A micro-grid charging pile fault positioning method and system based on HHT and improved BP neural network

Abstract

This invention discloses a method and system for fault location of charging piles in microgrids based on HHT and an improved BP neural network, belonging to the field of microgrid fault diagnosis technology. The method includes: acquiring transient current signals from each charging pile node; obtaining intrinsic mode function components (IMFs) through empirical mode decomposition; selecting the first IMF component and performing Hilbert transform to calculate its instantaneous amplitude vector sum as a single-dimensional fault feature; inputting the normalized feature into an improved BP neural network for fault location; and introducing a momentum term in the weight update of the improved BP neural network to optimize the training process. This invention solves the technical problems of insufficient feature extraction capability for non-stationary fault signals and slow and unstable convergence of the location model in traditional methods, achieving fast, accurate, and adaptive fault location of charging piles in microgrids. Simulation verification shows a location accuracy of 100%, significantly improving the safety and reliability of the microgrid system.

Description

Technical Field

[0001] This invention belongs to the field of fault diagnosis and location technology for electric vehicle charging piles in microgrids. Specifically, it relates to a fault location method and system for charging piles in microgrids based on Hilbert-Huang Transform (HHT) and an improved BP neural network. It is particularly suitable for feature extraction and intelligent identification of nonlinear and non-stationary fault signals of charging piles in microgrids, and can achieve fast, accurate and adaptive location of charging pile faults in microgrids. Background Technology

[0002] As microgrid systems rapidly evolve towards higher power density, integration, and intelligence, the scale of electric vehicle charging stations connected to microgrids continues to expand. Charging loads are exhibiting centralized and fluctuating characteristics, placing higher demands on the safety and reliability of microgrids. When a microgrid charging station or its distribution lines experience a fault, rapid and accurate fault location will significantly improve the stability and power quality of the microgrid system.

[0003] Currently, fault location for microgrid charging piles is typically divided into two stages: fault feature extraction and fault location identification. In terms of feature extraction, traditional methods mainly rely on Fourier transform, short-time Fourier transform, and wavelet transform. Among these, Fourier transform is suitable for stationary signals but has limited processing capabilities for non-stationary signals; while short-time Fourier transform can provide both time and frequency information, its fixed window width limits its time and frequency resolution; and although wavelet transform can handle non-stationary signals, the selection of its wavelet basis functions is subjective and fixed, lacking adaptive capabilities and making it difficult to adapt to complex and ever-changing fault signal characteristics.

[0004] In fault localization, common methods include Petri nets, expert systems, and various neural networks. Expert systems and other methods rely on manually constructed rule bases, which have poor fault tolerance and self-learning capabilities, making it difficult to adapt to changes in system structure or operating conditions. Although traditional BP neural networks have certain nonlinear mapping and adaptive capabilities, they suffer from slow convergence speed, susceptibility to getting trapped in local minima, and oscillations during training, which affect the real-time performance and stability of fault localization.

[0005] In summary, existing technologies generally suffer from insufficient adaptive feature extraction capabilities, poor robustness of positioning models, and complex system maintenance when processing nonlinear and non-stationary transient signals such as microgrid charging piles. Therefore, there is an urgent need for a fault diagnosis method that can adaptively extract fault features, possess strong learning and generalization capabilities, and achieve high-precision real-time positioning. Summary of the Invention

[0006] To address the problems of the existing technologies, this invention proposes a microgrid charging pile fault location method based on HHT and an improved BP neural network. First, the HHT method is used to perform adaptive time-frequency analysis on the non-stationary transient current signals of each node, extracting feature quantities that can effectively distinguish different fault locations. Then, these optimized feature quantities are input into a BP neural network improved with a momentum term, utilizing its powerful nonlinear mapping and self-learning capabilities to achieve intelligent identification and precise location of fault nodes.

[0007] The technical solution of the present invention is as follows: A fault location technology for microgrid charging piles based on Hilbert-Huang transform and an improved BP neural network with the introduction of momentum terms. First, fault transient current signals are collected at each charging pile node of the microgrid. Empirical mode decomposition (EMD) is performed on the signals of each node. The decomposition terminates when the residual components tend to be monotonic functions or constants. Cubic spline interpolation is used to construct upper and lower envelopes to obtain intrinsic mode functions arranged from high to low frequency. IMF1, which contains the most significant high-frequency impact transient information, is selected as the representative feature component. Hilbert transform is then performed to obtain the instantaneous amplitude vector. The element arithmetic sum is used to construct the node's single-dimensional fault feature quantity, which is then normalized to form a normalized feature vector. Subsequently, this vector is input into a pre-trained three-layer improved BP neural network. The convergence speed and stability are improved by using a weight update mechanism with momentum terms. During the training phase, overfitting is suppressed by dividing the training set / validation set and an early stopping strategy where the validation error no longer decreases. Finally, the network output enables intelligent identification and accurate location of fault nodes. This scheme can also be packaged into a positioning system consisting of signal acquisition, feature extraction, neural network processing, and result display modules. It supports periodic retraining using newly added fault data through a model update module to maintain long-term adaptability. The specific steps are as follows: Step 1: Adaptive Fault Feature Extraction Based on HHT S11. Collect transient current signals of each charging pile node in the microgrid; S12. Perform empirical mode decomposition on the current signal of each node to obtain multiple intrinsic mode function components arranged in descending order of frequency; S13. Select the first intrinsic mode function component of the current signal at each node as a representative feature component; S14. Perform Hilbert transform on the representative characteristic components to obtain the corresponding instantaneous amplitude vector; S15. Calculate the sum of all elements of the instantaneous amplitude vector of each node to obtain the single-dimensional fault characteristic of that node. S16. Normalize the single-dimensional fault features of all nodes to form a normalized feature vector. Step 2: Intelligent identification of fault nodes based on an improved BP neural network: S21. Input the normalized feature vector obtained in step one into the pre-trained improved BP neural network model, wherein the weight update formula of the improved BP neural network model includes a momentum term. S22. Obtain the fault node location result through the output of the improved BP neural network model.

[0008] Furthermore, the termination condition of the empirical mode decomposition in step S12 is that the residual components become monotonic functions or constants, and the empirical mode decomposition uses cubic spline interpolation to fit the upper and lower envelopes of the signal.

[0009] Furthermore, the basis for selecting the first intrinsic mode function component in step S13 is that this component contains the most significant high-frequency impulse characteristics in the original transient current signal, and can reflect the transient characteristics of the fault to the greatest extent.

[0010] Furthermore, the single-dimensional fault characteristic quantity mentioned in step S15 is obtained by calculating the arithmetic sum of all elements in the instantaneous amplitude vector.

[0011] Furthermore, the improved BP neural network model adopts a three-layer structure, including an input layer, a hidden layer, and an output layer, wherein the number of nodes in the input layer is equal to the number of charging pile nodes in the microgrid, and the number of nodes in the output layer corresponds to the number of possible faulty nodes.

[0012] Furthermore, a training sample construction step is included before step one: S01. Establish a microgrid model containing multiple charging pile nodes in the simulation environment; S02. Simulate different fault types at each node and collect the corresponding transient current signals; S03. Process the acquired signals using the methods in steps S11-S16 of claim 1 to construct a training sample set.

[0013] Furthermore, the training process of the improved BP neural network model also includes: S31. Divide the training sample set into a training set and a validation set; S32. Train the network using the training set and monitor the network performance on the validation set. S33. When the validation set error no longer decreases after multiple consecutive iterations, stop training to prevent overfitting.

[0014] Second, the present invention also provides a microgrid charging pile fault location system for implementing the above method, characterized in that it includes: The signal acquisition module is used to acquire the current signals of each charging pile node in the microgrid; The feature extraction module is used to extract feature quantities from the acquired current signal; The neural network processing module has an embedded improved BP neural network model that has been trained, which is used to receive the feature output from the feature extraction module and identify fault nodes. The results display module is used to output the fault location results.

[0015] Furthermore, it also includes a model update module, which is used to periodically retrain the improved BP neural network model using newly collected fault data to maintain the model's accuracy and adaptability.

[0016] Third, the present invention also provides a computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the steps of the above-described method.

[0017] Compared with the prior art, the technical effects of the present invention are as follows: 1) By adopting HHT adaptive time-frequency analysis technology, the complex fault signal is adaptively decomposed into IMF components with clear physical meaning through empirical mode decomposition (EMD). Then, the instantaneous amplitude features of the IMF1 component are extracted by Hilbert transform. This solves the problem that traditional Fourier transform, wavelet transform and other methods are insufficient for processing non-stationary signals, and realizes high-precision extraction of fault transient features, laying a data foundation for accurate fault location.

[0018] 2) An improved BP neural network technique with a momentum term is adopted. In the weight update process of the standard BP algorithm, a momentum term is introduced, so that the adjustment of network weights not only considers the current gradient direction, but also accumulates the historical adjustment trend. This effectively suppresses the oscillation phenomenon during training, speeds up the convergence speed, enhances the stability of model training, and solves the technical defects of traditional BP neural networks that are prone to getting trapped in local minima and have slow convergence.

[0019] 3) The single-dimensional feature extraction technique based on the IMF1 component is adopted. The first IMF component (IMF1) after EMD decomposition is selected for feature extraction. Only its instantaneous amplitude vector is calculated as the node feature quantity. While retaining the most critical fault information, the multi-dimensional time-frequency features are compressed into single-dimensional features, which greatly reduces the feature dimension, reduces the computational burden, and improves the real-time performance of the system. Attached Figure Description

[0020] Figure 1 Schematic diagram of a three-layer BP network structure; Figure 2 Fault location flowchart; Figure 3 BP network training flowchart; Figure 4 Schematic diagram of a microgrid charging pile structure; Figure 5 Current signals of each node when node 1 fails; Figure 6 The instantaneous amplitude waveforms of the IMF1 component at each node, where (a) the instantaneous amplitude of the IMF1 component at node 1, (b) the instantaneous amplitude of the IMF1 component at node 2, (c) the instantaneous amplitude of the IMF1 component at node 3, (d) the instantaneous amplitude of the IMF1 component at node 4, and (e) the instantaneous amplitude of the IMF1 component at node 5. Figure 7 BP neural network training error variation curve Figure 8 Confusion matrix of BP neural network training results Detailed Implementation To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention.

[0021] HHT transform signal processing: First, the signal is subjected to Empirical Mode Decomposition (EMD) to obtain Intrinsic Mode Function (IMF) components. Then, the Hilbert transform is performed on the IMF components obtained by decomposition to obtain the Hilbert spectrum.

[0022] The basic steps of the EMD algorithm are as follows: Step 1: Use cubic spline interpolation to fit all local maxima and minima of the original signal to obtain the upper and lower envelopes of the signal.

[0023] Step 2: Calculate the average value of the upper and lower envelopes of the signal, and record the difference between the original signal and the envelope mean as .

[0024] Step 3: When the IMF condition is met, the first IMF component is obtained; if not, Step 1 and Step 2 will be repeated as a new original signal until the first IMF component can be obtained, which is denoted as IMF1.

[0025] Step 4: Let be the new original signal, repeat the above steps to obtain the second IMF component, repeat n times to obtain n IMF components, and stop iterating when the function satisfies the condition of being a monotonic function or a constant. At this point: (1) Where is the i-th IMF component, and is the residual component after decomposition.

[0026] By performing a Hilbert transform on the IMF components obtained from EMD decomposition, their instantaneous amplitude and frequency can be obtained. Performing a Hilbert transform on any IMF component in equation (1) yields: (2) This is how to construct an analytical signal: (3) Its instantaneous amplitude and frequency are: (4) Let the instantaneous amplitude vector of the first IMF component after Hilbert transform be , and calculate the sum of the amplitude values ​​as: (5) Backpropagation (BP) is a widely used neural network training algorithm. To overcome its tendency to get trapped in local minima and its slow convergence speed, many improvements have been proposed. Adding a momentum term to the changes in network weights is an effective and practical strategy. Typically, the momentum factor is experimentally set to a constant within the (0,1) interval. If the momentum factor is too small, the effect of adding the momentum term is not significant; if it is too close to 1, the algorithm will become unstable. Therefore, the appropriate selection of the momentum factor is crucial for network training. This paper proposes an adaptive method for determining the momentum factor. Numerical experiments show that the BP algorithm using the momentum factor outperforms the traditional BP algorithm in both convergence speed and stability. A BP neural network consists of an input layer, hidden layers, and an output layer. Layers are fully interconnected, but neurons within the same layer are not connected. Figure 1 This is a schematic diagram of a typical three-layer neural network structure.

[0027] in x k is the input to the BP neural network. y m is the output of the hidden layer of the BP network. z n is the actual output of the network. T n represents the network's expected output.

[0028] The BP algorithm, employing a backpropagation learning rule, is a supervised learning algorithm. It consists of two parts: forward propagation of the signal and backward propagation of the error. Forward propagation involves feeding samples into the input layer, where they are processed by the hidden layer, and then the output layer outputs the network's actual output. When the error between the actual output and the desired output does not meet the requirements, backpropagation uses the steepest descent method to continuously modify the network's weights, minimizing the error between the actual and desired output.

[0029] Suppose a neuron in the output layer of a BP neural network j The actual network output is zj, expected output is T If j, then the network's error objective function is: (6) Because the standard backpropagation (BP) algorithm adjusts weights only according to the gradient descent direction of the time-major error, without considering the gradient direction before time t, it often causes oscillations during training and slow convergence. To improve training speed, a momentum term can be added to the weight formula: (7) Where W is the weight matrix of a certain layer, and z is the output vector of a certain layer. For learning rate, This is called the momentum factor. , This represents the accumulated experience in adjustments.

[0030] Based on the fundamental principles of HHT and BP neural networks, a fault location method for microgrid charging piles is proposed, the basic process of which is as follows: Figure 2 As shown.

[0031] The basic steps of fault location methods: Step 1: Collect current signals from each node of the microgrid charging pile.

[0032] Steps: Perform EMD decomposition on the acquired current signal to obtain several IMF components. Denote the first IMF component as IMF1.

[0033] Step 3: Perform Hilbert transform on the decomposed IMF components to obtain their instantaneous amplitude vectors.

[0034] Step 4: Calculate the sum of the instantaneous amplitude vectors according to equation (5) and use it as the characteristic quantity S of the acquired current signal.

[0035] Step 5: Normalize the obtained feature values. The normalization formula is as follows: (8) Wherein, represents the normalized sample data.

[0036] Step 6: Input the normalized sample data into the trained BP neural network to obtain the fault location result.

[0037] The BP neural network is trained using features of typical faults as samples, and its training process is as follows: Figure 3 As shown.

[0038] A microgrid charging pile model was built using the MATLAB / SIMULINK simulation platform, and experimental data were obtained. Its structural diagram is shown below. Figure 4As shown. The system has a total of 5 nodes. Taking the current signal collected when node 1 experiences an open circuit fault as an example, the fault duration is 5.8~6s, the sampling frequency is 1000Hz, and the data acquisition time is 5~7s.

[0039] Taking node 1 as an example, fault feature extraction is performed. Since the IMF components obtained after EMD decomposition are arranged in descending order of frequency, the IMF1 component best represents the characteristics of the original signal among all the IMF components obtained after EMD decomposition. Therefore, when an open-circuit fault occurs at node 1, the IMF1 components of each node are as follows: Figure 5 As shown, the Hilbert transform is then applied to the IMF1 components of each node to obtain their respective instantaneous amplitude vectors, as follows. Figure 5 As shown.

[0040] from Figure 6 As can be seen from the figures, when a single-phase ground fault occurs at node 1, the instantaneous amplitudes of nodes 1, 4, and 5 are all different. Observing Figures (a) to (d), it can be seen that although the instantaneous amplitude waveforms of nodes 2 and 3 are not significantly different, their minimum amplitude values ​​are different. This paper uses the sum of the instantaneous amplitude vectors of the IMF1 components of the collected fault current signal as the feature vector of the fault signal, and the feature vectors are statistically summarized in Table 1, with units of MA. In the table, represents the i-th node experiencing an open-circuit fault, and represents the i-th node.

[0041] The BP network has 5 input nodes (k=5), representing the characteristic vector of the current in each node when a single-phase ground fault occurs at a given node; and 5 output nodes (n=5), representing each node. Based on empirical formulas and through multiple experiments, the network achieves the best training performance when the number of hidden layer nodes is set to m=12.

[0042] The output vector of a BP network is represented by numbers, which correspond to faulty nodes. For example, when the network outputs 1, it represents the faulty node n1.

[0043] The target error for training the BP network was set to 0.001, the number of iterations to 5000, the learning rate to n=0.8, and the momentum factor to α=0.5. The transfer function between layers was a hyperbolic tangent sigmoid function, and the network algorithm was gradient descent with momentum BP. The BP network training error curve and confusion matrix are shown below. Figure 7 and Figure 8 As shown in Table 1, the fault identification results are as follows.

[0044] Table 1. Sum of instantaneous amplitudes of the IMF1 component of the fault current signal Depend on Figure 7 As can be seen, after 5000 training iterations, the set training error requirement was achieved. Figure 8As can be seen from the results, the fault node location method proposed in this paper achieves 100% accuracy in locating fault nodes of microgrid charging piles, thus achieving the goal of precise fault location for microgrid charging piles.

Claims

1. A fault location method for microgrid charging piles based on HHT and an improved BP neural network, characterized in that, Includes the following steps: Step 1: Adaptive Fault Feature Extraction Based on HHT S11. Collect transient current signals of each charging pile node in the microgrid; S12. Perform empirical mode decomposition on the current signal of each node to obtain multiple intrinsic mode function components arranged in descending order of frequency; S13. Select the first intrinsic mode function component of the current signal at each node as a representative feature component; S14. Perform Hilbert transform on the representative characteristic components to obtain the corresponding instantaneous amplitude vector; S15. Calculate the sum of all elements of the instantaneous amplitude vector of each node to obtain the single-dimensional fault characteristic of that node. S16. Normalize the single-dimensional fault features of all nodes to form a normalized feature vector. Step 2: Intelligent identification of fault nodes based on an improved BP neural network: S21. Input the normalized feature vector obtained in step one into the pre-trained improved BP neural network model, wherein the weight update formula of the improved BP neural network model includes a momentum term. S22. Obtain the fault node location result through the output of the improved BP neural network model.

2. The microgrid charging pile fault location method based on HHT and improved BP neural network according to claim 1, characterized in that, The termination condition for the empirical mode decomposition in step S12 is that the residual components become monotonic functions or constants. The empirical mode decomposition uses cubic spline interpolation to fit the upper and lower envelopes of the signal.

3. The microgrid charging pile fault location method based on HHT and improved BP neural network according to claim 1, characterized in that, The basis for selecting the first intrinsic mode function component in step S13 is that this component contains the most significant high-frequency impulse characteristics in the original transient current signal, and can reflect the transient characteristics of the fault to the greatest extent.

4. The microgrid charging pile fault location method based on HHT and improved BP neural network according to claim 1, characterized in that, The single-dimensional fault characteristic quantity mentioned in step S15 is obtained by calculating the arithmetic sum of all elements in the instantaneous amplitude vector.

5. The microgrid charging pile fault location method based on HHT and improved BP neural network according to claim 1, characterized in that, The improved BP neural network model adopts a three-layer structure, including an input layer, a hidden layer, and an output layer. The number of nodes in the input layer is equal to the number of charging pile nodes in the microgrid, and the number of nodes in the output layer corresponds to the number of possible faulty nodes.

6. The microgrid charging pile fault location method based on HHT and improved BP neural network according to claim 1, characterized in that, The training sample construction step is included before step one: S01. Establish a microgrid model containing multiple charging pile nodes in the simulation environment; S02. Simulate different fault types at each node and collect the corresponding transient current signals; S03. Process the acquired signals using the methods in steps S11-S16 of claim 1 to construct a training sample set.

7. The microgrid charging pile fault location method based on HHT and improved BP neural network according to claim 1, characterized in that, The training process of the improved BP neural network model also includes: S31. Divide the training sample set into a training set and a validation set; S32. Train the network using the training set and monitor the network performance on the validation set. S33. When the validation set error no longer decreases after multiple consecutive iterations, stop training to prevent overfitting.

8. A microgrid charging pile fault location system, used to implement the method according to any one of claims 1-7, characterized in that, include: The signal acquisition module is used to acquire the current signals of each charging pile node in the microgrid; The feature extraction module is used to extract feature quantities from the acquired current signal; The neural network processing module has an embedded improved BP neural network model that has been trained, which is used to receive the feature output from the feature extraction module and identify fault nodes. The results display module is used to output the fault location results.

9. The microgrid charging pile fault location system according to claim 8, characterized in that, It also includes a model update module, which is used to periodically retrain the improved BP neural network model using newly collected fault data to maintain the model's accuracy and adaptability.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1-7.

Images