Method and system for filling dissolved gas data in transformer oil

Abstract

The invention discloses a method and a system for filling data of dissolved gas in transformer oil, and aims to improve the integrity and the accuracy of detection data of dissolved gas in oil of power equipment. The method comprises the following steps: constructing a data filling data set of gas dissolved in oil, dividing a training set and a test set according to a time sequence, and carrying out standardization processing on the data; dynamically generating a missing mask matrix based on the training set, and realizing missing data prediction through multi-scale time sequence feature extraction by using a TimesNet model; and performing sliding slicing processing on the test data in a reasoning stage, and recovering the original scale of the data and outputting a final filling value through weighted fusion of multiple prediction results. According to the method, the filling precision and the model generalization ability are effectively improved, the reliability of the state monitoring data of the power equipment is enhanced, and the method is suitable for missing value complementation of the oil dissolved gas data in an actual power system and has a wide application prospect.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power equipment data analysis, and particularly relates to a transformer oil dissolved gas data filling method and system. BACKGROUND

[0002] Oil-immersed transformers are core equipment for energy transmission and power conversion in new power systems, and guarantee large-scale new energy grid connection and continuous stable operation. Oil dissolved gas is an important indicator of the health status of oil-immersed transformers. Existing oil dissolved gas collection often has misdiagnosis or false / missing alarms due to sensor abnormal failure, communication system defects, signal interference and other reasons, affecting the safe and stable operation of equipment.

[0003] Oil dissolved gas data filling essentially belongs to time series data filling and follows spatial and temporal dual dependence. Traditional statistical filling methods, such as mean filling, interpolation method, regression filling, etc., mainly rely on the overall distribution characteristics or adjacent values of data, are simple to operate, and are suitable for scenes with less data missing and obvious rules. However, in the scene of oil dissolved gas data filling, the traditional statistical filling method is difficult to fully consider the time sequence characteristics of data and the correlation between data, resulting in low accuracy of time series data filling.

[0004] Therefore, there is an urgent need for a new method for filling transformer oil dissolved gas data. SUMMARY

[0005] The application provides a transformer oil dissolved gas data filling method and system to solve the problems in the prior art and improve the data accuracy of power equipment.

[0006] TECHNICAL SOLUTION

[0007] The application provides a transformer oil dissolved gas data filling method, which comprises the following steps:

[0008] An oil dissolved gas data matrix is constructed according to actual data collected by a sensor, the data matrix is divided into a training set and a test set, corresponding mask matrices are generated for the training set and the test set, and the training set and the test set are standardized.

[0009] A filling model for oil dissolved gas data is constructed, which comprises a plurality of TimeNet sub-models, each sub-model has the same structure and is a neural network stacked with a plurality of TimesBlock modules, and is used to extract time sequence characteristics of input data and output filling values.

[0010] Training each sub-model in the filling model by using the training set, comprising: defining a training sequence slice for the training set and its mask matrix, and inputting the training sequence slice into the sub-model; setting a loss function, and updating the parameters of the sub-model by using an Adam optimization algorithm; iteratively training until the current sub-model converges;

[0011] Defining a test sequence slice for the test set and its mask matrix, the step length of the test sequence slice being less than or equal to the step length of the training sequence slice; inputting the test sequence slice into the trained filling model to obtain an output corresponding to each test sequence slice; and aggregating all the outputs to obtain the filling values of the oil dissolved gas data through inverse normalization.

[0012] The actual data is multivariate time series data, which is arranged into an oil dissolved gas data matrix:

[0013]

[0014] wherein, is an oil dissolved gas input data matrix, is a corresponding element, is a data index, is a variable index, is a total length of a sequence, is a variable dimension;

[0015] The variable is the type of oil dissolved gas, including hydrogen, methane, acetylene, ethylene, ethane, carbon monoxide, carbon dioxide, oxygen and micro water.

[0016] Further, the basis for dividing the data matrix into a training set and a test set is time sequence, comprising:

[0017] Let the training set be and the test set be :

[0018]

[0019] wherein, is a training set time length, is a test set time length.

[0020] Further, the respective generation of mask matrices for the training set and the test set comprises:

[0021] (1) generating a dynamic missing mask matrix for the training set, which is represented as:

[0022]

[0023] For each variable, the dynamic missing rate is randomly selected in the interval , denoted as:

[0024]

[0025] wherein, is the corresponding dynamic missing rate, is the lower limit of probability, is the upper limit of probability, is a uniform distribution operator;

[0026] (2) Generate a fixed missing rate mask matrix for the test set , denoted as:

[0027]

[0028] wherein, is the corresponding test set mask matrix, is the corresponding element.

[0029] Further, the time sequence feature of the extracted input data is outputted, including:

[0030] The TimesBlock module automatically determines the convolution kernel size by using frequency domain analysis, encodes the multi-scale periodicity characteristics of the dissolved gas in oil sequence, and the calculation of the corresponding multi-scale time sequence feature is represented as:

[0031]

[0032] wherein, is the corresponding multi-scale time sequence feature, is the total number of TimesBlock layers, is a multi-layer network forward transmission operator, is a TimesBlock module operator;

[0033] The corresponding multi-scale time sequence feature of the dissolved gas in oil is mapped back to the variable space through linear mapping:

[0034]

[0035] wherein, is the calculation result of linear mapping, is the corresponding weight, , is the corresponding bias, ; the variable space mapping result is outputted :

[0036]

[0037] The calculated The padding value.

[0038] Further, the definition training sequence slice includes:

[0039] The definition model input is a fixed batch size training sequence slice, denoted as:

[0040]

[0041] Wherein, The dissolved gas data slice in oil, The corresponding mask matrix slice, The batch index, , The total batch number, The sequence length index, , The corresponding sub-model sequence length, The sub-model index;

[0042] Statistics of the number of non-missing data of each batch sample :

[0043]

[0044] Calculate the batch non-missing data mean :

[0045]

[0046] In the formula The positive number;

[0047] Calculate the batch non-missing data standard deviation :

[0048]

[0049] In the formula The positive number;

[0050] The standardization calculation is denoted as:

[0051]

[0052] Wherein, The normalized result of the dissolved gas data in the slice.

[0053] Further, the setting loss function includes:

[0054]

[0055] Wherein, The loss value of the corresponding sub-model missing point mean square error is The training sequence slice value is defined.

[0056] Further, the definition test sequence slice includes:

[0057] The test set data is represented as:

[0058]

[0059] For the first Sub-model, the calculation formula of slice step is:

[0060]

[0061] Where is the floor operator; for the part of the slice length, zero padding is used to fill up; the test set sequence slice and the corresponding mask slice are represented as:

[0062]

[0063] Wherein, is the corresponding sequence slice, is the corresponding mask slice.

[0064] Further, the aggregation of all outputs includes:

[0065] The output corresponding to the test sequence slice with overlapping area is weighted and averaged to be fused, represented as:

[0066]

[0067] The aggregation of all sub-model oil dissolved gas fusion results is represented as:

[0068]

[0069] Wherein is the number of TimesNet sub-models;

[0070] The original data scale of the oil dissolved gas is restored by inverse normalization, represented as:

[0071]

[0072] Wherein, is the corresponding inverse normalization calculation result.

[0073] The present application also proposes a transformer oil dissolved gas data filling system, comprising:

[0074] A data matrix module is configured to construct a data matrix of dissolved gas in oil according to actual data collected by a sensor, divide the data matrix into a training set and a test set, generate a corresponding mask matrix for the training set and the test set respectively, and normalize the training set and the test set;

[0075] A modeling module is configured to construct a filling model of the data of the dissolved gas in oil, and the filling model includes a plurality of TimeNet sub-models, each of which has the same structure and is a neural network stacked with a plurality of TimeBlock modules, and is configured to extract time sequence features of input data and output filling values;

[0076] A training module is configured to train each of the sub-models in the filling model by using the training set, including: defining a training sequence slice for the training set and the mask matrix thereof, inputting the training sequence slice into the sub-model; setting a loss function, and updating parameters of the sub-model by using an Adam optimization algorithm; and iteratively training until the current sub-model converges;

[0077] An implementation module is configured to define a test sequence slice for the test set and the mask matrix thereof, the step length of the test sequence slice is less than or equal to the step length of the training sequence slice; input the test sequence slice into the trained filling model to obtain an output corresponding to each test sequence slice; and aggregate all the outputs to obtain filling values of the data of the dissolved gas in oil through inverse normalization.

[0078] Beneficial effects:

[0079] In the present application, dynamic missing rate mask generation and standardization processing are adopted when constructing a data set, which effectively adapts to diversified and complex missing distribution in actual power equipment data, and improves the generalization ability of the model.

[0080] In the present application, different input data scales are set for different sub-models, multi-scale time sequence feature learning and frequency domain analysis mechanism are introduced, high-precision missing value filling of the data of the dissolved gas in oil is realized, and the data integrity and reliability are significantly improved.

[0081] In the present application, the TimesBlock module in the TimesNet model is used to fully mine the periodicity and multivariate correlation characteristics of the time sequence data of the dissolved gas in oil, different slice step lengths are set in the reasoning stage and the training stage, and based on the sliding window and the overlapping fusion strategy, the description of the data law is enhanced, and the filling error is reduced. The stability and accuracy of the filling result are further improved. BRIEF DESCRIPTION OF DRAWINGS

[0082] Figure 1 A flowchart of the method of the present application;

[0083] Figure 2 is a comparison chart of the data filling results of the method of the present application and the linear interpolation algorithm under the same conditions for different dissolved gases in oil, wherein (a) represents hydrogen, (b) represents methane, (c) represents acetylene, (d) represents ethylene, (e) represents ethane, (f) represents carbon monoxide, (g) represents carbon dioxide, (h) represents oxygen, and (i) represents micro water. DETAILED DESCRIPTION

[0084] The present application is further illustrated below in conjunction with the accompanying drawings and specific embodiments.

[0085] As shown in the figure, a data filling method for dissolved gases in transformer oil comprises the following steps: Figure 1

[0086] Step 1: Construction of a data set for filling data of dissolved gases in oil.

[0087] Data is collected by sensors and uploaded to a database for storage. Each row of data in the database represents a sampling time, which is set to be sampled every 8 hours in this embodiment. Each table represents a dissolved gas in oil, including hydrogen, methane, acetylene, ethylene, ethane, carbon monoxide, carbon dioxide, oxygen, and micro water. Based on the data collected through the above process, an input data matrix of dissolved gases in oil is constructed. The data is divided into a training set and a test set in chronological order. A dynamic mask matrix is generated for the training set, and a fixed mask matrix is generated for the test set. Based on the non-missing data of the training set, the mean and standard deviation of each variable are calculated, and the training set and test set data are standardized respectively.

[0088] Specifically, the following sub-steps are included:

[0089] (1) Construction of an input data matrix, which is represented as:

[0090]

[0091] wherein, is the input data matrix of dissolved gases in oil, is the corresponding element, is the data index, is the variable index, is the total length of the sequence, is the variable dimension, and R is the real number field.

[0092] (2) The input data matrix is divided into a training set and a test set in chronological order, which is specifically represented as:

[0093]

[0094] wherein, is the training set,​ is the length of the training set, is the test set, is the length of the test set.

[0095] (3) Generating a dynamic mask matrix for the training set, denoted as:

[0096]

[0097] wherein, is the mask matrix of the training set, is an element in the matrix, and the subscript f represents the variable (the type of dissolved gas in oil) corresponding to the element. For each variable, the dynamic missing rate of the mask matrix is randomly selected within the interval , and the corresponding definition is denoted as:

[0098]

[0099] wherein, is the corresponding dynamic missing rate, is the lower limit of probability, is the upper limit of probability, is a uniform distribution operator. The probability upper / lower limit is specified according to expert experience, and in the present embodiment, the lower limit is 0.1 and the upper limit is 0.6.

[0100] (4) Generating a fixed mask matrix for the test set, denoted as:

[0101]

[0102] wherein, is the mask matrix corresponding to the test set, is the corresponding element.

[0103] (5) Calculating the mean and standard deviation of each variable according to the non-missing data of the training set, and using the mean and standard deviation as a reference to standardize the training set and the test set.

[0104]

[0105]

[0106] wherein, is the corresponding mean, is the corresponding standard deviation, is the corresponding standardized calculation result of the training set, is the corresponding standardized calculation result of the test set.

[0107] Step 2, constructing and training the dissolved gas data filling model.

[0108] The application discloses a dissolved gas in oil data filling model, specifically a deep neural network TimesNet stacked by multiple TimesBlock layers.

[0109] The input training sequence slice of the dissolved gas in oil data filling model is defined, non-missing data statistics are calculated for each sample and variable, batch standardization data input is performed, data embedding and position coding are performed, multi-scale time sequence features are extracted through a TimesBlock module, mapping is performed back to a variable space, model output is inversely standardized, a missing point mean square error loss value is calculated, and an Adam optimization algorithm is used to update model parameters, and iterative training is performed until convergence.

[0110] Specifically, the following sub-steps are included:

[0111] (1) In order to reduce the training amount and extract as many data time sequence features as possible, the model input is defined as a training sequence slice of a fixed batch size, and is represented as:

[0112]

[0113] wherein, is a training sequence slice corresponding to the sub-model i, is a mask matrix slice thereof, and the slice step is 1; represents dissolved gas in oil data in the slice, represents a corresponding mask element; is a batch index, wherein is a total batch number; is a sequence length index, wherein is the length of the training sequence slice, is a sub-model index.

[0114] The value of is selected according to the performance of the running platform, and is set to 32 in the embodiment; the training sequence slice length is an integer set by a person, and the corresponding of each sub-model is different, and is set to [10, 20, 30, 40, 50, 60, 70] in the embodiment, that is, corresponding to 7 sub-models.

[0115] (2) For each batch of samples, the number of non-missing data is counted , the mean and the standard deviation are calculated, and the standardization is performed based on the mean and the standard deviation:

[0116] ​

[0117]

[0118] wherein, is the mean of corresponding non-missing data, is a very small positive number.

[0119]

[0120] wherein, is the standard deviation of corresponding non-missing data, is a very small positive number.

[0121]

[0122] wherein, is the standardized result of oil dissolved gas data within batch.

[0123] (3) Linear embedding and position coding are performed on the standardized data, and are represented as:

[0124]

[0125] wherein, is the calculation result of linear embedding and position coding, is the linear embedding weight, , is the model hidden layer dimension, is the linear embedding bias, , is the position coding of time step .

[0126] (4) Multi-scale time sequence features are extracted through a multi-layer TimesBlock module. The TimesBlock module automatically determines the convolution kernel size by using frequency domain analysis, and encodes the multi-scale time sequence features of the oil dissolved gas sequence. The calculation method of the multi-scale time sequence features is represented as:

[0127]

[0128] wherein, is the corresponding multi-scale time sequence feature, is the total number of TimesBlock layers, is a multi-layer network forward transmission operator, is a TimesBlock module operator.

[0129] (5) The multi-scale time sequence features of the oil dissolved gas are mapped back to the variable space through linear mapping, and are represented as:

[0130] ​

[0131] wherein, is the linear mapping result, is the corresponding weight, , is the corresponding bias, .

[0132] To obtain the filling value of the dissolved gas in oil, the variable space mapping result is outputted by inverse standardization, denoted as:

[0133]

[0134] wherein, is the corresponding inverse standardization output result.

[0135] (6) The missing point mean square error loss function is calculated, denoted as:

[0136]

[0137] wherein, is the corresponding sub-model missing point mean square error loss value, is the sequence slice value corresponding to the training set real data. The training set real data is the region corresponding to the 0 value of the mask, denoted as:

[0138]

[0139] According to the training process of (2)-(6) above, the weights and biases of the model are trained, the loss value is calculated and iteratively updated, and the Adam stochastic gradient descent algorithm is used. According to the loss value update the first sub-model parameters, and repeat the training until convergence.

[0140] Step 3, using the trained dissolved gas in oil data filling model for data filling.

[0141] The standardized test set matrix obtained in step 1 and the corresponding missing mask matrix are input into the filling model, and the filling output is obtained according to the trained parameters. The overlapping slice outputs are weighted and averaged, the outputs of all sub-models are aggregated, and the original data scale is recovered by inverse standardization. The final filling result is outputted.

[0142] The standardized test set matrix and the corresponding missing mask matrix in step 1 are input:

[0143]

[0144] According to the TimesNet sub-model, the dissolved gas sequence (the test set described above) is slidingly sliced, and the first The calculation formula of the slice step length of the TimesNet sub-model is represented as:

[0145]

[0146] wherein, is the corresponding slice step length, i.e., the slice operation is performed on each incremental position, is the floor operator. Zero padding is used to fill the part of the slice length.

[0147] The slice of the dissolved gas sequence in oil and the corresponding mask slice are represented as:

[0148]

[0149] wherein, is the corresponding sequence slice, is the corresponding mask slice.

[0150] The trained parameters are input into the model for calculation:

[0151]

[0152] wherein, is the corresponding slice filling result, is the corresponding TimesNet sub-model operator.

[0153] Because the slice step length of step 3 for the test set is smaller than the slice step length set in step 2, there is an overlapping area between the obtained slices. The overlapping area of the dissolved gas slice in oil is weighted and averaged to be fused, which is represented as:

[0154]

[0155] wherein, is the corresponding fusion result.

[0156] All sub-model outputs are aggregated:

[0157]

[0158] wherein, is the corresponding fusion result of all sub-models, is the number of TimesNet sub-models.

[0159] The original data scale of the dissolved gas in oil is restored by inverse normalization, and the corresponding calculation method is represented as:

[0160]

[0161] wherein, is the corresponding inverse standardization calculation result.

[0162] The final filling result of the dissolved gas in oil is expressed as:

[0163]

[0164] wherein, is the corresponding final filling result.

[0165] The present application proposes a multi-scale TimesNet filling model on the basis of TimesNet. The multi-scale method is proposed according to the time sequence characteristics of the dissolved gas in oil. Step 1 proposes a dynamic random mask construction method, step 2 proposes a multi-scale TimesNet model and a training method, and step 3 proposes a multi-scale inference result fusion method. The three parts are organically combined to form the innovative method of the present application, which improves the filling accuracy of the dissolved gas in oil.

[0166] To further verify the performance of the method of the present application, a linear interpolation algorithm with better filling effect is selected as a comparison algorithm. The missing rate of the test set data is set to eight missing degrees from 0.1 to 0.8. In this embodiment, the filling effect of each dissolved gas in oil under a missing rate of 0.1 is displayed. The filling results corresponding to each dissolved gas in oil are shown in FIG. 2, wherein (a) represents hydrogen, (b) represents methane, (c) represents acetylene, (d) represents ethylene, (e) represents ethane, (f) represents carbon monoxide, (g) represents carbon dioxide, (h) represents oxygen, and (i) represents micro water. As can be seen from the figure, most of the test data of the gas fluctuates up and down within a certain range, and the filling effects of the two algorithms are close.

[0167] The comparison results of the method of the present application and the linear interpolation algorithm under each missing degree are shown in Table 1 and Table 2.

[0168] Table 1 Comparison of filling results under multiple missing rates (0.1-0.4)

[0169]

[0170] Table 2 Comparison of filling results under multiple missing rates (0.5-0.8)

[0171]

[0172] For the case where the missing rate is 0.1, the filling error of the method of the present application is slightly higher than that of the comparison algorithm for acetylene, ethylene and micro water, and the filling error of the remaining gas is significantly lower than that of the comparison algorithm. Compared with the total MAE, the method of the present application is significantly lower than the comparison algorithm, and the value is about 74.73% of the comparison algorithm.

[0173] By analyzing the other missing rate filling results provided in Table 1 and Table 2, it can be seen that the total MAE minimum value corresponds to a missing rate of 0.1, the total MAE maximum value corresponds to a missing rate of 0.8, and generally shows a gradual upward trend with the missing rate. The total MAE minimum value of the comparative algorithm corresponds to a missing rate of 0.7, and the total MAE maximum value corresponds to a missing rate of 0.2, which shows that the higher the missing rate, the smaller the filling error, indicating that most of the test set gas data is closer to linear distribution.

[0174] For hydrogen, ethylene and carbon monoxide, the filling error of both algorithms increases with the increase of the missing rate, indicating that the missing rate has a greater impact on the filling of these features. For methane, the filling error of the present method increases with the increase of the missing rate, and the comparative algorithm has no obvious trend. For acetylene, the filling error of the present method decreases with the increase of the missing rate, and the comparative algorithm has no obvious trend. This shows that the comparative algorithm is not sensitive to the missing rate when filling methane and acetylene data. For ethane, the present method has no obvious trend, and the comparative algorithm decreases with the increase of the missing rate. For oxygen, the filling error of the present method increases with the increase of the missing rate, and the filling error of the comparative algorithm decreases with the increase of the missing rate. This shows that ethane and oxygen are more consistent with the linear distribution rule. For carbon dioxide and micro water, the filling error of the present method increases with the increase of the missing rate, while the comparative algorithm remains basically unchanged, indicating that carbon dioxide and micro water are more adaptable when using the comparative algorithm for filling.

[0175] In summary, for each missing rate, the present method is significantly lower than the comparative algorithm in total MAE, showing good filling performance.

Claims

1. A method for filling dissolved gas data in transformer oil, characterized in that, include: A dissolved gas data matrix in oil is constructed based on the actual data collected by the sensor. The data matrix is ​​divided into a training set and a test set. A corresponding mask matrix is ​​generated for the training set and the test set respectively. The training set and the test set are then standardized. A model for filling in dissolved gas data in oil is constructed, which includes several TimeNet sub-models. Each sub-model has the same structure, which is a neural network with multiple layers of TimeBlock modules stacked together. It is used to extract the temporal features of the input data and output the filled values. Training each sub-model in the imputation model using the training set includes: defining training sequence slices for the training set and its mask matrix, and inputting the training sequence slices into the sub-model; setting a loss function and updating the parameters of the sub-model using the Adam optimization algorithm; and iteratively training until the current sub-model converges. For the test set and its mask matrix, a test sequence slice is defined, with the step size of the test sequence slice being less than or equal to the step size of the training sequence slice. The test sequence slice is input into the trained imputation model to obtain the output corresponding to each test sequence slice. All outputs are aggregated and denormalized to obtain the imputation value of dissolved gas data in oil.

2. The filling method according to claim 1, characterized in that, The construction of the dissolved gas data matrix in oil includes: The actual data is multivariate time series data, which is organized into a dissolved gas data matrix in oil: in, Input data matrix for dissolved gases in oil. For the corresponding element, For data indexing, For variable index, The total length of the sequence. For the dimension of the variable; The variables are the types of dissolved gases in the oil, including hydrogen, methane, acetylene, ethylene, ethane, carbon monoxide, carbon dioxide, oxygen, and trace amounts of water.

3. The filling method according to claim 2, characterized in that, The basis for dividing the data matrix into training and testing sets is chronological order, including: Let the training set be... The test set is : in, The length of the training set. The duration of the test set.

4. The filling method according to claim 3, characterized in that, The generation of mask matrices for the training set and the test set respectively includes: (1) Generate a dynamic missing mask matrix for the training set. , represented as: For each corresponding element, the dynamic missing rate is in the range [missing information]. Random selection within, represented as: in, To correspond to the dynamic missing rate, This is the lower bound of probability. This is the upper limit of probability. Operators that conform to uniform distribution; (2) Generate a fixed missing rate mask matrix for the test set. , represented as: in, For the corresponding test set mask matrix, For the corresponding element.

5. The filling method according to claim 4, characterized in that, The step of extracting temporal features from the input data and outputting imputed values ​​includes: The TimesBlock module automatically determines the convolution kernel size using frequency domain analysis to encode the multi-scale periodic features of the dissolved gas sequence in oil. The calculation representation of the corresponding multi-scale time series features is as follows: in, To correspond to multi-scale temporal features, This represents the total number of levels in TimesBlock. For multi-layer network forward pass operator, For TimesBlock module operators; The multi-scale time-series characteristics of dissolved gases in the oil are mapped back to the variable space through linear mapping: in, The result of the linear mapping calculation. For the corresponding weights, , For the corresponding bias, The variable space mapping result is denormalized and output. : Calculated The fill value is the value mentioned above.

6. The filling method according to claim 5, characterized in that, The definition of training sequence slices includes: Define the model input as slices of training sequences with a fixed batch size, represented as: in, Slicing data of dissolved gases in oil. Slice the corresponding mask matrix. For batch indexing, , Total number of batches For sequence length index, , For the corresponding sub-model sequence length, Index for sub-models; Count the number of non-missing data in each batch of samples. : Calculate the mean of non-missing data within a batch. : In the formula It is a positive number; Calculate the standard deviation of non-missing data within a batch. : In the formula It is a positive number; Standardized calculations are expressed as follows: in, This is the standardized result of the dissolved gas data in the oil within this slice.

7. The filling method according to claim 6, characterized in that, The setting of the loss function includes: in, This represents the mean squared error loss value for the missing points in the corresponding sub-model. These are slice values ​​for the training sequence.

8. The filling method according to claim 7, characterized in that, The definition of the test sequence slice includes: The test set data is represented as follows: Regarding the first Sub-model, slice step size The calculation formula is: in The floor function is used for rounding down; zero-padding is used to fill any parts of the slice that are not long enough; the test set sequence slice and the corresponding mask slice are represented as follows: in, Slice the corresponding sequence. This is the corresponding mask slice.

9. The filling method according to claim 8, characterized in that, The aggregation of all outputs includes: The outputs of test sequence slices with overlapping regions are weighted and averaged, and then fused as follows: The fusion result of dissolved gases in the oils of all sub-models is represented as follows: in This represents the number of TimesNet sub-models. The original scale of dissolved gas data in oil is restored by denormalization, and it is represented as follows: in, This corresponds to the inverse standardization calculation results.

10. A system for filling dissolved gas data in transformer oil, characterized in that, include: The data matrix module is used to construct a dissolved gas data matrix in oil based on the actual data collected by the sensor. The data matrix is ​​divided into a training set and a test set. A corresponding mask matrix is ​​generated for the training set and the test set respectively, and the training set and the test set are standardized. The modeling module is used to build a model for filling in dissolved gas data in oil. It includes several TimeNet sub-models. Each sub-model has the same structure and is a neural network with multiple layers of TimeBlock modules stacked together. It is used to extract the temporal features of the input data and output the filled values. The training module is used to train each sub-model in the imputation model using the training set, including: defining training sequence slices for the training set and its mask matrix, inputting the training sequence slices into the sub-model; setting a loss function, updating the parameters of the sub-model using the Adam optimization algorithm; and iteratively training until the current sub-model converges. The implementation module defines test sequence slices for the test set and its mask matrix, wherein the step size of the test sequence slices is less than or equal to the step size of the training sequence slices; inputs the test sequence slices into the trained imputation model to obtain the output corresponding to each test sequence slice; aggregates all outputs and performs inverse standardization to obtain the imputation values ​​of dissolved gas data in oil.

Images