A transformer oil dissolved gas content prediction method based on sequence decomposition

Abstract

The application discloses a transformer oil dissolved gas content prediction method based on sequence decomposition, which comprises the following steps: inputting an oil dissolved gas sequence to be predicted, optimizing a core hyperparameter of VMD through a RIME algorithm α and K; decomposing the oil dissolved gas sequence by using the VMD method with the optimal parameter to obtain IMF components and a residual component res; performing normalization processing on each subsequence component; constructing a transformer oil dissolved gas content prediction model based on TimesNet, converting one-dimensional subsequences to two-dimensional space, and obtaining prediction results of each subsequence component; finally, superimposing and reconstructing the prediction values of each subsequence to obtain the prediction results of the dissolved gas content; and inputting other gas to be predicted, the prediction task can be completed by the method. The application reduces the non-stationarity of the sequence by decomposition, models and predicts the sequence from the two-dimensional space, effectively improves the prediction accuracy of the dissolved gas, and provides a reference for the transformer oil dissolved gas content prediction work.

Description

Technical Field

[0001] This invention relates to the field of power equipment, and in particular to a method for predicting the dissolved gas content in transformer oil based on sequence decomposition. Background Technology

[0002] Oil-immersed transformers are crucial infrastructure equipment in power grids, and their safe operation is essential for grid security. Dissolved gas analysis in oil has been widely used in transformer condition assessment and latent fault diagnosis. Therefore, accurately predicting the trend of dissolved gas content changes can provide important reference value for assessing transformer operating conditions and detecting latent faults. Statistical prediction methods mainly include grey models and time series prediction models. However, these methods have poor fitting ability and significant limitations when predicting non-stationary gas content sequences. Machine learning-based prediction methods have good fitting ability for complex non-linear data; representative machine learning models include Extreme Learning Machines and Least Squares Support Vector Machines. However, traditional machine learning prediction models do not adequately consider the correlations between time series data, which limits the accuracy of predictions to some extent. With the development of artificial intelligence technology, deep learning methods have been widely used. Although the prediction accuracy of ordinary long short-term memory networks and bidirectional gated recurrent units has improved to some extent, they only model and predict time series from a one-dimensional perspective, making it difficult to effectively capture the dependencies between different periods in the time series. The dissolved gas content sequence in transformer oil exhibits non-stationarity; using the original sequence directly as input will affect the prediction accuracy. VMD can remove the non-stationarity of the original sequence without generating errors and can effectively suppress mode mixing and endpoint effects. However, subjectively setting the parameters of VMD can affect the decomposition effect. Summary of the Invention

[0003] To address the aforementioned technical challenges and improve the accuracy of predicting dissolved gas content in transformer oil, a sequence decomposition-based method for predicting dissolved gas content in transformer oil is proposed. The method involves inputting the dissolved gas sequence to be predicted, optimizing it using the RIME algorithm to determine the core hyperparameters α and K of VMD (Dynamic Variation of Dissolved Gases). The VMD method is then used to decompose the dissolved gas sequence, obtaining the IMF component and the residual component res. Each subsequence component is then normalized. A TimesNet-based prediction model for dissolved gas content in transformer oil is constructed, transforming the one-dimensional subsequence into a two-dimensional space to obtain the prediction results for each subsequence component. Finally, the predicted values ​​of each subsequence are superimposed and reconstructed to obtain the predicted dissolved gas content. This method can be applied to other gases to be predicted, and the prediction task can be completed by following the same approach. To achieve the above-mentioned technical effects, this invention provides a method for predicting the dissolved gas content in transformer oil based on sequence decomposition, comprising the following steps: S1: Input the sequence of dissolved gases in oil to be predicted, and use the RIME algorithm to optimize and determine the core hyperparameters α and K of VMD. S2, the dissolved gas sequence in the oil is decomposed using the VMD method with optimal parameters to obtain the IMF component and the residual component res. S3, normalize each subsequence component separately. S4. Construct a TimesNet prediction model to transform the one-dimensional subsequence into a two-dimensional space and obtain the prediction results of each subsequence component. S5. The predicted values ​​of each subsequence are superimposed and reconstructed to obtain the predicted result of dissolved gas content. Preferably, in step S1, applying VMD requires pre-setting two core hyperparameters: the penalty factor α and the number of decomposition layers K. Relying solely on subjective human settings cannot achieve the optimal decomposition effect. The RIME algorithm is used to find the optimal parameter combination of α and K, avoiding human intervention. The principle and specific steps of the RIME algorithm for optimizing VMD parameters are as follows: S101, randomly generate a set of frost individuals, each frost individual consists of n frost particles, as shown in the following expression: In the formula: For the i-th frost individual; Let j be the j-th frost particle in the i-th frost individual; S102, drawing inspiration from the significant randomness of soft frost formation under light wind conditions, where frost particles adhere widely and randomly to object surfaces, the algorithm mimics this random and comprehensive coverage capability of frost particles to rapidly explore the entire solution space in the early iteration stages. The expression is as follows: In the formula: The position of the updated particle; Let r1 be the j-th particle of the optimal frost individual in the frost swarm; r1 is a random number belonging to (-1, 1); r1 and cosθ are combined to adjust the search direction of the particle, and θ changes with the number of iterations; h is the adhesion degree, which is used to adjust the distance between two iterated particles. , These correspond to the upper and lower bounds of the particle diffusion space, respectively; β is the environmental factor, which simulates changes in the external environment as the number of iterations increases; r2 is a random number belonging to (0, 1); E is the adhesion coefficient, which increases with the cumulative number of iterations; t is the number of iterations already performed; T is the maximum number of iterations. For rounding; w can adjust the number of segments in the step function; S103, mimicking the growth of frost particles under strong wind conditions, is used to enhance the commutativity of algorithm particles during the optimization phase. Its expression is as follows: In the formula: r1 represents the normalized fitness of the individual frost; r3 is a random number in (-1, 1); S104. From the perspective of minimizing reconstruction error, the envelope entropy is determined as the fitness function. The formula for calculating the envelope entropy of signal x(j) is as follows: In the formula: N is the length of signal x(j); a(j) is the envelope signal of x(j) demodulated by Hilbert; p j This is the normalized form of a(j); S105, Input the dissolved gas sequence in oil to be predicted; Initialize RIME parameters, set the population size and number of iterations, and set the optimization range of VMD hyperparameters α and K; Select envelope entropy as the fitness function, perform VMD decomposition on a given set of α and K, calculate the sum of envelope entropies of each modal component, the smaller the envelope entropy, the higher the decomposition quality; Use the minimum envelope entropy as the optimization objective, use the RIME algorithm to generate new candidate parameter combinations α and K, calculate the fitness value and iteratively optimize; Output the optimal parameter combination. Preferably, in step S2, considering the nonlinear and non-stationary characteristics of the dissolved gas sequence in transformer oil, VMD is used to first stabilize the sequence, ensuring that the statistical characteristics of the sequence do not change over time, thereby improving the accuracy of subsequent predictions. The specific steps of the VMD algorithm are as follows: S201, construct a variational problem with constraints, deconstruct the original sequence into K IMFs, such that the bandwidth sum of each IMF is minimized, and the superposition of all modes equals the original sequence, as expressed below: In the formula: This represents taking the partial derivative with respect to time t; It is the k-th IMF component after decomposition; is the center frequency of the k-th mode; δ(t) is the Dirac function; * is the convolution operator; f(t) is the sequence of dissolved gases in the oil; S202 introduces the Lagrange multiplication operator λ(t) and the penalty factor α to integrate the constraints into the objective function, transforming the original problem into an unconstrained variational problem for solution, as shown in the following expression: S203, by iteratively updating the modal components and center frequency until convergence, obtains the saddle point of the function. The iterative update process is shown below: In the formula: For frequency; , , , These are the Fourier transforms of their respective original values, and n is the number of iterations. Preferably, in step S3, the dissolved gas content in the oil is normalized according to the following expression: In the formula: The data is after normalization; This is the original data; and These represent the maximum and minimum values ​​of the original data, respectively. Preferably, in step S4, for the time series prediction task, there is a close correlation between each input time point and its adjacent time points, and between time points in similar stages of different periods. Considering that the cyclic calculation of traditional one-dimensional models is one-dimensional and sequential, it is impossible to capture the relationships between periods in parallel. TimesNet overcomes the structural limitations of the aforementioned one-dimensional model. A TimesNet prediction model is constructed to extract multiple periods from the time series, reconstructing the one-dimensional time series into multiple two-dimensional tensors to extend it to a two-dimensional space, and predicting the series from a two-dimensional perspective. Its main implementation process is as follows: S401, performing a Fast Fourier Transform on a one-dimensional time series with a time length of T and a channel dimension of C to extract the periodic information of the series, the process is as follows: In the formula: The original one-dimensional time series; FFT(·) is the Fast Fourier Transform; Amp(·) is used to calculate the amplitude value; The amplitude is for each frequency; Avg(·) is the average value for calculating the C-dimensional channels; S402, according to the selected frequency and the corresponding period length One-dimensional time series The process of reconstructing multiple two-dimensional tensors can be formalized as follows: In the formula: This is the time series of the i-th reconstruction; and These represent the rows and columns of a two-dimensional tensor, respectively; Padding(·) extends the time series with zeros along the time dimension, making it consistent with the previous state. Compatible. Preferably, in step S5, the predicted values ​​of each subsequence are obtained through the TimesNet prediction model and then superimposed and reconstructed to obtain the final predicted result of dissolved gas content. The beneficial effects of the method for predicting dissolved gas content in transformer oil based on sequence decomposition and TimesNet model provided by this invention are as follows: 1. This invention uses the RIME algorithm to intelligently optimize the core hyperparameters of VMD, avoiding subjective parameter setting by humans and improving the effect of sequence decomposition; 2. VMD is used to decompose the dissolved gas sequence in oil, breaking down the non-stationary sequence into multiple weakly stationary subsequences, thereby reducing the non-stationarity of the sequence and improving the prediction accuracy of subsequent models. 3. This invention introduces the TimesNet prediction model to transform the one-dimensional sequence into a two-dimensional space, and explores the relationship between the changes in dissolved gas sequence at each time point within the period and during the period, resulting in better prediction accuracy. Attached Figure Description

[0004] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Figure 1 This is a flowchart illustrating a method for predicting dissolved gas content in transformer oil based on sequence decomposition. Figure 2 Schematic diagram of the TimesNet prediction model; Figure 3 The graph shows the VMD decomposition results of the gas sequence. Figure 4 This is a superimposed reconstruction of the prediction results for each subsequence. Detailed Implementation

[0005] To more clearly illustrate the technical solutions of the embodiments of the present invention, the present invention will be further described in conjunction with the accompanying drawings: Example 1: like Figure 1 As shown, this invention provides a method for predicting dissolved gas content in transformer oil based on sequence decomposition, or more specifically, a method for predicting dissolved gas content in transformer oil based on sequence decomposition and the TimesNet model, comprising the following steps: S1: Input the sequence of dissolved gases in oil to be predicted, and use the RIME algorithm to optimize and determine the core hyperparameters α and K of VMD. S2, the dissolved gas sequence in the oil is decomposed using the VMD method with optimal parameters to obtain the IMF component and the residual component res. S3, normalize each subsequence component separately. S4. Construct a TimesNet prediction model to transform the one-dimensional subsequence into a two-dimensional space and obtain the prediction results of each subsequence component. S5. The predicted values ​​of each subsequence are superimposed and reconstructed to obtain the predicted result of dissolved gas content. In step S1, the input dissolved gas data in the oil to be predicted must meet the following conditions: First, the total number of data points should exceed 100 to ensure sufficient data for model analysis; second, the data points should be evenly distributed over time, i.e., have a fixed sampling period, which can be hourly, daily, or other fixed time units. Applying VMD requires pre-setting two core parameters: the penalty factor α and the number of decomposition layers K. Existing technologies cannot achieve optimal decomposition results by relying solely on subjective human settings. To achieve the best decomposition results, this application uses the RIME algorithm to find the optimal parameter combination of α and K, avoiding human intervention. The principle and specific steps of the RIME algorithm for optimizing VMD parameters are as follows: S101, randomly generate a set of frost individuals, each frost individual consists of n frost particles, as shown in the following expression: In the formula: For the i-th frost individual; Let j be the frost particle in the i-th frost individual. S102, drawing inspiration from the significant randomness of soft frost formation under light wind conditions, where frost particles adhere widely and randomly to object surfaces, the algorithm mimics this random and comprehensive coverage capability of frost particles to rapidly explore the entire solution space in the early iteration stages. The expression is as follows: In the formula: The position of the updated particle; Let r1 be the j-th particle of the optimal frost individual in the frost swarm; r1 is a random number belonging to (-1, 1); r1 and cosθ are combined to adjust the search direction of the particle, and θ changes with the number of iterations; h is the adhesion degree, which is used to adjust the distance between two iterated particles. , These correspond to the upper and lower bounds of the particle diffusion space, respectively; β is the environmental factor, which simulates changes in the external environment as the number of iterations increases; r2 is a random number belonging to (0, 1); E is the adhesion coefficient, which increases with the cumulative number of iterations; t is the number of iterations already performed; T is the maximum number of iterations. For rounding; w can adjust the number of segments in the step function. S103, mimicking the growth of frost particles under strong wind conditions, is used to enhance the commutativity of algorithm particles during the optimization phase. Its expression is as follows: In the formula: r is the normalized fitness of the individual frost; r3 is a random number in (-1, 1). S104. From the perspective of minimizing reconstruction error, the envelope entropy is determined as the fitness function. The formula for calculating the envelope entropy of signal x(j) is as follows: In the formula: N is the length of signal x(j); a(j) is the envelope signal of x(j) demodulated by Hilbert; p j Let be the normalized form of a(j). S105, Input the dissolved gas sequence in oil to be predicted; Initialize RIME parameters, set the population size and number of iterations, and set the optimization range of VMD hyperparameters α and K; Select envelope entropy as the fitness function, perform VMD decomposition on a given set of α and K, calculate the sum of envelope entropies of each modal component, the smaller the envelope entropy, the higher the decomposition quality; Use the minimum envelope entropy as the optimization objective, use the RIME algorithm to generate new candidate parameter combinations α and K, calculate the fitness value and iteratively optimize; Output the optimal parameter combination. In step S2, considering the nonlinear and non-stationary characteristics of the dissolved gas sequence in transformer oil, VMD is used to first stabilize the sequence, ensuring that its statistical characteristics do not change over time, thereby improving the accuracy of subsequent predictions. The specific steps of the VMD algorithm are as follows: S201, construct a variational problem with constraints, deconstruct the original sequence into K IMFs, such that the bandwidth sum of each IMF is minimized, and the superposition of all modes equals the original sequence, as expressed below: In the formula: This represents taking the partial derivative with respect to time t; It is the k-th IMF component after decomposition; is the center frequency of the k-th mode; δ(t) is the Dirac function; * is the convolution operator; f(t) is the sequence of dissolved gases in the oil. S202 introduces the Lagrange multiplication operator λ(t) and the penalty factor α to integrate the constraints into the objective function, transforming the original problem into an unconstrained variational problem for solution, as shown in the following expression: S203, by iteratively updating the modal components and center frequency until convergence, obtains the saddle point of the function. The iterative update process is shown below: In the formula: For frequency; , , , These are the Fourier transforms of their respective original values, and n is the number of iterations. In step S3, the dissolved gas content in the oil is normalized according to the following expression: In the formula: The data is after normalization; This is the original data; and These represent the maximum and minimum values ​​of the original data, respectively.

[0006] like Figure 2 As shown, for time series prediction tasks, there are close correlations between each input time point and its adjacent time points, as well as between time points in similar stages of different periods. Considering that the cyclic computation of traditional one-dimensional models is sequential and cannot capture the relationships between periods in parallel, TimesNet overcomes the structural limitations of the aforementioned one-dimensional models. A TimesNet prediction model is constructed to extract multiple periods from the time series, reconstructing the one-dimensional time series into multiple two-dimensional tensors to expand it into a two-dimensional space, and then predicting the series from a two-dimensional perspective. The implementation process is as follows: S401, performing a Fast Fourier Transform on a one-dimensional time series with a time length of T and a channel dimension of C to extract the periodic information of the series, the process is as follows: In the formula: The original one-dimensional time series; FFT(·) is the Fast Fourier Transform; Amp(·) is used to calculate the amplitude value; The amplitude is for each frequency; Avg(·) is the average value for calculating the C-dimensional channels; S402, according to the selected frequency and the corresponding period length One-dimensional time series The process of reconstructing multiple two-dimensional tensors can be formalized as follows: In the formula: This is the time series of the i-th reconstruction; and These represent the rows and columns of a two-dimensional tensor, respectively; Padding(·) extends the time series with zeros along the time dimension, making it consistent with the previous state. Compatible. In step S5, the predicted values ​​of each subsequence are obtained through the TimesNet prediction model and then superimposed and reconstructed to obtain the final predicted result of dissolved gas content. The beneficial effects of the method for predicting dissolved gas content in transformer oil based on sequence decomposition and TimesNet model provided by this invention are as follows: 1. This invention uses the RIME algorithm to intelligently optimize the core hyperparameters of VMD, avoiding subjective parameter setting by humans and improving the effect of sequence decomposition; 2. VMD is used to decompose the dissolved gas sequence in oil, breaking down the non-stationary sequence into multiple weakly stationary subsequences, thereby reducing the non-stationarity of the sequence and improving the prediction accuracy of subsequent models. 3. This invention introduces the TimesNet prediction model to transform the one-dimensional sequence into a two-dimensional space, and explores the relationship between the changes in dissolved gas sequence at each time point within the period and during the period, resulting in better prediction accuracy. Example 2: In step S1, the collected methane (CH4) sequence to be predicted is input, and the optimal core hyperparameters α and K of VMD are determined by the RIME algorithm. The optimization results are shown in Table 1.

[0007] Table 1 Examples of Dissolved Gas Faults in Transformer Oil In step S2, the dissolved gas sequence in the oil is decomposed using the VMD method with optimal parameters to obtain the IMF component and the residual component res. The decomposition results are as follows: Figure 3 As shown.

[0008] In step S3, each subsequence is normalized and divided into a training set and a test set in a 3:1 ratio to verify the advantages of the proposed method.

[0009] In step S4, a TimesNet prediction model is established for each subsequence. The model is trained using the training set divided in step S3, and the prediction results for each subsequence are output.

[0010] In step S5, the predicted values ​​of each subsequence are superimposed and reconstructed to obtain the predicted result of the dissolved gas content, which is then compared with the original sequence. Figure 4 As shown in Table 2, to verify the advantages of the proposed method, VMD-TCN, TimesNet, TCN, and LSTM methods were selected to perform prediction experiments on the collected CH4 sequences. The prediction performance comparison of each method is shown in Table 2.

[0011] Table 2 Comparison of prediction performance of different methods As shown in Table 2, compared with VMD-TCN, TimesNet, TCN, and LSTM methods, the RMSE index of the method in this invention is reduced by 60.49%, 75.13%, 76.42%, and 80.63%, respectively; the MAE index is reduced by 69.55%, 76.87%, 77.78%, and 81.85%, respectively; and the R² is increased by 5.88%, 15.78%, 20.71%, and 30.39%, respectively, showing the best prediction performance.

[0012] The various sections in this manual are described in a progressive manner, with each section focusing on the differences from the others. Similar or identical sections can be referred to each other.

[0013] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for predicting dissolved gas content in transformer oil based on sequence decomposition, characterized in that, include: Step 1: Input the dissolved gas sequence in the oil to be predicted, and use the RIME algorithm to optimize and determine the core hyperparameters α and K of VMD; Step 2: The dissolved gas sequence in the oil is decomposed using the VMD method with optimal parameters to obtain the IMF component and the remaining component res; Step 3: Normalize each subsequence component separately; Step 4: Construct a TimesNet prediction model to transform the one-dimensional subsequence into a two-dimensional space and obtain the prediction results for each subsequence component; Step 5: Superimpose and reconstruct the predicted values ​​of each subsequence to obtain the predicted result of dissolved gas content; In step 1, applying VMD requires pre-setting two core hyperparameters: the penalty factor α and the number of decomposition layers K. The RIME algorithm is used to find the optimal combination of α and K to avoid human intervention. The principle and specific steps of the RIME algorithm for optimizing VMD parameters are as follows: S101, randomly generate a set of frost individuals, each frost individual consists of n frost particles, as shown in the following expression: In the formula: For the i-th frost individual; Let j be the j-th frost particle in the i-th frost individual; S102, drawing inspiration from the significant randomness of soft frost formation under light wind conditions, where frost particles adhere widely and randomly to object surfaces, the algorithm mimics this random and comprehensive coverage capability of frost particles to rapidly explore the entire solution space in the early iteration stages. The expression is as follows: In the formula: The position of the updated particle; Let r1 be the j-th particle of the optimal frost individual in the frost swarm; r1 is a random number belonging to (-1, 1); r1 and cosθ are combined to adjust the search direction of the particle, and θ changes with the number of iterations; h is the adhesion degree, which is used to adjust the distance between two iterated particles. , These correspond to the upper and lower bounds of the particle diffusion space, respectively; β is the environmental factor, which simulates changes in the external environment as the number of iterations increases; r2 is a random number belonging to (0, 1); E is the adhesion coefficient, which increases with the cumulative number of iterations; t is the number of iterations already performed; T is the maximum number of iterations. For rounding; w can adjust the number of segments in the step function; S103, mimicking the growth of frost particles under strong wind conditions, is used to enhance the commutativity of algorithm particles during the optimization phase. Its expression is as follows: In the formula: r1 represents the normalized fitness of the individual frost; r3 is a random number in (-1, 1); S104. From the perspective of minimizing reconstruction error, the envelope entropy is determined as the fitness function. The formula for calculating the envelope entropy of signal x(j) is as follows: In the formula: N is the length of signal x(j); p is the envelope signal of x(j) demodulated by Hilbert; j for The normalized form; S105, Input the dissolved gas sequence in oil to be predicted; Initialize RIME parameters, set the population size and number of iterations, and set the optimization range of VMD hyperparameters α and K; Select envelope entropy as the fitness function, perform VMD decomposition on a given set of α and K, calculate the sum of envelope entropies of each modal component, the smaller the envelope entropy, the higher the decomposition quality; Use the minimum envelope entropy as the optimization objective, use the RIME algorithm to generate new candidate parameter combinations α and K, calculate the fitness value and iteratively optimize; Output the optimal parameter combination.

2. The method for predicting dissolved gas content in transformer oil based on sequence decomposition according to claim 1, characterized in that: In step 2, considering the nonlinear and non-stationary characteristics of the dissolved gas sequence in transformer oil, VMD is used to first stabilize the sequence, ensuring that the statistical characteristics of the sequence do not change over time, thereby improving the accuracy of subsequent predictions. The specific steps of the VMD algorithm are as follows: S201, construct a variational problem with constraints, deconstruct the original sequence into K IMFs, such that the bandwidth sum of each IMF is minimized, and the superposition of all modes equals the original sequence, as expressed below: In the formula: This represents taking the partial derivative with respect to time t; It is the k-th IMF component after decomposition; It is the center frequency of the k-th mode; It is the Dirac function; * is the convolution operator; It is a sequence of dissolved gases in oil; S202 introduces the Lagrange multiplication operator λ(t) and the penalty factor α to integrate the constraints into the objective function, transforming the original problem into an unconstrained variational problem for solution, as shown in the following expression: S203, by iteratively updating the modal components and center frequency until convergence, the saddle point of the function is obtained. The iterative update process is as follows: In the formula: For frequency; , , , These are the Fourier transforms of their respective original values, and n is the number of iterations.

3. The method for predicting dissolved gas content in transformer oil based on sequence decomposition according to claim 1, characterized in that: In step 3, the dissolved gas content in the oil is normalized according to the following expression: In the formula: The data is after normalization; This is the original data; and These represent the maximum and minimum values ​​of the original data, respectively.

4. The method for predicting dissolved gas content in transformer oil based on sequence decomposition according to claim 1, characterized in that: In step 4, a TimesNet prediction model is constructed to extract multiple periods from the time series, and the one-dimensional time series is reconstructed into multiple two-dimensional tensors to expand into a two-dimensional space. The series is then predicted from the perspective of the two-dimensional space. The process is as follows: S401, performing a Fast Fourier Transform on a one-dimensional time series with a time length of T and a channel dimension of C to extract the periodic information of the series, the process is as follows: In the formula: The original one-dimensional time series; FFT(·) is the Fast Fourier Transform; Amp(·) is used to calculate the amplitude value; The amplitude is for each frequency; Avg(·) is the average value for calculating the C-dimensional channels; S402, according to the selected frequency and the corresponding period length One-dimensional time series The process of reconstructing multiple two-dimensional tensors can be formalized as follows: In the formula: This is the time series of the i-th reconstruction; and These represent the rows and columns of a two-dimensional tensor, respectively; Padding(·) extends the time series with zeros along the time dimension, making it consistent with the previous state. Compatible.

5. The method for predicting dissolved gas content in transformer oil based on sequence decomposition according to claim 1, characterized in that: In step 5, the predicted values ​​of each subsequence are obtained through the TimesNet prediction model and then superimposed and reconstructed to obtain the final predicted result of dissolved gas content.

Images