A partial discharge quantity prediction method based on improved GWO optimization VMD and differential deep learning model

Abstract

The application discloses a partial discharge quantity prediction method based on improved GWO optimization VMD and a differential deep learning model, relates to the field of power equipment state monitoring and fault prediction, and comprises the following steps: firstly, collecting tunnel cable partial discharge signals, and adopting a pixel extraction method to construct a partial discharge quantity time sequence; subsequently, performing self-adaptive decomposition on the time sequence by using an improved GWO-VMD algorithm, performing global search on VMD parameters by using an improved grey wolf optimization algorithm, and obtaining optimal decomposition parameters; on the basis, dividing the modal components into low-frequency trend components and high-frequency complex components according to the frequency characteristics of the modal components, and modeling by using differential prediction models for different frequency characteristics; and finally, superimposing and reconstructing the prediction results of the modal components to obtain a partial discharge quantity prediction result. The application can effectively improve the feature extraction capability and prediction accuracy of the partial discharge signal, enhance the stability and robustness of the model, and has good engineering application value.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring and fault prediction, specifically to a method for predicting partial discharge based on an improved GWO-optimized VMD and a differentiated deep learning model. Background Technology

[0002] Partial discharge is an important characteristic parameter reflecting the insulation state of power equipment. Accurate prediction of it is crucial for ensuring the safe operation of power equipment. With the development of online monitoring technology, methods based on time series analysis and machine learning are gradually being applied to the prediction of partial discharge signals. However, existing methods still have the following significant technical problems: First, traditional partial discharge signal processing methods often directly utilize raw monitoring data or simple statistical features to construct time series, which are easily affected by environmental noise and random interference, resulting in large data fluctuations and making it difficult to accurately reflect the true development trend of partial discharge, thus affecting the stability and accuracy of subsequent prediction models. Secondly, in terms of signal decomposition, existing methods usually adopt variational mode decomposition (VMD) or empirical mode decomposition (EMD) methods with empirically set parameters. The parameter selection lacks adaptability, which can easily lead to mode aliasing or insufficient decomposition, making it difficult to accurately extract feature information of different frequency components, thus affecting the feature representation ability. Furthermore, in the process of modal component processing, most methods adopt a unified modeling approach for each modal component, failing to fully consider the differences in physical and statistical properties of different frequency components. This makes it difficult for the model to simultaneously take into account both low-frequency trend information and high-frequency complex fluctuation information, thus limiting the prediction accuracy. In addition, existing deep learning prediction models mostly adopt a single structure and lack the ability to extract complex nonlinear features at multiple levels. Especially when processing high-frequency complex signals, they are difficult to effectively capture local mutation features and key time series information, thus affecting the model's ability to characterize complex dynamic processes. In summary, existing technologies still have shortcomings in partial discharge signal processing, adaptive parameter decomposition, modal component classification, and differential modeling. There is an urgent need to propose a partial discharge quantity prediction method that can improve signal processing quality, enhance feature representation capabilities, and improve prediction accuracy and stability. Summary of the Invention

[0003] This invention addresses the shortcomings of existing technologies by proposing a partial discharge quantity prediction method based on an improved GWO-optimized VMD and a differentiated deep learning model, comprising the following steps: Step 1: Acquire partial discharge signals of tunnel cables using HFCT sensors, and process the partial discharge signals using pixel extraction method to obtain a partial discharge quantity sequence. ; Step 2: Sequence of partial discharge quantities As the input signal, an improved GWO-VMD (Grey Wolf Optimization Algorithm and Variational Mode Decomposition Algorithm) optimization algorithm and decomposition processing are used to process the partial discharge quantity sequence. The initial comprehensive fitness value is obtained by decomposing the VMD parameters. An improved GWO optimization algorithm is then used for global search and iteration to find the minimum comprehensive fitness value, thus obtaining the optimal decomposition parameters (k, ); Step 3: Use the optimal decomposition parameters (k, Partial discharge quantity sequence VMD decomposition yields k modal components IMF1, IMF2, ..., IMF k ; Step 4: Using the frequency feature selection method, the k modal components are divided into low-frequency trend components and high-frequency complex components; Step 5: Use different prediction models to predict different categories of trend components. Low-frequency trend components are predicted using the LSTM model, while high-frequency complex components are predicted using a combination of CNN-LSTM-Attention model. Step 6: Superimpose and reconstruct the prediction results of each modal component to obtain the prediction result of partial discharge.

[0004] In step 1, the partial discharge signal of the tunnel cable is processed using a pixel extraction method to obtain a partial discharge quantity sequence. This includes the following steps: The partial discharge monitoring system records data at 3-minute intervals, totaling 1700 time windows. Within each time window, 300 partial discharge data points are collected, each including the discharge magnitude and corresponding phase information. Within each time window, the partial discharge data with the largest discharge magnitude is first identified, and its phase value is recorded. Using this phase value as the center, partial discharge data within a 30° range before and after that phase are extracted to form a candidate dataset. The discharge magnitude data in the candidate dataset are then sorted in descending order of numerical value, and the top 100 discharge magnitude data points are selected for aggregation calculation. This process is repeated over 1700 consecutive time windows to construct the partial discharge magnitude time series f(t).

[0005] Step 2 involves using an improved GWO-VMD (Grey Wolf Optimization Algorithm and Variational Mode Decomposition Algorithm) optimization algorithm and decomposition process, including the following steps: Step 21: Initialize the relevant parameters of the improved GWO-VMD algorithm, including setting the gray wolf population size N=40, the maximum number of iterations L=15, and the VMD parameter search range, where the data range of the number of modes k is [3,500], and the data range of the quadratic penalty factor α is [10,3000]. A chaotic mapping method is used to initialize the gray wolf population to improve the uniformity of the initial population distribution in the search space. Preferably, a Logistic chaotic mapping is used to generate a chaotic sequence, the expression of which is:

[0006] In the formula: is a chaotic variable with an optimal value of 0.6, and μ is a control parameter with an optimal value of 3.99; Map the generated chaotic sequence to the VMD parameter space:

[0007] In the formula: and These represent the lower and upper bounds of the parameters, respectively, thus obtaining the initial position of the individual gray wolf; Step 22: After initialization, perform VMD decomposition on the partial discharge time series f(t) based on the parameters (k, α). The specific algorithm is described below: Partial discharge time series The modal components are decomposed into a set of modal components with different bandwidths, each with a different center frequency, thereby constructing the corresponding VMD-constrained variational model, which is expressed as:

[0008] In the formula: f(t) represents the time series of partial discharge. This represents the k-th modal component. Indicates the modal center frequency; To solve the VMD-constrained variational problem described above, we introduce Lagrange multipliers λ and a penalty term α, transforming it into an unconstrained variational problem, whose Lagrangian function is expressed as:

[0009] The k modal components under different bandwidths are obtained by iterative solution; Step 23: Using the VMD parameters (k, α) as individual gray wolves in the GWO algorithm, calculate the fitness value of each individual gray wolf; The fitness function is a dynamic weighted fitness function, the expression of which is:

[0010] In the formula, This represents the normalized average permutation entropy. Indicates the normalized reconstruction error. For the number of iterations Dynamically changing weighting coefficients; By using a dynamic weight adjustment strategy, reconstruction error weights are added in the early stage of the algorithm to ensure signal decomposition accuracy, and permutation entropy weights are added in the later stage of the algorithm to improve the mode decomposition quality. Step 24: Determine whether the fitness of the current gray wolf individual meets the update conditions; If the current gray wolf meets the fitness update conditions, then update the individual's fitness; otherwise, retain the current gray wolf's fitness.

[0011] Step 25: Update the location of individual gray wolves; The positions of the remaining gray wolves are updated based on the three optimal individuals (α, β, and δ) in the gray wolf optimization algorithm, thereby obtaining a new combination of parameters. Meanwhile, a nonlinear convergence factor strategy is used to update the parameter α, thereby improving the algorithm's search capability.

[0012] in, This indicates the current iteration number. Indicates the maximum number of iterations; This nonlinear decreasing strategy can enhance the global search capability in the early stage of the algorithm and improve the local search accuracy in the later stage of the algorithm. Step 26: Introduce a perturbation mechanism to prevent the algorithm from getting trapped in local optima; When the optimal fitness value does not change significantly in several consecutive iterations, a random perturbation is introduced into the current optimal gray wolf individual, and the update method is as follows:

[0013] in, Indicates the disturbance amplitude parameter. () represents a standard normal random variable; Step 27: Determine whether the current iteration count is less than the maximum iteration count.

[0014] like < , then let = +1, continue with steps 22 to 26, if ≥ Then the iteration process ends; Finally, the parameters (k, α) corresponding to the current optimal gray wolf individual position are output, which are the optimal VMD decomposition parameters of the partial discharge time series f(t), thus completing the parameter optimization of the improved GWO-VMD algorithm. The optimal VMD decomposition parameters are used to decompose the partial discharge time series f(t).

[0015] In step 4, the frequency feature selection method is used to divide the k modal components into low-frequency trend components and high-frequency trend components, including the following steps: Step 41: Perform a Fourier transform on each modal component to obtain the corresponding spectrum. And calculate the center frequency of the i-th modal component based on the spectrum, the expression of which is:

[0016] in This represents the center frequency of the i-th modal component. Let f represent the spectrum of the i-th modal component, and let f represent the frequency variable. Step 42: Calculate the center frequency of each mode to obtain the set of center frequencies of all modal components: { , , ,

[0017] Step 43: Construct the center frequency classification threshold Preferably, the mean center frequency is used as the classification threshold, and its expression is:

[0018] in: The center frequency classification threshold is represented by k, which represents the number of modal components. Step 44: Compare the center frequency of each modal component with the center frequency classification threshold. When the center frequency is less than or equal to the classification threshold, the modal component is determined to be a low-frequency trend component. When the center frequency is greater than the classification threshold, the modal component is determined to be a high-frequency complex component.

[0019] Step 5 involves using different prediction models to predict different categories of trend components, including the following steps: The modal components are classified according to their frequency characteristics, and different prediction models are used for different categories of modal components. The low-frequency trend components are modeled using a prediction model suitable for stationary sequences, while the high-frequency complex components are modeled using a deep learning ensemble model that can extract nonlinear features. Finally, the prediction results of each modal component are superimposed and reconstructed to obtain the final prediction result.

[0020] In step 5, the high-frequency trend component is predicted using a CNN-LSTM-Attention combined model, including the following steps: Step 51: Set the number of input channels of the one-dimensional convolutional layer to 1, the number of output channels to 32, the kernel size to 3, the stride to 1, use padding to keep the sequence length unchanged, use the ReLU activation function to achieve non-linear mapping, and use a max pooling layer to downsample the features with a kernel size of 2. Step 52: Set up a three-layer long short-term memory layer structure with 64 hidden units in each layer and 32 input feature dimensions. Introduce a dropout mechanism with a dropout rate of 0.2 in the multi-layer structure to prevent overfitting. Step 53: The multi-head attention layer adopts a multi-head self-attention mechanism, with an embedding dimension of 64 and an attention head of 4. By weighting the temporal features output by the long short-term memory layer, the expressive power of the key time step features is enhanced. Step 54: Finally, the features output by the attention layer are mapped to one-dimensional predicted values ​​through a fully connected output layer. The input dimension is 64 and the output dimension is 1, thereby realizing the prediction of the partial discharge quantity time series.

[0021] Compared with the prior art, the present invention has the following beneficial effects: (1) The partial discharge quantity sequence construction method based on pixel extraction improves the effectiveness and stability of the original signal; In existing technologies, raw partial discharge signals or simple statistical features are often used as input, which are easily affected by noise and result in large data fluctuations. This invention processes partial discharge signals using a pixel extraction method, selects the phase corresponding to the maximum discharge amount within each time window, and extracts partial discharge data within a certain phase range before and after it. By sorting, filtering, and aggregating the data, a time series of partial discharge amounts is constructed. This method can effectively highlight the main characteristics of partial discharge activity, reduce the influence of random noise, and improve the stability and representativeness of the time series, thereby providing a higher quality data foundation for subsequent modeling.

[0022] (2) The improved GWO-VMD algorithm achieves adaptive parameter optimization, thereby improving the accuracy of signal decomposition; Existing VMD methods typically rely on empirical parameter selection, which can easily lead to unstable decomposition results or poor decomposition performance. This invention introduces the Grey Wolf Optimization Algorithm to optimize VMD parameters, and further introduces a chaotic initialization strategy, a dynamic weight fitness function, a nonlinear convergence factor, and a perturbation mechanism to enhance the algorithm's global search and local optimization capabilities. Specifically, chaotic initialization improves the uniformity of the initial population distribution, the dynamic weight strategy achieves a balance between decomposition accuracy and quality, the nonlinear convergence factor improves the algorithm's convergence efficiency, and the perturbation mechanism effectively avoids getting trapped in local optima. Through these improvements, better VMD decomposition parameters can be obtained, thereby significantly improving the accuracy and stability of mode decomposition.

[0023] (3) A frequency feature selection method based on center frequency is used to achieve adaptive classification of IMF components; In existing technologies, the classification of IMF components largely relies on manual experience or fixed rules, lacking adaptability. This invention calculates the center frequency of each IMF component and uses the average center frequency as a classification threshold to automatically classify the IMF components into low-frequency trend components and high-frequency complex components. This method can objectively reflect the frequency characteristics of each modal component, achieve data-driven adaptive classification, improve the rationality and stability of component classification, and provide an effective basis for subsequent differentiated modeling.

[0024] (4) Improve the prediction accuracy of high-frequency complex components based on the combined CNN-LSTM-Attention model; To address the issues of strong fluctuations and significant nonlinear characteristics in high-frequency complex components, this invention constructs a CNN-LSTM-Attention combined model for prediction. In this model, one-dimensional convolutional layers are used to extract local temporal features, a long short-term memory network is used to model temporal dependencies, and a multi-head attention mechanism is used to enhance information at key time steps, thereby improving the model's ability to express complex dynamic features. Compared to traditional single prediction models, this combined model can more fully mine deep features in high-frequency signals, improving prediction accuracy and model robustness.

[0025] (5) Differentiated modeling strategies improve overall prediction performance; This invention employs different prediction models for IMF components with different frequency characteristics. The low-frequency trend component is predicted using an LSTM model, while the high-frequency complex component is predicted using a CNN-LSTM-Attention combined model. This differentiated modeling strategy fully considers the characteristics of signals in different frequency bands, enabling the model to learn in a targeted manner. This avoids the problem that a unified model cannot take into account different features, and ultimately significantly improves the overall prediction accuracy. Attached Figure Description

[0026] Figure 1This is a flowchart illustrating the partial discharge quantity prediction method in this embodiment; Figure 2 This is a schematic diagram of the improved GWO-VMD algorithm in this embodiment; Figure 3 This is a schematic diagram of the differentiated deep learning model in this embodiment; Figure 4 This is a schematic diagram of the VMD decomposition of the optimal parameters in this embodiment; Figure 5 This is a schematic diagram comparing the prediction results with the original data in this embodiment. Detailed Implementation

[0027] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Example 1

[0028] like Figure 1-5 As shown, this invention discloses a method for predicting partial discharge, comprising the following steps: Step 1: Acquire partial discharge signals of tunnel cables using HFCT sensors, and process the partial discharge signals using pixel extraction method to obtain a partial discharge quantity sequence. This method can effectively extract the main discharge features in partial discharge signals, reduce noise interference, and improve the stability and representativeness of time series. Step 2: Sequence of partial discharge quantities As the input signal, an improved GWO-VMD (Grey Wolf Optimization Algorithm and Variational Mode Decomposition Algorithm) optimization algorithm and decomposition processing are adopted; the partial discharge quantity sequence is then processed. The initial comprehensive fitness value is obtained by decomposing the VMD parameters. An improved GWO optimization algorithm is then used for global search and iteration to find the minimum comprehensive fitness value, thus obtaining the optimal decomposition parameters (k, This improves the adaptability and decomposition accuracy of VMD parameter selection.

[0029] Step 3: Use the optimal decomposition parameters (k, Partial discharge quantity sequence VMD decomposition yields k modal components IMF1, IMF2, ..., IMF k This enables effective separation of different frequency characteristics of partial discharge signals; Step 4: Using the frequency feature selection method, the k modal components are divided into low-frequency trend components and high-frequency complex components, so as to achieve adaptive classification of signals with different frequency characteristics and improve the rationality of component division. Step 5: Use different prediction models to predict different categories of trend components. Low-frequency trend components are predicted using the LSTM model, while high-frequency complex components are predicted using a CNN-LSTM-Attention combined model. This differentiated modeling strategy is adopted for different frequency characteristics to improve the overall prediction accuracy.

[0030] Step 6: Superimpose and reconstruct the prediction results of each mode component to obtain the partial discharge quantity prediction result. The above methods can effectively improve the prediction accuracy of partial discharge signals.

[0031] In step 1, the partial discharge signal of the tunnel cable is processed using a pixel extraction method to obtain a partial discharge quantity sequence. This includes the following steps: The partial discharge monitoring system records data at 3-minute intervals, totaling 1700 time windows. Within each time window, 300 partial discharge data points are collected, each including the discharge magnitude and corresponding phase information. Within each time window, the partial discharge data with the largest discharge magnitude is first identified, and its phase value is recorded. Using this phase value as the center, partial discharge data within a 30° range before and after that phase are extracted to form a candidate data set. The discharge magnitude data in the candidate data set are then sorted in descending order of numerical value, and the top 100 discharge magnitude data points are selected for aggregation calculation. This process is repeated over 1700 consecutive time windows to construct the partial discharge magnitude time series f(t). This pixel extraction method effectively highlights the key features of partial discharge activity, suppresses random noise interference, and improves data quality and subsequent modeling results.

[0032] Step 2 involves using an improved GWO-VMD (Grey Wolf Optimization Algorithm and Variational Mode Decomposition Algorithm) optimization algorithm and decomposition process, including the following steps: Step 21: Initialize the relevant parameters of the improved GWO-VMD algorithm, including setting the gray wolf population size N=40, the maximum number of iterations L=15, and the VMD parameter search range. The data range for the number of modes k is [3, 500], and the data range for the secondary penalty factor α is [10, 3000]. A chaotic mapping method is used to initialize the gray wolf population, improving the uniformity of the initial population distribution in the search space and enhancing the global search capability. Preferably, a Logistic chaotic mapping is used to generate a chaotic sequence, the expression of which is:

[0033] In the formula: is a chaotic variable with an optimal value of 0.6, and μ is a control parameter with an optimal value of 3.99; Map the generated chaotic sequence to the VMD parameter space:

[0034] In the formula: and These represent the lower and upper bounds of the parameters, respectively, thus obtaining the initial position of the individual gray wolf; Step 22: After initialization, perform VMD decomposition on the partial discharge time series f(t) based on the parameters (k, α). The specific algorithm is described below: The partial discharge time series f(t) is decomposed into a set of modal components with different bandwidths, each with a different center frequency, thereby constructing a corresponding VMD-constrained variational model, which is expressed as:

[0035] In the formula: f(t) represents the time series of partial discharge. This represents the k-th modal component. Indicates the modal center frequency; To solve the VMD-constrained variational problem described above, we introduce Lagrange multipliers λ and a penalty term α, transforming it into an unconstrained variational problem, whose Lagrangian function is expressed as:

[0036] The k modal components under different bandwidths are obtained by iterative solution; Step 23: Using the VMD parameters (k, α) as individual gray wolves in the GWO algorithm, calculate the fitness value of each individual gray wolf; The fitness function is a dynamic weighted fitness function, the expression of which is:

[0037] In the formula, Represents the normalized average permutation entropy. Indicates the normalized reconstruction error. For the number of iterations Dynamically changing weighting coefficients; By using a dynamic weight adjustment strategy, reconstruction error weights are added in the early stage of the algorithm to ensure signal decomposition accuracy, and permutation entropy weights are added in the later stage of the algorithm to improve the mode decomposition quality. Step 24: Determine whether the fitness of the current gray wolf individual meets the update conditions; If the current gray wolf meets the fitness update conditions, then update the individual's fitness; otherwise, retain the current gray wolf's fitness.

[0038] Step 25: Update the location of individual gray wolves; The positions of the remaining gray wolves are updated based on the three optimal individuals (α, β, and δ) in the gray wolf optimization algorithm, thereby obtaining a new combination of parameters. Meanwhile, a nonlinear convergence factor strategy is used to update the parameter α, thereby improving the algorithm's search capability.

[0039] in, This indicates the current iteration number. Indicates the maximum number of iterations; This nonlinear decreasing strategy can enhance the global search capability in the early stages of the algorithm and improve the local search accuracy in the later stages.

[0040] Step 26: Introduce a perturbation mechanism to prevent the algorithm from getting trapped in local optima.

[0041] When the optimal fitness value does not change significantly in several consecutive iterations, a random perturbation is introduced into the current best gray wolf individual to effectively avoid the algorithm getting trapped in local optima. The update method is as follows:

[0042] in, Indicates the disturbance amplitude parameter. () represents a standard normal random variable; Step 27: Determine whether the current iteration count is less than the maximum iteration count.

[0043] like < , then let = +1, continue with steps 22 to 26, if ≥ Then the iteration process ends; Finally, the parameters (k, α) corresponding to the current optimal gray wolf individual position are output, which are the optimal VMD decomposition parameters of the partial discharge time series f(t), thus completing the parameter optimization of the improved GWO-VMD algorithm. The optimal VMD decomposition parameters are used to decompose the partial discharge time series f(t).

[0044] In step 4, the frequency feature selection method is used to divide the k modal components into low-frequency trend components and high-frequency complex components, including the following steps: Step 41: Perform a Fourier transform on each modal component to obtain the corresponding spectrum. And calculate the center frequency of the i-th modal component based on the spectrum, the expression of which is:

[0045] in This represents the center frequency of the i-th modal component. Let f represent the spectrum of the i-th modal component, and let f represent the frequency variable. Step 42: Calculate the center frequency of each mode to obtain the set of center frequencies of all modal components: { , , ,

[0046] Step 43: Construct the center frequency classification threshold Preferably, the mean center frequency is used as the classification threshold, and its expression is:

[0047] in: The center frequency classification threshold is represented by k, which represents the number of modal components. Step 44: Compare the center frequency of each modal component with the center frequency classification threshold. When the center frequency is less than or equal to the classification threshold, the modal component is determined to be a low-frequency trend component. When the center frequency is greater than the classification threshold, the modal component is determined to be a high-frequency complex component. Compared with the traditional method that relies on experience for classification, this method can achieve adaptive classification based on the characteristics of the data itself, thereby improving the objectivity and stability of component classification.

[0048] Step 5 involves using different prediction models to predict different categories of trend components, including the following steps: Modal components are classified based on their frequency characteristics, and different prediction models are used for different categories of modal components. Low-frequency trend components are modeled using a prediction model suitable for stationary sequences, while high-frequency complex components are modeled using a deep learning ensemble model capable of extracting nonlinear features. Finally, the prediction results of each modal component are superimposed and reconstructed to obtain the final prediction result. In step 5, the high-frequency trend component is predicted using a CNN-LSTM-Attention ensemble model, which includes the following steps: Step 51: Set the number of input channels of the one-dimensional convolutional layer to 1, the number of output channels to 32, the kernel size to 3, the stride to 1, use padding to keep the sequence length unchanged, use the ReLU activation function to achieve non-linear mapping, and use a max pooling layer to downsample the features with a kernel size of 2. Step 52: Set up a three-layer long short-term memory layer structure with 64 hidden units in each layer and 32 input feature dimensions. Introduce a dropout mechanism with a dropout rate of 0.2 in the multi-layer structure to prevent overfitting. Step 53: The multi-head attention layer adopts a multi-head self-attention mechanism, with an embedding dimension of 64 and an attention head of 4. By weighting the temporal features output by the long short-term memory layer, the expressive power of the key time step features is enhanced. Step 54: Finally, the features output by the attention layer are mapped to one-dimensional predicted values ​​through a fully connected output layer. The input dimension is 64 and the output dimension is 1, thereby realizing the prediction of the partial discharge time series. The model can more fully explore the nonlinear features in high-frequency complex signals, thereby improving the prediction accuracy and model robustness. This invention discloses a partial discharge quantity prediction method based on an improved GWO-optimized VMD and differentiated deep learning model. Partial discharge signals from tunnel cables are acquired using an HFCT sensor, and the original signals are processed using a pixel extraction method to construct a partial discharge quantity time series. This time series is then used as the input signal, and the improved GWO-VMD algorithm is used for parameter optimization and mode decomposition. The VMD parameters are then globally searched and iterated using the Grey Wolf optimization algorithm to obtain the optimal decomposition parameters, thereby achieving adaptive signal decomposition. Based on this, the partial discharge time series is decomposed using optimal parameters to obtain multiple modal components with different frequency characteristics. Furthermore, each modal component is classified based on a frequency feature selection method, dividing them into low-frequency trend components and high-frequency complex components. Differential prediction models are constructed for different frequency characteristics. Long Short-Term Memory (LSTM) networks are used to predict the low-frequency trend components, while a combined model of convolutional neural networks, LTM networks, and attention mechanisms is used to predict the high-frequency complex components, thereby enhancing the model's ability to extract nonlinear features and key time-series information. Finally, the prediction results of each modal component are superimposed and reconstructed to obtain the final prediction result of the partial discharge.

[0049] This invention integrates multiple technologies to construct a complete process from data processing, signal decomposition, feature classification to prediction and reconstruction, effectively reducing noise interference, improving signal decomposition accuracy, achieving adaptive modeling of different frequency features, significantly improving the accuracy and stability of partial discharge quantity prediction, and has good engineering application value.

[0050] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principle of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for predicting partial discharge based on an improved GWO-optimized VMD and a differentiated deep learning model, characterized in that, Includes the following steps: Step 1: Acquire partial discharge signals from tunnel cables, process the partial discharge signals using pixel extraction method, and obtain the partial discharge quantity sequence. ; Step 2: Sequence of partial discharge quantities As the input signal, an improved GWO-VMD optimization algorithm and decomposition processing are used to process the partial discharge quantity sequence. The initial comprehensive fitness value is obtained by decomposing the VMD parameters. An improved GWO optimization algorithm is then used for global search and iteration to find the minimum comprehensive fitness value, thus obtaining the optimal decomposition parameters (k, ); Step 3: Use the optimal decomposition parameters (k, Partial discharge quantity sequence VMD decomposition yields k modal components IMF1, IMF2, ..., IMF k ; Step 4: Using the frequency feature selection method, the k modal components are divided into low-frequency trend components and high-frequency complex components; Step 5: Use different prediction models to predict different categories of trend components. Low-frequency trend components are predicted using the LSTM model, while high-frequency complex components are predicted using a combination of CNN-LSTM-Attention model. Step 6: Superimpose and reconstruct the prediction results of each modal component to obtain the prediction result of partial discharge.

2. The partial discharge quantity prediction method based on improved GWO optimized VMD and differentiated deep learning model as described in claim 1, characterized in that, In step 1, the partial discharge signal of the tunnel cable is processed using a pixel extraction method, which includes the following steps: The partial discharge monitoring system records data at time intervals of K, totaling M time windows. Within each time interval, N partial discharge data points are collected. Each data point includes the discharge quantity and corresponding phase information. Within each time window, the partial discharge data with the largest discharge quantity is first determined, and the phase value of the discharge data is recorded. Using this phase value as the center, partial discharge data within a range of p degrees before and after this phase are extracted to form a candidate data set. The discharge quantity data in the candidate data set are sorted in descending order of numerical value, and the top X discharge quantity data points are selected for aggregation calculation. The above processing steps are repeated within M consecutive time windows to construct the partial discharge quantity time series f(t).

3. The partial discharge prediction method based on an improved GWO-optimized VMD and differentiated deep learning model as described in claim 1, characterized in that, Step 2 employs an improved GWO-VMD optimization algorithm and decomposition process, including the following steps: The gray wolf population is initialized by setting the population size, maximum number of iterations, and VMD parameter search range, and the individual gray wolves are initialized using a chaotic mapping method. The partial discharge quantity time series is decomposed into multiple modal components with different center frequencies based on the parameters obtained from initialization using VMD. Using VMD parameters as individual gray wolves, a fitness function is constructed and the fitness value of each individual gray wolf is calculated. The fitness function is a dynamic weighted comprehensive fitness function based on normalized average permutation entropy and normalized reconstruction error, wherein the weight coefficients are dynamically adjusted with the iteration process. The individual positions of gray wolves are updated based on the gray wolf optimization algorithm, and a nonlinear convergence factor is introduced to enhance the search capability of the algorithm at different stages. When the optimal fitness value remains unchanged within several iterations during the optimization process, a perturbation mechanism is introduced to randomly perturb the current best gray wolf individual in order to avoid the algorithm getting trapped in a local optimum. Repeat the above process until the maximum number of iterations is reached, and output the parameters corresponding to the optimal gray wolf individual as the optimal VMD decomposition parameters for the partial discharge time series.

4. The partial discharge prediction method based on an improved GWO-optimized VMD and differentiated deep learning model as described in claim 1, characterized in that, In step 4, the frequency feature selection method is used to divide the k modal components into low-frequency trend components and high-frequency trend components, including the following steps: Perform a Fourier transform on each modal component to obtain the corresponding spectrum. And calculate the center frequency of the i-th modal component based on the spectrum, the expression of which is: ； in This represents the center frequency of the i-th modal component. Let f represent the spectrum of the i-th modal component, and let f represent the frequency variable. Calculate the center frequency of each mode to obtain the set of center frequencies of all modal components: { , , , ； Construct center frequency classification threshold The mean center frequency is used as the classification threshold, and its expression is: ； in: The center frequency classification threshold is represented by k, which represents the number of modal components. The center frequency of each modal component is compared with the center frequency classification threshold. When the center frequency is less than or equal to the classification threshold, the modal component is determined to be a low-frequency trend component. When the center frequency is greater than the classification threshold, the modal component is determined to be a high-frequency complex component.

5. The partial discharge quantity prediction method based on improved GWO optimized VMD and differentiated deep learning model as described in claim 1, characterized in that, Step 5 involves using different prediction models to predict different categories of trend components, including the following steps: The modal components are classified according to their frequency characteristics, and different prediction models are used for different categories of modal components. The low-frequency trend components are modeled using a prediction model suitable for stationary sequences, while the high-frequency complex components are modeled using a deep learning ensemble model that can extract nonlinear features. Finally, the prediction results of each modal component are superimposed and reconstructed to obtain the final prediction result.

6. The partial discharge prediction method based on improved GWO optimized VMD and differentiated deep learning model as described in claim 5, characterized in that, in step 5, a CNN-LSTM-Attention combined model is used to predict high-frequency trend components, which includes the following steps: A one-dimensional convolutional layer is used to extract features from the input partial discharge time series. Local temporal features of the sequence are extracted through convolution operations, and nonlinear mapping is performed through activation functions. At the same time, pooling operations are used to downsample the features to reduce the feature dimensionality and improve the feature expressive power. The features output from the convolutional layer are input into a long short-term memory network. The sequence features are modeled through a multi-layer long short-term memory structure to extract long-term dependencies in the time series. The generalization ability of the model is improved by introducing a dropout mechanism. The output of the Long Short-Term Memory network is input into a multi-head attention layer. The features at different time steps are weighted through a self-attention mechanism to enhance the expressive power of the features at key time steps. The features output by the attention layer are then input into a fully connected layer for mapping to obtain the corresponding prediction results, thereby enabling the prediction of high-frequency complex components.

Images