Power distribution network fault detection method and apparatus based on traveling wave signal

Abstract

Disclosed in the present invention are a power distribution network fault detection method and apparatus based on a traveling wave signal. The method comprises the following steps: acquiring a fault traveling wave signal of a power distribution network, and determining a fault location of the power distribution network; on the basis of an optimized decomposition algorithm, performing frequency decomposition on the fault traveling wave signal, so as to obtain an instantaneous frequency; on the basis of a pre-constructed optimized fault detection model, performing feature extraction on the instantaneous frequency, so as to provide fault characteristics of the power distribution network; and on the basis of the fault characteristics of the power distribution network, providing a fault category corresponding to the fault traveling wave signal. The present invention can accurately determine the specific location and category of a fault in a power distribution network, thereby improving the accuracy and efficiency of fault detection.

Description

A method and device for distribution network fault detection based on traveling wave signals Technical Field

[0001] This invention belongs to the field of power distribution network fault detection technology, specifically relating to a power distribution network fault detection method and device based on traveling wave signals. Background Technology

[0002] With the large-scale deployment of power cables in power distribution networks, the resulting difficulties in cable fault location and classification have become particularly prominent.

[0003] Traditional cable fault location primarily relies on offline methods. However, offline methods suffer from limitations such as complex external equipment, limited applicability, and low accuracy, making them unsuitable for the demands of smart grid development. Furthermore, cable fault classification mainly employs the Hilbert-Huang Transform (HHT), which typically combines Empirical Mode Decomposition (EMD) with Hilbert transform. For example, patent CN106501668A discloses a method for selecting a single-phase open-circuit fault in a traditional distribution network, comprising: calculating the negative-sequence current of each feeder using the symmetrical component method; performing EMD decomposition on the negative-sequence current of each feeder when its amplitude exceeds the negative-sequence setting value to obtain the intrinsic mode components (IMFs) of each order; performing Hilbert transform on each IMF to obtain its corresponding instantaneous amplitude waveform; calculating the change in the sum of the instantaneous amplitudes of the IMFs of each feeder's set order over m cycles before and after the fault; and selecting the open-circuit fault line based on the magnitude of the change. EMD adaptively decomposes complex signals into a finite number of instantaneously meaningful, amplitude- or frequency-modulated high-frequency and low-frequency Intrinsic Mode Functions (IMFs). The frequency components of each IMF component are related to the sampling frequency and also vary with the signal itself. However, the HHT algorithm can exhibit mode aliasing during mode decomposition. Mode aliasing refers to the fact that the same IMF component contains different scale components, which directly leads to the lack of sufficient physical meaning in the aliased IMFs, resulting in confusion in the subsequent time-frequency distribution.

[0004] Therefore, how to provide a high-precision fault detection method is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] To address the shortcomings of the existing technologies, this invention provides a distribution network fault detection method based on traveling wave signals. This method can accurately determine the specific location and type of distribution network faults, improving the accuracy and efficiency of fault detection.

[0006] In a first aspect, the present invention provides a method for detecting faults in a distribution network based on traveling wave signals, specifically including the following steps: acquiring a fault traveling wave signal of the distribution network and determining the fault location of the distribution network; performing frequency decomposition on the fault traveling wave signal based on an optimized decomposition algorithm to obtain the instantaneous frequency; extracting features from the instantaneous frequency based on a pre-constructed optimized fault detection model to give the fault characteristics of the distribution network; and giving the fault category corresponding to the fault traveling wave signal based on the fault characteristics of the distribution network.

[0007] Furthermore, determining the fault location in the distribution network includes: identifying the corresponding power cable based on the acquired fault traveling wave signal; providing the time difference between the transmission of the fault traveling wave signal to both ends of the corresponding power cable based on the identified power cable; and determining the fault location on the power cable based on the time difference between the transmission of the fault traveling wave signal to both ends of the corresponding power cable and in conjunction with the length of the corresponding power cable.

[0008] Furthermore, the fault traveling wave signal is frequency decomposed based on an optimized decomposition algorithm to obtain the instantaneous frequency, including: deriving the noise figure based on the fault traveling wave signal and generating multiple first white noises accordingly; combining the multiple first white noises with the fault traveling wave signal respectively; performing empirical mode decomposition on each combined fault traveling wave signal to obtain multiple first intrinsic mode function components; performing mean processing on all first intrinsic mode function components, and performing Hilbert transform on the mean processing result to obtain the instantaneous frequency.

[0009] Furthermore, based on the fault traveling wave signal, a noise figure is given and multiple first white noises are generated accordingly, including: determining the range of noise figure values ​​based on the standard deviation of the fault traveling wave signal; using the noise figure as particles and combining the range of noise figure values, the particles are initialized to give a particle swarm composed of multiple random particles.

[0010] The fitness values ​​of the particle swarm are calculated, and the local optimal fitness value of each particle is given. The minimum value of all local optimal fitness values ​​is taken as the global optimal fitness value. The positions of the particle swarm are updated based on the evolution function and using extreme value perturbation. The fitness value calculation and position update steps are repeated iteratively until convergence, and the final global optimal fitness value is given. Based on the final global optimal fitness value, the corresponding noise coefficient value is determined, and multiple first white noise values ​​of the noise coefficient are obtained by combining different Gaussian white noise values.

[0011] Furthermore, the fitness value of the particle swarm is calculated to give the local optimal fitness value of each particle, including: combining the noise figure value of each particle in the particle swarm with the same Gaussian white noise to obtain multiple second white noises; combining the multiple second white noises with the fault traveling wave signal to obtain multiple signals to be decomposed.

[0012] Empirical mode decomposition is performed on multiple signals to be decomposed to obtain multiple sets of second intrinsic mode function components. Based on the fitness value function, the multiple sets of second intrinsic mode function components and the fault traveling wave signal are analyzed and processed to give the local optimal fitness value of each particle.

[0013] Furthermore, the fitness value function satisfies the following relationship:

[0014]

[0015] In the formula, RRMSE is the fitness value, N1 is the total number of components of the second intrinsic mode function, x0(k) is the fault traveling wave signal, and c max (k) is the second intrinsic mode function component with the highest correlation coefficient with the fault traveling wave signal among the k-th group of second intrinsic mode function components obtained by empirical mode decomposition, k∈[1,K], where K is the number of groups of second intrinsic mode function components, i.e. the number of signals to be decomposed;

[0016] The evolution function satisfies the following relationship:

[0017]

[0018] In the formula, Let be the position of the particle, representing the position of the i-th particle in the d-th dimension during the (t+1)-th iteration, where i = 1, 2, ..., m, d = 1, 2, ..., D, m is the number of particles in the swarm, and D is the set search dimension; ω is the momentum-inertia coefficient; c1 and c2 are non-negative learning factors; r1, r2, r3, and r4 are all random numbers following a U(0, 1) distribution; p is the extreme point of the individual particle and the swarm, t0 and t... g These represent the evolutionary stagnation steps for the individual particle extremum and the global extremum, respectively; T0 and T g These are the threshold numbers of stagnation steps required to disturb the individual particle extrema and the global extrema, respectively.

[0019] Furthermore, multiple instances of first white noise are combined with the fault traveling wave signal respectively, satisfying the following relationship:

[0020]

[0021]

[0022] In the formula: x m x(t) is the fault traveling wave signal after incorporating the first white noise, and n is the fault traveling wave signal. m (t) represents the first white noise, A noise Noise figure It is Gaussian white noise.

[0023] Furthermore, the pre-construction of the optimized fault detection model includes: acquiring historical fault traveling wave signals of the distribution network and providing the corresponding historical instantaneous frequencies based on an optimized decomposition algorithm; normalizing the historical instantaneous frequencies to obtain normalized data and dividing it into training, validation, and test sets; training the neural network model using the training set to obtain the trained neural network model; validating the trained neural network model using the validation set until the training termination condition is met, and then verifying the performance of the neural network model using the test set to obtain the optimized fault detection model.

[0024] Furthermore, the neural network model is trained using the training set, resulting in the trained neural network model, including: S31, initializing the neural network model and loss function, providing initial model parameters and loss function coefficients; where the model parameters include the weight matrix and bias vector; S32, performing forward propagation of the neural network model on the training data of the training set and calculating the value of the loss function; S33, calculating the gradient of the loss function with respect to the model parameters of the neural network model based on the backpropagation algorithm, and using the gradient descent algorithm to update the model parameters and loss function coefficients of the neural network model according to the gradient, so as to minimize the value of the loss function; S34, repeating S32 and S33 iteratively until convergence, resulting in the trained neural network model.

[0025] The loss function satisfies the following relationship:

[0026]

[0027]

[0028]

[0029] In the formula, E is the loss function, α and β are the coefficients of the loss function, n is the total number of network weights in the neural network model, l is the average number of network weights in the neural network model, and w i W represents the network weights of the i-th neural network model. k Let J be the network weight vector for the k-th iteration, and ε(W) be the Jacobian matrix. k ) is based on W k The error vector is given by μ, which is the step size control parameter with a value of 0.001, and γ, which is the number of effective parameters of the neural network model, γ = N - 2β'tr(H). -1 β' is the β value of the previous iteration, N is the total number of model parameters of the neural network model; H is the Hessian matrix of the loss function, and tr(H) represents the trace of matrix H.

[0030] Furthermore, the fault detection model is optimized to satisfy the following relationship:

[0031]

[0032]

[0033]

[0034]

[0035]

[0036]

[0037] In the formula, H s,t To optimize the input normalized data of the fault detection model at time t, It is the current input cell state, W f This is the weight matrix of the forget gate, b f It is the bias of the forgetting gate, W i It is the weight matrix of the input gate, b i It is the input gate bias, W c It is the weight matrix of cell states, b c It is a bias in cell state, W o It is the weight matrix of the output gate, b o It is the output gate bias, tanh is the activation function, and h is the output gate bias. t- 1 represents the fault characteristic output at time t-1, f t i is the proportionality coefficient that controls the amount of information forgotten in the forgetting gate. t c is the proportional coefficient that controls the input information in the input gate. t It is the state information after the cell state is updated, o t σ is the proportional coefficient that controls the output information of the output gate, and C is the sigmoid activation function. t h is the cell state at time t. t The output is the fault characteristics.

[0038] Furthermore, based on the fault characteristics of the distribution network, the fault category corresponding to the fault traveling wave signal is given, including: inputting the fault characteristics of the distribution network into a pre-built classifier, giving the probability of mapping the fault characteristics to each fault category; and taking the fault category with the highest probability as the fault category corresponding to the fault traveling wave signal.

[0039] Secondly, the present invention also provides a distribution network fault detection device based on traveling wave signals. The distribution network fault detection device using the above-mentioned method for detecting distribution network faults based on traveling wave signals includes: a signal acquisition and positioning module, which is used to acquire the fault traveling wave signal of the distribution network and determine the fault location of the distribution network.

[0040] The frequency decomposition module is used to decompose the frequency of the fault traveling wave signal based on the optimized decomposition algorithm to obtain the instantaneous frequency; the feature extraction module is used to extract features from the instantaneous frequency based on the pre-built optimized fault detection model to give the fault characteristics of the distribution network; the fault discrimination module is used to give the fault category corresponding to the fault traveling wave signal according to the fault characteristics of the distribution network.

[0041] Furthermore, the signal acquisition and positioning module includes: determining the corresponding power cable based on the acquired fault traveling wave signal; providing the time difference between the transmission of the fault traveling wave signal to both ends of the corresponding power cable based on the determined power cable; and providing the fault location on the power cable based on the time difference between the transmission of the fault traveling wave signal to both ends of the corresponding power cable and in combination with the length of the corresponding power cable.

[0042] Furthermore, the frequency decomposition module includes: providing noise figures based on the fault traveling wave signal and generating multiple first white noises accordingly; combining the multiple first white noises with the fault traveling wave signal respectively; and performing empirical mode decomposition on each combined fault traveling wave signal to obtain multiple first intrinsic mode function components.

[0043] The instantaneous frequency is obtained by averaging all components of the first intrinsic mode function and performing a Hilbert transform on the averaging result.

[0044] Furthermore, the frequency decomposition module includes: determining the range of noise figure values ​​based on the standard deviation of the fault traveling wave signal; initializing the particles using the noise figure as particles and combining the range of noise figure values ​​to give a particle swarm composed of multiple random particles; calculating the fitness value of the particle swarm, giving the local optimal fitness value of each particle, and taking the minimum of all local optimal fitness values ​​as the global optimal fitness value; updating the position of the particle swarm based on the evolution function and using extreme value perturbation; repeating the fitness value calculation and position update steps until convergence, giving the final global optimal fitness value; determining the corresponding noise figure value based on the final global optimal fitness value, and combining it with different Gaussian white noise to obtain multiple first white noises with noise figures.

[0045] Furthermore, the frequency decomposition module also includes: combining the noise figure of each particle in the particle swarm with the same Gaussian white noise to obtain multiple second white noises; combining the multiple second white noises with the fault traveling wave signal to obtain multiple signals to be decomposed; performing empirical mode decomposition on the multiple signals to be decomposed to obtain multiple sets of second intrinsic mode function components; and analyzing and processing the multiple sets of second intrinsic mode function components and the fault traveling wave signal based on the fitness value function to give the local optimal fitness value of each particle.

[0046] Furthermore, the feature extraction module includes a model building module, which includes: acquiring historical fault traveling wave signals of the distribution network and providing the corresponding historical instantaneous frequencies based on an optimized decomposition algorithm; normalizing the historical instantaneous frequencies to obtain normalized data and dividing it into training set, validation set, and test set; training the neural network model using the training set to provide the trained neural network model; validating the trained neural network model using the validation set until the training termination condition is met, and then verifying the performance of the neural network model using the test set to obtain the optimized fault detection model.

[0047] Furthermore, the model building module also includes: S31, initializing the neural network model and loss function, given initial model parameters and loss function coefficients; wherein, the model parameters include the weight matrix and bias vector; S32, the neural network model performs forward propagation on the training data of the training set and calculates the value of the loss function;

[0048] S33. Calculate the gradient of the loss function with respect to the model parameters of the neural network model based on the backpropagation algorithm, and use the gradient descent algorithm to update the model parameters and loss function coefficients according to the gradient to minimize the value of the loss function; S34. Repeat S32 and S33 until convergence, and give the trained neural network model.

[0049] Furthermore, the fault discrimination module includes: inputting the fault characteristics of the distribution network into a pre-built classifier, giving the probability of mapping the fault characteristics to each fault category; and taking the fault category with the highest probability as the fault category corresponding to the fault traveling wave signal.

[0050] The present invention provides a method and apparatus for detecting distribution network faults based on traveling wave signals, which has at least the following beneficial effects:

[0051] (1) After extracting the instantaneous frequency from the fault traveling wave signal, the fault feature is extracted from the instantaneous frequency using the pre-built optimized fault detection model. Finally, the fault classification of the fault traveling wave signal is completed, which can improve the accuracy of fault feature extraction and ultimately improve the accuracy of fault classification, i.e., improve the accuracy of fault detection in the distribution network.

[0052] (2) By optimizing the existing empirical mode decomposition, the numerical parameters of the empirical mode decomposition can be optimized more accurately according to the characteristics of the signal itself, thereby improving the accuracy of the empirical mode decomposition and obtaining more repeatable results with higher computational efficiency according to the characteristics of the signal.

[0053] (3) By optimizing the neural network model, overfitting can be effectively prevented, and the accuracy and robustness of the model can be balanced, improving the generalization ability and making it more suitable for fault feature extraction in the current application scenario.

[0054] (4) By combining optimized empirical mode decomposition with optimized neural network model, it can adapt to the complex and ever-changing environment of distribution network. By continuously learning and optimizing model parameters, it can improve the ability to identify and diagnose different types of faults. Attached Figure Description

[0055] Figure 1 is a flowchart of a power distribution network fault detection method based on traveling wave signals provided by the present invention;

[0056] Figure 2 is a flowchart of obtaining instantaneous frequency according to a certain embodiment of the present invention;

[0057] Figure 3 is a flowchart of a first white noise provided in a certain embodiment of the present invention;

[0058] Figure 4 is a flowchart illustrating the local optimal fitness value according to a certain embodiment of the present invention;

[0059] Figure 5 is a flowchart of a pre-built optimized fault detection model provided in a certain embodiment of the present invention;

[0060] Figure 6 is a flowchart of a trained neural network model provided in a certain embodiment of the present invention;

[0061] Figure 7 is a schematic diagram of a power distribution network fault detection device based on traveling wave signals provided by the present invention. Detailed Implementation

[0062] To better understand the above technical solutions, a detailed description of the solutions will be provided below in conjunction with the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0063] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “a,” “the,” and “the” as used in the embodiments of this invention and the appended claims are also intended to include the plural forms, and “multiple” generally includes at least two unless the context clearly indicates otherwise.

[0064] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that an article or device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such an article or device. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or device that includes said element.

[0065] To suppress mode aliasing, Ensemble Empirical Mode Decomposition (EEMD) is employed. EEMD is a noise-assisted data analysis method that effectively restores the essence of the signal. Furthermore, traditional RNNs often struggle to capture long-range information dependencies when processing long sequences due to vanishing or exploding gradients. To address this issue, the LSTM neural network algorithm is used. LSTM is a special neural network structure specifically designed to handle sequence data with long-term dependencies. LSTM effectively solves this problem by introducing a gating mechanism. However, when directly applied to distribution network fault diagnosis, LSTM neural networks struggle to directly capture the physical characteristics of fault signals.

[0066] Based on this, as shown in Figure 1, the present invention provides a distribution network fault detection method based on traveling wave signals, which specifically includes the following steps:

[0067] Acquire the fault traveling wave signal of the distribution network and determine the fault location in the distribution network;

[0068] Frequency decomposition of the fault traveling wave signal is performed based on an optimized decomposition algorithm to obtain the instantaneous frequency;

[0069] Based on a pre-built optimized fault detection model, features of instantaneous frequency are extracted to provide fault characteristics of the distribution network.

[0070] Based on the fault characteristics of the distribution network, the fault category corresponding to the fault traveling wave signal is given.

[0071] This invention enables fault detection in the distribution network after determining the fault location and fault type. Determining the fault location includes:

[0072] Based on the acquired fault traveling wave signal, determine the corresponding power cable;

[0073] Based on the given power cable, provide the time difference between the transmission of the fault traveling wave signal to the two ends of the corresponding power cable;

[0074] Based on the time difference between the transmission of the fault traveling wave signal to both ends of the corresponding power cable, and combined with the length of the corresponding power cable, the location of the fault on the power cable is determined, satisfying the following relationship:

[0075]

[0076] Where t1 and t2 are the times when the signal arrives at the two measurement ends of the line, v is the line mode wave velocity, l is the total length of the line, and x is the calculated fault distance.

[0077] As shown in Figure 2, this embodiment uses an optimized decomposition algorithm to perform frequency decomposition on the fault traveling wave signal to obtain the instantaneous frequency, including:

[0078] The noise figure is given based on the fault traveling wave signal, and multiple first white noises are generated accordingly.

[0079] Multiple first white noises are combined with the fault traveling wave signal respectively, satisfying the following relationship:

[0080]

[0081]

[0082] In the formula: x m x(t) is the fault traveling wave signal after incorporating the first white noise, and n is the fault traveling wave signal. m (t) represents the first white noise, A noise Noise figure It is Gaussian white noise;

[0083] Empirical mode decomposition (EMD) is performed on each combined fault traveling wave signal to obtain multiple first intrinsic mode function components, which satisfy the following relationship:

[0084] In the formula, r n (t) represents the residual component, and c i (t) represents the first intrinsic mode function component, and n represents the number of the first intrinsic mode function components;

[0085] The instantaneous frequency is obtained by averaging all components of the first intrinsic mode function and performing a Hilbert transform on the averaging result.

[0086] Among them, the mean value of all first intrinsic mode function components is applied, satisfying the following relationship:

[0087] In the formula, Let M be the mean of all first intrinsic mode function components, M be the order of empirical mode decomposition, and c be the mean of all first intrinsic mode function components. ij(t) is the sum of all first intrinsic mode function components corresponding to the j-th decomposition.

[0088] As shown in Figure 3, this embodiment generates multiple different first white noises with different noise figures based on the fault traveling wave signal. This can include: determining the range of noise figures based on the standard deviation of the fault traveling wave signal; initializing the particles using the noise figures as particles and combining the range of noise figures to generate a particle swarm composed of multiple random particles; calculating the fitness value of the particle swarm, giving the local optimal fitness value for each particle, and taking the minimum of all local optimal fitness values ​​as the global optimal fitness value; updating the position of the particle swarm using the evolution function of the time-series particle swarm optimization algorithm and extreme value perturbation; repeating the fitness value calculation and position update steps iteratively until convergence, giving the final global optimal fitness value; wherein, the position of the particle swarm represents the position of each particle, and the position of each particle is the value of the noise figure; the convergence condition is reaching the maximum number of iterations.

[0089] Based on the final global optimal fitness value, the corresponding noise figure value is determined, and combined with different Gaussian white noise, multiple first white noises with noise figures are obtained.

[0090] Among them, the standard deviation of the fault traveling wave signal It satisfies the following relationship:

[0091]

[0092] In the formula, p is the mean of the fault traveling wave signal, N2 is the number of time windows, and x i The fault traveling wave signal data is within the i-th time window.

[0093] After obtaining the standard deviation of the fault traveling wave signal, the noise figure can be categorized into the following ranges: .

[0094] As shown in Figure 4, the fitness value of the particle swarm is calculated, and the local optimal fitness value of each particle is given. This includes: combining the noise figure value of each particle in the particle swarm with the same Gaussian white noise to obtain multiple second white noises; and combining the multiple second white noises with the fault traveling wave signal to obtain multiple signals to be decomposed.

[0095] Empirical mode decomposition is performed on multiple signals to be decomposed to obtain multiple sets of second intrinsic mode function components. Based on the fitness value function, the multiple sets of second intrinsic mode function components and the fault traveling wave signal are analyzed and processed to give the local optimal fitness value of each particle.

[0096] The fitness value function satisfies the following relationship:

[0097]

[0098] In the formula, RRMSE is the fitness value, N1 is the total number of components of the second intrinsic mode function, x0(k) is the fault traveling wave signal, and c max (k) is the second intrinsic mode function component with the highest correlation coefficient with the fault traveling wave signal among the second intrinsic mode function components in the k-th group obtained by empirical mode decomposition;

[0099] In this embodiment, when updating the position of the particle swarm, the evolution function of the time-series particle swarm optimization algorithm needs to be initialized with parameters. Specifically, based on the characteristics of historical fault traveling wave signals, the following parameters are set: the number of individuals in the population m, the dimension of the target search space D, the total number of iterations N, the momentum-inertia coefficient ω, the learning factors c1 and c2, and the initial stagnation step thresholds T0 and T1 for perturbation of individuals and the global extremum. g The particle swarm positions are updated iteratively based on the initialization parameters of the evolution function, where the evolution function satisfies the following relationship:

[0100]

[0101] In the formula, Let be the position of the particle, representing the position of the i-th particle in the d-th dimension during the (t+1)-th iteration, where i = 1, 2, ..., m, d = 1, 2, ..., D, m is the number of particles in the swarm, and D is the set search dimension; ω is the momentum-inertia coefficient; c1 and c2 are non-negative learning factors; r1, r2, r3, and r4 are all random numbers following a U(0, 1) distribution; p is the extreme point of the individual particle and the swarm, t0 and t... g These represent the evolutionary stagnation steps for the individual particle extremum and the global extremum, respectively; T0 and T g These are the threshold numbers of stagnation steps required to disturb the individual particle extrema and the global extrema, respectively.

[0102] As shown in Figure 5, the pre-construction of the optimized fault detection model can include: acquiring historical fault traveling wave signals of the distribution network and providing the corresponding historical instantaneous frequencies based on an optimization decomposition algorithm; normalizing the historical instantaneous frequencies to obtain normalized data and dividing it into training, validation, and test sets; training the neural network model (LSTM model) using the training set to obtain the trained neural network model; validating the trained neural network model using the validation set until the training termination condition is met, and then verifying the performance of the neural network model using the test set to obtain the optimized fault detection model; specifically, the trained neural network model is verified using the validation set to determine whether the training termination condition is met. If it is met, training is stopped, and the performance of the neural network model is verified using the test set to obtain the optimized fault detection model; otherwise, the internal parameters of the neural network model are adjusted and retrained. The training termination condition is that the network generalization ability of the neural network model reaches its maximum and the network performance meets the requirements. The normalization of the historical instantaneous frequencies includes: vectorizing the historical instantaneous frequencies into fault feature vectors H=[H1,H2,...,H...] n Then, the fault feature vector is linearly normalized so that its amplitude is within [0,1], satisfying the following relationship:

[0103]

[0104] In the formula, H s For normalized data, H is the fault feature vector. min H is the smallest fault feature vector data. max This represents the largest fault feature vector data.

[0105] Meanwhile, in this embodiment, when extracting features from instantaneous frequency based on a pre-built optimized fault detection model, the instantaneous frequency is first processed using the same normalization method as described above, and then the normalized instantaneous frequency is input into the pre-built optimized fault detection model for detection to give the fault characteristics of the distribution network.

[0106] As shown in Figure 6, the neural network model is trained using a training set, resulting in the trained neural network model, including: S31, initializing the neural network model and loss function, providing initial model parameters and loss function coefficients; where the model parameters include the weight matrix and bias vector; S32, performing forward propagation of the neural network model on the training data of the training set, calculating the value of the loss function; S33, calculating the gradient of the loss function with respect to the model parameters of the neural network model based on the backpropagation algorithm, and using the gradient descent algorithm to update the model parameters and loss function coefficients according to the gradient to minimize the value of the loss function; S34, repeating S32 and S33 iteratively until convergence, resulting in the trained neural network model. The optimization of the fault detection model satisfies the following relationship:

[0107]

[0108]

[0109]

[0110]

[0111]

[0112] In the formula, To optimize the input normalized data of the fault detection model at time t, It is the current input cell state. This is the weight matrix of the forget gate, b f It's a bias in the forgetting gate. It is the weight matrix of the input gate. It is the bias of the input gate. It is the weight matrix of cell states. It is a bias in the cell state. It is the weight matrix of the output gate. tanh is the output gate bias, and tanh is the activation function. It is the fault characteristic output at time t-1. It is the proportional coefficient that controls the amount of information forgotten in the forgetting gate. It is the proportional coefficient that controls the input information in the input gate. It is the state information after the cell state is updated. It is the proportional coefficient for controlling the output information of the output gate. It is the Sigmoid activation function. It is the cell state at time t. The output is the fault characteristics.

[0113] An optimization algorithm is used to optimize the neural network model from two aspects: weight optimization and structure adjustment. This includes adjusting the loss function of the neural network model to satisfy the following relationship:

[0114]

[0115]

[0116]

[0117] In the formula, E is the loss function, α and β are the coefficients of the loss function, n is the total number of network weights in the neural network model, l is the average number of network weights in the neural network model, and w i W represents the network weights of the i-th neural network model. k Let J represent the network weight vector for the k-th iteration, J be the Jacobian matrix, μ be the step size control parameter with a value of 0.001, and γ be the number of effective parameters in the neural network model, γ = N-2β'tr(H). -1 β' is the β value of the previous iteration, N is the total number of model parameters of the neural network model; H is the Hessian matrix of the loss function, tr(H) represents the trace of matrix H, and W k For w i The matrix formed, ε(W) k ) indicates based on the current weight W k The error vector.

[0118] In this embodiment, after pre-constructing an optimized fault detection model, the instantaneous frequency is input into the optimized fault detection model to perform corresponding processing and analysis, and finally outputs the fault characteristics of the distribution network.

[0119] Based on the fault characteristics of the distribution network, the fault category corresponding to the fault traveling wave signal is given, including: inputting the fault characteristics of the distribution network into the classifier and giving the probability of mapping the fault characteristics to each fault category; taking the fault category with the highest probability as the fault category corresponding to the fault traveling wave signal.

[0120] The fault features ht extracted by the neural network model are mapped to the probability distributions y=[y1,y2,y3,...,y4] of each fault cause using a softmax classifier. n The category corresponding to the highest calculated probability is selected as the classification result, thereby achieving the classification of the cause of the fault.

[0121]

[0122] In the formula, W is the weight of the output layer, b is the bias, and n is the total number of fault cause categories.

[0123] As shown in Figure 7, the present invention also provides a distribution network fault detection device based on traveling wave signals. This device employs the aforementioned distribution network fault detection method based on traveling wave signals. The distribution network fault detection device includes: a signal acquisition and location module, used to acquire the fault traveling wave signal of the distribution network and determine the fault location; a frequency decomposition module, used to perform frequency decomposition on the fault traveling wave signal based on an optimized decomposition algorithm to obtain the instantaneous frequency; a feature extraction module, used to extract features from the instantaneous frequency based on a pre-built optimized fault detection model to provide the fault characteristics of the distribution network; and a fault discrimination module, used to provide the fault category corresponding to the fault traveling wave signal based on the fault characteristics of the distribution network.

[0124] Furthermore, the signal acquisition and positioning module includes: determining the corresponding power cable based on the acquired fault traveling wave signal; providing the time difference between the transmission of the fault traveling wave signal to both ends of the corresponding power cable based on the determined power cable; and providing the fault location on the power cable based on the time difference between the transmission of the fault traveling wave signal to both ends of the corresponding power cable and in combination with the length of the corresponding power cable.

[0125] Furthermore, the frequency decomposition module includes: providing noise figures based on the fault traveling wave signal and generating multiple first white noises accordingly; combining the multiple first white noises with the fault traveling wave signal respectively; performing empirical mode decomposition on each combined fault traveling wave signal to obtain multiple first intrinsic mode function components; performing mean processing on all first intrinsic mode function components, and performing Hilbert transform on the mean processing result to obtain the instantaneous frequency.

[0126] Furthermore, the frequency decomposition module includes: determining the range of noise figure values ​​based on the standard deviation of the fault traveling wave signal; initializing the particles using the noise figure as particles and combining the range of noise figure values ​​to give a particle swarm composed of multiple random particles; calculating the fitness value of the particle swarm, giving the local optimal fitness value of each particle, and taking the minimum of all local optimal fitness values ​​as the global optimal fitness value; updating the position of the particle swarm based on the evolution function and using extreme value perturbation; repeating the fitness value calculation and position update steps until convergence, giving the final global optimal fitness value; determining the corresponding noise figure value based on the final global optimal fitness value, and combining it with different Gaussian white noise to obtain multiple first white noises with noise figures.

[0127] Furthermore, the frequency decomposition module also includes: combining the noise figure of each particle in the particle swarm with the same Gaussian white noise to obtain multiple second white noises; combining the multiple second white noises with the fault traveling wave signal to obtain multiple signals to be decomposed; performing empirical mode decomposition on the multiple signals to be decomposed to obtain multiple sets of second intrinsic mode function components; and analyzing and processing the multiple sets of second intrinsic mode function components and the fault traveling wave signal based on the fitness value function to give the local optimal fitness value of each particle.

[0128] Furthermore, the feature extraction module includes a model building module, which includes: acquiring historical fault traveling wave signals of the distribution network and providing the corresponding historical instantaneous frequencies; normalizing the historical instantaneous frequencies to obtain normalized data and dividing it into training set, validation set, and test set; training the neural network model using the training set to provide the trained neural network model; validating the trained neural network model using the validation set until the training termination condition is met, and then verifying the performance of the neural network model using the test set to obtain the optimized fault detection model.

[0129] Furthermore, the model building module also includes: S31, initializing the neural network model and loss function, providing initial model parameters and loss function coefficients; wherein, the model parameters include the weight matrix and bias vector; S32, the neural network model performs forward propagation on the training data of the training set, and calculates the value of the loss function; S33, based on the backpropagation algorithm, calculates the gradient of the loss function with respect to the model parameters of the neural network model, and uses the gradient descent algorithm to update the model parameters and loss function coefficients according to the gradient, so as to minimize the value of the loss function; S34, repeating S32 and S33 iteratively until convergence, giving the trained neural network model.

[0130] Furthermore, the fault discrimination module includes: inputting the fault characteristics of the distribution network into a pre-built classifier, giving the probability of mapping the fault characteristics to each fault category; and taking the fault category with the highest probability as the fault category corresponding to the fault traveling wave signal.

[0131] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and modifications.

Claims

1. A method of fault detection in a power distribution network based on travelling wave signals, characterized by, Specifically comprising the following steps: obtaining a fault traveling wave signal of a power distribution network, and determining a fault position of the power distribution network; performing frequency decomposition on the fault traveling wave signal based on an optimization decomposition algorithm to obtain an instantaneous frequency; performing feature extraction on the instantaneous frequency based on a pre-constructed optimization fault detection model to give a fault feature of the power distribution network; and giving a fault category corresponding to the fault traveling wave signal according to the fault feature of the power distribution network.

2. The power distribution network fault detection method of claim 1, wherein, The fault position of the power distribution network is determined by: determining a corresponding power cable according to the obtained fault traveling wave signal; giving a time difference of the fault traveling wave signal transmitted to both ends of the corresponding power cable according to the determined power cable; and giving a fault position on the power cable based on the time difference of the fault traveling wave signal transmitted to both ends of the corresponding power cable and in combination with the length of the corresponding power cable.

3. The power distribution network fault detection method of claim 1, wherein, The frequency decomposition on the fault traveling wave signal based on the optimization decomposition algorithm to obtain the instantaneous frequency comprises: giving a noise coefficient based on the fault traveling wave signal and correspondingly generating a plurality of first white noises; and combining the plurality of first white noises with the fault traveling wave signal respectively. The empirical mode decomposition is performed on each combined fault traveling wave signal to obtain a plurality of first intrinsic mode function components; the mean value processing is performed on all the first intrinsic mode function components, and the Hilbert transform is performed on the mean value processing result to obtain the instantaneous frequency.

4. The power distribution network fault detection method of claim 3, wherein, The noise coefficient is given based on the fault traveling wave signal and a plurality of first white noises are correspondingly generated, which comprises: determining a value range of the noise coefficient based on a standard deviation of the fault traveling wave signal; initializing the particles with the noise coefficient and in combination with the value range of the noise coefficient to give a particle swarm composed of a plurality of random particles; calculating the fitness value of the particle swarm to give a local optimal fitness value of each particle, and taking the minimum value of all the local optimal fitness values as a global optimal fitness value; updating the position of the particle swarm based on an evolution function and using an extreme value disturbance; repeating the steps of the fitness value calculation and the position updating until convergence to give a final global optimal fitness value; and determining the corresponding noise coefficient value based on the final global optimal fitness value and in combination with different Gaussian white noises to obtain a plurality of first white noises with the noise coefficient.

5. The power distribution network fault detection method of claim 4, wherein, The fitness value of the particle swarm is calculated to give a local optimal fitness value of each particle, which comprises: combining the noise coefficient value of each particle in the particle swarm with the same Gaussian white noise to obtain a plurality of second white noises; combining the plurality of second white noises with the fault traveling wave signal to obtain a plurality of decomposed signals; performing empirical mode decomposition on the plurality of decomposed signals to obtain a plurality of second intrinsic mode function components; and analyzing and processing the plurality of second intrinsic mode function components and the fault traveling wave signal based on a fitness value function to give the local optimal fitness value of each particle.

6. The power distribution network fault detection method of any one of claims 1-5, wherein, The pre-construction of the optimized fault detection model comprises: acquiring historical fault traveling wave signals of the power distribution network, and giving corresponding historical instantaneous frequencies based on an optimized decomposition algorithm; normalizing the historical instantaneous frequencies to obtain normalized data and dividing the normalized data into a training set, a validation set and a test set; training a neural network model through the training set to give a trained neural network model; verifying the trained neural network model through the validation set until a training end condition is met, and using the test set to verify the performance of the neural network model to obtain the optimized fault detection model.

7. The power distribution network fault detection method of claim 6, wherein, The training of the neural network model through the training set to give the trained neural network model comprises: S31, initializing the neural network model and a loss function, and giving initial model parameters and loss function coefficients; wherein the model parameters comprise a weight matrix and a bias vector; S32, the neural network model performs forward propagation on training data of the training set to calculate the value of the loss function; S33, the gradient of the loss function with respect to the model parameters of the neural network model is calculated based on a back propagation algorithm, and the model parameters and the loss function coefficients are updated according to the gradient using a gradient descent algorithm to minimize the value of the loss function; S34, repeating and iterating S32 and S33 until convergence, and giving the trained neural network model.

8. The power distribution network fault detection method of claim 7, wherein, The loss function satisfies the following relationship: In the formula, E is a loss function, α and β are loss function coefficients, n is the total number of network weights of the neural network model, l is the average number of network weights of the neural network model, w i is the network weight of the i-th neural network model, W k represents the network weight vector of the k-th iteration, J is a Jacobian matrix, ε(W k ) is an error vector based on W k , μ is a step control parameter, γ is the number of effective parameters of the neural network model, γ = N - 2β'tr(H) -1 , β' is the β value of the previous iteration, N is the total number of model parameters of the neural network model; H is a Hessian matrix of the loss function, and tr(H) represents the trace of the matrix H.

9. The power distribution network fault detection method of claim 7 or 8, wherein, According to the fault characteristics of the power distribution network, the fault class corresponding to the fault traveling wave signal is given, which comprises: inputting the fault characteristics of the power distribution network into the pre-constructed classifier to give the probability of mapping the fault characteristics to each fault class; and taking the fault class with the maximum probability as the fault class corresponding to the fault traveling wave signal.

10. A power distribution network fault detection apparatus based on travelling wave signals, characterised in that, The power distribution network fault detection device adopts the power distribution network fault detection method based on the traveling wave signal according to any one of claims 1-9, and comprises: a signal acquisition and positioning module configured to acquire the fault traveling wave signal of the power distribution network and determine the fault position of the power distribution network; a frequency decomposition module configured to perform frequency decomposition on the fault traveling wave signal based on an optimized decomposition algorithm to obtain the instantaneous frequency; a feature extraction module configured to perform feature extraction on the instantaneous frequency based on the pre-constructed optimized fault detection model to give the fault characteristics of the power distribution network; and a fault discrimination module configured to give the fault class corresponding to the fault traveling wave signal according to the fault characteristics of the power distribution network.

Images