Width learning-based power distribution network grounding fault diagnosis method

Abstract

The invention belongs to the technical field of power system protection, and discloses a power distribution network ground fault diagnosis method based on width learning, aiming at solving the problems that multiple types of power distribution network ground faults may exist and the fault types may overlap in electrical characteristics and are easy to confuse. The method comprises the following steps: adaptively adjusting an inertia weight and a learning factor through a nonlinear dynamic decline strategy, optimizing a BLS hyper-parameter, carrying out data preprocessing, initializing a particle swarm parameter, dynamically updating a particle position and fitness, and outputting an optimal parameter to train a BLS model after iterative optimization. The method is driven by the electric quantity data, judges the arc grounding fault, the DC system grounding fault and the single-phase grounding fault in the power distribution network, and is suitable for realizing the grounding fault diagnosis of the power distribution network.

Description

Technical Field

[0001] The present application belongs to the technical field of power system protection, and in particular relates to a distribution network grounding fault diagnosis method based on width learning. Background Art

[0002] Ground faults in power systems mainly include arc ground faults, DC system ground faults and single-phase ground faults.

[0003] An arc grounding fault occurs when the neutral point is ungrounded or grounded via an arc suppression coil. When insulation damage forms a path to the ground in the system, the fault current returns to the neutral point of the power supply through the ground, generating an arc at the fault point. In a 10kV system with an ungrounded neutral point, arcing often occurs when a phase-to-ground short circuit occurs. This is an arc grounding fault. An arc grounding fault has the following characteristics: (1) A sudden drop in phase voltage will trigger a discharge capacitor current, which flows through the busbar to the fault point and decays rapidly. The oscillation frequency can reach tens of kilohertz or even hundreds of kilohertz, which mainly depends on the grid line parameters, the location of the fault point and the transition resistance value.

[0004] (2) A sudden increase in the non-fault phase voltage will cause a charging capacitor current, which forms a loop through the transformer coil. Due to the large inductance of the circuit, the charging current decays slowly and the oscillation frequency is low. The high frequency and fast decay of the discharge current make it less useful for ground line selection. However, the large amplitude, low frequency, and slow decay of the charging current make it easier to measure and play a major role in ground line selection.

[0005] (3) The characteristics of the transient component are basically not affected by the neutral point grounding method. The zero-sequence current of each line is mainly composed of high-frequency attenuation transient components, which can reach several times, dozens of times, or even hundreds of times the power frequency steady-state component.

[0006] (4) The transient component frequency during arc grounding is related to many factors such as the grid structure, transformer parameters, and fault location, and its value is uncertain. However, the zero-sequence transient currents of the fault and non-fault lines have the same characteristics in terms of frequency and attenuation speed. Moreover, the zero-sequence transient current of the fault line is equal to the sum of the zero-sequence transient currents of all non-fault lines, but in the opposite direction, which is consistent with the zero-sequence steady-state current characteristics in a neutral point ungrounded system.

[0007] (5) When the grounding resistance is small, the transient component is much larger than the steady-state component. The magnitude of the zero-sequence transient current decreases exponentially with the increase of arc resistance, and the attenuation rate increases rapidly.

[0008] In the prior art, faults are generally determined by measuring phase current or zero-sequence current. In some cases where the current value is not large, strong electromagnetic noise may cause inaccurate data and lead to misjudgment.

[0009] A DC system ground fault is a fault in which the insulation of one pole in the DC system is damaged and forms a path to the ground. Currently, the following methods are used to detect DC system ground faults: (1) Pull-out method. This method involves instantaneously grounding the DC circuit and observing the voltage between the positive and negative poles of the DC busbar and the ground. However, the "pull-out method" often causes accidents such as tripping of the control circuit or protection circuit.

[0010] (2) Portable instrument location and manual troubleshooting methods. Use a portable DC ground fault finder to locate DC ground. Use a segmented processing method to measure the voltage between the positive and negative poles to the ground to determine whether it is the positive or negative pole grounded; use segmented inspection to determine which section the DC ground occurs in. However, both methods use manual methods, which are very tedious, time-consuming and labor-intensive.

[0011] A single-phase grounding fault occurs when the insulation of a phase in a power system is damaged, forming a path to the ground. It is the most common type of grounding fault in power systems. Single-phase grounding faults are a common fault in low-current grounding systems. After a single-phase grounding fault occurs, the phase voltages of the two non-faulty phases increase, but the line voltages remain symmetrical, which does not affect power supply. However, long-term operation can cause weak links in the insulation to break down, leading to faults such as phase-to-phase short circuits and voltage transformer burnout. Faults are generally diagnosed based on the voltage at the open delta of the voltage transformer. When a single-phase grounding fault occurs on a distribution line, the insulation monitoring device, operating through the voltage transformers on the substation busbar, detects the grounding fault and issues a grounding signal, allowing on-duty personnel to promptly determine and address the fault. First, comprehensively analyze the voltmeter indication, weather conditions, operating mode, etc. to distinguish the true and false signals; then, in order to reduce the scope of the power outage, carry out sub-grid operation, and at the same time, check whether the electrical equipment is operating normally, such as: whether the equipment is damaged, whether the distribution line is broken or grounded, whether the transformer fuse is blown, whether the cable head is broken, etc.; finally, after confirming that there is no problem with the equipment in the distribution station, use the instantaneous stop and turn off the power in sequence to search.

[0012] The existing method for determining single-phase grounding fault is: (1) When the grounding is incomplete, the voltage of the fault phase decreases, the voltage of the non-fault phase increases but does not reach the line voltage, and the voltage at the open delta of the voltage transformer reaches the set value, sending a grounding signal.

[0013] (2) When fully grounded, the fault phase voltage is zero, the non-fault phase voltage rises to the line voltage, and a three-fold phase voltage appears at the open triangle, sending a grounding signal.

[0014] (3) One phase on the high-voltage side of the voltage transformer is broken, the fault phase indication is not zero, the non-fault phase is still the phase voltage, and there is a voltage value of about 35V at the open triangle, sending a grounding signal.

[0015] The above three methods have defects. Either they may cause misjudgment due to inaccurate data due to strong electromagnetic or environmental noise, or they require traditional manual judgment, which is very cumbersome, time-consuming and labor-intensive.

[0016] Traditional distribution network fault diagnosis relies primarily on manual inspections, empirical judgment, and some basic electrical testing methods, including the following four methods: (1) Inspection is a common fault diagnosis method. It involves manually inspecting distribution equipment, cable lines, switches, etc., observing the operating status of the equipment, and checking for hot spots, looseness, corrosion, and other problems. This method can detect some visible problems, but it cannot detect hidden faults.

[0017] (2) Electrical parameter measurement: Use traditional electrical test instruments such as ammeters, current meters, and voltmeters to measure the electrical parameters of the power distribution system. For example, by measuring parameters such as current, voltage, and resistance, it is possible to identify whether there is an electrical fault, such as a short circuit or ground fault. This method requires manual judgment and execution, which affects the time it takes for protection to operate. It is generally used for fault analysis or fault simulation.

[0018] (3) Relay protection device alarm: Distribution systems are usually equipped with relay protection devices that monitor the operating status of the power grid and issue alarm signals if abnormal conditions such as current overload or short circuit are detected. This method can quickly respond to electrical faults, but it can only provide limited information.

[0019] (4) Infrared thermal imaging: Using an infrared thermal imager to detect the temperature distribution on the surface of the equipment, it can identify possible hot spots, which can help detect potential electrical component failures or connection problems. However, it is only sensitive to abnormal temperatures.

[0020] Although these traditional methods are still effective in some cases, they usually require professional operation and judgment, and there are still major problems in judgment accuracy and timeliness. Moreover, these methods are difficult to cope with complex power systems, large-scale distribution networks and high-density equipment. Summary of the Invention

[0021] The main purpose of this application is to provide a distribution network grounding fault diagnosis method based on width learning, which solves the problem that there may be multiple types of distribution network grounding faults and the fault types may overlap in electrical characteristics, and traditional methods are easily confused. This application can accurately identify the fault type.

[0022] Another purpose of the present application is to provide a distribution network grounding fault diagnosis method based on width learning to solve the problem that traditional fault diagnosis methods may produce false alarms due to factors such as signal interference and measurement errors. The present application can reduce the false alarm rate and missed alarm rate.

[0023] In order to achieve the above objectives, this application provides the following technical solutions: A distribution network grounding fault diagnosis method based on width learning, comprising: Step 1: Perform preliminary cleaning and dimensionality reduction on the collected data to determine the model input variables; Step 2: Build a distribution network fault diagnosis model, adopt a wide learning system framework, use wide shallow neural networks to process high-dimensional power data, improve the particle swarm algorithm, and adaptively adjust the inertia weight and learning factor through a nonlinear dynamic reduction strategy. The core of this strategy is to dynamically adjust the inertia weight and learning factor in the particle swarm algorithm to adapt to different search stages. Specifically, the nonlinear dynamic reduction strategy will gradually reduce the inertia weight according to the global and local search requirements during the iteration process, thereby enhancing local development capabilities and avoiding falling into inefficient search areas in the later stages. At the same time, the learning factor will also be dynamically adjusted to better balance global search and local development capabilities. Step 3: Optimize BLS hyperparameters through the improved particle swarm algorithm, perform data preprocessing, initialize particle swarm parameters, dynamically update particle positions and fitness, and iteratively search for optimal parameters to train the BLS model.

[0024] In some embodiments, step 1 includes: Step 1.1: Preprocess the distribution network monitoring data, fill in missing values, unify the data dimensions through normalization, and balance the feature weights; Step 1.2: Use the Pearson correlation coefficient method to reduce the dimension of the electrical characteristics of the distribution network, analyze the linear correlation between variables, and eliminate redundant features with an absolute value of the correlation coefficient higher than 0.8.

[0025] In some embodiments, in step 1.1, for missing values ​​caused by sensor failure, extreme environment or human error, a method of deleting missing items is used for preprocessing, and records containing missing items are directly deleted.

[0026] In some embodiments, in step 1.1, interpolation is used to pre-process missing values ​​caused by sensor failure, extreme environment or human error. For continuous data, interpolation is used to fill missing items.

[0027] In some embodiments, in step 1.1, mode, median or mean filling is used for preprocessing. For numerical data, the mode, median or mean of the attribute is used to fill in the missing items.

[0028] In some embodiments, step 2 includes: Step 2.1: Select a width learning system as the machine learning framework for distribution network relay protection fault diagnosis technology, and process the measurement data in the power system by building a neural network; Step 2.2: For the inertia weight ω1 and learning factor in the particle swarm algorithm c 1. c 2. Update the three weights using a nonlinear dynamic decreasing method.

[0029] In some embodiments, in step 2.1, it is assumed that the model input training set is X , the output is Y , input the model X Do mapping, enhance the input characteristics of features through mapping, i The expression for group mapping is: Where, W ei and β ei are the weights and biases initially randomly generated from the input to the feature node group; Merge the generated feature node groups into Z=[Z1,Z2,Z3…Z n ]; connect it with the enhanced node group in sequence to generate the corresponding enhanced node group H=[H1,H2,H3…H n ]; All mapped feature nodes Z 1, Z 2,…, Z n Combined into a new feature space, the output of the enhancement layer node is as follows: Where, H i For the i The output of the group enhancement node, γ is the activation function, W hj From the enhanced node H i To the weight matrix of the new feature space, β hj is the bias term of this layer; Set Z=[Z1,Z2,Z3…Z n ] Perform nonlinear transformation and merge H Matrix and Z Matrix, denoted as matrix A ,as follows: in, W ei The calculation formula is as follows: The optimization problem formula is as follows: Where, W is the weight matrix, η is the regularization parameter, a and b is a norm type (usually a positive integer), which represents the measurement method of error and regularization term. u and v is the power parameter, which controls the sensitivity of the error term and the regularization term; Among them, the weight matrix W as follows: Solve the matrix A Pseudo-inverse A + for: Where, I is the identity matrix.

[0030] In some embodiments, in step 2.2, the inertia weight ω1 is improved. During the optimization process, the particle obtains different inertia capabilities as the number of iterations increases, and a dynamically changing inertia weight is adopted: Where, is the initial inertia weight, To meet the final inertia weight when the convergence condition is met, k is the current algebra, T max is the maximum algebra.

[0031] In some embodiments, in step 2.2, the improved learning factor c 1. c 2 is as follows: Where, k is the current algebra, T max is the maximum search algebra.

[0032] In some embodiments, step 3 includes: Step 3.1: Collect electrical variable data and environmental parameters, and perform data preprocessing on model input variables; Step 3.2: Set the fitness function f ( x ), particle swarm population size N , and the upper and lower limits of each parameter; Step 3.3: Initialize the initial distribution position of the particle swarm x i With random initialization speed v i ; Step 3.4: According to f ( x ) Find the individual optimal P best and population optimality G best ; Step 3.5: Update the optimal particle position using nonlinear weight changes; Step 3.6: Recalculate the fitness of each particle; Step 3.7: Update individual optimal P best and population optimality G best ; Step 3.8: Determine whether the upper limit of the number of iterations has been reached. If not, jump to step 3.5. If the upper limit has been reached, output the coordinates of the optimal particle as the hyperparameters of the BLS model.

[0033] Compared with the prior art, the distribution network grounding fault diagnosis method based on width learning provided by this application has the following beneficial effects: This application is driven by electrical quantity data to make judgments on arc grounding faults, DC system grounding faults and single-phase grounding faults in the distribution network. This application is suitable for realizing distribution network grounding fault diagnosis.

[0034] This application can accurately identify fault types and, by studying a large number of different fault samples, uncover the unique characteristics of each fault type. For example, using a cluster analysis algorithm, fault data can be classified according to their characteristics, distinguishing various ground faults. By analyzing characteristics such as voltage and current waveform changes and harmonic content, the model can accurately identify fault types, significantly improving over traditional methods.

[0035] This application can reduce false alarm and missed alarm rates by learning from a large amount of normal operation data and fault data to establish a more accurate fault diagnosis model. By using the autoencoder in width learning to perform feature extraction and noise reduction on distribution network monitoring data, it can extract weak signal characteristics at the initial stage of a fault and reduce the occurrence of missed alarms.

[0036] This application can fill in missing data, compensate for abnormal data with background noise, and ensure the accuracy of input data by processing the collected data.

[0037] This application performs breadth-based learning (BLS), which can handle large-scale, high-dimensional feature datasets.

[0038] Furthermore, this application proposes a method for optimizing hyperparameters in BLS using a particle swarm algorithm with nonlinear weight changes (ISPO-BLS algorithm), which can reduce the number of iterations, improve accuracy, and accurately, efficiently and conveniently judge grounding faults in distribution networks and analyze the fault type.

[0039] This application uses artificial intelligence to replace manual on-site diagnosis, thereby improving fault diagnosis efficiency and saving manpower and material resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the technical solution of this application, the following is a brief introduction to the drawings required for the technical description.

[0041] Figure 1 A flow chart of a distribution network grounding fault diagnosis method based on width learning provided in an embodiment of the present application; Figure 2 Flowchart for determining and preprocessing model input variables provided in an embodiment of the present application; Figure 3 A schematic diagram of width learning (BLS) provided in an embodiment of the present application; Figure 4 A schematic diagram of the inertia weight change law provided in an embodiment of the present application; Figure 5 A schematic diagram of the change pattern of the learning factor provided in the embodiment of the present application; Figure 6 Flowchart of data preprocessing for the ISPO-BLS algorithm provided in an embodiment of the present application; Figure 7 This is a flowchart of the ISPO-BLS algorithm provided in the embodiments of this application. DETAILED DESCRIPTION

[0042] The following is further explained in detail through specific implementation methods.

[0043] like Figure 1 As shown, the embodiment of the present application provides a distribution network grounding fault diagnosis method based on width learning, including: Step 1: Perform preliminary cleaning and dimensionality reduction on the collected data to determine the final input variables for the fault diagnosis model. Preprocess the distribution network monitoring data and use normalization to eliminate feature dimensional differences and balance feature weights. Analyze linear correlations between features, remove redundant variables whose absolute values ​​exceed the set value, reduce multicollinearity interference, and optimize model stability and generalization capabilities.

[0044] Step 2: Build a distribution network fault diagnosis model, employing a wide learning system framework and utilizing wide-shallow neural networks to process high-dimensional power data, balancing computational efficiency and generalization capabilities. Improve the particle swarm algorithm by adaptively adjusting the inertia weight and learning factor through a nonlinear dynamic reduction strategy, balancing global search and local development capabilities, and preventing the algorithm from falling into inefficiency in the later stages. This improves the accuracy and convergence speed of fault feature optimization, and optimizes the performance of the diagnostic model. The core of this strategy is to dynamically adjust the inertia weight and learning factor in the particle swarm algorithm to adapt to different search stages. Specifically, the nonlinear dynamic reduction strategy gradually reduces the inertia weight based on the global and local search requirements during the iteration process, thereby enhancing local development capabilities and avoiding falling into inefficient search areas in the later stages. Simultaneously, the learning factor is dynamically adjusted to better balance global search and local development capabilities.

[0045] Step 3: Optimize BLS hyperparameters (number of feature nodes and enhanced node groups) through an improved particle swarm algorithm (nonlinear weight change), perform data preprocessing (feature extraction, principal component analysis dimensionality reduction, exception handling, etc.), initialize particle swarm parameters, dynamically update particle positions and fitness, and iteratively search for the optimal parameters to train the BLS model, ultimately improving classification accuracy and detection efficiency. This solves the problem of low efficiency of hyperparameter tuning and easy falling into local optimality in traditional methods.

[0046] Step 1 is the preprocessing step, which is the basis of this method and provides high-quality input data for steps 2 and 3. The quality of preprocessing affects the training effect and diagnostic accuracy of the model. Step 2 is the model building step, which is the core of this method. By constructing and optimizing the ISPO-BLS model, it provides an efficient fault diagnosis tool for step 3. The ISPO optimization algorithm ensures the optimality of the model parameters, thereby improving the model performance. Step 3 is the model training and verification step, in which the optimized model is applied to the actual distribution network relay protection fault diagnosis task and its performance is verified. Based on the verification results, the model parameters or optimization strategy can be further adjusted to improve the accuracy and reliability of the model. Through the close cooperation of these three steps, efficient and accurate distribution network fault diagnosis can be achieved.

[0047] In one embodiment, step 1 includes: Step 1.1: Outlier Processing and Normalization. Preprocess the distribution network monitoring data (three-phase voltage / current, zero-sequence / negative-sequence current). Missing values ​​due to sensor failure, extreme environments, or human error are filled by deletion, interpolation, or statistical analysis. Normalization is then used to unify the data dimensions, balance feature weights, and optimize model accuracy and algorithm performance.

[0048] Specifically, the three-phase voltage, three-phase current, zero-sequence current, and negative-sequence current collected by the measurement unit are used as the initial input variables of the model. Due to sensor failure or communication failure, the measured data of electrical quantities may be missing, which has a certain adverse impact on the safety and stability of the power system and the operation of relay protection devices. At the same time, environmental factors such as extreme weather and natural disasters may cause damage to the data acquisition system or sensors, resulting in the loss of measured data. The station operation and maintenance personnel may also cause the data acquisition system to stop or data to be lost by mistakenly deleting data or failing to back up data in time. Therefore, in order to ensure the efficiency and accuracy of the proposed distribution network relay protection fault diagnosis technology, the data should be preprocessed first to fill in the missing data and repair the abnormal data.

[0049] Afterward, data normalization is performed. The purpose of data normalization is to scale the data of different features to the same range, ensuring that each feature has roughly equal weight influencing the model. Data normalization helps improve the performance of machine learning models, especially for algorithms that rely on distance metrics or weights. After data normalization, the data is in the interval [0, 1]. Normalization can minimize regression error and improve prediction accuracy. It is expressed as follows: Where, y is the result after data normalization, x is the original data, x min and x max are the maximum and minimum values ​​in the original data respectively.

[0050] In one embodiment, preprocessing is performed by deleting missing items. Directly deleting records containing missing items is suitable for situations where the proportion of missing data is very small and deletion will not significantly affect the analysis results.

[0051] In one embodiment, interpolation is used for preprocessing. For continuous data, interpolation methods such as linear interpolation, polynomial interpolation, and time series interpolation can be used to fill in missing items. This method estimates missing values ​​based on information from known data points.

[0052] In one embodiment, mode, median, or mean filling is used for preprocessing. For numerical data, the missing items can be filled using the mode (most common value), median (middle value), or mean (average) of the attribute, which are statistical characteristics of known data.

[0053] Step 1.2: Feature variable dimensionality reduction. Dimensionality reduction is performed on the distribution network electrical characteristics (three-phase voltage / current, zero-sequence / negative-sequence current) using the Pearson correlation coefficient method. Linear correlations between variables are analyzed, and highly correlated redundant features are removed. The correlation coefficients of these redundant features typically exceed a certain threshold, such as 0.8. Removing these redundant features helps reduce multicollinearity interference, thereby optimizing the model's stability and generalization capabilities.

[0054] Through step 1.2, we quantify the feature interactions and screen key variables to reduce computational complexity, improve algorithm convergence speed and prediction accuracy, and ultimately determine a streamlined and efficient model input set.

[0055] Specifically, interactions between feature variables refer to the mutual influence or interaction between different features (or attributes) in machine learning and statistical modeling. These interactions can significantly impact model performance, sometimes changing feature importance and critically affecting the model's classification performance. Therefore, they need to be considered in feature engineering and model selection.

[0056] In order to ensure the accuracy, convergence speed and calculation time of the algorithm, it is necessary to reduce the dimensionality of the target samples, discard one or several features with high correlation coefficients, use the electrical data detected by the measurement unit as the set input, and introduce the Pearson correlation coefficient method to realize correlation analysis and dimensionality reduction.

[0057] The Pearson correlation coefficient is a statistical indicator used to measure the strength and direction of the linear relationship between two continuous variables. The Pearson correlation coefficient is used to calculate the correlation between feature variables, and its absolute value ranges from [-1, 1]. The larger the absolute value, the stronger the correlation. When the correlation coefficient is greater than 0.8, it indicates that there is a strong positive linear relationship between the two variables. Highly correlated variables may contain duplicate information, which does not necessarily provide additional characterization capabilities. Therefore, during feature selection or modeling, consider removing one of them or combining them into new features to reduce redundant information. Pearson correlation coefficient R The calculation formula is as follows: Where, X i For the i data points X the value of the variable, For variables X The sample mean of Y i For the i data points Y the value of the variable, for Y The sample mean of the variable,n is the total number of sample data.

[0058] After performing Pearson correlation analysis, the Pearson correlation coefficient between two variables can be obtained, providing a method to quantify the relationship between the two variables. It can also help identify redundant or irrelevant features in the data, allowing for feature dimensionality reduction or elimination of irrelevant features to simplify the model and improve its generalization ability. Finally, it can also be used to detect multicollinearity between features. Multicollinearity can lead to model instability and reduced predictive performance. Therefore, correlation analysis can be used to filter out highly correlated features to reduce the impact of collinearity on the model.

[0059] In step 1, the collected data is preliminarily cleaned and dimensionally reduced to determine the final model input variables. The process is shown in the following figure: Figure 2 shown.

[0060] In one embodiment, step 2 includes: Step 2.1: Select the Broad Learning System (BLS) as the machine learning framework for distribution network relay protection fault diagnosis technology. By constructing a wide and shallow neural network to process large-scale, high-dimensional measurement data in the power system, BLS has good generalization ability and learning effect. Figure 3 shown.

[0061] Assume that the input training set of the fault diagnosis model is X , the output is Y The first step of BLS is to input the model X Do mapping, try to enhance the input characteristics of the feature through mapping, where i The expression for group mapping is: Where, W ei and β ei are the weights and biases initially randomly generated from the input to the feature node group.

[0062] Then combine the generated feature nodes into Z=[Z1,Z2,Z3…Z n ]. Connect it with the enhanced node group in sequence to generate the corresponding enhanced node group H=[H1,H2,H3…H n ].

[0063] H i For the output of the enhancement layer node, all the mapped feature nodes Z 1, Z 2,…,Z n Combined into a new feature space for further processing, as follows: Where, H i For the i The output of the group enhancement node, γ Is an activation function, usually used to further improve nonlinear expression capabilities, W hj From the enhanced node H i To the weight matrix of the new feature space, β hj is the bias term for this layer, used to adjust the output of the node. A layer refers to a hidden layer or output layer in a neural network. Specifically, it is a layer in a multi-layer network structure built through width learning or other deep learning methods.

[0064] Set Z=[Z1,Z2,Z3…Z n ] to perform nonlinear transformation. Finally, merge H Matrix and Z Matrix, denoted as matrix A ,as follows: To correct the initial random W ei , since we already know Y The label is a classification problem. Therefore, the weight W ei The calculation method is: In actual calculations, the matrix A There is no actual inverse operation, so the ridge regression method is introduced to find the matrix A Pseudo-inverse A + .

[0065] Among them, the following formula is the optimization problem formula: This formula is an optimization problem, the goal is to minimize the bi-norm of a weighted error. Specifically, the goal is to adjust the matrix W To make Minimum. Here W is the desired weight matrix, A and Y is a known matrix, and η is the regularization parameter, which controls the balance between error and weight. a andb is the norm type, which represents the measurement method of error and regularization term. u and v is a power parameter that controls the sensitivity of the error term and the regularization term.

[0066] The matrix is ​​given by W The solution of η is a regularization coefficient, I is the identity matrix, A T is a matrix A The transpose of . By this formula, the matrix W The solution can be expressed as: Where, η is the regularization parameter I is the identity matrix, W is the weight matrix of the model.

[0067] Solve the matrix A Pseudo-inverse A + for: Calculate the matrix A Pseudo-inverse A + It can make the calculation process of the optimization problem more stable and avoid adverse effects in the calculation.

[0068] This demonstrates that BLS training does not rely on gradient descent and is less prone to falling into local optima. Furthermore, when adding new nodes, the pseudo-inverse does not need to be repeatedly calculated, ensuring efficient computational speed while expanding nodes horizontally. Therefore, this embodiment selects width learning as the basis for the fault diagnosis model and employs both input feature enhancement and parameter optimization to further enhance the accuracy of the algorithm model.

[0069] Step 2.2: For the inertia weight ω1 in the particle swarm algorithm, and the learning factor c 1. c 2. The three weights are updated in a nonlinear dynamic decreasing manner, so that they can always maintain an efficient optimization rhythm throughout the search, and the low efficiency of the later search will not be caused by the fixed weights.

[0070] Step 2.2 improves the particle swarm optimization algorithm. Particle swarm optimization (PSO) is a typical swarm intelligence optimization algorithm. Its basic concept is inspired by collective behaviors in nature, such as animal predation. In this algorithm, the flight behavior of each particle in space simulates the hunting motion of birds in nature. By observing the collaborative system of individual birds during the hunting process, the algorithm uses the particle motion and mathematical formulas to describe it, and then uses it to solve complex optimization problems.

[0071] In the particle swarm algorithm, each particle has its own corresponding position and velocity at each moment. After the algorithm is started, each particle is randomly assigned a velocity and position, and the particles of the pre-set group size are indiscriminately scattered in the solution vector space. Then, the position and velocity are updated through the free search of each individual and the continuous iteration. In each iteration, the individual's optimal point is calculated by the fitness function. P best and population optimality G best size for replacement.

[0072] Individual optimal point P best Represents the optimal point found by a single particle during iteration, the optimal point of the population G best It represents the optimal point found by all particles in this search. With each iteration, the individual optimal point needs to be updated accordingly. P best and population optimality G best .

[0073] However, it is inevitable that the algorithm has the disadvantages of being prone to premature convergence, being prone to local extreme values ​​for multi-peak functions, decreasing search efficiency in the later stages, and having random and uncontrollable initial states. To address this, this embodiment proposes a nonlinear decreasing dynamic weight update strategy for the inertia weight ω1 in the particle swarm algorithm and the learning factor c 1. c 2. The three weights are updated in a nonlinear dynamic decreasing manner, so that they can always maintain an efficient optimization rhythm throughout the search, and the low efficiency of the later search will not be caused by the fixed weights.

[0074] Inertia weight ω1, learning factor c 1 and c 2 are the three most important parameters in the particle swarm optimization algorithm. This embodiment proposes a method for dynamically determining the values ​​of these three parameters to improve the global search ability and convergence speed of particles, while effectively preventing the particle swarm from falling into local optimality.

[0075] In one embodiment, the inertia weight ω1 is improved. The inertia weight ω1 represents the ability to maintain the previous state of motion, and the size of the inertia weight value is very important to the overall search ability of the particle. In the early stage of the search, an attempt is made to maintain the inertia weight value at a high value to ensure that it can explore the unknown area efficiently and quickly in large steps in the early stage, and traverse all areas within the specified search range as much as possible in the early stage of the iteration; as the number of searches increases, the particles are usually gathered near the optimal point. At this time, it is hoped that the particles can slow down and explore the global optimal point position more finely with smaller steps, rather than still maintaining a high inertia to fly to other areas. In order to allow particles to obtain different inertia capabilities as the number of iterations increases during the optimization process, a dynamically changing inertia weight is designed: Where, is the initial inertia weight, To meet the final inertia weight when the convergence condition is met, k is the current algebra, T max is the maximum number of generations. In the case of standard particle swarm optimization, the inertia weight is usually set to =0.9, =0.4, the change rule of inertia weight is as follows Figure 4 Shown as a solid curve.

[0076] Furthermore, the learning factor c 1. c 2 for improvement. Learning factor c 1 represents the degree of individual cognition of the particle, that is, the degree to which the particle approaches the optimal point in its own motion process, which in individual cognition means the proportion that affects the next speed; learning factor c 2 represents the degree of group cognition of particles, that is, the degree to which particles approach the optimal point in the group motion process. In group cognition, it means the proportion that affects the next speed.

[0077] In the early stage of PSO particle search, the particles are distributed in the entire sample space. Each particle does not go through many iterations, and the point found is likely not the optimal value point of the final objective function. Therefore, in the early stage of search, try to c 1 maintains a high level, while c 2. Keep the operation low, so that particles can find the optimal point based on their own experience as much as possible in the early stage without being disturbed by immature ideas of the group. When the search enters the middle and late stages, many iterations have been passed and the particles have gradually approached the optimal value point. In this case, try to c 1 gradually pulls to a lower level, while c2 is raised to a higher level. Because at this point the particles have basically traversed the sample space, and the level of cognition of the entire space has reached a high level.

[0078] Using learning factors c 1 and c 2. Dynamically changing method, in order to better balance the local search and global search capabilities of particles: Where, k is the current algebra, T max is the maximum search algebra.

[0079] Learning Factor c 1. c The change curve of 2 is as follows Figure 5 As shown, the solid line curve in the figure is c 1. The dotted curve is c 2.

[0080] In one embodiment, in order to improve the classification ability and detection efficiency of the width learning system in the distribution network relay protection fault diagnosis, the particle swarm algorithm with nonlinear weight change proposed in step 2 is used to optimize the hyperparameters in BLS. The main hyperparameters involved in BLS are the number of characteristic node groups m and enhanced node group number n , the specific algorithm fusion process is as follows Figure 6 and Figure 7 shown.

[0081] In step 3, the ISPO-BLS (nonlinear weight-varying particle swarm optimization width learning) distribution network relay protection fault diagnosis model is trained, which includes: Step 3.1: Collect data on electrical variables such as current, voltage, and other environmental parameters such as temperature and humidity; perform data preprocessing on the model input variables, including feature extraction, data dimensionality reduction based on principal component analysis, abnormal data processing, normalization, and other existing technical operations; Step 3.2: Set the fitness function f ( x ), particle swarm population size N , and the upper and lower limits of each parameter; Step 3.3: Initialize the initial distribution position of the particle swarm x i With random initialization speed v i ; Step 3.4: According to f ( x ) Find the individual optimalP best and population optimality G best ; Step 3.5: Update the optimal particle position using nonlinear weight changes. In the BLS model, the optimal point usually refers to the optimal configuration of model parameters (such as weights and biases) to achieve the best model performance. The optimal point is usually defined as minimizing the loss function, that is, minimizing the mean squared error between the actual value and the model prediction value. This is different from the optimal fitness point in the particle swarm algorithm. Step 3.6: Recalculate the fitness of each particle. Step 3.7: Update individual optimal P best and population optimality G best ; Step 3.8: Determine whether the upper limit of the number of iterations has been reached. If not, jump to step 3.5. If the upper limit has been reached, output the coordinates of the optimal particle as the hyperparameters of the BLS model.

[0082] Through the above steps of training the BLS model, and based on the output of step 3.8, the distribution network fault diagnosis is finally completed.

[0083] Therefore, this application processes outliers and normalizes the collected data. Then, a particle swarm optimization algorithm with nonlinear weight variation is used to optimize hyperparameters in the BLS algorithm (ISPO-BLS algorithm). This reduces the number of iterations and improves accuracy. The training width learning model is then used to perform fault classification, enabling accurate, efficient, and convenient determination of ground faults in the distribution network and analysis of the fault type.

[0084] The above description is only a specific implementation method of the present application, but the scope of protection of the present application is not limited thereto. Any changes or replacements that can be easily thought of by any technician familiar with this technical field within the technical scope disclosed in this application should be covered by the scope of protection of the present application.

Claims

1. A distribution network grounding fault diagnosis method based on width learning, characterized in that: include: Step 1: Perform preliminary cleaning and dimensionality reduction on the collected data to determine the model input variables; Step 2: Build a distribution network fault diagnosis model, adopt a wide learning system framework, use a wide shallow neural network to process high-dimensional power data, improve the particle swarm algorithm, and adaptively adjust the inertia weight and learning factor through a nonlinear dynamic reduction strategy; Step 3: Optimize BLS hyperparameters through the improved particle swarm algorithm, perform data preprocessing, initialize particle swarm parameters, dynamically update particle positions and fitness, and iteratively search for optimal parameters to train the BLS model.

2. The distribution network grounding fault diagnosis method based on width learning according to claim 1 is characterized in that: Step 1 includes: Step 1.1: Preprocess the distribution network monitoring data, fill in missing values, unify the data dimensions through normalization, and balance the feature weights; Step 1.2: Use the Pearson correlation coefficient method to reduce the dimension of the electrical characteristics of the distribution network, analyze the linear correlation between variables, and eliminate redundant features with an absolute value of the correlation coefficient higher than 0.

8.

3. The distribution network grounding fault diagnosis method based on width learning according to claim 2, characterized in that: In step 1.1, for missing values ​​caused by sensor failure, extreme environment or human error, the method of deleting missing items is used for preprocessing, and the records containing missing items are directly deleted.

4. The distribution network grounding fault diagnosis method based on width learning according to claim 2, characterized in that: In step 1.1, interpolation is used to preprocess missing values ​​caused by sensor failure, extreme environment or human error. For continuous data, interpolation is used to fill missing items.

5. The distribution network grounding fault diagnosis method based on width learning according to claim 2, characterized in that: In step 1.1, mode, median, or mean filling is used for preprocessing. For numerical data, the mode, median, or mean of the attribute is used to fill the missing items.

6. The distribution network grounding fault diagnosis method based on width learning according to claim 1, characterized in that: Step 2 includes: Step 2.1: Select a width learning system as the machine learning framework for distribution network relay protection fault diagnosis technology, and process the measurement data in the power system by building a neural network; Step 2.2: For the inertia weight ω1 and learning factor in the particle swarm algorithm c 1. c 2. Update the three weights using a nonlinear dynamic decreasing method.

7. The distribution network grounding fault diagnosis method based on width learning according to claim 6, characterized in that: In step 2.1, assume that the model input training set is X , the output is Y , input the model X Do mapping, enhance the input characteristics of features through mapping, i The expression for group mapping is: Where, W ei and β ei are the weights and biases initially randomly generated from the input to the feature node group; Merge the generated feature node groups into Z=[Z1,Z2,Z3…Z n ]; connect it with the enhanced node group in sequence to generate the corresponding enhanced node group H=[H1,H2,H3…H n ]; All mapped feature nodes Z 1, Z 2,…, Z n Combined into a new feature space, the output of the enhancement layer node is as follows: Where, H i For the i The output of the group enhancement node, γ is the activation function, W hj From the enhanced node H i To the weight matrix of the new feature space, β hj is the bias term of this layer; Set Z=[Z1,Z2,Z3…Z n ] Perform nonlinear transformation and merge H Matrix and Z Matrix, denoted as matrix A ,as follows: in, W ei The calculation formula is as follows: The optimization problem formula is as follows: Where, W is the weight matrix, η is the regularization parameter, a and b is the norm type, u and v is the power parameter; Among them, the weight matrix W as follows: Solve the matrix A Pseudo-inverse A + for: Where, I is the identity matrix.

8. The method for diagnosing ground faults in a distribution network based on width learning according to claim 6, characterized in that: In step 2.2, the inertia weight ω1 is improved. During the optimization process, the particle obtains different inertia capabilities as the number of iterations increases, and a dynamically changing inertia weight is adopted: Where, is the initial inertia weight, To meet the final inertia weight when the convergence condition is met, k is the current algebra, T max is the maximum algebra.

9. The distribution network grounding fault diagnosis method based on width learning according to claim 6, characterized in that: In step 2.2, the improved learning factor c 1. c 2 is as follows: Where, k is the current algebra, T max is the maximum search algebra.

10. The distribution network grounding fault diagnosis method based on width learning according to claim 1, characterized in that: Step 3 includes: Step 3.1: Collect electrical variable data and environmental parameters, and perform data preprocessing on model input variables; Step 3.2: Set the fitness function f ( x ), particle swarm population size N , and the upper and lower limits of each parameter; Step 3.3: Initialize the initial distribution position of the particle swarm x i With random initialization speed v i ; Step 3.4: According to f ( x ) Find the individual optimal P best and population optimality G best ; Step 3.5: Update the optimal particle position using nonlinear weight changes; Step 3.6: Recalculate the fitness of each particle; Step 3.7: Update individual optimal P best and population optimality G best ; Step 3.8: Determine whether the upper limit of the number of iterations has been reached. If not, jump to step 3.

5. If the upper limit has been reached, output the coordinates of the optimal particle as the hyperparameters of the BLS model.

Images