Power distribution network fault classification method based on VMD decomposition and CNN-BiLSTM
By optimizing VMD decomposition parameters and building CNN-BiLSTM network model, the problem of rapid and accurate identification of distribution network fault types is solved, and efficient fault classification is achieved in complex environments.
Patent Information
- Application Number
- CN202510564002.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-30
- Publication Date
- 2025-08-26
AI Technical Summary
The prior art is difficult to quickly and accurately identify fault types in complex distribution network environments, especially under non-stationary, nonlinear signal and noise interference. The fault classification accuracy and robustness of traditional methods are insufficient.
Using the sparrow optimization algorithm that optimizes VMD decomposition parameters combined with Osprey and Cauchy mutation, a CNN-BiLSTM network model is constructed, and the time-frequency characteristics and timing characteristics of the fault signal are deeply mined to achieve accurate classification of fault types.
It improves the accuracy and robustness of fault classification in distribution networks, especially in complex operating conditions, and improves the accuracy of fault classification.
Smart Images
Figure CN120541594A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of safe operation and analysis of distribution networks, and in particular to a distribution network fault classification method based on VMD decomposition and CNN-BiLSTM. Background Art
[0002] As a vital component of the power system, the distribution network undertakes the critical task of distributing electricity to end users. According to statistics, approximately 90% of faults occurring in power system operations occur within the distribution network. When a distribution network fault occurs, the ability to quickly and accurately identify the fault type directly impacts the efficiency of fault location, isolation, and power restoration. However, due to the complex structure and volatile operating environment of the distribution network, fault signals often exhibit non-stationary and nonlinear characteristics and are susceptible to noise interference, posing significant challenges to traditional fault classification methods.
[0003] Distribution network fault classification methods are mainly divided into two categories: those based on traditional signal processing and those based on machine learning. Among signal processing methods, the Fourier transform is used due to its advantages in steady-state signal analysis, but its ability to capture transient fault characteristics is limited; although the wavelet transform can provide time-frequency analysis capabilities, its basis function selection has a significant impact on the results, and it is difficult to adaptively process complex fault signals; the S transform has high computational complexity, which is not conducive to real-time engineering applications. In terms of signal preprocessing, empirical mode decomposition and its improved algorithms have been used for fault signal analysis, but these methods have problems such as modal aliasing and endpoint effects, resulting in poor accuracy in identifying distribution network fault types. Among machine learning methods, algorithms such as support vector machines and random forests rely on manual feature extraction, which is not only inefficient but also lacks generalization capabilities when faced with new fault modes.
[0004] The present invention constructs a distribution network fault classification and identification model based on VMD decomposition combined with CNN-BiLSTM. First, the zero-sequence voltage of the distribution network bus and the current signal of the low-voltage side of the transformer are collected as subsequent analysis signals. The Osprey and Cauchy variant sparrow optimization algorithms are used to optimize the mode number and bandwidth constraint parameters of the VMD decomposition to form a time-frequency feature matrix. Secondly, the CNN-BiLSTM recognition network model is used to deeply mine the extracted video feature vectors from two aspects: local spatial features and forward and reverse direction temporal characteristics, ultimately achieving accurate classification of distribution network fault types. Summary of the Invention
[0005] 1. Technical problems to be solved:
[0006] In response to the above technical problems, the present invention provides a method for realizing signal adaptive decomposition by optimizing VMD decomposition parameters, and combining the CNN-BiLSTM network to deeply mine fault characteristics from both spatial and temporal dimensions, thereby effectively improving the accuracy and robustness of distribution network fault classification, especially the classification performance under complex working conditions such as different fault locations, transition resistances, fault initial phase angles, and strong noise.
[0007] 2. Technical solution:
[0008] In order to solve the above problems, the present invention provides a distribution network fault classification method based on VMD decomposition and CNN-BiLSTM, comprising the following steps:
[0009] Step 1: Collect the busbar zero-sequence voltage and transformer low-voltage side three-phase current signals, and intercept the four cycles before the fault and the five cycles after the fault as signals for subsequent analysis.
[0010] Step 2: Apply the sparrow optimization algorithm combined with the Osprey algorithm and Cauchy mutation to optimize the VMD decomposition parameters of various fault signals.
[0011] Step 3: Perform VMD decomposition on the fault signal and construct the time-frequency feature matrix of the fault data.
[0012] Step 4: Build a CNN-BiLSTM recognition network model and input the fault dataset to train the CNN-BiLSTM model. Deeply mine the extracted time-frequency feature vectors from the perspectives of spatial features and forward and reverse time series characteristics to achieve accurate classification of distribution network fault types.
[0013] Step 5: Validate the model using the test dataset.
[0014] Furthermore, in step 2, the application of the Osprey algorithm in the global exploration strategy is described by the following formula (1):
[0015]
[0016] Where: In the first stage Osprey location; is a random number between [1,2]; It is The location of the fish determined by the osprey; The value of is 1 or 2.
[0017] Furthermore, in step 2, the Cauchy mutation algorithm is described by the following formula (2) by perturbing the individuals in the sparrow position update:
[0018]
[0019] Where: It is the position of the sparrow after replacement; is the current optimal solution; is the standard Cauchy distribution.
[0020] The distribution network fault classification method based on VMD decomposition and CNN-BiLSTM is characterized in that: in order to optimize the performance of CNN, the fully connected layer uses the ReLU function as the activation function of the neuron. For one-dimensional input data, assuming the length is , the fully connected layer contains neurons, and the output formula of each neuron is:
[0021]
[0022] Where: Indicates the Layer The neuron and The first layer The connection weights between neurons; Representative Layer The input of each neuron; For the Layer The bias term of a neuron,
[0023] In the output layer, after processing by the softmax function, we get:
[0024]
[0025] Where: Indicates the The original output value of each neuron is classified through the softmax function of the output layer to obtain the final classification result.
[0026] BiLSTM is an improved version of LSTM. The gating mechanism of LSTM consists of three parts: forget gate, input gate, and output gate. At the same time, a state unit is introduced to control the operation of the entire network. The calculation method of LSTM is as follows:
[0027]
[0028] in, It is the Gate of Oblivion; is the forget gate parameter; express Input at the moment; yes The hidden layer output at time t; is the input gate; is the input gate parameter; yes The input state at the moment; tanh is the activation function; yes The internal state at a moment; is the state parameter; is the output gate; is the output gate parameter; is the hidden layer output,
[0029] BiLSTM contains two independent hidden layers. There is no direct interaction between the hidden layer of forward propagation and the hidden layer of backward propagation. They are regarded as two independently running networks with opposite data flows:
[0030]
[0031] Where: yes The hidden layer state that is always positive; yes The hidden layer state that is reversed at that moment.
[0032] 3.Beneficial effects:
[0033] This paper proposes a comprehensive fault classification method based on the Osprey and Cauchy variation sparrow optimization algorithm (OCSSA), variational mode decomposition (VMD), convolutional neural network (CNN), and bidirectional long short-term memory (BiLSTM) networks. This method optimizes VMD parameter settings by improving the sparrow optimization algorithm. It then combines the CNN and BiLSTM models, extracting local spatial features of faults through the CNN and capturing the forward and reverse temporal characteristics of fault signals using the BiLSTM, enabling accurate and efficient fault classification. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] Figure 1 It is a flow chart of the present invention.
[0035] Figure 2 This is the structure diagram of the CNN-BiLSTM model.
[0036] Figure 3 It is a simulation model diagram of a 10kV distribution network in an embodiment.
[0037] Figure 4 Schematic diagram of VMD decomposition in the embodiment (taking AB two-phase short circuit as an example).
[0038] Figure 5 This is the comparison curve of fault classification accuracy between OCSSA-VMD-CNN-BiLSTM and traditional methods DETAILED DESCRIPTION
[0039] The present invention is described in detail below with reference to the accompanying drawings and embodiments.
[0040] like Figure 1 As shown in FIG, a distribution network fault classification method based on VMD decomposition and CNN-BiLSTM includes the following steps:
[0041] Step 1: Collect the busbar zero-sequence voltage and transformer low-voltage side three-phase current signals, and intercept the four cycles before the fault and the five cycles after the fault as signals for subsequent analysis.
[0042] Step 2: Apply the sparrow optimization algorithm combined with the Osprey algorithm and Cauchy mutation to optimize the VMD decomposition parameters of various fault signals.
[0043] Step 3: Perform VMD decomposition on the fault signal and construct the time-frequency feature matrix of the fault data.
[0044] Step 4: Build a CNN-BiLSTM recognition network model and input the fault dataset to train the CNN-BiLSTM model. Deeply mine the extracted time-frequency feature vectors from the perspectives of spatial features and forward and reverse time series characteristics to achieve accurate classification of distribution network fault types.
[0045] Step 5: Validate the model using the test dataset.
[0046] The training of CNN-BiLSTM model is based on a large number of data sets. Figure 3 The 10kV distribution network model shown in Figure 2 As shown, the sampling frequency is set to 100kHz. The distribution network structure model in the figure adopts the operation mode of the transformer neutral point not being grounded, the transformer ratio is 110 / 10kV, and the capacity is 80MVA. L1 to L6 represent six feeders, and F11 to F64 are the fault locations of each line. Figure 4 shown.
[0047] Considering the randomness of fault location, type, initial angle, and transition resistance in actual distribution network operation, this paper used this model to traverse all simulation parameters listed in Table 1, generating 15 × 10 × 6 × 4 = 3600 samples. Under each operating condition, four sets of fault data were collected, including three-phase currents on the low-voltage side of the main transformer and busbar zero-sequence voltage. Each set of data consisted of four cycles before the fault and five cycles after the fault.
[0048] Table 1 Simulation parameters
[0049] .
[0050] The present invention is explained by taking the short circuit fault of the two phases AB of the distribution network as an example. Figure 5 As shown in the figure, the key parameters of the improved sparrow optimization algorithm are set as: population size 30 and Cauchy coefficient of variation 1.49. In VMD decomposition, when the number of modes K is set to 7, a large number of redundant modes with flat amplitudes are generated, increasing the difficulty of feature extraction; K=3 leads to under-decomposition and insufficient separation of key features. Experimental verification shows that K=5 ensures modal completeness while avoiding invalid modal interference, achieving optimal decomposition. Similarly, when choosing bandwidth constraints, when α=500, insufficient frequency band constraints can easily lead to modal aliasing, while when α=2000, the problem of over-segmentation of transient features occurs. Through system optimization, α was selected as 1100.
[0051] This paper uses a one-dimensional convolutional neural network architecture consisting of a single convolutional layer and a downsampling layer. The convolutional layer is configured with 64 1×1 convolution kernels and uses the ReLU activation function for nonlinear transformations. The pooling layer uses a 1×1 unit window size and a parameter combination of 1 step size to maintain the temporal integrity of the feature sequence. The output layer uses a 10-neuron fully connected structure to connect with the subsequent BiLSTM module. The BiLSTM network adopts a single-layer architecture with 20 LSTM units in each direction, for a total of 40 hidden units in both directions.
[0052] We randomly separated 30% of the 360 training samples for each type of fault in a specific application instance as test samples to test the CNN-BILSTM. The test results are shown in Table 2.
[0053] Table 2 Test results
[0054] .
[0055] To verify the effectiveness of the OCSSA-VMD-CNN-BILSTM method in fault classification, two different model architectures, VMD-CNN and OCSSA-VMD-CNN, were compared on the same test set. Figure 5As shown in the results, the proposed OCSSA-VMD-CNN-BILSTM method improved fault classification accuracy by 2.47% compared to the OCSSA-VMD-CNN model and by 4.19% compared to the VMD-CNN model. These results demonstrate that the proposed method significantly improves performance over both VMD-CNN and OCSSA-VMD-CNN, demonstrating its superiority in processing.
[0056] Conclusion: The present invention adopts a serial distribution network fault classification and identification network combining CNN and BiLSTM. While extracting the local features of the fault signal, it captures the positive and negative temporal characteristics of the fault signal, comprehensively reflects the characteristics of the fault signal, improves the accuracy of fault classification, and shows good adaptability and robustness.
[0057] Although the present invention has been disclosed above in terms of preferred embodiments, they are not intended to limit the present invention. Anyone skilled in the art can make various changes or modifications without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection defined by the claims of this application.
Claims
1. A distribution network fault classification method based on VMD decomposition and CNN-BiLSTM, characterized by: The following steps are involved: Step 1: Collect the busbar zero-sequence voltage and transformer low-voltage side three-phase current signals, and intercept the four cycles before the fault and the five cycles after the fault as signals for subsequent analysis; Step 2: Apply the sparrow optimization algorithm combined with the Osprey algorithm and Cauchy mutation to optimize the VMD decomposition parameters of various fault signals; Step 3: Perform VMD decomposition on the fault signal and construct the time-frequency feature matrix of the fault data; Step 4: Build a CNN-BiLSTM recognition network model and input the fault dataset to train the CNN-BiLSTM model. The model then conducts in-depth mining of the extracted time-frequency feature vectors from the perspectives of spatial features and forward and reverse time series characteristics to accurately classify distribution network fault types. Step 5: Validate the model using the test dataset.
2. The distribution network fault classification method based on VMD decomposition and CNN-BiLSTM according to claim 1, characterized in that: In step 2, the application of the Osprey algorithm in the global exploration strategy is described by the following formula (1): Where: is the position of the i-th osprey in the first stage; r i,j Is a random number between [1,2]; SF i,j is the position of the fish determined by the i-th osprey; I i,j The value of is 1 or 2.
3. The distribution network fault classification method based on VMD decomposition and CNN-BiLSTM according to claim 1, characterized in that: In step 2, the Cauchy mutation algorithm perturbs the individuals in the sparrow position update, which is described by the following formula (2): x best (t+1)=x best (t)+x best (t)×Cauchy(0,1) (2) Where: is the position of the sparrow after replacement; x best (t) is the current optimal solution; Cauchy(0,1) is the standard Cauchy distribution.
4. The distribution network fault classification method based on VMD decomposition and CNN-BiLSTM according to claim 3 is characterized in that: In order to optimize the performance of CNN, the fully connected layer uses the ReLU function as the activation function of the neuron. For one-dimensional input data, assuming the length is M, the fully connected layer contains N neurons, and the output formula of each neuron is: Where: represents the connection weight between the i-th neuron in the L-1th layer and the j-th neuron in the L-th layer; Represents the input of the i-th neuron in the L-1 layer; is the bias term of the j-th neuron in the L-th layer, In the output layer, after processing by the softmax function, we get: Where: u j Represents the original output value of the jth neuron, which is classified by the softmax function of the output layer to obtain the final classification result.
5. The distribution network fault classification method based on VMD decomposition and CNN-BiLSTM according to claim 1, characterized in that: BiLSTM is an improved version of LSTM. The gating mechanism of LSTM consists of three parts: forget gate, input gate, and output gate. At the same time, a state unit is introduced to control the operation of the entire network. The calculation method of LSTM is as follows: Among them, f t is the forget gate; W f 、U f 、b f is the forget gate parameter; x t represents the input at time t; h t-1 is the hidden layer output at time t-1; i t is the input gate; W i , U i , b i is the input gate parameter; is the input state at time t; tanh is the activation function; C t is the internal state at time t; W o 、U o 、b o is the state parameter; o t is the output gate; W o 、U o 、b o is the output gate parameter; h t is the hidden layer output, BiLSTM contains two independent hidden layers. There is no direct interaction between the hidden layer of forward propagation and the hidden layer of backward propagation. They are regarded as two independently running networks with opposite data flows: Where: is the positive hidden layer state at time t; is the reverse hidden layer state at time t.