; S46. Output the optimized parameter vector and generate the final improved naive Bayes classifier: S optimal =[P optimal (C),λ optimal ]。 6. A distribution network fault calibration and early warning method based on data analysis according to claim 1, characterized in that: The S5 includes: S51. Input the preprocessed distribution network operation data set into the optimized improved naive Bayes classifier S optimal , extract the running data feature X; S52. Calculating the posterior probability of the fault level according to the optimized improved naive Bayes classifier; S53. Compare the posterior probabilities of all fault levels and determine the most likely fault level C best ; S54. According to the fault level C output by the optimized improved naive Bayes classifier best , combined with the characteristic value X, determine the fault type T, which includes: When x C ≥I threshold It is judged as an overcurrent fault, where I threshold is the threshold for overcurrent determination; When x V ≤V threshold It is judged as undervoltage fault, where V threshold It is the threshold value for undervoltage determination; When x H ≥H threshold It is judged to be a harmonic distortion fault, where H threshold is the threshold for judging harmonic distortion; S55. Fault level C of output distribution network best and fault type T, and calibrate the specific category v of the fault.

7. A distribution network fault calibration early warning system based on data analysis, used to execute a distribution network fault calibration early warning method based on data analysis according to any one of claims 1 to 6, characterized in that: The system comprises: A data acquisition module is used to collect distribution network operation data in real time through multi-point sensors arranged in the distribution network, wherein the distribution network operation data includes current, voltage, load fluctuation, harmonic distortion rate and ambient temperature; and to preliminarily store the collected data; The data preprocessing module is used to normalize, fill missing values ​​and remove noise on the collected distribution network operation data, segment the data according to the time series characteristics, and generate a preprocessed distribution network operation data set; The feature extraction and selection module is used to extract features related to distribution network faults from the preprocessed distribution network operation data set, including harmonic distortion rate, fluctuation rate, overcurrent ratio and differential features, and screen the extracted features based on the feature importance sorting algorithm to generate a feature subset that is highly related to distribution network faults; The naive Bayes classifier module is used to construct an improved naive Bayes classifier based on multi-scale features, combine short-term features and long-term features for classification, complete the posterior probability calculation of the fault level based on the optimized parameters, and output the optimal judgment result of the distribution network fault level; Kepler optimization module, used to optimize the prior probability and conditional probability smoothing factor of the naive Bayes classifier, uses the Kepler optimization algorithm to iteratively optimize the classifier parameters, and outputs the optimized parameters through population initialization, fitness calculation and position update; The fault calibration module is used to input the pre-processed distribution network operation data into the optimized naive Bayes classifier, determine the fault level of the distribution network according to the posterior probability calculation result, and calibrate the fault type, including overcurrent fault, undervoltage fault and harmonic distortion fault, by combining the classifier output and the characteristic value range; An early warning information generation module is used to generate early warning information according to the fault calibration results. The early warning information includes fault level, fault type, fault location and recommended response strategy, and sends the early warning information to maintenance personnel through sound and light alarm equipment, mobile terminals or distribution network dispatching center system; The model update module is used to regularly update the training data set of the naive Bayes classifier based on the distribution network operation history data and new fault data, retrain the classifier model, and optimize the parameters through the Kepler optimization algorithm.

Images