Power distribution network fault line selection method based on spwvd time-frequency analysis and deep belief network

Abstract

This invention relates to the field of distribution network fault detection technology and discloses a method for fault line selection in distribution networks based on SPWVD time-frequency analysis and deep belief networks. The method includes the following steps: collecting fault transient current signals of each line in a new distribution network; preprocessing the fault transient current signals; extracting time-frequency features using the optimal SPWVD time-frequency analysis method and generating a time-frequency spectrum matrix; based on the time-frequency spectrum matrix, using a waveform similarity algorithm to measure the similarity of each candidate line in the distribution network, completing the preliminary identification of the faulty line, and obtaining the preliminary identification result; performing normalization preprocessing on the time-frequency spectrum matrix and inputting it into a trained deep belief network; mining the deep nonlinear fault features of the time-frequency spectrum matrix through the deep belief network and outputting the fault probability distribution of each line to obtain a refined identification result; fusing the preliminary identification result and the refined identification result to determine the faulty line in the new distribution network; this scheme achieves accurate and rapid identification of faulty lines in new distribution networks.

Description

Technical Field

[0001] This invention relates to the field of distribution network fault detection technology, specifically to a distribution network fault selection method based on SPWVD time-frequency analysis and deep belief network. Background Technology

[0002] Against the backdrop of deepening energy transition, the penetration rate of distributed energy has significantly increased, driving the evolution of distribution networks towards a more complex form with multiple power sources and branches. The fault characteristics of distribution networks are exhibiting new features such as multi-source, randomness, and nonlinearity. Fault location, as a core technology for refined operation and maintenance of distribution networks, directly affects fault isolation efficiency and outage loss control in terms of accuracy and real-time performance. New distribution network models place higher demands on the adaptability, reliability, and accuracy of fault location technology.

[0003] Current fault location technologies for distribution networks mainly revolve around three main directions: steady-state quantities, transient quantities, and intelligent algorithms. However, all of them have significant drawbacks and cannot adapt to the fault location requirements of new distribution networks. Steady-state quantity technology is simple in principle and low in deployment cost, and is widely used in traditional distribution networks. However, this technology relies on power frequency electrical quantity criteria under the assumption of a single power source. After the integration of distributed energy resources, the fault current exhibits bidirectionality and amplitude uncertainty, and is affected by arc suppression coil compensation and harmonic interference, causing the power frequency electrical quantity criteria to generally fail. Transient quantity technology focuses on the high-frequency components of fault transients and shows certain advantages in some complex fault scenarios. However, the fault signal is easily distorted by the control strategy of distributed energy resources. Transient processes are characterized by short duration and indistinct features. Furthermore, the complex calculations required for feature extraction methods such as time-frequency analysis conflict with the real-time requirements of fault line selection. Fixed analysis parameters are also difficult to adapt to diverse fault conditions. Intelligent algorithm technology has become a research hotspot due to its strong pattern recognition capabilities through machine learning and deep learning. However, its performance is highly dependent on a massive amount of comprehensive fault samples. In contrast, samples for rare scenarios such as high-resistance grounding and multiple fault superposition in distribution networks are scarce, resulting in insufficient model generalization ability. Moreover, the "black box" nature of intelligent algorithms leads to poor decision interpretability. Dynamic changes in distribution network topology and fluctuations in distributed energy output further increase the operation and maintenance costs of the model.

[0004] Traditional time-frequency analysis methods struggle with adaptive parameter optimization. Given the non-stationary and nonlinear characteristics of transient signals from distribution network faults, they often suffer from low time-frequency plot resolution and significant cross-term interference, failing to accurately characterize fault features and directly impacting the accuracy of subsequent fault location. Furthermore, traditional intelligent algorithms struggle to distinguish the electrical characteristics of faulty and non-faulty lines in new distribution network scenarios with multiple power sources. They also lack generalization ability when facing diverse fault scenarios, dynamically changing topologies, and randomly fluctuating distributed power output, and their decisions lack physical meaning, making it difficult to trace the causes of misjudgments. Therefore, there is an urgent need to propose a fault location technology adapted to the new distribution network structure and fault characteristics to improve the accuracy, reliability, and adaptability of fault location. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a novel fault line selection method for distribution networks based on SPWVD time-frequency analysis and deep belief networks. This method solves the problems of poor adaptability, low selection accuracy, and insufficient generalization ability of existing fault line selection methods in new distribution networks, and achieves accurate and rapid identification of faulty lines in new distribution networks.

[0006] To achieve the above objectives, the present invention employs the following technical solution: A method for fault location in distribution networks based on SPWVD time-frequency analysis and deep belief networks, comprising the following steps: Step S1: Collect fault transient current signals of each line in the new distribution network. After preprocessing the fault transient current signals, extract time-frequency features and generate a time-frequency matrix using the optimal SPWVD time-frequency analysis method. The optimal SPWVD time-frequency analysis method is based on the WVD distribution and introduces windowing smoothing to obtain SPWVD. The frequency smoothing window and time smoothing window parameters of SPWVD are globally optimized using the particle swarm optimization algorithm. The fault transient current signal is then analyzed using the SPWVD with optimized window parameters. Step S2: Based on the time-frequency matrix, a waveform similarity algorithm is used to measure the similarity of each candidate line in the distribution network, to complete the preliminary identification of the faulty line and obtain the preliminary identification result; Step S3: After normalizing the time-frequency spectrum matrix, input it into the trained deep belief network. The deep belief network will mine the deep nonlinear fault features of the time-frequency spectrum matrix and output the fault probability distribution of each line to obtain a refined identification result. Step S4: Combine the preliminary judgment results with the fine identification results to determine the faulty lines in the new distribution network.

[0007] Furthermore, in step S1, the window parameters of SPWVD are globally optimized using the particle swarm optimization algorithm, including the following steps: Step S11: Initialize the particle swarm optimization (PSO) algorithm parameters, including the number of particles, number of iterations, learning factor, upper and lower limits of inertia weight, and search range of SPWVD window parameters, and randomly generate the initial position and initial velocity of the particle swarm. Step S12: Construct the velocity and position update formulas for the particle swarm optimization algorithm. The velocity and position of the particles in the solution space are iteratively updated according to the following formulas: ; ; In the formula, V id X represents the flight velocity of particle i in the d-th dimension of the decision variable at the t-th or t+1-th iteration; id V represents the spatial position of particle i in the d-th dimension of the decision variable at the t-th or t+1-th iteration; i X is the velocity of particle i; i Let i be the spatial position of particle i; , ω represents the individual optimal solution and the global optimal solution at time t, respectively; ω is the inertia weight; d is the dimension of the decision variable, consistent with the number of parameters in the SPWVD window; c1 and c2 are learning factors; r1 and r2 are random numbers uniformly distributed between (0, 1); Step S13: Use a nonlinear adaptive dynamic adjustment strategy to determine the inertia weight ω, and dynamically adjust the magnitude of the inertia weight according to the fitness value of the particle under the current iteration number to achieve a balance between the global search and local search capabilities of the algorithm; Step S14: Construct a fitness function based on time-frequency clustering, and calculate the fitness value of each particle according to the following formula: ; Where: Fitness (J), Here, N represents the fitness function value, and N represents the signal sample parameters of the fault transient current signal. The value of SPWVD is the time-frequency distribution value under the corresponding window parameters of the particle. The larger the value of this function, the higher the energy concentration of the time-frequency distribution and the better the time-frequency resolution. Step S15: Update the individual optimal solution of the particle and the global optimal solution of the particle swarm according to the fitness value, and determine whether the iteration termination condition has been met. If not, return to step S12 to continue the iteration. If it has been met, use the current global optimal solution as the optimal solution of the SPWVD window parameters.

[0008] Furthermore, the inertia weight of the nonlinear adaptive dynamic adjustment strategy described in step S13 is specifically formulated as follows: ; In the formula: f i w represents the fitness value of the i-th particle. max and wmin These are the maximum and minimum values ​​of the inertia weight, respectively, usually taken as w. max =0.9, w min =0.4; f min and f a These are the minimum and average particle fitness values ​​for the current iteration number, respectively.

[0009] Furthermore, the update rule for the optimal solution of an individual particle in step S15 is as follows: ; In the formula, Let be the individual optimal solution for particle i at time t+1. Let be the individual optimal solution for particle i at time t. Let be the spatial position of particle i at time t+1. This is the fitness function.

[0010] Furthermore, the specific process of generating the time-frequency spectrum matrix in step S1 is as follows: First, the time-frequency distribution signal output by SPWVD after optimizing the window parameters is divided into M non-folding frequency bands. , (j=1,2,...M), the complex signals of each frequency band are obtained by Fast Fourier Transform (FFT). ; Then, for each frequency band complex signal, sample at N time-domain points t. i The real part amplitude is extracted at (i=1,2,...N), and the time-frequency energy block at the i-th time frequency of the j-th frequency band is defined as follows: ; In the formula, The local energy intensity of the j-th frequency band during the i-th time period, where Re[] represents the real part extraction operation. The center frequency of the j-th frequency band; Then, organize all the time-frequency energy blocks according to the dimension of "frequency band as row and time domain as column" to form an M×N dimensional time-frequency spectrum matrix. It fully characterizes the time-frequency panoramic features of the fault transient current signal.

[0011] Furthermore, the time-spectrum matrix Specifically as follows: ; In the matrix, row vector E p (j,:) reflects the time-domain variation trend of the traveling wave energy in the j-th frequency band, and the column vector E p (:,i) reflects the frequency domain distribution characteristics of the traveling wave energy in the i-th time period.

[0012] Furthermore, in step S2, a waveform similarity algorithm is used to perform preliminary identification of the faulty line, including the following steps: Step S21: Suppose there are n candidate lines in the distribution network. Perform pairwise similarity calculations on the time-frequency matrix of any two lines, La and Lb. First, calculate the squared Euclidean distance using the following formula to quantify the absolute difference in waveforms between the two lines: ; In the formula, Let be the square of the Euclidean distance between lines La and Lb. , These are the elements in the i-th row and j-th column of the time-frequency spectrum matrix for lines La and Lb, respectively, where M is the number of frequency bands and N is the number of time-domain sampling points; Step S22: Normalize the squared Euclidean distance and calculate the similarity coefficient using the following formula to characterize the morphological similarity of the two line waveforms: ; In the formula, Rab is the similarity coefficient between lines La and Lb, where Rab∈[0,1]. The closer the value is to 1, the more similar the waveforms of the two lines are; the closer the value is to 0, the greater the difference in waveform. Step S23: Based on the similarity coefficients among all routes, construct an n×n dimensional global correlation coefficient matrix R. Summate the elements of each row of the global correlation coefficient matrix to obtain the sum R of the morphological similarity between a single route and all other routes. i The line with the lowest sum of morphological similarity is identified as the faulty line in the preliminary judgment results: ; In the formula, R i R is the sum of the correlation coefficients of the i-th line. ij This represents the correlation coefficient between the various lines.

[0013] Furthermore, the deep belief network described in step S3 adopts a hierarchical architecture of "visible layer + multi-layer Restricted Boltzmann Machine (RBM) + Softmax classifier", specifically as follows: Step S31: The visible layer is the input layer, and the number of nodes is exactly matched with the dimension M×N of the time-spectrum matrix. It is used to receive the normalized time-spectrum matrix data. Step S32: The intermediate layer is a multi-layer stacked RBM. The layers are fully connected and there are no connections within the layers. The number of hidden layer nodes of each RBM is adjusted step by step according to the data complexity of the time-frequency matrix, so as to realize the layer-by-layer abstraction and extraction of time-frequency features. Step S33: The top layer is a Softmax classifier, which serves as the output layer. The number of nodes is the same as the number of candidate lines n in the distribution network. It is used to output the fault probability distribution of each candidate line.

[0014] Furthermore, the training process of the deep belief network includes two stages: unsupervised layer-by-layer pre-training and supervised reverse fine-tuning, specifically: Unsupervised layer-by-layer pre-training: A greedy layer-by-layer training strategy is adopted to optimize the parameters of each layer of RBM. The time-spectrum matrix samples generated under different fault scenarios of the new distribution network are input into the first layer of RBM. The weights and bias parameters of the first layer of RBM are optimized by the contrastive divergence algorithm. The hidden layer output of the first layer of RBM is used as the input of the second layer of RBM. All RBMs are trained layer by layer in this way, so that the deep belief network has the initial fault feature extraction capability. Supervised backpropagation fine-tuning: The hidden layer output of the top-level RBM is connected to the Softmax classifier to build a complete deep belief network model. The prediction error of the model is calculated using the cross-entropy loss function. The parameters of the entire deep belief network are adjusted layer by layer from the top to the bottom through the backpropagation algorithm combined with gradient descent, thereby reducing the model prediction error and improving the model's fault classification accuracy.

[0015] Furthermore, the time-spectrum matrix samples cover fault scenarios of different fault locations, different transition resistances, different distributed energy outputs, and different fault types in the new distribution network. The samples are divided into training set, validation set, and test set according to a preset ratio. During the training process, the model hyperparameters are optimized through the validation set, and the fault identification accuracy of the model is verified through the test set.

[0016] Compared with existing technologies, the beneficial effects of the proposed optimal SPWVD time-frequency analysis method are as follows: Based on the traditional SPWVD, the window parameters are optimized by the particle swarm algorithm, and the optimization efficiency of the particle swarm algorithm is improved by adopting a nonlinear adaptive inertia weight adjustment strategy. The global optimization of the window parameters is achieved by using time-frequency clustering as the fitness function. This effectively solves the problems of low resolution and prominent cross-term interference in traditional time-frequency analysis methods. It can accurately characterize the non-stationary and nonlinear features of fault transient signals in new distribution networks, and provide high-quality time-frequency feature data for subsequent fault line selection. This invention integrates waveform similarity algorithms and deep belief networks for fault line selection. First, the waveform similarity algorithm is used to initially identify faulty lines. This algorithm measures similarity based on the time-frequency matrix, which has clear physical meaning and can quickly screen out suspected faulty lines, reducing the computational load of the subsequent deep belief network. Then, the deep belief network is used to mine deep nonlinear information of time-frequency features, which makes up for the shortcomings of the waveform similarity algorithm in capturing the nonlinear distortion features of fault signals, thus improving the identification accuracy in complex fault scenarios. The two work together to achieve accurate identification of faulty lines. The fault location technology of this invention is adapted to the complex structure of new distribution networks with multiple power sources and multiple branches. It overcomes the problems of failure of traditional steady-state quantity technology criteria, fixed parameters of transient quantity technology, and insufficient generalization ability of intelligent algorithms. It has good adaptability to complex fault scenarios such as distributed power source output fluctuations, dynamic changes in distribution network topology, and high-resistance grounding, and has high accuracy and reliability in fault location. The technical solution of this invention has a clear process and a reasonable connection logic between feature extraction and fault identification. It ensures the accuracy of line selection while taking into account real-time performance, enabling rapid location of faulty lines, improving the efficiency of fault isolation in the distribution network, reducing power outage losses, and providing technical support for the refined operation and maintenance of new distribution networks. Attached Figure Description

[0017] Figure 1 This is an overall architecture diagram of a distribution network fault location method based on SPWVD time-frequency analysis and deep belief network, according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the optimal SPWVD time-frequency analysis process according to an embodiment of the present invention; Figure 3 This is a flowchart illustrating the fault selection process of one embodiment of the present invention. Detailed Implementation

[0018] To facilitate understanding of the present invention by those skilled in the art, specific embodiments of the present invention are described below with reference to the accompanying drawings. Preferred embodiments of the present invention are shown in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described in this specification. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the present invention.

[0019] It should be noted that when a component is said to be "fixed to" another component, it can be directly attached to the other component or there may be an intervening component. When a component is said to be "connected to" another component, it can be directly connected to the other component or there may be an intervening component.

[0020] Unless otherwise defined, all technical and scientific terms used in this specification have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention.

[0021] One embodiment of the present invention is as follows: Figure 1 , 2As shown in Figure 3, this distribution network fault selection method based on SPWVD time-frequency analysis and deep belief network is applied to a new type of distribution network with multiple power sources and multiple branches containing distributed energy resources. It enables accurate identification of lines with faults such as single-phase grounding and phase-to-phase short circuits in the distribution network. The overall architecture is as follows: Figure 1 As shown, it includes a signal acquisition module, an optimal SPWVD time-frequency feature extraction module, a waveform similarity preliminary discrimination module, a deep belief network fine recognition module, and a result fusion module. The signal acquisition module acquires the transient current signal of each line fault and transmits it to the optimal SPWVD time-frequency feature extraction module to generate a time-frequency spectrum matrix. The time-frequency spectrum matrix is ​​input to the waveform similarity preliminary discrimination module and the deep belief network fine recognition module respectively to obtain preliminary discrimination results and fine recognition results. Finally, the result fusion module combines the two types of results to determine the faulty line. The optimal SPWVD time-frequency analysis process is as follows: Figure 2 As shown, firstly, the transient current signal of the fault is collected, the SPWVD window parameters and particle swarm algorithm parameters are initialized, the particle swarm velocity and position update formula and the nonlinear adaptive inertial weight strategy are constructed, and the particle swarm iterative optimization is performed with time-frequency aggregation as the fitness function. It is determined whether the iteration termination condition has been met. If not, the iteration continues. If it is met, the optimal solution of the window parameters is obtained. The optimized SPWVD is used to perform time-frequency analysis on the fault signal to generate the time-frequency spectrum matrix. The fault selection process is as follows: Figure 3 As shown, firstly, a time-frequency spectrum matrix is ​​generated using optimal SPWVD. Based on the time-frequency spectrum matrix, a waveform similarity algorithm is used to calculate the similarity coefficients between lines. A correlation coefficient matrix is ​​constructed and summed to obtain a preliminary discrimination result. Simultaneously, the time-frequency spectrum matrix is ​​input into a trained deep belief network, and a refined recognition result is obtained through multi-layer RBM feature extraction and Softmax classification. Finally, the preliminary discrimination result and the refined recognition result are fused to output the final fault line judgment result.

[0022] This distribution network fault selection method based on SPWVD time-frequency analysis and deep belief network first employs the optimal SPWVD time-frequency analysis method: based on WVD, windowing smoothing is introduced to obtain SPWVD suppressed cross terms, and then the window parameters are optimized through particle swarm optimization. A nonlinear adaptive inertial weight adjustment strategy is adopted, and time-frequency clustering is used as the fitness function to achieve global optimization and improve time-frequency resolution. Then, a fusion algorithm is used to achieve fault selection: the waveform similarity algorithm divides the time-frequency spectrum matrix generated by SPWVD into energy blocks, constructs a matrix, and calculates the similarity by combining Euclidean distance and normalized correlation coefficient. The summation criterion is used for preliminary discrimination. The deep belief network takes this time-frequency spectrum matrix as input, and through the "visible layer + multi-layer RBM + Softmax" architecture, it mines deep nonlinear features through unsupervised pre-training and supervised fine-tuning. The two work together to achieve accurate fault selection, thus systematically breaking through the bottlenecks of adaptability, reliability and accuracy of fault selection in new distribution networks.

[0023] (1) SPWVD time-frequency analysis was adopted, and windowing and smoothing were introduced into the WVD distribution calculation model. The calculation model is as follows: ; In the formula: x * (t) is the conjugate of signal x(t), τ is the signal time difference of SPWVD distribution, u is the frequency difference of SPWVD, h(τ) is the frequency smoothing window, and g(u) is the time smoothing window.

[0024] While SPWVD can suppress multi-component signal cross-terms in time-frequency analysis of new distribution networks, it suffers from problems such as decreased time-frequency resolution and difficulty in parameter selection. Particle swarm optimization (PSO) has the advantages of fast convergence and strong global optimization ability, which can effectively make up for the shortcomings of SPWVD and adapt to the complex signal analysis needs of new distribution networks.

[0025] Particle swarm optimization is a typical swarm intelligence optimization algorithm. It typically initializes a set of solutions in a random manner and then updates these solutions iteratively to adjust the entire population towards a better fitness value. Ultimately, it aims to find the optimal solution to the problem within a finite number of iterations.

[0026] The corresponding velocity and position of a particle in space can be expressed as: ; ; In the formula, V i X is the velocity of particle i; i Let i be the spatial position of particle i; , Let be the individual optimal solution and the global optimal solution at time t, respectively; ω be the inertia weight; d be the dimension of the decision variable; c1 and c2 be the learning factors; and r1 and r2 be random numbers uniformly distributed between (0 and 1).

[0027] After each iteration, the optimal solution for each particle needs to be updated according to the following rules: ; in The objective function is the desired fitness value at the corresponding position.

[0028] In the application scenario of SPWVD video analysis for new distribution networks optimized by particle swarm optimization, considering the characteristics of multi-component, significant time-varying characteristics of new distribution network signals and the high timeliness requirements for capturing transient fault signals, a nonlinear adaptive dynamic adjustment strategy is adopted to determine the inertia weight. The specific formula is as follows: ; In the formula: f i w represents the fitness value of the i-th particle.max and w min These are the maximum and minimum values ​​of the inertia weight, respectively, usually taken as w. max =0.9, w min =0.4; f min and f a These are the minimum and average particle fitness values ​​for the current iteration number, respectively.

[0029] The particle swarm optimization algorithm is used to optimize the parameters of the smoothing window function. The optimization process aims at improving the clarity of the time-frequency distribution, and the time-frequency clustering index defines a fitness function to guide the search. The expression of this function is as follows: ; In the formula: The fitness function value is denoted by N, which represents the signal sample parameters. This function measures the degree of energy concentration in the time-frequency distribution; the more concentrated the energy, the greater the fitness value.

[0030] To accurately quantify the time-frequency characteristics of traveling wave signals in distribution network faults and adapt to the multi-component, time-varying characteristics of fault features under distributed power source access, a waveform similarity algorithm achieves preliminary identification of faulty lines through time-frequency energy modeling and similarity measurement. The specific design is as follows: Time-frequency energy block division and definition of broadband traveling wave signals: For distribution network line L p Fault traveling wave current signal i p (t), local energy features are extracted by frequency domain decomposition and time domain sampling.

[0031] Frequency domain decomposition: Dividing a broadband traveling wave signal into M non-folding frequency bands. , (j=1,2,...M), obtain the complex signals of each frequency band through FFT. .

[0032] Time-domain sampling: For each frequency band complex signal, at N time-domain sampling points t i Extract the real part amplitude at (i=1,2,...N), and define the time-frequency energy block at the i-th time frequency of the j-th frequency band as: ; The local energy intensity of the j-th frequency band during the i-th time period, where Re[] represents the real part extraction operation. This is the center frequency of the frequency band.

[0033] Time-frequency matrix construction: Organizing the time-frequency energy blocks according to the dimension of "frequency band as rows and time domain as columns" forms an M×N dimensional time-frequency spectrum matrix, which fully characterizes the time-frequency panoramic features of the traveling wave: ; In the matrix, row vector E p (j,:) reflects the time-domain variation trend of the traveling wave energy in the j-th frequency band, and the column vector E p (:,i) reflects the frequency domain distribution characteristics of the traveling wave energy in the i-th time period.

[0034] ③ Calculation of waveform similarity: Define line L a With L b Squared Euclidean distance of the time-spectrum matrix, absolute difference of the quantized waveform: ; In the formula, , Line L a With L b The element in the i-th row and j-th column of the time-frequency spectrum matrix; The larger the value, the more significant the absolute difference between the two line waveforms.

[0035] Correlation coefficient normalization: right After normalization, the similarity coefficients reflecting morphological similarity are obtained: ; In this formula, the numerator is the inner product of the two matrices, which characterizes the synchronicity of element changes; the denominator is the product of the L2 norms of the two matrices, which achieves amplitude-independent normalization. ∈[0,1], the closer to 1, the more similar the waveform shape, and the closer to 0, the greater the difference in shape.

[0036] Construction of the correlation coefficient matrix: Perform pairwise similarity calculations on the n candidate routes to construct an n×n dimensional global correlation coefficient matrix: ; Fault selection criteria: Summing the elements of each row of the correlation coefficient matrix yields the "sum of morphological similarity" between a single route and all other routes: ; In the formula, R i R is the sum of the correlation coefficients of the i-th line. ij This represents the correlation coefficient between the various lines.

[0037] (2) Although waveform similarity algorithms can provide basic features with clear physical meaning, they are difficult to capture the nonlinear distortion features of fault signals under distributed power supply access. Deep belief networks (DBNs), as a deep learning model of "unsupervised pre-training + supervised fine-tuning", can abstract time-frequency features layer by layer through multi-layer restricted Boltzmann machines (RBMs) to improve the fault identification accuracy in complex scenarios. The connection logic between DBN and waveform similarity algorithms is as follows: taking the time-frequency matrix generated by the waveform similarity algorithm as input, DBN is used to mine deep features and output the fault probability. The specific design is as follows: Layered construction of network structure: The architecture adopted is "visible layer + multi-layer RBM + Softmax". The number of nodes in the visible layer matches the dimension of the input data to receive data; the middle layers are stacked RBMs, with full connections between layers and no connections within layers; the number of nodes in the hidden layer is adjusted according to the data complexity to abstract features layer by layer; the top layer, the Softmax classifier, is the output layer, and the number of nodes matches the number of classification categories.

[0038] Unsupervised layer-by-layer pre-training: A greedy layer-by-layer training strategy is adopted to optimize the parameters of each RBM layer. The original data is input into the first layer, and the parameters are optimized by the contrastive divergence algorithm. The output of the first hidden layer is used as the input of the second layer. The feature extraction capability is obtained by training layer by layer.

[0039] Supervised reverse fine-tuning: The top-level RBM output is fed into Softmax to build the model. The error is calculated using the cross-entropy loss function. The parameters of the entire network are adjusted through backpropagation and gradient descent to improve classification accuracy.

[0040] Reasoning applications and result output: After preprocessing new data, it is input into the model. Through layer-by-layer feature extraction, the Softmax output probability distribution is used to select the category with the highest probability as the result. Multiple rounds of verification can be performed to improve reliability.

[0041] In this embodiment, as Figure 3 As shown, the specific method for obtaining the refined identification result in step S3 is as follows: sort the fault probability distribution of each candidate line output by the deep belief network, and take the line corresponding to the maximum fault probability as the fault line in the refined identification result; the rule for fusing the preliminary judgment result and the refined identification result in step S4 is as follows: if the preliminary judgment result and the refined identification result are the same line, then the line is directly determined to be the fault line; if the two results are different, then the refined identification result output by the deep belief network shall be used to determine the fault line.

[0042] For example, this embodiment provides a novel distribution network fault location method based on optimal SPWVD time-frequency analysis and deep belief network. The specific implementation steps are as follows: Step S1: Collect fault transient current signals of each line in the new distribution network, and generate a time-frequency matrix through optimal SPWVD time-frequency analysis: S111: Collect fault transient current signals using current transformers on each line of the distribution network. p (t), performs preprocessing such as denoising and filtering on the signal to eliminate environmental interference and measurement errors; S112: Construct the SPWVD time-frequency analysis model, and introduce windowing and smoothing processing into the WVD distribution calculation model. * (t) is the conjugate of signal x(t), τ is the signal time difference of SPWVD distribution, u is the frequency difference of SPWVD, h(τ) is the frequency smoothing window, and g(u) is the time smoothing window. S12: The window parameters of SPWVD are optimized using the particle swarm optimization algorithm, specifically: S121: Initialize particle swarm optimization parameters: set the number of particles to 50, the number of iterations to 100, the learning factor c1=c2=2, the maximum value of the inertia weight wmax=0.9, the minimum value wmin=0.4, and the number of parameters of the decision variable dimension matching SPWVD window. S122: Construct the particle swarm velocity and position update formula, and perform iterative updates of particle velocity and position according to the following formula. After each iteration, update the individual optimal solution and the global optimal solution according to the fitness value. ; ; S13: Adjust the inertia weight using a nonlinear adaptive dynamic adjustment strategy: ; Based on the minimum particle fitness value f at the current iteration number min Average value f a The inertia weight is dynamically changed based on the fitness value fi of the i-th particle to achieve a balance between global and local search in the algorithm. S14: Construct a fitness function based on time-frequency clustering, and calculate the fitness value of each particle according to the following formula: ; With the fitness value as the optimization objective, the optimal solution of the SPWVD window parameters is obtained through particle swarm optimization. S15: The update rule for the optimal solution of an individual particle is: ; Let be the individual optimal solution for particle i at time t+1. Let be the individual optimal solution for particle i at time t. Let be the spatial position of particle i at time t+1. This is the fitness function.

[0043] Step S16: Perform time-frequency analysis on the preprocessed fault transient current signal using the optimized SPWVD. First, divide the time-frequency distribution signal output by the SPWVD after optimizing the window parameters into M mutually non-folding frequency bands. , (j=1,2,...M), the complex signals of each frequency band are obtained by Fast Fourier Transform (FFT). ; Then, for each frequency band complex signal, sample at N time-domain points t. i The real part amplitude is extracted at (i=1,2,...N), and the time-frequency energy block at the i-th time frequency of the j-th frequency band is defined as follows: ; Then, organize all the time-frequency energy blocks according to the dimension of "frequency band as row and time domain as column" to form an M×N dimensional time-frequency spectrum matrix. The signal is divided into M=20 non-foldable frequency bands, and the real part amplitude is extracted from N=50 time-domain sampling points. Time-frequency energy blocks are defined and organized by frequency bands as rows and time domain as columns to generate a 20×50 dimensional time-spectrum matrix. .

[0044] Step S2: Based on the time-frequency matrix, perform preliminary fault identification of the circuit using a waveform similarity algorithm. S21: Perform pairwise similarity calculations on the time-spectrum matrices of n=10 candidate lines in the distribution network, calculate the squared Euclidean distance Dab between any two lines using the following formula, and quantify the absolute difference in waveforms: ; S22: Normalize the squared Euclidean distance and calculate the similarity coefficient using the following formula. This yields the morphological similarity measurement results for any two lines: ; S23: Construct a 10×10 dimensional global correlation coefficient matrix R based on the similarity coefficients, where the elements of the matrix R... ij This represents the similarity coefficient between the i-th route and the j-th route; S24: Sum the elements of each row of the correlation coefficient matrix to obtain the sum of the morphological similarity Ri between a single line and the other 9 lines. The line with the smallest sum of morphological similarity is denoted as the line with the preliminary fault diagnosis, L1.

[0045] Specifically, based on the similarity coefficients between all routes, an n×n dimensional global correlation coefficient matrix R is constructed as follows: .

[0046] Step S3: Construct and train a deep belief network to achieve precise identification of faulty lines: S31, the deep belief network adopts a hierarchical architecture of "visible layer + multi-layer restricted Boltzmann machine RBM + Softmax classifier". The visible layer is the input layer, and the number of nodes is completely matched with the dimension M×N of the time-spectrum matrix, which is used to receive the normalized time-spectrum matrix data. S32, the middle layer is a multi-layer stacked RBM, with full connection between layers and no connection within layers. The number of hidden layer nodes in each RBM is adjusted step by step according to the data complexity of the time-frequency matrix, so as to realize the layer-by-layer abstraction and extraction of time-frequency features. S33, the top layer is a Softmax classifier, which serves as the output layer. The number of nodes is the same as the number of candidate lines n in the distribution network, and it is used to output the fault probability distribution of each candidate line. S34, Constructing a Deep Belief Network Architecture: The architecture employs a "visible layer + 3-layer RBM + Softmax" structure. The visible layer has 1000 nodes, matching the dimensions of a 20×50-dimensional time-frequency spectrum matrix. The hidden layers of the 3-layer RBM have 500, 200, and 100 nodes respectively, with full connectivity between layers and no connectivity within layers. The Softmax classifier has 10 nodes, matching the number of candidate lines. S35, Create training samples for deep belief network: Collect fault transient current signals under different fault scenarios (different fault locations, transition resistance, distributed energy output) in the distribution network, generate time spectrum matrix through optimal SPWVD, divide it into training set, validation set and test set in a ratio of 7:2:1, and normalize the samples. S36, Unsupervised layer-by-layer pre-training of deep belief networks: A greedy layer-by-layer training strategy is adopted. The training set samples are input into the first layer RBM, and the weights and bias parameters of the first layer RBM are optimized by the contrastive divergence algorithm. The output of the first hidden layer is used as the input of the second layer RBM. The pre-training of the three layers of RBM is completed in sequence, so that the network has the initial feature extraction capability. S37, Supervised backpropagation fine-tuning of deep belief networks: The output of the top-level RBM is fed into the Softmax classifier to build a complete deep belief network model. The model prediction error is calculated using the cross-entropy loss function. The parameters of the entire network are adjusted through the backpropagation algorithm and the stochastic gradient descent method. The model hyperparameters are optimized using the validation set until the model's recognition accuracy on the test set reaches more than 98%, thus completing the model training. S38. Input the time-frequency matrix generated in step S1 into the trained deep belief network. After three layers of RBM to abstract the deep nonlinear features, the Softmax classifier outputs the fault probability distribution of each line. The line with the highest probability is denoted as the finely identified fault line L2.

[0047] Step S4: Combine the preliminary judgment results with the detailed identification results to determine the final faulty line; S41. Compare the initially identified faulty line L1 with the finely identified faulty line L2. If L1 and L2 are the same line, then directly determine that the line is the faulty line of the distribution network. S42, if L1 and L2 are different lines, then the detailed identification result of the deep belief network shall be used as the standard to determine that L2 is the faulty line; S43 outputs the final fault line determination result, providing a basis for distribution network fault isolation and operation and maintenance.

[0048] The fault location technology in this embodiment was tested in a 10kV new distribution network containing distributed photovoltaic and wind power. The test results show that the accuracy of fault location under different fault scenarios is over 97%. Compared with traditional steady-state quantity fault location technology and single intelligent algorithm fault location technology, the fault location accuracy and adaptability are significantly improved, and the overall fault location time is controlled within 50ms, which meets the real-time requirements of fault location in distribution networks.

[0049] It should be noted that the above-mentioned technical features can be combined with each other to form various embodiments not listed above, all of which are considered to be within the scope of this invention specification; and, for those skilled in the art, improvements or modifications can be made based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A method for fault location in distribution networks based on SPWVD time-frequency analysis and deep belief networks, characterized by: Includes the following steps: Step S1: Collect fault transient current signals of each line in the new distribution network. After preprocessing the fault transient current signals, extract time-frequency features and generate a time-frequency matrix using the optimal SPWVD time-frequency analysis method. The optimal SPWVD time-frequency analysis method is based on the WVD distribution and introduces windowing smoothing to obtain SPWVD. The frequency smoothing window and time smoothing window parameters of SPWVD are globally optimized using the particle swarm optimization algorithm. The fault transient current signal is then analyzed using the SPWVD with optimized window parameters. Step S2: Based on the time-frequency matrix, a waveform similarity algorithm is used to measure the similarity of each candidate line in the distribution network, to complete the preliminary identification of the faulty line and obtain the preliminary identification result; Step S3: After normalizing the time-frequency spectrum matrix, input it into the trained deep belief network. The deep belief network will mine the deep nonlinear fault features of the time-frequency spectrum matrix and output the fault probability distribution of each line to obtain a refined identification result. Step S4: Combine the preliminary judgment results with the fine identification results to determine the faulty lines in the new distribution network.

2. The method for fault location in distribution networks based on SPWVD time-frequency analysis and deep belief networks as described in claim 1, characterized in that, Step S1 involves globally optimizing the window parameters of SPWVD using the particle swarm optimization algorithm, including the following steps: Step S11: Initialize the particle swarm optimization (PSO) algorithm parameters, including the number of particles, number of iterations, learning factor, upper and lower limits of inertia weight, and the search range of the SPWVD window parameters, and randomly generate the initial position and initial velocity of the particle swarm. Step S12: Construct the velocity and position update formulas for the particle swarm optimization algorithm. The velocity and position of the particles in the solution space are iteratively updated according to the following formulas: ； ； In the formula, V id X represents the flight velocity of particle i in the d-th dimension of the decision variable at the t-th or t+1-th iteration; id V represents the spatial position of particle i in the d-th dimension of the decision variable at the t-th or t+1-th iteration; i X is the velocity of particle i; i Let i be the spatial position of particle i; , ω represents the individual optimal solution and the global optimal solution at time t, respectively; ω is the inertia weight; d is the dimension of the decision variable, consistent with the number of parameters in the SPWVD window; c1 and c2 are learning factors; r1 and r2 are random numbers uniformly distributed between (0, 1); Step S13: Use a nonlinear adaptive dynamic adjustment strategy to determine the inertia weight ω, and dynamically adjust the magnitude of the inertia weight according to the fitness value of the particle under the current iteration number to achieve a balance between the global search and local search capabilities of the algorithm; Step S14: Construct a fitness function based on time-frequency clustering, and calculate the fitness value of each particle according to the following formula: ； Where: Fitness (J), Here, N represents the fitness function value, and N represents the signal sample parameters of the fault transient current signal. The value of SPWVD is the time-frequency distribution value under the corresponding window parameters of the particle. The larger the value of this function, the higher the energy concentration of the time-frequency distribution and the better the time-frequency resolution. Step S15: Update the individual optimal solution of the particle and the global optimal solution of the particle swarm according to the fitness value, and determine whether the iteration termination condition has been met. If not, return to step S12 to continue the iteration. If it has been met, use the current global optimal solution as the optimal solution of the SPWVD window parameters.

3. The distribution network fault location method based on SPWVD time-frequency analysis and deep belief network as described in claim 2, characterized in that, The inertia weight of the nonlinear adaptive dynamic adjustment strategy described in step S13 is specifically formulated as follows: ； In the formula: f i w represents the fitness value of the i-th particle. max and w min These are the maximum and minimum values ​​of the inertia weight, respectively, usually taken as w. max =0.9, w min =0.4; f min and f a These are the minimum and average particle fitness values ​​for the current iteration number, respectively.

4. The distribution network fault location method based on SPWVD time-frequency analysis and deep belief network as described in claim 2, characterized in that, The update rule for the optimal solution of an individual particle in step S15 is as follows: ； In the formula, Let be the individual optimal solution for particle i at time t+1. Let be the individual optimal solution for particle i at time t. Let be the spatial position of particle i at time t+1. This is the fitness function.

5. The distribution network fault location method based on SPWVD time-frequency analysis and deep belief network as described in claim 1, characterized in that, The specific process of generating the time spectrum matrix in step S1 is as follows: First, the time-frequency distribution signal output by SPWVD after optimizing the window parameters is divided into M non-folding frequency bands. , (j=1,2,...M), the complex signals of each frequency band are obtained by Fast Fourier Transform (FFT). ; Then, for each frequency band complex signal, sample at N time-domain points t. i The real part amplitude is extracted at (i=1,2,...N), and the time-frequency energy block at the i-th time frequency of the j-th frequency band is defined as follows: ； In the formula, The local energy intensity of the j-th frequency band during the i-th time period, where Re[] represents the real part extraction operation. The center frequency of the j-th frequency band; Then, organize all the time-frequency energy blocks according to the dimension of "frequency band as row and time domain as column" to form an M×N dimensional time-frequency spectrum matrix. It fully characterizes the time-frequency panoramic features of the fault transient current signal.

6. The distribution network fault location method based on SPWVD time-frequency analysis and deep belief network as described in claim 5, characterized in that, The time-frequency matrix Specifically as follows: ； In the matrix, row vector E p (j,:) reflects the time-domain variation trend of the traveling wave energy in the j-th frequency band, and the column vector E p (:,i) reflects the frequency domain distribution characteristics of the traveling wave energy in the i-th time period.

7. The method for fault location in distribution networks based on SPWVD time-frequency analysis and deep belief networks according to claim 1, characterized in that, Step S2 uses a waveform similarity algorithm to perform preliminary identification of faulty lines, including the following steps: Step S21: Suppose there are n candidate lines in the distribution network. Perform pairwise similarity calculations on the time-frequency matrix of any two lines, La and Lb. First, calculate the squared Euclidean distance using the following formula to quantify the absolute difference in waveforms between the two lines: ； In the formula, Let be the square of the Euclidean distance between lines La and Lb. , These are the elements in the i-th row and j-th column of the time-frequency spectrum matrix for lines La and Lb, respectively, where M is the number of frequency bands and N is the number of time-domain sampling points; Step S22: Normalize the squared Euclidean distance and calculate the similarity coefficient using the following formula to characterize the morphological similarity of the two line waveforms: ； In the formula, Rab is the similarity coefficient between lines La and Lb, where Rab∈[0,1]. The closer the value is to 1, the more similar the waveforms of the two lines are; the closer the value is to 0, the greater the difference in waveform. Step S23: Based on the similarity coefficients among all routes, construct an n×n dimensional global correlation coefficient matrix R. Summate the elements of each row of the global correlation coefficient matrix to obtain the sum R of the morphological similarity between a single route and all other routes. i The line with the lowest sum of morphological similarity is identified as the faulty line in the preliminary judgment results: ； In the formula, R i R is the sum of the correlation coefficients of the i-th line. ij This represents the correlation coefficient between the various lines.

8. The method for fault location in distribution networks based on SPWVD time-frequency analysis and deep belief networks according to claim 1, characterized in that, The deep belief network described in step S3 adopts a hierarchical architecture of "visible layer + multi-layer restricted Boltzmann machine (RBM) + softmax classifier", specifically: Step S31: The visible layer is the input layer, and the number of nodes is exactly matched with the dimension M×N of the time-spectrum matrix. It is used to receive the normalized time-spectrum matrix data. Step S32: The intermediate layer is a multi-layer stacked RBM. The layers are fully connected and there are no connections within the layers. The number of hidden layer nodes of each RBM is adjusted step by step according to the data complexity of the time-frequency matrix, so as to realize the layer-by-layer abstraction and extraction of time-frequency features. Step S33: The top layer is a Softmax classifier, which serves as the output layer. The number of nodes is the same as the number of candidate lines n in the distribution network. It is used to output the fault probability distribution of each candidate line.

9. The distribution network fault location method based on SPWVD time-frequency analysis and deep belief network as described in claim 8, characterized in that, The training process of the deep belief network includes two stages: unsupervised layer-by-layer pre-training and supervised reverse fine-tuning, specifically: Unsupervised layer-by-layer pre-training: A greedy layer-by-layer training strategy is adopted to optimize the parameters of each layer of RBM. The time-spectrum matrix samples generated under different fault scenarios of the new distribution network are input into the first layer of RBM. The weights and bias parameters of the first layer of RBM are optimized by the contrastive divergence algorithm. The hidden layer output of the first layer of RBM is used as the input of the second layer of RBM. All RBMs are trained layer by layer in this way, so that the deep belief network has the initial fault feature extraction capability. Supervised backpropagation fine-tuning: The hidden layer output of the top-level RBM is connected to the Softmax classifier to build a complete deep belief network model. The prediction error of the model is calculated using the cross-entropy loss function. The parameters of the entire deep belief network are adjusted layer by layer from the top to the bottom through the backpropagation algorithm combined with gradient descent, thereby reducing the model prediction error and improving the model's fault classification accuracy.

10. The distribution network fault location method based on SPWVD time-frequency analysis and deep belief network as described in claim 9, characterized in that, The time-spectrum matrix samples cover fault scenarios of different fault locations, different transition resistances, different distributed energy outputs, and different fault types in the new distribution network. The samples are divided into training set, validation set, and test set according to a preset ratio. During the training process, the model hyperparameters are optimized through the validation set, and the fault identification accuracy of the model is verified through the test set.

Images