Prediction method for dissolved gas in transformer oil based on CEEMDAN-CNN-BIGRU

Abstract

The invention discloses a method for predicting dissolved gas in transformer oil based on CEEMDAN-CNN-BIGRU, and belongs to the technical field of power system detection, an original dissolved gas time sequence is decomposed through a CEEMDAN decomposition algorithm to obtain a plurality of stationary and non-stationary intrinsic mode components (IMFs), then a dissolved gas prediction model is used to train and predict each component, and a prediction result is obtained. The one-dimensional convolutional neural network (CNN) is used for extracting time sequence characteristics of original data, the gated recurrent neural network (GRU) captures a time sequence dependency relationship, and finally prediction results of all components are superposed, so that reconstruction and prediction of an original time sequence are realized, and a high-precision prediction result is obtained. According to the method, the non-stationary and nonlinear characteristics of the data of the dissolved gas in the transformer oil can be effectively handled, the highest prediction precision is shown through the CEEMDAN-CNN-BiGRU model, and the prediction result is closest to the change trend of the actual gas content.

Description

Technical Field

[0001] The invention relates to the technical field of power system detection, and in particular to a method for predicting dissolved gas in transformer oil based on CEEMDAN-CNN-BIGRU. Background Art

[0002] As the core equipment in the power system, the safe and reliable operation of the transformer is crucial to the stability of the power grid. Dissolved gas monitoring in oil is one of the most widely used transformer fault diagnosis methods. According to the change trend of dissolved gas in oil, it can effectively identify potential faults inside the transformer and further determine whether there are abnormalities inside the equipment, providing an important basis for status assessment and life prediction, thereby ensuring the reliable and stable operation of the power system.

[0003] As the transformer runs longer, the dissolved gas in the oil will show nonlinear and non-stationary change characteristics. Affected by complex operating conditions, measurement errors and other factors, the prediction accuracy is often difficult to meet actual needs. Domestic and foreign scholars have proposed a variety of methods for the prediction of dissolved gas in oil, such as radial basis neural network (RBFNN), least squares support vector machine (LSSVM), long short-term memory network (LSTM) and its combination model. These methods have their own characteristics: RBFNN can avoid local minima and has strong approximation ability; LSSVM is suitable for small sample and nonlinear data prediction, and LSTM has the advantage of processing long-term dependence characteristics and is widely used in time series prediction. However, the above methods still have many limitations in practical applications. RBFNN is sensitive to the selection of center points, which may affect the generalization performance of the model; LSSVM has a high computational complexity when processing large-scale data; GM model depends on the monotonicity of data and is difficult to adapt to non-stationary and nonlinear time series. In addition, traditional LSTM and GRU have problems such as complex parameter settings and large computational complexity, making it difficult to efficiently process large-scale non-stationary data.

[0004] To further improve the prediction accuracy, researchers have proposed a decomposition-reconstruction strategy, which combines signal decomposition technology with deep learning models and has achieved remarkable results. For example: variational mode decomposition (VMD) + gated recurrent unit (GRU), the VMD algorithm decomposes the original signal into multiple stationary components, and GRU is used to predict the trend of each component. Ensemble empirical mode decomposition (EEMD) + extreme learning machine (ELM), EEMD decomposes the signal and eliminates modal aliasing, and ELM quickly completes the prediction of the component. LSTM + particle swarm optimization (PSO), optimizes LSTM network parameters through PSO, and effectively improves the prediction accuracy. Although these combined methods can improve the prediction performance to a certain extent, there are still deficiencies in parameter setting, decomposition accuracy and model training efficiency. VMD and EWT are sensitive to parameter settings, EEMD may introduce residual noise, and models based on LSTM and GRU have difficulty capturing fine features when processing high-frequency signals. Summary of the invention

[0005] In order to solve the above technical problems, the present invention proposes a prediction method for dissolved gas in transformer oil based on CEEMDAN-CNN-BIGRU. First, CEEMDAN is used to decompose complex raw data into multiple intrinsic mode components (IMFs) to reduce the coupling effect between different components. Then, a one-dimensional convolutional neural network (CNN) is used to capture the local features in the components, and the bidirectional gated recurrent unit (BiGRU) is combined to deeply mine the time series dependency of the data to achieve trend prediction of each modal component. Finally, high-precision prediction results are obtained through reconstruction. It can effectively deal with the non-stationary and nonlinear characteristics of dissolved gas data in oil.

[0006] To achieve the above object, the technical solution adopted by the present invention is:

[0007] The method for predicting dissolved gas in transformer oil based on CEEMDAN-CNN-BIGRU is characterized by comprising the following steps:

[0008] S1. Construct CEEMDAN decomposition algorithm:

[0009] S11: Add white noise

[0010] Add a Gaussian white noise signal δ(t) to the original dissolved gas C(t) sequence:

[0011] S j (t)=C(t)+δ j (t)#(1)

[0012] Where j = 1, 2, ..., N, represents the number of times white noise is added.

[0013] S12: Extraction of the first-order modal components

[0014] For the noisy signal sequence S j (t) Perform EMD decomposition and extract the first modal component (IMF 1 ), and perform an ensemble average on them:

[0015]

[0016] Calculate the residual signal R 1 (y), for further decomposition:

[0017] R 1 (t) = C(t) - IMF 1 (t)#(3)

[0018] S13: Repeated decomposition

[0019] For the residual signal R k (t) Add white noise again to perform EMD decomposition and extract the next-order modal component IMF k+1 (t):

[0020]

[0021] The residual signal R k+1 (t):

[0022] R k+1 (t) = R k (t)-IMF k+1 (t)#(5)

[0023] When the residual signal R(t) becomes stable or cannot be further decomposed, the decomposition stops and the final original sequence can be expressed as:

[0024] S2. Decompose the original dissolved gas time series using the CEEMDAN decomposition algorithm:

[0025] The original dissolved gas sequence is decomposed into sub-modal components of different frequencies by CEEMDAN decomposition algorithm, thus obtaining stationary and non-stationary eigenmodal components;

[0026] S3. Constructing a dissolved gas prediction model:

[0027] The dissolved gas prediction model includes a convolutional neural network and a bidirectional gated recurrent unit, where:

[0028] Convolutional neural network: A one-dimensional convolutional neural network is selected to process the gas time series and extract the local features of the data and mine the potential features in the original sequence.

[0029] The one-dimensional convolutional neural network consists of an input layer, a hidden layer and an output layer, wherein the hidden layer includes a convolutional layer, a pooling layer and a fully connected layer.

[0030] The formula of one-dimensional convolutional neural network is expressed as:

[0031]

[0032] Among them, C is the output of the convolutional layer, is the activation function, W c is the convolution kernel weight, X is the input data, b c is bias;

[0033] Bidirectional Gated Recurrent Unit: It includes two parallel forward and reverse gated recurrent neural networks, which use the parallel forward and reverse gated recurrent neural networks to capture the forward and backward temporal dependency information at the same time.

[0034] The core formula of the gated recurrent neural network is:

[0035]

[0036] Among them, x t is the input vector, is the candidate state, h t is the hidden state, z t and r t are the update gate and reset gate respectively, W and b are the weight matrix and bias term of the network respectively, σ represents the Sigmoid activation function,

[0037] When two parallel forward and reverse gated recurrent neural networks are used to process the forward and reverse data of the input sequence, respectively, is the output of the positive sequence gated recurrent neural network at time t, is the output of the negative sequence gated recurrent neural network at time t,

[0038] The calculation formula of the dissolved gas prediction model can be expressed as:

[0039]

[0040] x t Input vector for the current moment, and The forward gated recurrent neural network and the reverse gated recurrent neural network output hidden states at time t-1, and They are the hidden states output by the forward and reverse gated recurrent neural networks at the current moment;

[0041] S4. Use the dissolved gas prediction model to train and predict each component:

[0042] S41, performing preprocessing operations such as missing value filling on the original data;

[0043] S42, decomposing the processed sequence using the CEEMDAN decomposition algorithm to obtain multiple IMF components;

[0044] S43, dividing each imf component into a training set and a test set to train the forward and reverse gated recurrent neural networks, iteratively optimizing the model parameters through the loss function, and establishing a corresponding prediction model;

[0045] S44, superimposing the prediction results of each component to obtain the final prediction result;

[0046] S5. Example analysis:

[0047] S51. Select the monitoring data of dissolved gas in oil of oil-immersed transformers as the research object, remove outliers and fill in missing values ​​on the original data, and divide the processed data into a training set and a test set in a ratio of 8:2;

[0048] S52, setting the signal-to-noise ratio in the CEEMDAN decomposition algorithm to 0.2, and setting the number of noise additions to 100;

[0049] S53. Formulate evaluation indicators: Use mean absolute error, root mean square error and coefficient of determination as evaluation indicators of prediction effect.

[0050] The calculation formulas are:

[0051]

[0052] Where, X i Indicates the actual value, represents the predicted value, n is the number of sampling points;

[0053] S54. Obtain the decomposition results and analyze them according to the evaluation indicators: It is observed that after the decomposition, the volatility of each generated sequence is reduced and the regularity is enhanced.

[0054] As a preferred technical solution of the present invention: in step S3, the gated recurrent neural network is a long short-term memory network, which simplifies the LSTM network structure and removes the explicit memory unit c t , the input gate and forget gate are merged into the update gate.

[0055] In the above structure: the method for predicting dissolved gas in transformer oil based on CEEMDAN-CNN-BiGRU proposed in the present invention firstly decomposes complex original data into multiple intrinsic mode components (IMFs) through CEEMDAN to reduce the coupling influence between different components, then uses a one-dimensional convolutional neural network (CNN) to capture local features in the components, and combines a bidirectional gated recurrent unit (BiGRU) to deeply mine the time series dependency of the data to achieve trend prediction of each modal component; finally, a high-precision prediction result is obtained through reconstruction, which can effectively deal with the non-stationary and nonlinear characteristics of dissolved gas data in oil.

[0056] Compared with the prior art, the present invention has the following beneficial effects:

[0057] The present invention processes the dissolved gas sequence by using the CEEMDAN decomposition algorithm. The CEEMDAN decomposition algorithm effectively improves the accuracy of decomposition and the integrity of the reconstructed signal by introducing adaptive Gaussian white noise and performing multiple iterations and averaging during the decomposition process, thereby effectively solving the modal aliasing problem existing in EMD and EEMD and improving the decomposition efficiency. In the prediction of dissolved gas in transformer oil, the CEEMDAN decomposition algorithm can decompose the original dissolved gas sequence into sub-modal components of different frequencies, effectively extract the stationary and non-stationary characteristic components of the sequence, and greatly improve the performance of the prediction model.

[0058] The present invention adopts the CNN-BiGRU prediction model to make predictions based on the decomposed sequence as input. Compared with the single BiGRU network, the combined model combines the advantages of CNN and can effectively capture temporal dependencies, thereby improving prediction accuracy.

[0059] The present invention uses the CEEMDAN decomposition algorithm and the CNN-BiGRU prediction model, and these two methods complement each other, effectively improving the accuracy of the prediction of dissolved gas in transformer oil, and providing an important basis for transformer state assessment and life prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 It is a one-dimensional convolutional neural network structure diagram;

[0061] Figure 2 This is a schematic diagram of the BiGRU structure;

[0062] Figure 3 It is the flow chart of the dissolved gas prediction model;

[0063] Figure 4 It is a line graph of CEEMDAN results of transformer H2 content;

[0064] Figure 5It is a table of prediction performance indicators of the dissolved gas prediction model;

[0065] Figure 6 It is a diagram of the prediction results of each model in the present invention. DETAILED DESCRIPTION

[0066] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments:

[0067] like Figure 1-6 As shown, the present invention proposes a method for predicting dissolved gas in transformer oil based on CEEMDAN-CNN-BIGRU, comprising the following steps:

[0068] S1. Construct CEEMDAN decomposition algorithm.

[0069] The CEEMDAN decomposition algorithm is improved on the basis of the EMD algorithm and the EEMD algorithm. It is an adaptive signal processing method proposed for the nonlinearity and non-stationarity of the signal. CEEMDAN effectively improves the accuracy of decomposition and the integrity of the reconstructed signal by introducing adaptive Gaussian white noise and performing multiple iterations and averaging during the decomposition process. It effectively solves the modal aliasing problem in EMD and EEMD and improves the decomposition efficiency. In the prediction of dissolved gas in transformer oil, CEEMDAN decomposition can decompose the original dissolved gas sequence into sub-modal components of different frequencies, effectively extract the stationary and non-stationary characteristic components of the sequence, and greatly improve the performance of the prediction model. The specific method is as follows:

[0070] S11: Add white noise

[0071] Add a Gaussian white noise signal δ(t) to the original dissolved gas C(t) sequence:

[0072] S j (t)=C(t)+δ j (t)#(1)

[0073] Where j = 1, 2, ..., N, represents the number of times white noise is added.

[0074] S12: Extraction of the first-order modal components

[0075] For the noisy signal sequence S j (t) Perform EMD decomposition and extract the first modal component (IMF 1 ), and perform an ensemble average on them:

[0076]

[0077] Calculate the residual signal R 1 (t), for further decomposition:

[0078] R 1 (t)=C(t)+IMF 1 (t)#(3)

[0079] S13: Repeated decomposition

[0080] For the residual signal R k (t) Add white noise again to perform EMD decomposition and extract the next-order modal component IMF k+1 (t):

[0081]

[0082] The residual signal R k+1 (t):

[0083] R k+1 (t) = R k (t) +IMF k+1 (t)#(5)

[0084] When the residual signal R(t) becomes stable or cannot be further decomposed, the decomposition stops and the final original sequence can be expressed as:

[0085] S2. Decompose the original dissolved gas time series using the CEEMDAN decomposition algorithm:

[0086] The original dissolved gas sequence is decomposed into sub-modal components of different frequencies by CEEMDAN decomposition algorithm, thus obtaining stationary and non-stationary eigenmodal components;

[0087] S3. Constructing a dissolved gas prediction model:

[0088] The dissolved gas prediction model includes a convolutional neural network and a bidirectional gated recurrent unit, where:

[0089] Convolutional neural network: A one-dimensional convolutional neural network is selected to process the gas time series and extract the local features of the data and mine the potential features in the original sequence.

[0090] The one-dimensional convolutional neural network consists of an input layer, a hidden layer and an output layer, wherein the hidden layer includes a convolutional layer, a pooling layer and a fully connected layer.

[0091] The formula of one-dimensional convolutional neural network is expressed as:

[0092]

[0093] Among them, C is the output of the convolutional layer, is the activation function, W c is the convolution kernel weight, X is the input data, b c is bias;

[0094] Convolutional neural network is a deep learning model widely used in computer vision, natural language processing and other fields. One-dimensional convolutional neural network (CNN) can efficiently extract local features of data and reduce model complexity, and mine potential features in the original sequence. Therefore, the present invention uses one-dimensional convolutional neural network (CNN) to process gas time series and enhance the model's ability to capture the law of gas changes. Its structure diagram is shown in the figure. Figure 1 shown.

[0095] Bidirectional Gated Recurrent Unit: It includes two parallel forward and reverse gated recurrent neural networks, which use the parallel forward and reverse gated recurrent neural networks to capture the forward and backward temporal dependency information at the same time.

[0096] The gated recurrent neural network is a long short-term memory network, which simplifies the LSTM network structure and removes the explicit memory unit c t , merge the input gate and forget gate into the update gate,

[0097] The core formula of the gated recurrent neural network is:

[0098]

[0099] Among them, x t is the input vector, is the candidate state, h t is the hidden state, z t and r t are the update gate and reset gate respectively, W and b are the weight matrix and bias term of the network respectively, σ represents the Sigmoid activation function,

[0100] When two parallel forward and reverse gated recurrent neural networks are used to process the forward and reverse data of the input sequence, respectively, is the output of the positive sequence gated recurrent neural network at time t, is the output of the negative sequence gated recurrent neural network at time t,

[0101] The calculation formula of the dissolved gas prediction model can be expressed as:

[0102]

[0103] x t Input vector for the current moment, and The forward gated recurrent neural network and the reverse gated recurrent neural network output hidden states at time t-1, and They are the hidden states output by the forward and reverse gated recurrent neural networks at the current moment;

[0104] The bidirectional gated recurrent unit (BiGRU) uses a scalable gated recurrent neural network (GRU). It uses two parallel forward and reverse gated recurrent neural networks GRU to capture the forward and backward temporal dependency information at the same time, thereby better capturing the dynamic changes in the sequence data. Its structure is shown in the figure below. Figure 2 shown.

[0105] The gated recurrent neural network is an improved long short-term memory network, which simplifies the LSTM network structure and removes the explicit memory unit c t , the input gate and forget gate are merged into the update gate. Therefore, the gated recurrent neural network has a simpler structure, fewer parameters, and higher computational efficiency.

[0106] S4. Use the dissolved gas prediction model to train and predict each component:

[0107] S41, performing preprocessing operations such as missing value filling on the original data;

[0108] S42, decomposing the processed sequence using the CEEMDAN decomposition algorithm to obtain multiple IMF components;

[0109] S43, dividing each imf component into a training set and a test set to train the forward and reverse gated recurrent neural networks, iteratively optimizing the model parameters through the loss function, and establishing a corresponding prediction model;

[0110] S44, superimposing the prediction results of each component to obtain the final prediction result;

[0111] The dissolved gas prediction model fully utilizes the feature extraction capability of CNN and deeply mines the connection between data. BiGRU processes both forward and reverse information at the same time, making up for the defect that the unidirectional network cannot obtain reverse dependency. The combination of the two can make full use of the spatial and temporal characteristics of dissolved gas data. The flow chart is as follows: Figure 3 shown.

[0112] S5. Example analysis:

[0113] S51. Select the monitoring data of dissolved gas in oil of oil-immersed transformers as the research object, remove outliers and fill in missing values ​​on the original data, and divide the processed data into a training set and a test set in a ratio of 8:2;

[0114] S52, setting the signal-to-noise ratio in the CEEMDAN decomposition algorithm to 0.2, and setting the number of noise additions to 100;

[0115] S53. Formulate evaluation indicators: Use mean absolute error, root mean square error and coefficient of determination as evaluation indicators of prediction effect.

[0116] The calculation formulas are:

[0117]

[0118] Where, X i Indicates the actual value, represents the predicted value, n is the number of sampling points;

[0119] S54. Obtain the decomposition results and analyze them according to the evaluation indicators: It is observed that after the decomposition, the volatility of each generated sequence is reduced and the regularity is enhanced.

[0120] In order to verify that the efficiency of the transformer oil dissolved gas prediction method based on CEEMDAN-CNN-BiGRU proposed in this invention is higher than other models, the oil dissolved gas monitoring data of an oil-immersed transformer (model TDS-50000 / 220) in operation of a power company was selected as the research object. The data used were collected by online monitoring equipment from January 1, 2016 to February 3, 2019, with a sampling time interval of 1 day, totaling 1495 data points. At the same time, in order to ensure the integrity and accuracy of the data, the original data was processed by outlier elimination and missing value completion. The processed data is divided into training set and test set in a ratio of 8:2.

[0121] Original H 2 The concentration series shows obvious nonlinear trends and local fluctuations, which makes the model prediction more difficult. To further verify the effectiveness of CEEMDAN decomposition, the series is decomposed into multiple intrinsic mode components (IMFs). In the experiment, the signal-to-noise ratio in the CEEMDAN decomposition method is set to 0.2, and the number of noise additions is set to 100. The decomposition results are shown in Figure 2. Figure 4 shown.

[0122] from Figure 4 From the decomposition results, it can be observed that after decomposition, the volatility of each generated sequence is significantly reduced and the regularity is enhanced. This process effectively suppresses the influence of residual white noise and provides a more stable and accurate data basis for subsequent prediction models.

[0123] In order to verify the advantages of the model proposed in the present invention, the present invention uses a variety of comparative models for comparison, including CNN-GRU, CNN-BiGRU and CEEMDAN-CNN-GRU, and conducts a detailed performance evaluation. The prediction performance indicators of each model are as follows: Figure 5 As shown in Figure 2, the prediction results of each model are compared. Figure 6 shown.

[0124] By observation Figure 5 as well as Figure 6It can be seen from the comparison of the evaluation indicators in that the four prediction models can accurately predict the changing trend of H2 content. In comparison, the prediction effect after decomposing the original data by the CEEMDAN method is better. The CEEMDAN-CNN-BiGRU model proposed in the present invention shows the highest prediction accuracy, and its prediction result is closest to the changing trend of the actual gas content.

[0125] Therefore, the combined model based on CEEMDAN decomposition method and CNN-BiGRU prediction proposed in the present invention is theoretically analyzed and compared with the other three models, which shows the superiority of the model of the present invention and draws the following conclusions:

[0126] By using the CEEMDAN decomposition method to process the dissolved gas sequence, the volatility of the original data is effectively reduced, and the quality of the input data used in the subsequent model is improved, so as to accurately capture the periodicity and trend information in the data.

[0127] The CNN-BiGRU combined model is used to make predictions based on the decomposed sequence as input. Compared with the single BiGRU network, this combined model combines the advantages of CNN and can effectively capture temporal dependencies, thereby improving prediction accuracy.

[0128] CEEMDAN decomposition and CNN-BiGRU prediction, these two methods complement each other, improve the accuracy of prediction of dissolved gas in transformer oil, and provide an important basis for transformer condition assessment and life prediction.

[0129] The above description is only a preferred embodiment of the present invention and does not constitute any other form of limitation to the present invention. Any modification or equivalent change made based on the technical essence of the present invention still falls within the scope of protection required by the present invention.

Claims

1. A method for predicting dissolved gas in transformer oil based on CEEMDAN-CNN-BIGRU, characterized by: The following steps are included: S1. Constructing CEEMDAN decomposition algorithm: S11: Add white noise Add a Gaussian white noise signal δ(t) to the original dissolved gas C(t) sequence: S j (t)=C(t)+δ j (t)#(1) Where j = 1, 2, ..., N, represents the number of times white noise is added. S12: Extraction of the first-order modal components For the noisy signal sequence S j (t) Perform EMD decomposition, extract the first modal component (IMF1), and perform ensemble averaging on it: Calculate the residual signal R1(t) for further decomposition: R1(t)=C(t)-IMF1(t)#(3) S13: Repeated decomposition For the residual signal R k (t) Add white noise again to perform EMD decomposition and extract the next-order modal component IMF k+1 (t): The residual signal R k+1 (t): R k+1 (t)=R k (t)-IMF k+1 (t)#(5) When the residual signal R(t) becomes stable or cannot be further decomposed, the decomposition stops and the final original sequence can be expressed as: S2. Decompose the original dissolved gas time series using the CEEMDAN decomposition algorithm: The original dissolved gas sequence is decomposed into sub-modal components of different frequencies by CEEMDAN decomposition algorithm, thus obtaining stationary and non-stationary eigenmodal components; S3. Constructing a dissolved gas prediction model: The dissolved gas prediction model includes a convolutional neural network and a bidirectional gated recurrent unit, where: Convolutional neural network: A one-dimensional convolutional neural network is selected to process the gas time series and extract the local features of the data and mine the potential features in the original sequence. The one-dimensional convolutional neural network consists of an input layer, a hidden layer and an output layer, wherein the hidden layer includes a convolutional layer, a pooling layer and a fully connected layer. The formula of one-dimensional convolutional neural network is expressed as: Among them, C is the output of the convolutional layer, is the activation function, W c is the convolution kernel weight, X is the input data, b c is bias; Bidirectional Gated Recurrent Unit: It includes two parallel forward and reverse gated recurrent neural networks, which use the parallel forward and reverse gated recurrent neural networks to capture the forward and backward temporal dependency information at the same time. The core formula of the gated recurrent neural network is: Among them, x t is the input vector, is the candidate state, h t is the hidden state, z t and r t are the update gate and reset gate respectively, W and b are the weight matrix and bias term of the network respectively, σ represents the Sigmoid activation function, When two parallel forward and reverse gated recurrent neural networks are used to process the forward and reverse data of the input sequence, respectively, is the output of the positive sequence gated recurrent neural network at time t, is the output of the negative sequence gated recurrent neural network at time t, The calculation formula of the dissolved gas prediction model can be expressed as: x t Input vector for the current moment, and The forward gated recurrent neural network and the reverse gated recurrent neural network output hidden states at time t-1, and They are the hidden states output by the forward and reverse gated recurrent neural networks at the current moment; S4. Use the dissolved gas prediction model to train and predict each component: S41, performing preprocessing operations such as missing value filling on the original data; S42, decomposing the processed sequence using the CEEMDAN decomposition algorithm to obtain multiple IMF components; S43, dividing each imf component into a training set and a test set to train the forward and reverse gated recurrent neural networks, iteratively optimizing the model parameters through the loss function, and establishing a corresponding prediction model; S44, superimposing the prediction results of each component to obtain the final prediction result; S5. Example analysis: S51. Select the monitoring data of dissolved gas in oil of oil-immersed transformers as the research object, remove outliers and fill in missing values ​​on the original data, and divide the processed data into a training set and a test set in a ratio of 8:2; S52, setting the signal-to-noise ratio in the CEEMDAN decomposition algorithm to 0.2, and setting the number of noise additions to 100; S53. Formulate evaluation indicators: Use mean absolute error, root mean square error and coefficient of determination as evaluation indicators of prediction effect, and their calculation formulas are: Where, X i Indicates the actual value, represents the predicted value, n is the number of sampling points; S54. Obtain the decomposition results and analyze them according to the evaluation indicators: It is observed that after the decomposition, the volatility of each generated sequence is reduced and the regularity is enhanced.

2. The method for predicting dissolved gas in transformer oil based on CEEMDAN-CNN-BIGRU according to claim 1, characterized in that: In step S3, the gated recurrent neural network is a long short-term memory network, which simplifies the LSTM network structure and removes the explicit memory unit c t , the input gate and forget gate are merged into the update gate.

Images