Line fault position prediction method

Abstract

The invention provides a line fault position prediction method, and belongs to the technical field of fault detection, and the method comprises the steps: carrying out the modeling of a transmission line, and carrying out the preprocessing of a feature signal; analyzing the frequency domain of the discrete signal in combination with discrete Fourier transform and fast Fourier transform; and constructing a dynamic shaping fusion model, and predicting a line fault position by using the dynamic shaping fusion model. According to the method, a dynamic shaping model is introduced, the feature extraction capability of CNN, the time sequence processing advantage of LSTM and the complex mode learning capability of ANN are combined, and an efficient algorithm framework suitable for transmission line multi-fault positioning is provided. On the basis of deep learning, the problems that an existing fault positioning method is inaccurate in fault positioning, low in actual use reliability, weak in fault signal detectability and weak in characterization capacity are solved, and accurate fault positioning of the energy transmission line is more effectively achieved.

Description

Technical Field

[0001] The present invention belongs to the field of fault detection technology, and particularly relates to a method for predicting the position of a line fault. Background Art

[0002] The quality and continuity of energy have become very important today. In particular, interruptions due to industrial and power plant failures cause difficulties for producers and consumers and result in losses of manpower and time. The most important cause of energy interruption is the short-circuit fault that occurs in the transmission line. From the perspective of energy sustainability, it is very important to intervene in these faults quickly and correctly. The types of faults that occur in the transmission line are divided into two categories: temporary faults and permanent faults. Although temporary faults can be cleared spontaneously, power supply can only be restored after a permanent fault if the fault location is detected quickly and accurately. On the contrary, if the fault point is not detected quickly, the entire production line must be inspected. Technicians have conducted many studies to determine the fault point. These studies usually use power frequency current and voltage phasors based on transient state waves, impedance-based traveling waves, and artificial intelligence methods. Impedance-based techniques are usually based on current and voltage values at the fundamental frequency, but for high anti-error values, inaccurate results are usually given. The prior art conducted a study on fault distance estimation using the reactance method and data at the terminal of the transmission line, and developed a method called digital fault recorder to achieve more accurate prediction. The prior art used the short-circuit fault conditions on a 380 kV, 360 km long transmission line for fault position prediction. Using the fault current and voltage information obtained from the simulation, the wavelet packet transform technique was used to decompose these signals, and these information were trained in the ANN to determine the fault position. The simulation results showed a high accuracy rate, and their fault location prediction was based on a recursive artificial neural network. This study observed the excellent performance of the recursive neural network model in fault location. The prior art proposed a method for predicting the position of a transmission line fault using the stationary wavelet method, support vector machine (SVM), and regression (SVR) for preprocessing and filtering, and concluded that the fault position in the transmission line can be accurately and quickly predicted regardless of the fault impedance. The prior art proposed a method for detecting and classifying transmission line faults using phasor measurement unit (PMU) data. The weighted extreme learning machine (WELM) was used to detect and classify the faults. For classification, the input feature data was obtained using an integrated feature extraction technique based on wavelet transform, and the WELM classifier was optimized using particle swarm optimization (PSO). These methods have problems such as inaccurate prediction of the line fault position, large prediction fault delay, large error, and large prediction calculation amount. Summary of the Invention

[0003] In view of the above-mentioned deficiencies in the prior art, a line fault location prediction method provided by the present invention solves the problems in the prior art of inaccurate line fault location prediction, large fault prediction delay, large error, and large prediction calculation amount. The prediction scheme based on dynamic fusion neural network proposed in the present invention can solve the above-mentioned problems to a large extent by learning a large amount of effective data.

[0004] In order to achieve the above object, the technical solution adopted by the present invention is: a line fault location prediction method, comprising the following steps:

[0005] S1. Model the transmission line and preprocess the characteristic signal;

[0006] S2. Based on the preprocessing results, the frequency domain of the discrete signal is analyzed by combining discrete Fourier transform and fast Fourier transform;

[0007] S3. Construct a dynamic shaping fusion model and use it to predict the line fault location based on the frequency domain analysis results.

[0008] The beneficial effects of the present invention are as follows: Based on the core theory of deep learning, the present invention proposes an innovative method combining multiple deep learning network models. Specifically, the method organically integrates convolutional neural networks (CNNs), long short-term memory networks (LSTMs) and artificial neural networks (ANNs), and by constructing a complex and efficient deep learning framework, systematically trains and learns the current and voltage signals after modal transformation matrix processing and fast Fourier transform (FFT) analysis. The goal of the present invention is to fully tap the deep features in the power system fault data, thereby achieving accurate prediction of fault signals, and ultimately providing more efficient and reliable technical support for the operation of smart grids.

[0009] Furthermore, the step S1 includes the following steps:

[0010] S101. Model the electrical characteristics and structure of the transmission line using electromagnetic transient programs;

[0011] S102. Based on the modeling results, the modal transformation matrix is applied to the current and voltage signals received from one end of the three-phase transmission line, and the modal transformation matrix is used to decompose and reorganize the current and voltage signals to obtain decoupled modal signals, thereby completing the preprocessing of the characteristic signals.

[0012] The beneficial effect of the above further scheme is that the current and voltage signals can be decomposed and reorganized through the modal transformation matrix, and the signals of different modes can be separated, thereby greatly improving the signal recognition and feature extraction capabilities.

[0013] Furthermore, the expression of the decoupled modal signal is as follows:

[0014]

[0015] V phase = M v V mod

[0016]

[0017] I phase = M i I mod

[0018] Wherein, V mod and I mod respectively represent the modal voltage and the modal current, that is, the decoupled modal signal, M v and M i respectively represent the modal transformation matrices of voltage and current, and respectively represent the inverse matrices of the modal transformation matrices of voltage and current, which are used to restore the original signal from the modal signal, V phase represents the collected phase voltage signal, and I phase represents the collected phase current signal.

[0019] The beneficial effects of the above further solution are as follows: By using the modal transformation matrix, the complex multivariable system is transformed into a simple single-variable system, which not only simplifies the calculation process, but also improves the accuracy of measurement and control, because this transformation can eliminate the mutual interference inside the system. The decoupled modal signal is more conducive to frequency-domain and time-domain analysis, which helps to more effectively diagnose faults and evaluate performance, and the control strategy based on these signals can better achieve the stability and dynamic performance optimization of the system. In addition, the modal transformation also reduces the amount of data to be processed, thereby reducing the demand for computing resources, and significantly improving the efficiency and effectiveness of the entire system.

[0020] Furthermore, the specific steps of step S2 are as follows:

[0021] Convert the preprocessed modal signal into a frequency-domain representation by using the Fourier transform;

[0022] Based on the frequency-domain representation, use the discrete Fourier transform to process the frequency-domain analysis of the discrete signal, and obtain the signal components in the discrete frequency domain. Among them, use the fast Fourier transform to decompose the N-point discrete Fourier transform into multiple small-scale discrete Fourier transforms through recursive decomposition, and combine the odd and even sampling points to repeatedly decompose the N / 2-point discrete Fourier transform.

[0023] The beneficial effects of the above further solution are as follows: Through the recursive decomposition method of the Fast Fourier Transform (FFT), the computational amount can be significantly reduced and the computational efficiency can be improved. At the same time, the signal is transformed into the frequency domain for processing, which not only simplifies the signal processing process but also makes the analysis of the frequency components of the signal more intuitive. In the frequency domain, it becomes easier to identify and separate signals with different frequency components, thus enhancing the signal analysis ability. In addition, decomposing the large-scale Discrete Fourier Transform into multiple small-scale transforms can better utilize parallel computing resources and optimize the overall performance of the algorithm. Signal processing in the frequency domain can also avoid the noise interference that may exist in time-domain processing and further improve the accuracy of data processing. In summary, the combined effect of these technical effects improves the efficiency, accuracy, and flexibility of signal processing, enabling the system to perform signal analysis and processing more efficiently.

[0024] Furthermore, the forward propagation of the dynamic shaping fusion model is calculated as follows:

[0025]

[0026] Where, represents the fault location detection confidence output by the dynamic shaping fusion model, F ANN represents the forward propagation function of the ANN module, F LSTM represents the forward propagation function of the LSTM module, F CNN represents the forward propagation function of the CNN module, X represents the input data, θ CNN represents the set of trainable parameters of the CNN module, θ LSTM represents the set of trainable parameters of the LSTM module, θ ANN represents the set of trainable parameters of the ANN module.

[0027] The beneficial effects of the above further solution are as follows: Through the multi-layer network structure, the model can learn deeper feature representations, thus improving the accuracy of fault location prediction. By fusing different types of neural networks, the model can process spatial and temporal features, thus more comprehensively capturing the complex characteristics of the signal. The model has stronger robustness to noise and interference and can work stably in complex environments. These technical effects make the model more reliable in practical applications.

[0028] Furthermore, the dynamic shaping fusion model includes:

[0029] A convolutional neural network layer for extracting local spatial features based on the frequency domain analysis results;

[0030] A long short-term memory network layer for obtaining the dynamic features of the line fault signal in the time dimension based on the extracted local spatial features;

[0031] An artificial neural network layer for predicting the location of line faults based on the obtained dynamic features.

[0032] The beneficial effects of the above further solution are as follows: Through the combination of the convolutional neural network layer and the long short-term memory network, the model can more comprehensively extract the spatial and temporal features of the input data, thereby improving the feature extraction ability. Introducing the Dropout layer effectively prevents the model from overfitting and enhances the robustness of the model when dealing with complex data. Further, the artificial neural network integrates and predicts the temporal features output by the LSTM, achieving accurate positioning of the fault location and improving the prediction accuracy. In addition, the reasonable network structure design and parameter configuration reduce the consumption of computing resources and optimize the computing efficiency. In summary, these improvements work together to significantly improve the model performance, enabling it to perform fault detection and location more efficiently and accurately.

[0033] Furthermore, the expression of the optimal hyperparameters of the dynamic shaping fusion model is as follows:

[0034] α * = g (T)

[0035]

[0036] where α * represents the optimal hyperparameter, g (T) represents the global optimal position, represents the hyperparameter vector of the i-th particle in the t-th generation, represents the fitness value of the i-th particle, represents the loss function, that is, the loss value calculated by the hyperparameters in the dynamic shaping fusion model.

[0037] The beneficial effects of the above further solution are as follows: By introducing the particle swarm optimization (PSO) algorithm to further optimize the hyperparameters of the dynamic shaping fusion model, the dynamic shaping fusion model has higher efficiency and performance during training, accelerating convergence while improving the stability of training. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 is a flowchart of the method of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0039] The following describes the specific embodiments of the present invention to facilitate those skilled in the art to understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions made using the concept of the present invention are within the scope of protection.

[0040] Example

[0041] Based on the core theory of deep learning, the present invention proposes an innovative method that combines multiple deep learning network models. Specifically, this method organically integrates convolutional neural networks (CNNs), long short-term memory networks (LSTMs), and artificial neural networks (ANNs), and systematically trains and learns the current and voltage signals after modal transformation matrix processing and fast Fourier transform (FFT) analysis by constructing a complex and efficient deep learning framework. The goal of this method is to fully tap the deep features in the power system fault data, so as to achieve accurate prediction of fault signals, and ultimately provide more efficient and reliable technical support for the operation of smart grids. The entire scheme design is based on the efficiency and accuracy of modern deep learning technology, while taking into account the complexity and real-time requirements of power system fault detection. In the signal processing stage, the collected original current and voltage signals are preprocessed by the modal transformation matrix, and the frequency domain features are extracted by combining FFT technology, which fundamentally enhances the prediction ability of fault signals and provides a high-quality data foundation for the training of subsequent deep learning models. Figure 1 As shown, the present invention provides a line fault location prediction method, and its implementation method is as follows:

[0042] S1. Model the transmission line and preprocess the characteristic signal. The implementation method is as follows:

[0043] S101. Model the electrical characteristics and structure of the transmission line using electromagnetic transient programs;

[0044] S102. Based on the modeling results, the modal transformation matrix is applied to the current and voltage signals received from one end of the three-phase transmission line, and the modal transformation matrix is used to decompose and reorganize the current and voltage signals to obtain decoupled modal signals, thereby completing the preprocessing of the characteristic signals.

[0045] In this embodiment, the transmission line is modeled and the characteristic signal is preprocessed: the electrical characteristics and structure of the transmission line are modeled in detail using the electromagnetic transient program (EMTP / ATP) to truly reproduce the dynamic behavior of the transmission line under fault conditions. After the modeling is completed, the modal transformation matrix is applied to the current and voltage signals received from one end of the transmission line. The modal transformation matrix can decompose and reorganize the current and voltage signals, and separate the signals of different modes, thereby greatly improving the signal recognition and feature extraction capabilities. This process not only effectively reduces the impact of external noise and interference on the signal, but also significantly enhances the harmonic characteristics caused by the fault, making these features more meaningful for detection and identification, and providing more accurate signal input for subsequent steps.

[0046] The current signal and voltage signal after modal transformation can better reflect the dynamic changes during fault occurrence. Combining the characteristics of the modal transformation matrix, this processing method can ensure that the fault feature signal will not be masked or weakened during transmission, and at the same time provides a clear and reliable data source for the training of subsequent deep learning models. The process of modal transformation is as follows:

[0047] In a three-phase transmission line system, due to the existence of electromagnetic coupling, the traveling wave signals between phases affect each other. This coupling effect increases the complexity of signal transmission and fault analysis. Therefore, in order to effectively extract the traveling wave characteristics, it is usually necessary to decompose the phase voltage and phase current signals into non-interfering modal signals. The key method to achieve this goal is to use a modal transformation matrix, which can linearly transform the coupled three-phase line signals into decoupled modal signals. Through modal decomposition, the characteristics of traveling waves in independent modes are clearer, facilitating subsequent feature extraction and fault detection analysis. At discrete frequencies, the steady-state behavior of a multi-conductor transmission line can be described by the following second-order differential equation:

[0048]

[0049] where z represents the series impedance matrix per unit length, which describes the resistance and inductance characteristics in the transmission line, y represents the shunt admittance matrix per unit length, which describes the conductance and capacitance characteristics in the transmission line, V phase represents the vector form of the phase voltage signal, and I phase represents the vector form of the phase current signal. The phase voltage signal V phase and the phase current signal I phase are collected from one end of the three-phase transmission line, and then the above-mentioned inter-phase coupled signals are transformed into single modal signals. The transformation matrix M is used to convert the phase voltage signal into a modal voltage signal and the phase current signal into a modal current signal. The modal transformation relationship is introduced as follows:

[0050] V phase = M v V mod ,

[0051] I phase = M i I mod ,

[0052] where V mod and I mod represent the modal voltage and modal current respectively, that is, the decoupled modal signals, and M v and M i represent the modal transformation matrices of voltage and current respectively, and They respectively represent the inverse matrices of the modal transformation matrices for voltage and current, which are used to restore the original signal from the modal signal, V phase represents the collected phase voltage signal, I phase represents the collected phase current signal.

[0053] For a three-phase transmission line with frequency division characteristics, the modal transformation matrix M is usually derived from the diagonalization conditions of the electrical characteristic matrices z and y of the transmission line. Under the condition of satisfying zy = yz, the following modal transformation matrix is adopted in the present invention:

[0054]

[0055] Through modal decomposition, the coupling characteristics in the three-phase signals are decoupled into three non-interfering modal components. The three column vectors of this matrix respectively correspond to the three independent modal components of the transmission line. Through modal decomposition, the coupling characteristics in the three-phase signals are decoupled into three non-interfering modal components. By applying the transformation matrix M to the original three-phase signals, the coupled voltage signal V phase can be converted into an independent modal signal V mod . The same method is applicable to the modal decomposition of current signals, that is, I phase is converted into I mod , and if necessary, the modal signal can be restored to the original signal through the inverse matrix.

[0056] The voltage and current signals after modal decomposition can be independently analyzed subsequently, avoiding the interference of inter-phase coupling on the signal characteristics. Through this method, not only the analysis of complex multi-phase systems is simplified, but also the accuracy and efficiency of signal feature extraction and fault detection are improved. The obtained preprocessed signal is subjected to Fourier transform in step S2.

[0057] S2. Based on the preprocessing result, the frequency domain of the discrete signal is analyzed by combining the discrete Fourier transform and the fast Fourier transform. The implementation method is as follows:

[0058] The modal signal after preprocessing is converted into a frequency domain representation by using Fourier transform;

[0059] Based on the frequency domain representation, the frequency domain analysis of the discrete signal is processed by using the discrete Fourier transform to obtain the signal components in the discrete frequency domain. Among them, the fast Fourier transform is used to decompose the N-point discrete Fourier transform into multiple small-scale discrete Fourier transforms through recursive decomposition, and the N / 2-point discrete Fourier transform is repeatedly decomposed by combining odd and even sampling points.

[0060] In this embodiment, the three-phase signals obtained by modal decomposition in step S1 are further transformed into the frequency domain by Fourier transform (FT) to extract the harmonic characteristics in the signals. The present invention combines the discrete Fourier transform (DFT) and the fast Fourier transform (FFT) for feature extraction, reducing the occupation of computing resources. Fourier transform (FT) is a technique for analyzing the harmonics of a steady-state signal that remains unchanged over time by decomposing the steady-state signal into various components. The core idea of the Fourier transform is to use the complex exponential function as the basis function to map the original signal into the frequency domain F(ω). The defined expression used is as follows:

[0061]

[0062] where t represents the time variable, i represents the imaginary unit, and ω represents the angular frequency.

[0063] Through the above formula, the preprocessed time-domain signal f(t) obtained in step S1 can be converted into the frequency-domain representation F(ω). To restore the signal from the frequency domain, the inverse Fourier transform is required, and its formula is:

[0064]

[0065] In practical applications, signals are usually discrete and of finite length, and directly using the Fourier transform does not work in practice. Therefore, the present invention introduces the discrete Fourier transform (Discrete Fourier Transform, DFT) to handle the frequency-domain analysis of discrete signals. The calculation formula of the discrete Fourier transform DFT is as follows:

[0066]

[0067] where N represents the number of sampling points of the signal, x(n) represents the discrete-time domain signal, X(k) represents the signal component in the discrete frequency domain, and k represents the frequency index.

[0068] The discrete Fourier transform DFT decomposes the discrete-time signal into frequency-domain components. However, its computational complexity is O(N 2 ), and as the number of signal points increases, the amount of calculation rises sharply, limiting its use efficiency in real-time and large-scale applications. To improve the computational efficiency, the present invention uses the fast Fourier transform (Fast Fourier Transform, FFT). The fast Fourier transform FFT is an efficient implementation of the discrete Fourier transform DFT. By using the recursive decomposition idea, the N-point discrete Fourier transform DFT is decomposed into multiple smaller-scale discrete Fourier transforms DFT, thereby reducing the computational complexity to O(NlogN). The basic principle of the fast Fourier transform FFT is to use the recursive decomposition formula of the discrete Fourier transform DFT:

[0069]

[0070] Among them, W N represents a rotation factor. By repeatedly decomposing the N / 2-point discrete Fourier transform (DFT) and combining the calculation formulas for even and odd sampling points:

[0071]

[0072] Among them, X1[k] represents the subsequence with even indices, and X2[k] represents the subsequence with odd indices. represents the rotation factor, and X[k] represents the amplitude and phase information of the original signal x[n] at the k-th frequency component.

[0073] Benefiting from the symmetry and periodicity characteristics of the discrete Fourier transform (DFT), this decomposition method can further reduce the computational amount and optimize the entire calculation process. The symmetry and periodicity of the discrete Fourier transform (DFT) are reflected by the following formulas:

[0074]

[0075] Among them, and represent the rotation factors when the frequency indices are k + N / 2 and k + N, respectively.

[0076] Utilizing this characteristic, the calculation process can be optimized through the sample sparsity method. The even and odd parts of the signal have specific symmetry in the frequency domain performance, which can reduce the need for repeated calculations.

[0077] In this embodiment, in this step, the three-phase signals after modal decomposition in step S1 are further transformed into the frequency domain to extract the harmonic characteristics in the signals. By combining the discrete Fourier transform (DFT) and the fast Fourier transform (FFT), not only the feature extraction efficiency is improved, but also the occupation of computing resources is reduced.

[0078] S3. Construct a dynamic shaping fusion model and use the dynamic shaping fusion model to predict the line fault location.

[0079] In this embodiment, the dynamic shaping fusion model includes:

[0080] A convolutional neural network layer for extracting local spatial features according to the frequency domain analysis results;

[0081] A long short-term memory network layer for obtaining the dynamic features of the line fault signal in the time dimension based on the extracted local spatial features;

[0082] An artificial neural network layer for predicting the line fault location based on the obtained dynamic features.

[0083] In this embodiment, the convolutional neural network layer serves as the spatial feature extraction module of the dynamic shaping fusion model, and is used to extract the local spatial features of the input data according to the transmission line data. This convolutional neural network layer includes two convolutional layers, one pooling layer, and one Flatten layer. The convolutional layers both use convolutional kernels of size 3×3, with a stride of 1, and the number of output channels is 16 and 32 respectively. After each convolutional layer, a ReLU activation function is connected for non-linearity. The pooling layer uses max pooling (MaxPooling), with a pooling window size of 2×2 and a stride of 2 for feature dimensionality reduction. The Flatten layer is used to flatten the multi-dimensional feature map into a one-dimensional vector as the input for the subsequent long short-term memory network (LSTM).

[0084] In this embodiment, the long short-term memory network layer serves as the temporal feature extraction module of the dynamic shaping fusion model, and is used to further process the spatial features extracted by the convolutional neural network (CNN), and capture the dynamic features of the fault signal in the time dimension. This module includes an LSTM layer and a Dropout layer. The LSTM layer includes two LSTM sub-layers. The first layer has 64 hidden units and returns the complete time series; the second layer has 32 hidden units and returns the output of the final time step. After the LSTM layer, a Tanh (hyperbolic tangent function) activation function is connected. The Dropout layer is used to prevent overfitting during the temporal feature extraction process, and the Dropout ratio is set to 0.2.

[0085] In this embodiment, the artificial neural network layer serves as the fault location prediction module of the dynamic shaping fusion model, and is used to integrate the dynamic temporal features output by the long short-term memory network layer for the final fault location prediction. The hidden layer contains two layers. The first hidden layer has 64 neurons, and the second hidden layer has 32 neurons. The activation function is the ReLU function.

[0086] In this embodiment, during the forward propagation process, the data stream sequentially passes through the three module layers of the dynamic shaping fusion model, namely the convolutional neural network layer (CNN), the long short-term memory network layer (LSTM), and the artificial neural network layer (ANN), and finally outputs a confidence sequence. The forward propagation calculation process is as follows:

[0087]

[0088] Among them, represents the fault location detection confidence output by the dynamic shaping fusion model, F ANN represents the forward propagation function of the ANN module, F LSTM represents the forward propagation function of the LSTM module, F CNN represents the forward propagation function of the CNN module, X represents the input data, N represents the number of samples, T represents the time step, C represents the number of channels, θCNN Denotes the set of trainable parameters of the CNN module, θ LSTM Denotes the set of trainable parameters of the LSTM module, θ ANN Denotes the set of trainable parameters of the ANN module Denotes the fault location detection confidence of each sample at each time step Denotes the fault location detection confidence of the i-th sample at the j-th time step

[0089] In this embodiment, the prediction process of predicting the line fault location using the dynamic shaping fusion model is as follows: The data passes through the forward propagation of the dynamic shaping fusion model and finally outputs the fault detection confidence Indicates that the probability of the i-th sample having a fault at the j-th time step is relatively high Indicates that the fault probability of the i-th sample at the j-th time step is relatively low. The output result can be further combined with a threshold and a post-processing strategy to convert the fault location detection confidence into a specific fault location or interval

[0090] In this embodiment, the training settings of the dynamic shaping fusion model are as follows

[0091] The loss function uses the binary cross-entropy function

[0092]

[0093] Among them Denotes the loss function (cross-entropy loss), N denotes the number of samples, T denotes the number of time steps Denotes the predicted fault location detection confidence of the i-th sample at the j-th time step, Y i,j ∈{0,1} denotes the true label of the i-th sample at the j-th time step, and θ denotes the set of all trainable parameters of the model

[0094] Optimizer: The present invention uses SGD (stochastic gradient descent) as the training optimizer for the dynamic shaping fusion model. In view of the fact that the fusion network is prone to gradient problems, a suitable learning rate scheduler and an adaptive optimizer need to be selected. At the same time, although the binary cross-entropy loss has good adaptability in probability prediction problems, it needs to be combined with an efficient gradient update strategy. The present invention improves the optimization effect by adding a momentum term (Momentum) and a weight decay (Weight Decay) to enhance the convergence speed and stability of network training. For a parameter θ in the dynamic shaping fusion model, the formula for gradient update is as follows

[0095]

[0096] Among them denotes the gradient of the loss function with respect to the current parameter θ at the t-th iteration. λ represents the weight decay term, which is taken as 0.0001 in the present invention, and v t and v t-1 denotes the momentum term, which is the exponentially weighted average of the gradient history. β represents the momentum coefficient, which is taken as 0.9 in the present invention.

[0097] The formula for parameter update is as follows:

[0098] θ t+1 = θ t - ηv t

[0099] where η represents the learning rate, which is set to 0.001 in the present invention. θ t denotes the current parameter value of the dynamic shaping fusion model, and θ t+1 denotes the updated parameter value of the dynamic shaping fusion model. t represents the number of iterations.

[0100] Combining dynamic shaping with the loss function: The particle swarm optimization (PSO) algorithm is used to optimize the hyperparameters of CNN, LSTM, and ANN (such as convolution kernel stride, Dropout ratio, etc.). The PSO algorithm iteration is performed after each round of parameter update. In each PSO algorithm iteration, the inference model is run and the loss function value is calculated. The fitness function Fitness of the PSO algorithm is the PSO loss function of the dynamic shaping fusion model

[0101]

[0102] The PSO algorithm will continuously adjust the hyperparameter combination to minimize the loss function The goal is:

[0103] α * = argmin α L(w; α)

[0104] where w represents the weight parameter, α represents the model hyperparameter, and α * represents the optimized hyperparameter.

[0105] In this embodiment, the process of selecting the optimal hyperparameters is as follows:

[0106] In the t-th generation, the fitness function of the i-th particle (i.e., hyperparameter) is defined as:

[0107]

[0108] where, denotes the loss function, representing the loss value calculated by the hyperparameter in the model (the smaller the better), Represents the fitness value of the i-th particle.

[0109] In each generation, the velocity is updated through the particle update formula, and the formula is as follows:

[0110]

[0111] Among them, Represents the hyperparameter vector of the i-th particle in the t-th generation, Represents the velocity vector of the i-th particle in the t-th generation, which is initialized to a zero vector, Represents the historical best position of the i-th particle in the t-th generation, g (t) Represents the global best position of all particles in the t-th generation, ω represents the inertia weight, which is set to 0.5 in the present invention, c1 and c2 represent the learning factors, which are set to 2.0 in the present invention, r1 and r2 represent uniformly distributed random numbers in the range of [0,1], and the position is updated by the updated velocity, and the formula is as follows:

[0112]

[0113] The PSO particle algorithm will continuously update the global optimal position g (t) , so that it satisfies:

[0114]

[0115] Final optimal hyperparameters:

[0116] α * = g (T)

[0117] Among them, T represents the maximum number of iterations, and α * Represents the optimized hyperparameter.

[0118] Apply the optimized hyperparameters to the dynamic shaping fusion model, that is, obtain the dynamic shaping fusion model optimized by PSO for the next round of training.

[0119] In this embodiment, to solve the problem of multi-fault location of the transmission line (TL), the present invention proposes a dynamic shaping fusion model (CLA) that combines a convolutional neural network (CNN), a long short-term memory network (LSTM), and an artificial neural network (ANN). This model significantly improves the accuracy and efficiency of fault detection and location by improving the working mode of the network, the characteristics of the input data, the training mechanism, and the output performance.

[0120] An artificial neural network (ANN) is a computational model inspired by the biological nervous system, mimicking the functions of neurons in the human brain and used to perform tasks such as data analysis, classification, and prediction. Its basic unit is the neuron, which receives input signals, processes them using mathematical models and activation functions, and generates corresponding outputs. The mathematical expression of the neuron is as follows:

[0121]

[0122] where x i represents the input signal, w i represents the weight corresponding to the input signal, b represents the bias term, f represents the activation function, i represents the index of the input signal, and n represents the number of input signals.

[0123] In this embodiment, a convolutional neural network (CNN) usually consists of three types of layers: the input layer is responsible for receiving the original data; the hidden layer is responsible for feature extraction and pattern learning; the output layer is responsible for generating prediction results. Through forward propagation, the input data is passed and processed layer by layer, and finally a predicted value is generated. Subsequently, the error is calculated by comparing the predicted value with the actual value, and the network weights are adjusted during the backpropagation process to improve the network performance.

[0124] Common activation functions include Sigmoid, Tanh, and ReLU, which can introduce non-linear characteristics, enabling the artificial neural network (ANN) to learn complex mapping relationships. The convolutional neural network (CNN) is mainly used to process data with spatial features. It extracts local features of the input data through convolutional operations, and then reduces the dimension and enhances the robustness of the features through the pooling layer, and is suitable for processing multi-dimensional signals, such as the time-frequency data of current and voltage in the power system. The long short-term memory network (LSTM) is a special recurrent neural network (RNN), which is good at processing time series data. By introducing memory units and gating mechanisms, it can capture the long-term dependencies of the input data and is very suitable for analyzing the dynamic characteristics of power signals.

[0125] The dynamic shaping fusion model (CLA) is a fusion model of the convolutional neural network (CNN), the long short-term memory network (LSTM), and the artificial neural network (ANN). To further improve its performance, the present invention uses a dynamic shaping mechanism, aiming to dynamically adjust the model structure according to the requirements of different tasks to achieve higher accuracy and efficiency.

[0126] In this embodiment, the shaping of the dynamic shaping fusion model (CLA) relies on the particle swarm optimization (PSO) algorithm. The particle swarm optimization PSO algorithm is an optimization method based on swarm intelligence, which finds the optimal solution of the objective function by simulating the movement of particles in the solution space. The specific steps are as follows:

[0127] Parameter initialization: Randomly select several hidden layer configurations (including the number of neurons);

[0128] Model training: Each hidden layer configuration corresponds to a dynamic shaping fusion model (CLA), and all models are trained in parallel;

[0129] Performance monitoring: The particle swarm optimization (PSO) algorithm dynamically adjusts the hidden layer configuration by monitoring the mean square error (MSE) of each dynamic shaping fusion model (CLA);

[0130] Parameter update: Based on the feedback of the optimal mean square error (MSE), the PSO algorithm updates the velocity and position of the particles, and finally selects the network configuration with the minimum mean square error (MSE).

[0131] In this embodiment, by dynamically adjusting the hidden layer size and the number of neurons, the dynamic shaping fusion model (CLA) can converge to the optimal solution faster, while improving the accuracy and robustness of fault location.

[0132] In summary, by introducing the dynamic shaping fusion model (CLA), combining the powerful feature extraction ability of CNN (Convolutional Neural Network), the advantages of LSTM (Long Short-Term Memory Network) in time series processing, and the excellent performance of ANN (Artificial Neural Network) in complex pattern learning, the present invention proposes an efficient algorithm framework for fault location prediction of power energy transmission lines. Compared with the prior art, the present invention has significant advantages, which are specifically reflected in the following aspects:

[0133] Optimization of the data preprocessing module: Through modal transformation and Fourier transformation, the present invention can effectively extract key features in current and voltage signals, significantly improving the sensitivity and robustness of the model to fault signals. Compared with traditional preprocessing methods, this solution can capture multi-dimensional signal features more comprehensively and avoid misjudgment problems caused by noise interference.

[0134] Dynamic optimization of network configuration: The present invention introduces the particle swarm optimization algorithm (PSO) to dynamically adjust the structural parameters of the CLA model (dynamic shaping fusion model) and optimize the network configuration. This technology greatly reduces the complexity of manual parameter tuning, can quickly find the optimal model parameters according to the actual application scenario, improves the training efficiency and performance stability of the model. Compared with the traditional method with a fixed network structure, dynamic adjustment makes the network more adaptable and has better generalization ability.

[0135] Efficient model training mechanism: Large-scale data training is carried out on the optimized dynamic shaping fusion model (CLA), enabling the dynamic shaping fusion model to have good fault classification ability and prediction accuracy. The present invention is particularly outstanding in the learning ability of complex non-linear fault patterns and can accurately locate complex fault situations including multi-fault resistors. Compared with existing models that rely on simple fault assumptions, the present invention is more reliable in dealing with diverse fault scenarios. Among them, the cross-entropy is used as the loss function and SGD is used as the optimizer in the training process, and the particle swarm optimization PSO algorithm is combined with the loss function to find the optimal hyperparameters in each round for iterative training.

[0136] Real-time fault prediction: By analyzing the input signal in real time, the present invention can quickly predict the abnormal situation of the energy transmission line and achieve early prevention of the fault location. Compared with traditional detection methods that rely on fixed rules, the present invention significantly improves the speed and accuracy of fault prediction through the adaptive characteristics of the deep learning model, providing more efficient technical support for the energy transmission system. Among them, the data passes through the forward propagation of the dynamic shaping fusion model and finally outputs the fault detection confidence. The time steps with higher confidence are the parts with higher fault prediction possibilities in the sample.

[0137] Improvement of data security: The present invention uses deep learning technology to automatically extract and process signal features, avoiding the potential risk of information leakage in the manual annotation process. At the same time, data augmentation technology is introduced in the model training process to effectively prevent the impact of data forgery on the system performance and further improve the security of data processing.

[0138] Saving network resources: The dynamically optimized dynamic shaping fusion model (CLA) significantly reduces the computational complexity of the model and the demand for hardware resources by streamlining the network structure and intelligently adjusting parameters. Compared with traditional models that rely on high-computing power devices, the present invention is more lightweight and suitable for practical scenarios with limited resources.

Claims

1. A method for predicting the location of a line fault, characterized in that, It includes the following steps: S1. Model the transmission line and preprocess the characteristic signals; S2. Based on the preprocessing results, analyze the frequency domain of the discrete signals by combining the discrete Fourier transform and the fast Fourier transform; S3. Construct a dynamic shaping fusion model, and based on the frequency domain analysis results, use the dynamic shaping fusion model to predict the line fault location.

2. The line fault location prediction method according to claim 1, characterized in that, The step S1 includes the following steps: S101. Use the electromagnetic transient program to model the electrical characteristics and structure of the transmission line; S102. Based on the modeling results, apply the modal transformation matrix to the current and voltage signals received from one end of the three-phase transmission line, and use the modal transformation matrix to decompose and recombine the current and voltage signals to obtain the decoupled modal signals, thus completing the preprocessing of the characteristic signals.

3. The line fault location prediction method according to claim 2, wherein The expression of the decoupled modal signals is as follows: V phase = M v V mod I phase = M i I mod Among them, V mod and I mod respectively represent the modal voltage and the modal current, that is, the decoupled modal signals. M v and M i respectively represent the modal transformation matrices of voltage and current. and respectively represent the inverse matrices of the modal transformation matrices of voltage and current, which are used to restore the original signals from the modal signals. V phase represents the collected phase voltage signal, and I phase represents the collected phase current signal.

4. The line fault location prediction method according to claim 1, characterized in that, The step S2 is specifically: Convert the preprocessed modal signals into the frequency domain representation by using the Fourier transform; Based on the frequency domain representation, use the discrete Fourier transform to process the frequency domain analysis of the discrete signals to obtain the signal components in the discrete frequency domain. Among them, use the fast Fourier transform to decompose the N-point discrete Fourier transform into multiple small-scale discrete Fourier transforms through recursive decomposition, and combine the odd and even sampling points to repeatedly decompose the N / 2-point discrete Fourier transform.

5. The method for predicting the location of a line fault according to claim 1, characterized in that, The forward propagation calculation of the dynamic shaping fusion model is as follows: Among them, represents the fault detection reliability output by the dynamic integer fusion model, F ANN represents the forward propagation function of the ANN module, F LSTM represents the forward propagation function of the LSTM module, F CNN represents the forward propagation function of the CNN module, X represents the input data, θ CNN represents the set of trainable parameters of the CNN module, θ LSTM represents the set of trainable parameters of the LSTM module, θ ANN represents the set of trainable parameters of the ANN module.

6. The method for predicting the location of a line fault according to claim 1, characterized in that, The dynamic shaping fusion model includes: A convolutional neural network layer, which is used to extract local spatial features based on the frequency domain analysis results; A long short-term memory network layer, which is used to obtain the dynamic features of the line fault signals in the time dimension based on the extracted local spatial features; An artificial neural network layer, which is used to predict the line fault location based on the obtained dynamic features.

7. The method for predicting the location of a line fault according to claim 6, wherein The expression of the optimal hyperparameters of the dynamic shaping fusion model is as follows: α * ＝g (T) Among them, α * represents the optimal hyperparameter, g (T) represents the global optimal position, represents the hyperparameter vector of the i-th particle in the t-th generation, represents the fitness value of the i-th particle, represents the loss function, that is, the loss value calculated by the hyperparameter in the dynamic shaping fusion model.

Images