Method for predicting dissolved gas in transformer oil based on univariate de-timing

Abstract

The invention discloses a method for predicting dissolved gas in transformer oil based on univariate de-timing. The method comprises the following steps: firstly, collecting a time sequence sample of online monitoring of the dissolved gas in the transformer oil, and carrying out data preprocessing; secondly, feature extraction is conducted on historical data of gas dissolved in oil through a sliding window technology, and gas time sequence sample data are converted into non-time-sequence feature matrix samples; and finally, constructing a deep forest prediction model, deeply mining information of the non-time-sequence characteristic matrix through a multi-granularity scanning module, and continuously training through a cascade forest module to realize accurate prediction of the content of the gas dissolved in the oil at the future moment. According to the method, the dependence on a specific time point or a time sequence relation can be further reduced under the condition that the rule of predicting future in the past of a time sequence is not damaged, the applicability of a traditional non-time-sequence prediction model is improved, and the prediction precision of online monitoring data of the dissolved gas in the transformer oil can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment monitoring, and in particular to a method for predicting dissolved gas in transformer oil based on single variable de-sequencing. Background Art

[0002] Online monitoring of dissolved gas in transformer oil is an important part of transformer equipment status monitoring and maintenance. By accurately predicting the dissolved gas in transformer oil, the operating status of the transformer can be predicted in advance and the internal defects of the transformer can be discovered in time, which is of great significance to improving the reliability of the equipment.

[0003] The current prediction methods for online monitoring data of dissolved gas in transformer oil are mainly based on time series analysis methods, such as long short-term memory networks, moving average models, and convolutional neural networks. The principle is to predict the numerical trend of future time points by analyzing and modeling historical time series data. The above methods have achieved certain results in improving the prediction accuracy of dissolved gas in transformer oil, but there are still many problems to be solved in the time series prediction method in field applications, mainly including: (1) The generation and release of dissolved gas in transformer oil are affected by multiple factors, including operating load, external environmental interference, etc., which makes the construction of time series prediction models too complex and uncertain. (2) The change of dissolved gas in transformer oil may have a time lag before the occurrence of internal defects in the equipment, and the prediction model is prone to ignore this time lag effect, resulting in the inability to accurately capture the correlation between dissolved gas in oil and equipment status. (3) Compared with traditional prediction models, deploying and maintaining highly complex time series prediction models will increase the cost of system maintenance, and there are also problems such as over-reliance on specific time points or time series relationships. Summary of the invention

[0004] In view of the problems existing in the prior art, the purpose of the present invention is to provide a method for predicting dissolved gas in transformer oil based on single variable de-sequencing. The method de-sequencing the online monitoring data of dissolved gas in transformer oil and realizes accurate prediction of dissolved gas in oil through a deep forest prediction model.

[0005] To achieve the purpose of the invention, the present invention adopts the following technical solution, a method for predicting dissolved gas in transformer oil based on single variable de-sequencing, comprising the following steps: S1: Collect online monitoring data of dissolved gases in transformer oil, including seven characteristic gases: H2, CH4, C2H6, C2H4, C2H2, CO and CO2, and perform outlier processing and missing data filling on the time series sample data of the above characteristic gases; S2: Use sliding window technology to extract features from historical data of dissolved gas in oil and convert time series sample data into non-time series feature matrix samples; S3: Divide the non-time series feature matrix samples into training sets and test sets, build a deep forest prediction model, and use the tree structure probability density estimation algorithm to optimize the parameters of the deep forest prediction model; S4: Evaluate the performance of the deep forest prediction model, predict the dissolved gas content in the oil at future times, and optimize the operation and maintenance strategy of the transformer in advance.

[0006] Furthermore, the outlier processing described in S1 uses a box plot to identify and process outliers in the time series sample data of dissolved gas in transformer oil, and the processing rules are as follows: , In the formula, T is the outlier in the time series sample data, Q u and Q L are the upper quartile and the lower quartile, respectively, Q c Q u and Q L The numerical difference.

[0007] Furthermore, the missing data filling described in S1 is performed on the time series sample data of dissolved gas in transformer oil based on the KNN interpolation method of neighboring points. The specific steps are as follows: S1.1: Use cross-validation method to select the number of neighboring points K; S1.2: Use the Euclidean distance to calculate the distance between the target point and all other points. The calculation formula is as follows: , In the formula, (x i , y i ) are the coordinates of the neighboring points and the target point in the i-th dimension, and m is the data dimension; S1.3: Select the K data points closest to the target point, assign a weight to each neighboring point in inverse proportion to the Euclidean distance, and calculate the weighted average of all neighboring points as the filling value for the missing data.

[0008] Furthermore, in S2, the sliding window technology is used to extract the features of the historical data of dissolved gas in oil, and the time series sample data is converted into a non-time series feature matrix. The specific steps are as follows: S2.1: Determine the sliding window size and stride based on the time series sample data of dissolved gas in transformer oil; S2.2: Perform “sliding window” processing on the time series samples of each characteristic gas, and calculate the statistical characteristics of the gas in each window, including the mean, maximum, minimum and standard deviation; S2.3: Integrate the features extracted from each window to generate a feature matrix, where each row represents the features extracted from a time window and each column represents a feature.

[0009] Furthermore, a deep forest prediction model is constructed in S3. The specific steps are as follows: S3.1: Based on the multi-granularity scanning module, feature extraction is performed on the input non-time-series feature matrix samples. Assuming that the feature dimension is l, the scanning size is q, and the step size is k, the multi-granularity scanning module extracts a sub-feature vector of (lq) / k+1 dimensions and trains it as the input of the random forest and completely random forest estimators. Each estimator outputs an α-dimensional probability vector. All output vectors are concatenated in series, and the dimension of the enhanced feature vector obtained is: , S3.2: The feature vector obtained by the multi-granularity scanning module is used as the input of the cascade forest, which is passed and trained step by step in each cascade layer, and the predicted probability vector and the input feature vector are combined in series; the determination coefficient R 2 The evaluation index for judging the expansion of cascade forest is calculated as follows: , In the formula, n is the number of prediction points; is the predicted value of gas; i is the true value of gas; is the average value; the coefficient of determination R 2 The closer the value is to 1, the better the deep forest prediction model fits the input feature data; S3.3: If , the cascade forest expansion is stopped, and the final prediction value of the deep forest prediction model for dissolved gas in transformer oil is output.

[0010] Furthermore, S3 uses a tree structure probability density estimation algorithm to optimize the parameters of the deep forest prediction model. The specific steps are as follows: S3-1: Determine the parameter space of the deep forest prediction model, including the number of trees, the depth of the trees, and the minimum number of samples per leaf node; S3-2: The improvement expectation is used to measure the extent to which the loss value corresponding to each selected parameter can be reduced. The improvement expectation calculation formula is as follows: , In the formula, is the conditional probability distribution; v is the sampling value; v′ is the optimal sampling threshold; u is any value in the configuration space parameter; S3-3: Select the parameter combination with the highest probability density and calculate the maximum expected value as the candidate for the next iteration; S3-4: Repeat steps S3-1 to S3-3 until the set number of iterations is reached or the stop condition is met.

[0011] Furthermore, the performance of the deep forest prediction model is evaluated in S4 using evaluation indicators including the determination coefficient R 2 and the root mean square error R MSE , where R MSE The calculation formula is as follows: , Based on the deep forest prediction model, the dissolved gas content in the oil is predicted at future times to achieve the goal of optimizing the transformer operation and maintenance strategy in advance.

[0012] Beneficial effects of the present invention: The present invention provides a method for predicting dissolved gas in transformer oil based on single variable de-sequencing. By de-sequencing the online monitoring data of dissolved gas in transformer oil, the complexity of the prediction model can be reduced and the applicability of the traditional non-series prediction model can be improved. This helps to improve the prediction accuracy and effect of the online monitoring data of dissolved gas in transformer oil, and also provides a reliable guarantee for early fault warning of the transformer. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 It is a flow chart of the method for predicting dissolved gas in transformer oil based on single variable de-sequencing in the present invention; Figure 2 It is a schematic diagram of the de-sequencing of the single variable of the dissolved gas in the oil in the present invention; Figure 3 It is the cascade forest structure diagram in the present invention; Figure 4 It is the prediction result of methane gas in the embodiment of the present invention. DETAILED DESCRIPTION

[0014] The present invention is further described below in conjunction with embodiments. It is necessary to point out that the following embodiments are only used to further illustrate the present invention and cannot be understood as limiting the scope of protection of the present invention. Some non-essential improvements and adjustments made by technical personnel in the field based on the above invention content still fall within the scope of protection of the present invention.

[0015] like Figure 1 As shown, a specific embodiment of the present invention provides a method for predicting dissolved gas in transformer oil based on single variable de-sequencing, comprising the following steps: S1: Collect online monitoring data of dissolved gases in transformer oil, mainly including seven characteristic gases: H2, CH4, C2H6, C2H4, C2H2, CO and CO2, and perform outlier processing and missing data filling on the time series sample data of the above characteristic gases.

[0016] In the present invention, outlier processing is performed on the time series sample data of dissolved gas in transformer oil, and a box plot is mainly used to identify and process outliers in the time series sample data of dissolved gas in transformer oil. The processing rules are as follows: ,

[0017] In the formula, T is the outlier in the time series sample data, Q u and Q L are the upper quartile and the lower quartile, respectively, Q c Q u and Q L The missing data of dissolved gas in transformer oil are filled mainly based on the KNN interpolation method of neighboring points to fill the missing data of the time series sample data of dissolved gas in transformer oil. The specific steps are as follows: S1.1: Use the cross-validation method to select the appropriate value of the number of neighboring points K.

[0018] S1.2: Use the Euclidean distance to calculate the distance between the target point and all other points. The calculation formula is as follows: ,

[0019] In the formula, (x i , y i ) are the coordinates of the neighboring points and the target point in the i-th dimension, and m is the data dimension.

[0020] S1.3: Select the K data points closest to the target point, assign a weight to each neighboring point in inverse proportion to the Euclidean distance, and calculate the weighted average of all neighboring points as the filling value for the missing data.

[0021] S2: Use sliding window technology to extract features from the historical data of dissolved gas in oil and convert the time series sample data into non-time series feature matrix samples. Figure 2 As shown, the specific steps are as follows: S2.1: Determine the sliding window size and stride based on the time series sample data of dissolved gas in transformer oil.

[0022] S2.2: Perform sliding window processing on the time series samples of each characteristic gas and calculate the statistical characteristics of the gas in each window, mainly including the mean, maximum, minimum and standard deviation.

[0023] S2.3: Integrate the features extracted from each window to generate a feature matrix, where each row represents the features extracted from a time window and each column represents a feature.

[0024] S3: Divide the non-time series feature matrix samples into training sets and test sets, build a deep forest prediction model, and use the tree structure probability density estimation algorithm to optimize the parameters of the deep forest prediction model.

[0025] In the present invention, feature extraction is performed on the input non-time-series feature matrix samples based on a multi-granularity scanning module. Assuming that the feature dimension is l, the scanning size is q, and the step size is k, the multi-granularity scanning module can extract a sub-feature vector of (lq) / k+1 dimensions and train it as the input of the two evaluators, random forest and completely random forest. Each evaluator outputs an α-dimensional probability vector. All output vectors are concatenated in series, and the dimension of the enhanced feature vector obtained is: , The feature vector obtained by the multi-granularity scanning module is used as the input of the cascade forest. The cascade forest structure is as follows: Figure 3 As shown. In each cascade layer, the predicted probability vector and the input feature vector are passed and trained step by step, and the predicted probability vector and the input feature vector are connected in series. 2 The evaluation index for judging the expansion of cascade forest is calculated as follows: , In the formula, n is the number of prediction points; is the predicted value of gas; i is the true value of gas; is the average value. 2 The closer the value is to 1, the better the deep forest prediction model fits the input feature data. , the cascade forest expansion is stopped, and the final prediction value of the deep forest prediction model for dissolved gas in transformer oil is output.

[0026] The tree structure probability density estimation algorithm is used to optimize the parameters of the deep forest prediction model. First, the parameter space of the deep forest prediction model is determined, including the number of trees, the depth of the tree, the minimum number of samples for each leaf node and other hyperparameters. Then, the lift expectation is used to measure the extent to which the corresponding loss value of each selected parameter can be reduced. The lift expectation calculation formula is as follows: , In the formula, is the conditional probability distribution; v is the sampling value; v′ is the optimal sampling threshold; u is any value in the configuration space parameter. Select the parameter combination with the highest probability density and calculate the maximum expected value as the candidate for the next iteration. Repeat the above steps until the set number of iterations is reached or the stop condition is met.

[0027] S4: Evaluate the performance of the deep forest prediction model using evaluation metrics including the coefficient of determination R 2and the root mean square error R MSE , where R MSE The calculation formula is as follows: , Based on the deep forest prediction model, the dissolved gas content in the oil is predicted at future times to achieve the goal of optimizing the transformer operation and maintenance strategy in advance.

[0028] Example

[0029] The time interval for online monitoring of dissolved gas in a transformer oil is 2 hours. The historical data of methane gas within 30 days is extracted to predict the gas change trend in the next three days. The sliding window size is set to 12, the sliding step number is 1, and the above methane online monitoring data is de-sequentialized. The average, maximum, minimum and standard deviation of each sliding window are calculated as the statistical characteristics of the change in methane gas content within one day. The data features and statistical features obtained by the sliding window are concatenated and spliced, and the samples with missing data are deleted. Finally, 348×15 non-time series feature matrix samples are obtained. The non-time series feature matrix samples are divided into training sets and test sets at a ratio of 7:3, and the deep forest prediction model is trained. The tree structure probability density estimation algorithm is used to simultaneously optimize the main parameters of the deep forest prediction model, including the number of trees, the maximum depth of the tree, and the minimum number of samples for each leaf node, and finally the optimal deep forest prediction model is obtained. The prediction results of the optimized deep forest prediction model for the methane gas content in the next three days are as follows Figure 4 As shown, the determination coefficient is 0.87, the root mean square error is 48.2, and the predicted value is basically consistent with the true value, which shows that the present invention can achieve the prediction accuracy of dissolved gas in transformer oil and effectively improve the prediction effect of traditional machine learning regression model.

[0030] The above only expresses the preferred embodiments of the present invention, and the description is relatively specific and detailed, but it cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that for ordinary technicians in this field, several modifications, improvements and substitutions can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be based on the attached claims.

Claims

1. A method for predicting dissolved gas in transformer oil based on single variable de-sequencing, characterized in that: The steps include: S1: Collect online monitoring data of dissolved gases in transformer oil, including seven characteristic gases: H2, CH4, C2H6, C2H4, C2H2, CO and CO2, and perform outlier processing and missing data filling on the time series sample data of the above characteristic gases; S2: Use sliding window technology to extract features from historical data of dissolved gas in oil and convert time series sample data into non-time series feature matrix samples; S3: Divide the non-time series feature matrix samples into training sets and test sets, build a deep forest prediction model, and use the tree structure probability density estimation algorithm to optimize the parameters of the deep forest prediction model; S4: Evaluate the performance of the deep forest prediction model, predict the dissolved gas content in the oil at future times, and optimize the operation and maintenance strategy of the transformer in advance.

2. The method for predicting dissolved gas in transformer oil based on single variable de-sequencing according to claim 1, characterized in that: The outlier processing described in S1 uses a box plot to identify and process outliers in the time series sample data of dissolved gas in transformer oil. The processing rules are as follows: ， In the formula, T is the outlier in the time series sample data, Q u and Q L are the upper quartile and the lower quartile, respectively, Q c Q u and Q L The numerical difference.

3. The method for predicting dissolved gas in transformer oil based on single variable de-sequencing according to claim 1, characterized in that: The missing data filling described in S1 is performed on the time series sample data of dissolved gas in transformer oil based on the KNN interpolation method of neighboring points. The specific steps are as follows: S1.1: Use cross-validation method to select the number of neighboring points K; S1.2: Use the Euclidean distance to calculate the distance between the target point and all other points. The calculation formula is as follows: ， In the formula, (x i , y i ) are the coordinates of the neighboring points and the target point in the i-th dimension, and m is the data dimension; S1.3: Select the K data points closest to the target point, assign a weight to each neighboring point in inverse proportion to the Euclidean distance, and calculate the weighted average of all neighboring points as the filling value for the missing data.

4. The method for predicting dissolved gas in transformer oil based on single variable de-sequencing as claimed in claim 1, characterized in that: S2 uses sliding window technology to extract features from historical data of dissolved gas in oil and converts time series sample data into non-time series feature matrix. The specific steps are as follows: S2.1: Determine the sliding window size and stride based on the time series sample data of dissolved gas in transformer oil; S2.2: Perform sliding window processing on the time series samples of each characteristic gas, and calculate the statistical characteristics of the gas in each window, including the mean, maximum, minimum and standard deviation; S2.3: Integrate the features extracted from each window to generate a feature matrix, where each row represents the features extracted from a time window and each column represents a feature.

5. The method for predicting dissolved gas in transformer oil based on single variable de-sequencing according to claim 1, characterized in that: The deep forest prediction model is constructed in S3. The specific steps are as follows: S3.1: Based on the multi-granularity scanning module, feature extraction is performed on the input non-time-series feature matrix samples. Assuming that the feature dimension is l, the scanning size is q, and the step size is k, the multi-granularity scanning module extracts a sub-feature vector of (lq) / k+1 dimensions and trains it as the input of the random forest and completely random forest estimators. Each estimator outputs an α-dimensional probability vector. All output vectors are concatenated in series, and the dimension of the enhanced feature vector obtained is: ， S3.2: The feature vector obtained by the multi-granularity scanning module is used as the input of the cascade forest, which is passed and trained step by step in each cascade layer, and the predicted probability vector and the input feature vector are combined in series; the determination coefficient R 2 The evaluation index for judging the expansion of cascade forest is calculated as follows: ， In the formula, n is the number of prediction points; is the predicted value of gas; i is the true value of gas; is the average value; the coefficient of determination R 2 The value is less than or equal to 1, R 2 The larger the value of indicates, the better the deep forest prediction model fits the input feature data; S3.3: If , the cascade forest expansion is stopped, and the final prediction value of the deep forest prediction model for dissolved gas in transformer oil is output.

6. The method for predicting dissolved gas in transformer oil based on single variable de-sequencing according to claim 1, characterized in that: In S3, the tree structure probability density estimation algorithm is used to optimize the parameters of the deep forest prediction model. The specific steps are as follows: S3-1: Determine the parameter space of the deep forest prediction model, including the number of trees, the depth of the trees, and the minimum number of samples per leaf node; S3-2: The improvement expectation is used to measure the extent to which the loss value corresponding to each selected parameter can be reduced. The improvement expectation calculation formula is as follows: ， In the formula, is the conditional probability distribution; v is the sampling value; v′ is the optimal sampling threshold; u is any value in the configuration space parameter; S3-3: Select the parameter combination with the highest probability density and calculate the maximum expected value as the candidate for the next iteration; S3-4: Repeat steps S3-1 to S3-3 until the set number of iterations is reached or the stop condition is met.

7. The method for predicting dissolved gas in transformer oil based on single variable de-sequencing as claimed in claim 5, characterized in that: In S4, the performance of the deep forest prediction model is evaluated using evaluation indicators including the determination coefficient R 2 and the root mean square error R MSE , where R MSE The calculation formula is as follows: ， Based on the deep forest prediction model, the dissolved gas content in the oil is predicted at future times to achieve the goal of optimizing the transformer operation and maintenance strategy in advance.

Images