Power distribution network grounding fault positioning method based on CEEMDAN-VMD and CNN-BiLSTM-Attention

Abstract

The invention discloses a power distribution network grounding fault positioning method based on CEEMDAN-VMD and CNN-BiLSTM-Attention, and the method comprises the steps: 1, carrying out the signal decomposition of a high-frequency disturbance component in a traveling wave signal through employing CEEMDAN, carrying out the processing of a high-frequency component through employing VMD, carrying out the classification of an IMF component through combining with K-means clustering, and screening out a high-frequency fault component; and step 2, designing a CNN-BiLSTM-Attention hybrid network architecture, extracting spatial features by using a CNN, capturing time sequence information by using BiLSTM, focusing a key mutation region by using an Attention mechanism, and positioning a fault mutation point. According to the method, the fault positioning precision and real-time performance of the power distribution network can be effectively improved, and the method has high engineering application value and wide popularization prospects.

Description

TECHNICAL FIELD

[0001] The application relates to the field of power systems, in particular to a power distribution network grounding fault positioning method based on CEEMDAN-VMD and CNN-BiLSTM-Attention. BACKGROUND

[0002] With the development of power systems towards "intelligentization, distribution and renewable", the grounding fault in the power distribution network has become an important problem restricting the reliability and safety of power supply. Accurate and rapid positioning of grounding faults not only helps to reduce equipment damage and user outage time, but also promotes the development of power distribution automation construction and fault self-healing system. Traditional grounding fault positioning mainly relies on steady-state analysis and protection devices, but its positioning accuracy will decrease significantly in complex fault scenarios such as high resistance and weak grounding. In contrast, the positioning method based on transient traveling wave signals has the characteristics of high time-frequency resolution and anti-interference ability, and has gradually become a research hotspot in recent years.

[0003] At present, the research on power distribution network fault positioning at home and abroad mainly focuses on the following aspects. First, the fault positioning based on signal processing technology. In the traditional signal processing method, EMD, CEEMDAN, VMD, wavelet transform and other methods are widely used to extract the frequency domain and time domain features of the traveling wave signal. Studies have shown that these methods not only can effectively detect faults, but also can combine the time difference method and waveform matching method to locate the fault position. For example, using the traveling wave signal decomposed by CEEMDAN and VMD, combined with the time difference method, the position of the fault can be accurately calculated between different monitoring points. In addition, the positioning method based on multi-point data fusion can reduce the single-point positioning error and improve the positioning accuracy by combining the signals of multiple sampling points. Although the above methods can provide relatively accurate fault time positioning under certain conditions, the positioning accuracy still has great challenges in complex network environment and multi-point fault conditions. Secondly, the fault time and position prediction based on deep learning. Deep learning methods, especially CNN, LSTM, BiLSTM, Transformer and other models, have shown great potential in power distribution network fault positioning in recent years. Through the above models, spatial and temporal features can be extracted from complex time series data, so as to not only realize accurate prediction of fault time, but also effectively locate the position of fault occurrence. For example, combined with the signal data preprocessed by VMD and CEEMDAN, the CNN-BiLSTM structure has shown excellent performance in fault positioning, especially in high noise environment, it can still maintain high positioning accuracy.

[0004] Although significant progress has been made in the field of fault location using the above-mentioned techniques, there are still some challenges. First, the complex network topology and non-uniform signal propagation speed in distribution networks make accurate positioning significantly more difficult. Especially under the conditions of multiple point faults and different types of ground faults, the signal may be disturbed and the signal propagation path may change, which affects the positioning accuracy. Second, the high complexity of deep learning models and their dependence on large-scale data still limit their computational resources and real-time performance. Finally, improving the real-time performance and accuracy of fault location remains an important challenge in current research. SUMMARY

[0005] To solve the problems in the prior art, the application provides a distribution network grounding fault location method based on CEEMDAN-VMD and CNN-BiLSTM-Attention.

[0006] To achieve the above-mentioned purposes, the technical scheme of the application is as follows: a distribution network grounding fault location method based on CEEMDAN-VMD and CNN-BiLSTM-Attention:

[0007] Step 1: For high-frequency disturbance components in the traveling wave signal, CEEMDAN is used for signal decomposition, VMD is used for high-frequency component processing, and K-means clustering is used for IMF component classification to screen out high-frequency fault components;

[0008] CEEMDAN stabilizes each modal component by introducing adaptive noise and multiple iterative decompositions, reduces modal aliasing and end effects in EMD, and decomposes the signal into several intrinsic modal functions (IMF). The intrinsic modal function reflects the local oscillation characteristics of the signal. For a nonlinear and non-stationary signal x(t), the EMD decomposition process is as follows:

[0009] (1)

[0010] (2)

[0011] In the formula, is the i-th intrinsic modal function, is the decomposition result of the i-th mode in the n-th iteration, and N represents the number of iterations. r(t) represents the residual signal.

[0012] CEEMDAN reduces modal aliasing and boundary effects in the EMD method by introducing adaptive noise for decomposition through multiple iterations,

[0013] VMD is a signal decomposition method based on variational optimization method, which decomposes the signal into modal signals with different center frequencies, minimizes the bandwidth of each mode by variational method, and keeps the sum of the bandwidths of all modes minimum. For a given non-stationary signal x(t), the purpose of VMD is to decompose the signal into several modal components by minimizing the bandwidth of each mode, which is specifically expressed as:

[0014] (3)

[0015] (4)

[0016] (5)

[0017] In the formula, is the kth mode, is the center frequency of the kth mode, K is the number of decomposed modes, H represents Hilber transform, is a weight factor;

[0018] By alternating direction multiplier method ADMM, VMD converts the optimization problem into multiple sub-problems, and further solves the center frequency and bandwidth of each mode by iteration,

[0019] In the traveling wave signal processing of power distribution network, CEEMDAN and VMD are combined, where CEEMDAN is used to decompose low and medium frequency components in the signal, and VMD is used to further process high frequency signal part. The decomposition process combining CEEMDAN and VMD is expressed as:

[0020] (6)

[0021] In the formula, is the IMF component after CEEMDAN decomposition, is the high frequency mode after VMD decomposition.

[0022] Step two, design CNN-BiLSTM-Attention hybrid network architecture, use CNN to extract spatial features, BiLSTM to capture time sequence information, and Attention mechanism to focus on key mutation area to locate fault mutation point;

[0023] The BiLSTM model combines the outputs of forward and reverse LSTM units in terms of deep network structure; the LSTM neural network is composed of an input layer, a hidden layer and an output layer, and the hidden layer of the LSTM contains a special memory module which controls the flow of information through a gating mechanism; the LSTM memory unit is the core part which precisely manages the storage and update of information by controlling the forget gate, input gate and output gate, the forget gate determines which historical information needs to be retained, the input gate controls whether new information should be written into the memory unit, and the output gate determines the output of the current state, wherein the calculation formula of the forget gate is:

[0024] (7)

[0025] wherein, is a sigmoid activation function; is a weight matrix of the forget gate; is the output at t-1 time; is the input at t time; is the bias of the forget gate;

[0026] The input gate generates a new candidate vector by a tanh function , and determines the new information that needs to be added to the cell state by a sigmoid function.

[0027] (8)

[0028] wherein, , are weight matrices of the input gate and the candidate vector, respectively; , are biases of the input gate and the candidate vector, respectively,

[0029] The above BiLSTM model is combined with CNN and Attention to form a CNN-BiLSTM-Attention model to identify the traveling wave jump point information.

[0030] Compared with existing technologies, the present invention offers significant advantages in fault location accuracy and noise immunity. Existing technologies, traditional traveling wave fault location methods primarily rely on signal processing methods such as wavelet transform, EMD, or CEEMDAN. While these methods can effectively extract signal features under certain conditions, significant noise interference remains due to signal complexity and high-frequency disturbances, leading to large deviations in location accuracy. Furthermore, traditional methods rely on specific wavefront features for identifying fault abrupt changes, making it difficult to accurately identify high-frequency abrupt changes in complex power distribution network environments. The present invention, by combining CEEMDAN and VMD decomposition techniques, can better extract high-frequency and low-frequency features from current signals, especially effectively suppressing noise in the signal, thus improving the identifiability of fault features. In addition, a CNN-BiLSTM-Attention network architecture is employed. This architecture not only extracts spatial and temporal features of the signal but also focuses on key abrupt change regions through the Attention mechanism, further improving the accuracy of abrupt change point location. Meanwhile, although similar methods exist in stock price prediction, they address the non-stationarity of stock price fluctuations but do not involve high-frequency signal extraction or specific abrupt change point identification. Furthermore, stock price analysis falls under the field of statistics, thus lacking relevance. This innovative technical solution effectively improves the accuracy and real-time performance of fault location in power distribution networks, possessing strong engineering application value and broad prospects for widespread adoption. Attached Figure Description

[0031] Figure 1 The diagram shows the intrinsic mode functions in Example 1;

[0032] Figure 2 This is a graph showing the K-means clustering results in Example 1;

[0033] Figure 3 The spectrum diagrams for different frequency components in Example 1 are shown below.

[0034] Figure 4 The waveform diagram is a real-world test field waveform of the distribution network in Example 1.

[0035] Figure 5 This is a multi-noise level analysis diagram based on the CEEMDAN-VMD+GCN-LSTM algorithm in Example 1;

[0036] Figure 6 This is a flowchart of the method of the present invention. Detailed Implementation

[0037] With reference to the accompanying drawings, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by a person of ordinary skill in the art without creative effort belong to the scope of the present application.

[0038] A power distribution network grounding fault positioning method based on CEEMDAN-VMD and CNN-BiLSTM-Attention:

[0039] Step one: for the high-frequency disturbance component in the traveling wave signal, CEEMDAN is used for signal decomposition, VMD is used for processing the high-frequency component, K-means clustering is used for classifying the IMF component, and the high-frequency fault component is screened out;

[0040] CEEMDAN introduces adaptive noise and multiple iterative decompositions, so that each modal component is stable. The modal aliasing and end effect in EMD are reduced. EMD decomposes the signal into several intrinsic modal functions (IMF). The intrinsic modal function reflects the local oscillation characteristics of the signal. For a nonlinear and non-stationary signal x(t), the EMD decomposition process is:

[0041] (1)

[0042] (2)

[0043] In the formula, is the i-th intrinsic modal function, is the decomposition result of the i-th mode in the n-th iteration, and N represents the number of iterations. r(t) represents the residual signal.

[0044] CEEMDAN reduces the modal aliasing and boundary effect in the EMD method by introducing adaptive noise for decomposition through multiple iterations,

[0045] VMD is a signal decomposition method based on variational optimization method. The signal is decomposed into modal signals with different center frequencies. The bandwidth of each mode is minimized by variational method, and the sum of the bandwidths of all modes is minimized. For a given non-stationary signal x(t), the purpose of VMD is to minimize the bandwidth of each mode, and decompose the signal into several modal components. The specific representation is:

[0046] (3)

[0047] (4)

[0048] (5)

[0049] wherein, is the kth mode, is the center frequency of the kth mode, K is the number of decomposed modes, H represents the Hilber transform, is the weight factor;

[0050] By the alternating direction multiplier method (ADMM), VMD converts the optimization problem into multiple sub-problems, and further solves the center frequency and bandwidth of each mode by iteration,

[0051] In the traveling wave signal processing of power distribution network, CEEMDAN and VMD are combined, wherein CEEMDAN is used to decompose low and medium frequency components in the signal, and VMD is used to further process high frequency signal part, and the decomposition process combining CEEMDAN and VMD is expressed as:

[0052] (6)

[0053] wherein, is the IMF component after CEEMDAN decomposition, is the high frequency mode after VMD decomposition.

[0054] Step two, design the CNN-BiLSTM-Attention hybrid network architecture, use CNN to extract spatial features, BiLSTM to capture time sequence information, and Attention mechanism to focus on key mutation area to locate fault mutation point;

[0055] In terms of deep network structure, the BiLSTM model combines the outputs of forward and reverse LSTM units; wherein the LSTM neural network is composed of an input layer, a hidden layer and an output layer, and the hidden layer of LSTM contains a special memory module which controls the flow of information through a gating mechanism; the LSTM memory unit is the core part which accurately manages the storage and update of information by controlling the forget gate, input gate and output gate, the forget gate decides which historical information needs to be retained, the input gate controls whether new information should be written into the memory unit, and the output gate decides the output of the current state, wherein the calculation formula of the forget gate is:

[0056] (7)

[0057] wherein, is a sigmoid activation function; is the weight matrix of the forget gate; is the output at t-1 time; is the input at t time; is the bias of the forget gate;

[0058] The input gate generates a new candidate vector from a tanh function and the sigmoid function determines the new information that needs to be added to the cell state.

[0059] (8)

[0060] wherein, , are the weight matrices of the input gate and the candidate vector, respectively; , are the biases of the input gate and the candidate vector, respectively,

[0061] The above BiLSTM model is combined with CNN and Attention to form a CNN-BiLSTM-Attention model to identify traveling wave jump point information.

[0062] Embodiment 1: The specific scheme of using the power distribution network grounding fault positioning method based on CEEMDAN-VMD and CNN-BiLSTM-Attention is as follows:

[0063] 1. Signal preprocessing and high-frequency component extraction:

[0064] First, the current signal of the power distribution network is preprocessed to remove interference information such as background noise. Then, the CEEMDAN (Complete Ensemble Empirical Mode Decomposition with Adaptive Noise) algorithm is used to decompose the original current signal to obtain multiple intrinsic mode functions (IMF). The extracted high-frequency components are further decomposed by VMD (Variational Mode Decomposition) to enhance the recognizability of high-frequency features. As shown in Figure 1 , the original signal is decomposed into each IMF component for visualization processing, and it can be seen that the high-frequency period is mainly concentrated in the range of 50-150 Hz.

[0065] 2. Sample entropy and K-means clustering:

[0066] The sample entropy of each IMF component is calculated to quantify its complexity and irregularity, and the K-means clustering algorithm is used to classify the IMF components to distinguish high-frequency, medium-frequency and low-frequency components, and filter out high-frequency components as fault features. As shown in Figure 2 , the high-frequency components in Co-IMF1 are more obvious after summing the same class of IMF components, containing fault jump point information. At the same time, the frequency components of each component are calculated using Fourier transform to verify whether the above method has realized the division of high, medium and low frequencies. As shown in Figure 3As shown in Fig. 2, it can be seen that IMF1 and IMF2 complete the decomposition of high-frequency signals, which confirms the feasibility of the above method.

[0067] 3. CNN-BiLSTM-Attention deep learning network:

[0068] A deep learning model composed of a convolutional neural network (CNN), a bidirectional long short-term memory network (BiLSTM), and an attention mechanism (Attention) is constructed. CNN is used to extract spatial features of time series signals, BiLSTM captures the time series dependence of signals, and Attention mechanism focuses on key fault mutation areas. The output of the network is the time delay of the fault time relative to the start time of the sliding window.

[0069] 4. Multi-terminal synchronous measurement and double-ended positioning:

[0070] A double-ended ranging method is used to calculate the time difference of traveling wave propagation by synchronously collecting fault traveling wave signals at both ends of the distribution network, and then to realize accurate positioning of the fault point. The time delay is converted into spatial distance by formula, and error correction is performed to further improve the positioning accuracy.

[0071] 5. Field test verification and performance evaluation:

[0072] The performance of the method is evaluated by considering various fault scenarios such as different neutral grounding methods and high resistance grounding on the actual distribution network platform. Among them Figure 4 are the actual waveforms measured in the distribution network test field. The experimental results show that under the condition of 5dB noise, the fault positioning error can be controlled within 50 meters, and the method can effectively deal with complex noise environment. In order to further prove the adaptability of the algorithm in complex environment, this paper further verifies that the fault traveling wave energy is mainly concentrated in the high-frequency components (d1-d3) by the energy proportion of each layer (a3 low-frequency trend, d1-d3 high-frequency details) of wavelet decomposition. As Figure 5 shown, with the increase of noise (decrease of SNR), the proportion of high-frequency detail energy increases significantly, and CEEMDAN-VMD can still accurately extract effective features. At the same time, the wave head position marked by the red vertical line is always close to the true value, even under the condition of high noise with SNR=5dB, the traveling wave point error is not more than ±10ns, corresponding to the traveling wave velocity positioning error less than 30m, and the d1 coefficient still maintains clear pulse characteristics, which proves that CEEMDAN-VMD has significant suppression ability to high-frequency noise.

[0073] The above shows and describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above-mentioned embodiments, and the above-mentioned embodiments and descriptions in the specification are only preferred examples of the present application and are not intended to limit the present application. Various changes and improvements can be made to the present application without departing from the spirit and scope of the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for locating grounding faults in distribution networks based on CEEMD-VMD and CNN-BiLSTM-Attention, characterized in that: Step 1: For the high-frequency disturbance components in the traveling wave signal, CEEMD is used for signal decomposition, VMD is used to process the high-frequency components, and K-means clustering is combined to classify the IMF components and screen out the high-frequency fault components. Step 2: Design a CNN-BiLSTM-Attention hybrid network architecture. CNN is used to extract spatial features, BiLSTM is used to capture temporal information, and the Attention mechanism is used to focus on key mutation regions to locate fault mutation points.

2. The method for locating grounding faults in a distribution network based on CEEMD and CNN-BiLSTM-Attention as described in claim 1, characterized in that: The specific method in step one is as follows: CEEMDAN stabilizes each modal component by introducing adaptive noise and multi-round iterative decomposition. This reduces mode aliasing and endpoint effects in EMD. EMD decomposes the signal into several Intrinsic Mode Functions (IMFs), which reflect the local oscillation characteristics of the signal. For a nonlinear and non-stationary signal x(t), the EMD decomposition process is as follows: （1） （2） In the formula, Let i be the i-th intrinsic mode function. Let be the decomposition result of the i-th mode in the nth iteration, where N represents the iteration number and r(t) represents the residual signal; CEEMDAN introduces adaptive noise through multiple iterations for decomposition, reducing mode mixing and boundary effects in the EMD method. Variational Dynamics (VMD) is a signal decomposition method based on variational optimization. It decomposes a signal into modal signals with different center frequencies, minimizing the bandwidth of each mode while keeping the sum of the bandwidths of all modes to a minimum. For a given non-stationary signal x(t), the goal of VMD is to decompose the signal into several modal components by minimizing the bandwidth of each mode, specifically as follows: （3） （4） （5） In the formula, For the k-th mode, Let H be the center frequency of the k-th mode, K be the number of modes in the decomposition, and H represent the Hilber transform. As a weighting factor; By employing the Alternating Direction Multiplier Method (ADMM), VMD transforms the optimization problem into multiple subproblems, and then iteratively solves for the center frequency and bandwidth of each mode. In distribution network traveling wave signal processing, CEEMD and VMD are combined. CEEMD is used to decompose the low-frequency and mid-frequency components of the signal, while VMD is used to further process the high-frequency signal portion. The decomposition process combining CEEMD and VMD is expressed as follows: （6） In the formula, The IMF components after CEEMDAN decomposition. These are the high-frequency modes after VMD decomposition.

3. The method for localizing grounding faults in a distribution network based on CEEMD and CNN-BiLSTM-Attention as described in claim 1, characterized in that: The specific method in step two is as follows: In terms of deep network structure, the BiLSTM model combines the outputs of forward and backward LSTM units. The LSTM neural network consists of an input layer, hidden layers, and an output layer. The hidden layers of the LSTM contain a special memory module that controls the flow of information through a gating mechanism. The LSTM memory unit is the core component, precisely managing the storage and updating of information by controlling the forget gate, input gate, and output gate. The forget gate determines which historical information needs to be retained, the input gate controls whether new information should be written to the memory unit, and the output gate determines the output of the current state. The calculation formula for the forget gate is: （7） In the formula, It is the sigmoid activation function; Here is the weight matrix for the forget gate; This is the output at time t-1; This is the input at time t; For the offset of the forget gate; The input gate generates new candidate vectors using the tanh function. At the same time, the sigmoid function is used to determine the new information about the cell state that needs to be added. （8） In the formula, , These are the weight matrices for the input gate and the candidate vector, respectively; , These are the biases for the input gate and the candidate vector, respectively. The BiLSTM model described above is combined with CNN and Attention to form a CNN-BiLSTM-Attention model, which identifies the information of the transition points of traveling waves.

Images