Method for predicting concentration of dissolved gas in transformer oil based on mixed decomposition and BiTCN-BiGRU

Abstract

The invention provides a method for predicting the concentration of gas dissolved in transformer oil based on mixed decomposition and Bi TCN-Bi GRU, and is applied to the field of prediction of the concentration of gas dissolved in oil. The method comprises the following steps: firstly, carrying out preliminary decomposition on a gas concentration sequence by utilizing improved self-adaptive noise complete set empirical mode decomposition; secondly, reconstructing a sub-sequence in combination with a sample entropy, and performing secondary decomposition on a high-entropy sub-sequence by using variational mode decomposition; and finally, constructing a Bi TCN-B GRU prediction model, simultaneously capturing forward and reverse features of time series data by using bidirectional convolution and a bidirectional gating mechanism, predicting subsequences of primary decomposition and secondary decomposition, and reconstructing a gas concentration sequence. Compared with the prior art, the method is higher in prediction precision and better in robustness and generalization.

Description

Technical Field

[0001] The present invention relates to the field of predicting the concentration of dissolved gases in oil, and particularly to a method for predicting the concentration of dissolved gases in transformer oil based on hybrid decomposition. Background Art

[0002] In recent years, with the rapid development of the power system and the increasing importance of transformers as key equipment, the fault diagnosis and preventive maintenance of transformers have become the core issues for ensuring the safety and stability of the power system. Transformer faults often lead to system outages, seriously affecting power supply. Currently, dissolved gas analysis (DGA) in insulating oil is the most widely used method in transformer fault diagnosis, mainly including gases such as hydrogen, methane, ethane, acetylene, and ethylene. When a fault or abnormality occurs inside the transformer, the content of some components in these gases will increase rapidly. Therefore, developing a method for predicting the concentration of dissolved gases based on early data can provide strong support for transformer preventive maintenance and has important practical significance. Researchers mainly divide the prediction methods for the concentration of dissolved gases in transformer oil into three categories: data statistical methods, artificial intelligence methods, and combined prediction methods. Although data statistical methods have achieved prediction through improving grey theory, autoregressive integrated moving average, etc., their single fitting function and simplified model structure are difficult to accurately reflect the complex characteristics of time series and have certain limitations. With the development of artificial intelligence, more and more intelligent methods have been applied to the prediction of dissolved gases in transformer oil. Methods such as support vector regression models, random forests, and backpropagation neural networks have been widely used. However, when dealing with time series data, it is difficult for the above methods to deeply explore the complex time series relationships between data. Then researchers proposed recurrent neural networks to capture the time dependence and sequence features in data, improving the accuracy of time series prediction. However, recurrent neural networks are prone to the problem of gradient disappearance when capturing long-term dependencies. The combined prediction model usually decomposes the time series of dissolved gases in oil using various algorithms, and then inputs each subsequence into the prediction model for prediction. The empirical mode decomposition method is used for prediction, but when dealing with complex signals, modal aliasing may occur; the complete ensemble empirical mode decomposition with adaptive noise adds paired white noise to suppress modal aliasing to a certain extent but is prone to pseudo-modalities; variational mode decomposition is used for prediction, effectively solving the problem of modal aliasing and having a small reconstruction error, but variational mode decomposition is sensitive to noise and is not conducive to processing over-noisy data; the improved complete ensemble empirical mode decomposition with adaptive noise effectively suppresses modal aliasing and end effects, improving the prediction accuracy, but the prediction effect of high-entropy components is poor. Summary of the Invention

[0003] In view of the above problems, the present invention proposes a method for predicting the concentration of dissolved gases in transformer oil based on hybrid decomposition and BiTCN-BiGRU. First, the improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) is used to preliminarily decompose the gas concentration sequence; secondly, the sample entropy is combined to reconstruct the subsequences, and the variational mode decomposition (VMD) is applied to perform secondary decomposition on the high-entropy subsequences; then, a bidirectional time convolutional network-bidirectional gated recurrent unit (BiTCN-BiGRU) prediction model is constructed, and the forward and backward features of the time series data are captured simultaneously by using bidirectional convolution and bidirectional gating mechanisms; furthermore, the constructed prediction model is used to predict the subsequences obtained from the primary decomposition and the secondary decomposition, and the gas concentration sequence is reconstructed.

[0004] The present invention is implemented by the following technical solutions: The prediction of the concentration of dissolved gases in transformer oil based on hybrid decomposition and BiTCN-BiGRU is realized by the following steps:

[0005] S1: Collect the publicly available chromatographic time series data of transformer oil.

[0006] S2: After the collected chromatographic time series data of transformer oil is subjected to hybrid decomposition, the subsequences are divided into a training set and a test set according to a ratio of 7:3.

[0007] S3: Establish a hybrid neural network model of a bidirectional time convolutional network-bidirectional gated recurrent unit, use the bidirectional time convolutional network for feature extraction, and use the bidirectional gated recurrent unit to capture the forward and backward time-dependent information simultaneously.

[0008] S4: Establish an evaluation index system to comprehensively judge the prediction effect of the model.

[0009] S5: Input the processed subsequences into the model for prediction evaluation.

[0010] Furthermore, the S2 includes:

[0011] Establish a hybrid decomposition algorithm model, use the ICEEMDAN decomposition algorithm to decompose the gas concentration sequence, and reconstruct the subsequences according to the sample entropy value into a primary high-entropy sequence, a primary medium-entropy sequence, and a primary low-entropy sequence. Use the variational mode decomposition algorithm to decompose the primary high-entropy sequence and reconstruct it into secondary high-entropy subsequences, secondary medium-entropy subsequences, and secondary low-entropy subsequences according to the sample entropy value. Divide the five groups of subsequences into a training set and a test set according to a ratio of 7:3, and the steps are as follows:

[0012] Step 1: Perform primary decomposition using the ICEEMDAN decomposition algorithm. The detailed steps of the ICEEMDAN algorithm are as follows:

[0013] 1) Add a specific amount of white noise to the initial sequence set x to construct a new sequence:

[0014] x(i) = x + β 0 E(w (i) )

[0015] where: β 0 is the Gaussian white noise weight coefficient; w i represents the i-th group of white noise added, and E(w (i) ) represents the EMD component of the white noise.

[0016] 2) Calculate the residual component R 1 :

[0017] R 1 = (M(x (t) ))

[0018] where: M represents the local average envelope for determining the margin of the Intrinsic Mode Function (IMF).

[0019] 3) Calculate the first mode component IMF of the original sequence using the difference between the original sequence x and the first residual component 1 :

[0020] IMF 1 = x - R 1

[0021] 4) Determine the second residual component R 2 and the second mode component IMF 2 using the formula:

[0022]

[0023] 5) The k-th residual component R k and the k-th mode component IMF k can be expressed as:

[0024]

[0025] 6) Continue to repeat step five until the residual is indecomposable, and then stop this process.

[0026] Step two: Based on the sample entropy value, perform sequence reconstruction, and divide the first decomposition subsequence into a first high-entropy subsequence, a first medium-entropy subsequence, and a first low-entropy subsequence. The sample entropy can be represented by SampEn(m, r, N), where m is the embedding dimension, r is the similarity tolerance, and N is the length of the time series. It is defined as shown in the following formula, where A(m) is the number of pairs of similar subsequences of length m.

[0027]

[0028] SE is mainly used to quantify the complexity and irregularity of time series. Therefore, the decomposed subsequences are divided into three groups: high-entropy sequences, medium-entropy sequences, and low-entropy sequences according to the following formula using SE, and the subsequences within each group are reconstructed.

[0029]

[0030] In the formula: min is the minimum sample entropy value of the subsequence, and Range is the difference between the maximum subsequence entropy value and min.

[0031] Step 3: Decompose the high-entropy subsequences once using the variational mode decomposition algorithm, and reconstruct the subsequences obtained from the secondary decomposition using the method shown in Step 2. The detailed steps of the variational mode decomposition algorithm are as follows:

[0032] 1) Construct a constrained variational model:

[0033]

[0034] In the formula: K represents K IMFs obtained by decomposition; represents the corresponding frequency center of each component; δ(t) represents the Dirac function and the convolution operator; f(t) is the original input signal.

[0035] 2) Make the constrained variational problem into an unconstrained variational problem by increasing the penalty coefficient α and the Lagrange multiplier λ. Use the formula to expand the Lagrangian expression to obtain the augmented Lagrangian function:

[0036]

[0037] 3) Use the multiplier alternating method shown in the following formula to solve this constrained problem. Until the constraint conditions are met, the adaptive decomposition of the signal is realized.

[0038]

[0039] Step 4: Divide the five groups of subsequences obtained from the first decomposition and the second decomposition, namely the first decomposition medium-entropy sequences, the first decomposition low-entropy sequences, the second decomposition high-entropy sequences, the second decomposition medium-entropy sequences, and the second decomposition low-entropy sequences, into a training set and a test set according to a ratio of 7:3.

[0040] Furthermore, the step S3 includes:

[0041] Build a bidirectional temporal convolutional network - bidirectional gated recurrent unit hybrid neural network model. The BiTCN model extracts forward and backward data through one-dimensional causal convolution. Residual connections accelerate the convergence speed, and dilated convolution realizes temporal feature extraction. BiGRU combines the advantages of bidirectional processing and GRU, can capture forward and backward time-dependent information simultaneously in time series modeling tasks, can effectively extract data features, and obtains a faster convergence speed while ensuring a prediction effect similar to that of LSTM. The steps are as follows:

[0042] Step 1: TCN is a special convolutional network. It is a new type of convolutional network through causal convolution and residual connections. Compared with traditional CNN, TCN can capture the dependencies of time series through convolutional layers and is more suitable for time series data processing. BiTCN aims to capture both forward and backward features in temporal data to build a higher-precision model. Causal convolution is a strict time-constrained model. The output at time t only depends on the values before time t, that is, future information will not interfere with past information, making BiTCN have strong time information processing capabilities. For a one-dimensional sequence x and a convolutional kernel f, the dilated convolution F on element s is as follows:

[0043]

[0044] In the formula, d is the causal coefficient; K is the size of the convolutional kernel; s - d·i represents the (s - d·i)-th element of the previous layer. Residual connections ensure the accuracy of deep networks by coupling the input and output information of convolutional operations. A residual block with input x and output y is defined as follows:

[0045] y = H(x, W) + x

[0046] In the formula, H represents a series of convolutional operation functions; W is the weight of the convolutional kernel.

[0047] Each residual block contains two layers of causal convolution, uses a Dropout layer to prevent overfitting, the rectified linear unit (ReLU) is used as the activation function to improve the nonlinearity of the network, and WeightNorm is used to normalize the weight values of the convolutional kernel.

[0048] Step 2: As a variant of the LSTM network, GRU has a simpler structure and faster running speed, and can overcome the deficiencies of recurrent neural networks (RNN) in processing long time series data. GRU combines the "forget gate" and "input gate" of LSTM into an "update gate", and at the same time removes the "output gate" and adds a "reset gate". The "reset gate" is used to judge the information that needs to be ignored at the previous moment, and the "update gate" is used to judge whether the information at the current moment is passed to the next moment. The GRU calculation process is as follows:

[0049] zt = σ(W z x t + U z h t-1 )

[0050] r t = σ(W r x t + U r h t-1 )

[0051] h t = (1 - z t ) * h t-1 + z t * h t

[0052] h t = tanh(W h x t + U h (r t * h t-1 ))

[0053] where z t and r t are the "update gate" and "reset gate" respectively; h t and h t are the state vector and candidate state vector at time t respectively; σ and tanh are activation functions; W z , W, W h , U z , U r , U h are model parameters.

[0054] The BiGRU combines the advantages of bidirectional processing and GRU, and can capture both forward and backward time-dependent information in time series modeling tasks.

[0055] Furthermore, in step S4, five evaluation metrics are used to measure the degree of agreement between the predicted value and the actual value: mean squared error (MSE), mean absolute percentage error (MAPE), mean absolute error (MAE), root mean squared error (RMSE), and coefficient of determination (R 2 ). The calculation formulas for the above metrics are as follows:

[0056]

[0057] where y t is the actual value, is the predicted value.

[0058] Beneficial effects

[0059] Compared with the prior art, the present invention has the following beneficial effects:

[0060] 1) The proposed hybrid decomposition method effectively reduces the complexity and non-linearity of the dissolved gas concentration sequence in oil, solves the problem of low prediction accuracy of the first decomposition for high-entropy components, and significantly improves the prediction accuracy of the model for the time-series data of the dissolved gas concentration in oil.

[0061] 2) The constructed hybrid neural network combines a bidirectional temporal convolutional network and a bidirectional gated recurrent unit, and has excellent ability to capture and process long-term dependence relationships. By introducing a bidirectional structure, the problem of insufficient attention to reverse features in traditional TCN is overcome, and at the same time, combined with the modeling ability of BiGRU for complex temporal dependencies, the prediction accuracy of the time-series data of the dissolved gas concentration in oil is effectively improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 is a flowchart of the method of the present invention;

[0063] Figure 2 is a schematic diagram of a bidirectional dilated convolution structure;

[0064] Figure 3 is a schematic diagram of a GRU structure;

[0065] Figure 4 is a schematic diagram of a BiGRU structure; DETAILED DESCRIPTION OF THE EMBODIMENTS

[0066] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0067] The present invention is implemented by adopting the following technical solutions:

[0068] Figure 1 shows a method for predicting the dissolved gas concentration in transformer oil based on hybrid decomposition and BiTCN-BiGRU of the present application. Refer to Figure 1 , and is implemented by the following steps:

[0069] S1: Collect the publicly available time-series data of transformer oil chromatography;

[0070] S2: After performing hybrid decomposition on the collected time-series data of transformer oil chromatography, divide the subsequences into a training set and a test set according to a ratio of 7:3;

[0071] Further, the S2 includes:

[0072] Establish a hybrid decomposition algorithm model, decompose the gas concentration sequence using the ICEEMDAN decomposition algorithm, and reconstruct the subsequences according to the sample entropy values into a first high-entropy sequence, a first medium-entropy sequence, and a first low-entropy sequence. Decompose the first high-entropy sequence using the variational mode decomposition algorithm and reconstruct it into a second high-entropy subsequence, a second medium-entropy subsequence, and a second low-entropy subsequence according to the sample entropy values. Divide the five groups of subsequences into a training set and a test set according to a ratio of 7:3, and the steps are as follows:

[0073] Step 1: Perform the first decomposition using the ICEEMDAN decomposition algorithm. The detailed steps of the ICEEMDAN algorithm are as follows:

[0074] 1) Add a specific amount of white noise to the initial sequence set x to construct a new sequence:

[0075] x(i) = x + β 0 E(w (i) )

[0076] In the formula: β 0 is the Gaussian white noise weight coefficient; w i is the i-th group of white noise added, and E(w (i) ) represents the EMD component of the white noise.

[0077] 2) Calculate the residual component R 1 :

[0078] R 1 = (M(x (t) ))

[0079] In the formula: M represents the local average envelope for determining the edge of the intrinsic mode function (IMF).

[0080] 3) Calculate the first mode component IMF of the original sequence using the difference between the original sequence x and the first residual component 1 :

[0081] IMF 1 = x - R 1

[0082] 4) Use the formula to determine the second residual component R 2 and the second mode component IMF 2 :

[0083]

[0084] 5) The k-th residual component R k and the k-th mode component IMF k can be expressed as:

[0085]

[0086] 6) Continue to repeat step five until the residual is indecomposable, and stop this process.

[0087] Step 2: Based on the sample entropy value, perform sequence reconstruction, and divide the first-stage decomposed subsequences into high-entropy subsequences, medium-entropy subsequences, and low-entropy subsequences of the first stage. The sample entropy can be expressed as SampEn(m, r, N), where m is the embedding dimension, r is the similarity tolerance, and N is the length of the time series. It is defined as shown in the following formula, where A(m) is the number of pairs of similar subsequences of length m.

[0088]

[0089] SE is mainly used to quantify the complexity and irregularity of time series. Therefore, use SE to divide the decomposed subsequences into three groups: high-entropy sequences, medium-entropy sequences, and low-entropy sequences according to the following formula, and reconstruct the subsequences within each group.

[0090]

[0091] In the formula: min is the minimum sample entropy value of the subsequences, and Range is the difference between the maximum subsequence entropy value and min.

[0092] Step 3: Use the variational mode decomposition algorithm to decompose the high-entropy subsequences of the first stage, and use the method shown in step 2 to reconstruct the second-stage decomposed subsequences. The detailed steps of the variational mode decomposition algorithm are as follows:

[0093] 1) Construct a constrained variational model:

[0094]

[0095] In the formula: K represents the K IMFs obtained by decomposition; represents the corresponding frequency center of each component; δ(t) represents the Dirac function and the convolution operator; f(t) is the original input signal.

[0096] 2) By increasing the penalty coefficient α and the Lagrange multiplier λ, transform the constrained variational problem into an unconstrained variational problem. Use the formula to expand the Lagrangian expression to obtain the augmented Lagrangian function:

[0097]

[0098] 3) Use the multiplier alternating method shown in the following formula to solve this constrained problem. Until the constraint conditions are met, realize the adaptive decomposition of the signal.

[0099]

[0100] Step 4: Divide the five groups of subsequences, namely the entropy sequence in the first decomposition, the low-entropy sequence in the first decomposition, the high-entropy sequence in the second decomposition, the medium-entropy sequence in the second decomposition, and the low-entropy sequence in the second decomposition, obtained after the first decomposition and the second decomposition, into a training set and a test set according to a ratio of 7:3.

[0101] S3: Establish a hybrid neural network model of a bidirectional temporal convolutional network - bidirectional gated recurrent unit. Use the bidirectional temporal convolutional network for feature extraction, and use the bidirectional gated recurrent unit to capture time-dependent information in both the forward and backward directions simultaneously;

[0102] Further, the step S3 includes:

[0103] Establish a hybrid neural network model of a bidirectional temporal convolutional network - bidirectional gated recurrent unit. The BiTCN model extracts data in both the forward and backward directions through one-dimensional causal convolution. Residual connections accelerate the convergence speed, and dilated convolution realizes temporal feature extraction. The bidirectional dilated convolution structure is as shown in the appendix Figure 2 shown. BiGRU combines the advantages of bidirectional processing and GRU, can capture time-dependent information in both the forward and backward directions simultaneously in time series modeling tasks, can effectively extract data features, and obtains a faster convergence speed while ensuring a prediction effect similar to that of LSTM. The steps are as follows:

[0104] Step 1. TCN is a special convolutional network. It is a new type of convolutional network through causal convolution and residual connections. Compared with traditional CNN, TCN can capture the dependencies of time series through convolutional layers and is more suitable for time series data processing. BiTCN aims to capture both forward and backward features in time series data to build a higher-precision model. Causal convolution is a strict time-constrained model. The output at time t only depends on the values before time t, that is, future information will not interfere with past information, making BiTCN have strong time information processing capabilities. For a one-dimensional sequence x and a convolutional kernel f, the dilated convolution F on element s is as follows:

[0105]

[0106] where d is the causal coefficient; K is the size of the convolutional kernel; s - d·i represents the (s - d·i)-th element of the previous layer. Residual connections ensure the accuracy of deep networks by coupling the input and output information of convolutional operations. The residual block with input x and output y is defined as follows:

[0107] y = H(x, W) + x

[0108] where H represents a series of convolutional operation functions; W is the weight of the convolutional kernel.

[0109] Each residual block contains two layers of causal convolution, uses a Dropout layer to prevent overfitting, the rectified linear unit (ReLU) as the activation function to improve the nonlinearity of the network, and WeightNorm is used to normalize the weight values of the convolution kernels.

[0110] Step 2: As a variant of the LSTM network, GRU has a simpler structure and faster running speed, and can overcome the deficiencies of the recurrent neural network (RNN) when processing long time series data. GRU combines the "forget gate" and "input gate" of LSTM into an "update gate", and at the same time removes the "output gate" and adds a "reset gate". The "reset gate" is used to determine the information that needs to be ignored at the previous moment, and the "update gate" is used to determine whether the information at the current moment is passed to the next moment. The schematic diagram of the GRU structure is as shown in the appendix Figure 3 as follows. The calculation process of GRU is as follows:

[0111] z t = σ(W z x t + U z h t-1 )

[0112] r t = σ(W r x t + U r h t-1 )

[0113] h t = (1 - z t ) * h t-1 + z t * h t

[0114] h t = tanh(W h x t + U h (r t * h t-1 ))

[0115] where z t and r t are the "update gate" and "reset gate" respectively; h t and h t are the state vector and candidate state vector at time t respectively; σ and tanh are activation functions; W z , W, W h , U z , U r , U h are model parameters.

[0116] BiGRU combines the advantages of bidirectional processing and GRU, and can capture both forward and backward time-dependent information in time series modeling tasks. The schematic diagram of the BiGRU structure is shown in the appendix Figure 4 as follows

[0117] S4: Establish an evaluation index system to comprehensively judge the prediction effect of the model;

[0118] Furthermore, in step S4, five evaluation indicators are used to measure the degree of agreement between the predicted value and the actual value: mean squared error (MSE), mean absolute percentage error (MAPE), mean absolute error (MAE), root mean squared error (RMSE), and coefficient of determination (R 2 ). The calculation formulas of the above indicators are as follows:

[0119]

[0120] where y t is the actual value, is the predicted value.

[0121] S5: Input the processed subsequence into the model for prediction evaluation.

Claims

1. A method for predicting dissolved gas concentration in transformer oil based on hybrid decomposition and BiTCN-BiGRU, characterized by: This is accomplished by the following steps: S1: Collect publicly available transformer oil chromatogram time series data; S2: After the collected transformer oil chromatogram time series data is mixed and decomposed, the subsequences are divided into a training set and a test set in a ratio of 7:3; S3: Establish a hybrid neural network model of bidirectional temporal convolutional network and bidirectional gated recurrent unit, use the bidirectional temporal convolutional network for feature extraction, and use the bidirectional gated recurrent unit to simultaneously capture the forward and reverse time dependency information; S4: Establish an evaluation index system to comprehensively judge the prediction effect of the model; S5: Input the processed subsequence into the model for prediction evaluation.

2. A method for predicting dissolved gas concentration in transformer oil based on hybrid decomposition and BiTCN-BiGRU as claimed in claim 1, characterized in that: The proposed mixed decomposition of transformer oil chromatogram gas concentration time series data in step S2 has the following specific steps: Step 1: Use the ICEEMDAN decomposition algorithm to decompose the gas concentration sequence. The detailed steps of the ICEEMDAN algorithm are as follows: 1) Add a certain amount of white noise to the initial sequence set x to construct a new sequence: x(i)=x+β0E(w (i) ) Where: β0 is the Gaussian white noise weight coefficient; w i is the number of white noise groups added, E(w (i) ) represents the EMD component of white noise; 2) Calculate the residual component R1: R1=(M(x (t) )) Where: M represents the local average envelope used to determine the edge of the intrinsic mode function (IMF); 3) Calculate the first modal component IMF1 of the original sequence using the difference between the original sequence x and the first residual component: IMF1=x-R1 4) Determine the second residual component R2 and the second modal component IMF2 using the formula: 5) The kth residual component R k and the kth modal component IMF k It can be expressed as: 6) Continue to repeat step 5 until the residual cannot be decomposed, and then stop this process; Step 2: Reconstruct the sequence based on the sample entropy value, and divide the decomposed subsequence into a high entropy subsequence, a medium entropy subsequence, and a low entropy subsequence. The sample entropy can be represented by SampEn(m,r,N), where m is the embedding dimension, r is the similarity tolerance, and N is the length of the time series; it is defined as shown in the following formula, where A(m) is the logarithm of similar subsequences with length m: Step 3: Decompose the high entropy subsequence using the variational mode decomposition algorithm, and reconstruct the secondary decomposition subsequence using the method shown in step 2; the detailed steps of the variational mode decomposition algorithm are as follows: 1) Construct constrained variational model: Where: K represents the K IMFs obtained by decomposition; represents the frequency center corresponding to each component; δ(t) represents the Dirac function and the convolution operator; f(t) is the original input signal; 2) By increasing the penalty coefficient α and the Lagrange multiplier λ, the constrained variational problem is transformed into an unconstrained variational problem; Use the formula to expand the Lagrangian expression and obtain the augmented Lagrangian function: 3) Use the following multiplier alternation method to solve the constraint problem until the constraint conditions are met and the adaptive decomposition of the signal is achieved: Step 4: Divide the five groups of subsequences obtained after the first decomposition and the second decomposition into training set and test set in a ratio of 7:3, namely, the first decomposition medium entropy sequence, the first decomposition low entropy sequence, the second decomposition high entropy sequence, the second decomposition medium entropy sequence, and the second decomposition low entropy sequence.

3. A method for predicting dissolved gas concentration in transformer oil based on hybrid decomposition and BiTCN-BiGRU as claimed in claim 1, characterized in that: The proposed hybrid neural network model combining a bidirectional temporal convolutional network and a bidirectional gated recurrent unit in step S3 is established, and its specific steps are: Step 1: Build a TCN model, which uses a new convolutional network with causal convolution and residual connection. Compared with traditional CNN, TCN can capture the dependency of time series through convolution layer, which is more suitable for time series data processing. BiTCN aims to capture both positive and negative features in time series data at the same time, so as to build a more accurate model. Causal convolution is a strict time constraint model. The output at time t depends only on the value before time t, that is, future information will not interfere with past information, which makes BiTCN have a strong ability to process time information. For a one-dimensional sequence x and convolution kernel f, the dilated convolution F on element s is as follows: Where d is the causal coefficient; K is the size of the convolution kernel; sd·i represents the sd·ith element of the previous layer. The residual connection ensures the accuracy of the deep network by coupling the input and output information of the convolution operation. The residual block with input x and output y is defined as follows: y=H(x,W)+x Where H represents a series of convolution operation functions, and W is the weight of the convolution kernel; Each residual block contains two layers of causal convolution, using the Dropout layer to prevent overfitting, the Rectified Linear Unit (ReLU) as the activation function to improve the nonlinearity of the network, and WeightNorm to normalize the convolution kernel weights; Step 2: Build a BiGRU model. Use the built GRU model to add bidirectional processing to build a BiGRU model. GRU merges the "forget gate" and "input gate" of LSTM into an "update gate", removes the "output gate", and adds a "reset gate". The GRU calculation process is as follows: zt=σ(W z x t +U z h t-1 ) rt=σ(W r x t +U r h t-1 ) h t =(1-z t )*h t-1 +z t *h t ht=tanh(W h x t +U h (rt*ht-1)) Among them, z t and r t They are "update gate" and "reset gate" respectively; h t and h t are the state vector and candidate state vector at time t respectively; σ and tanh are activation functions; W z , W, W h , U z , U r , U h are model parameters.

4. A method for predicting dissolved gas concentration in transformer oil based on hybrid decomposition and BiTCN-BiGRU as claimed in claim 1, characterized in that: The specific steps in step S4 are: Five evaluation indicators are used to measure the consistency between the predicted value and the actual value: mean square error (MSE), mean absolute percentage error (MAPE), mean absolute error (MAE), root mean square error (RMSE) and determination coefficient (R 2 ); The calculation formula of the above indicators is as follows: Where y t is the actual value, is the predicted value.

Images