A transformer fault early warning method based on multivariate time series

Abstract

The application discloses a transformer fault early warning method based on multivariate time series. First, data preprocessing is performed on the online monitoring time series sample of dissolved gas in transformer oil; then, the anti-short-circuit current and load time series are taken as exogenous variables, a vector autoregressive moving average model based on the multivariate time series of the dissolved gas in the oil is constructed, and model parameters are estimated to predict the content of the dissolved gas in the transformer oil at a future time; finally, the predicted dissolved gas in oil data is input into a transformer fault diagnosis model based on gradient boosting tree to realize early fault warning of the transformer. The application can reflect the time sequence characteristics of the dissolved gas in the oil, fully consider the connection between feature gases and the influence of external factors, and make a more explicit statistical test on the effectiveness of the prediction model, thereby improving the prediction accuracy of the online monitoring data of the dissolved gas in the transformer oil and providing reliable guarantee for the fault warning of the transformer.

Description

Technical Field

[0001] This invention relates to the field of power equipment monitoring technology, specifically to a transformer fault early warning method based on multivariate time series. Background Technology

[0002] Early prediction of internal faults in oil-filled electrical equipment such as transformers, preventing accidents before they occur, is crucial for the safe operation of the power system and the stable development of the national economy. Based on dissolved gas analysis methods in transformer insulating oil, periodic diagnosis of internal faults in operating transformers has become an important means of transformer condition assessment and maintenance. Simultaneously, predicting the changing trends of dissolved gas content in the transformer oil monitored online can help determine the transformer's operating conditions in advance, which is of great significance for promoting the shift from routine maintenance to predictive maintenance of internal defects.

[0003] For the problem of predicting dissolved gases in transformer oil, existing methods can be roughly divided into the following categories: (1) Oil dissolved gas prediction models based on machine learning or deep learning, such as least squares method, support vector machine, neural network, etc. The biggest feature of these methods is that they continuously optimize and iterate the model based on the posterior results to achieve the prediction of the content of characteristic gases. However, their mathematical models have a "black box" problem and are difficult to explain through clear physical interpretation and hypothesis testing; (2) Based on univariate time series models, this method extracts information from the historical data and sample autocorrelation of the original sequence to establish a prediction model and achieve the trend prediction of single gases. Due to the complex relationship between different dissolved gases in oil and the fact that the gas content is easily affected by the on-site environmental factors, the model is difficult to guarantee the accuracy and stability of gas prediction. At the same time, the commonly used transformer fault diagnosis methods on-site, such as the three ratios or David's triangle, also have a large diagnostic bias in the diagnostic results of the predicted oil dissolved gas content.

[0004] With the increasing number of transformers and the growing demand for higher reliability of power equipment, there is a need for an accurate and reliable online monitoring and prediction method for dissolved gases in transformer oil to accurately predict transformer faults. Summary of the Invention

[0005] To address the problems existing in the prior art, the present invention aims to provide a transformer fault early warning method based on multivariate time series. This method constructs a multivariate time series prediction model based on dissolved gases in oil and performs fault diagnosis on the gas prediction results, thereby accurately realizing transformer fault early warning.

[0006] To achieve this objective, the present invention employs the following technical solution: a transformer fault early warning method based on multivariate time series analysis, comprising the following steps:

[0007] S1: Convert online monitoring data such as dissolved gas in transformer oil, short-circuit current, and load into time series samples and perform data preprocessing.

[0008] S2: Construct a multivariate time series VARMA model based on dissolved gases in oil, and use short-circuit current and load as exogenous variables of the model.

[0009] S3: Identify the VARMA model to which the sample time series belongs, estimate and statistically test the hyperparameters of the model, and predict the content of dissolved gas in transformer oil at future times.

[0010] S4: Establish a transformer fault diagnosis model based on gradient boosting tree, provide early fault warning based on the predicted dissolved gas content in the oil, and strengthen corresponding maintenance measures based on the diagnosis results.

[0011] Furthermore, in S1, the time series samples undergo data preprocessing, specifically as follows:

[0012] S1.1: Collect and organize synchronous historical data of transformer online monitoring, such as dissolved gases in oil, short-circuit current, and load, for at least one month to obtain time series samples. The dissolved gases in transformer oil mainly include seven characteristic gases: H2, CH4, C2H6, C2H4, C2H2, CO, and CO2.

[0013] S1.2: Use box plots to identify and process outliers in time series samples. If the feature value is greater than Q... U +1.5Q C The eigenvalue is Q. U If the eigenvalue is less than Q L +1.5Q C The eigenvalue is Q. L Among them, Q U Q is the upper quartile. L Q is the lower quartile. C For Q U and Q L The numerical difference, i.e., the quartiles.

[0014] S1.3: Based on the distribution of missing data, the sample data is piecewise fitted using a Lagrange interpolation polynomial to estimate the missing values. The calculation formula is as follows:

[0015]

[0016] In the formula, x represents missing data in the time series sample; (x0,y0), (x1,y1), ..., (x n ,y n (n+1) data points selected for adjacent sample points of missing values.

[0017] Furthermore, the VARMA model based on multivariate time series constructed in S2, compared to the univariate time series model ARIMA, fully considers the relationships between dissolved gases in different oils and the influence of load and short-circuit current on the gases. The VARMA model combines the characteristics of vector autoregression (VAR) and vector moving average (VMA), and the calculation formula is as follows:

[0018] Y t =v+φ1Y t-1 +…+φ p Y t-p +Θ1ε t-1 +…+Θ q ε q-1 +ε t +BX t

[0019] In the formula, Y t ε represents the dissolved gas in the oil at time t; v is a constant vector; ε t The white noise error at time t; φ1, φ2, ..., φ p For autoregressive parameters; Θ1,Θ2,…,Θ q q represents the moving average parameters; p is the autoregression order, q is the moving average order; X t B represents the exogenous variable at time t; B is the effect of the exogenous variable on the gas.

[0020] Furthermore, in S3, the specific steps for identifying the VARMA model to which the sample time series belongs are as follows:

[0021] S3.1: The autocorrelation coefficient measures the correlation between the dissolved gas in the oil at the current time point and the observed values ​​at any historical time point. The calculation formula is as follows:

[0022]

[0023] S3.2: The partial correlation coefficient measures the magnitude of the direct correlation between dissolved gas in oil at the current time point and the observed values ​​at any historical time point. The calculation formula is as follows:

[0024]

[0025]

[0026] In the formula, It is the variance function. Let φ be the autocovariance function. kk This is the partial correlation coefficient.

[0027] S3.3: Input the dissolved gas time series in transformer oil into the VARMA model and plot the autocorrelation (ACF) and partial correlation (PACF) plots. When the ACF plot shows a tail and the PACF plot shows a truncated state, the current time series is suitable for the VAR model, and the lag order of the truncated PACF plot is the ideal value of the hyperparameter p; when the PACF plot shows a tail and the ACF plot shows a truncated state, the current time series is suitable for the VMA model, and the lag order of the truncated ACF plot is the ideal value of the hyperparameter q; when neither the ACF nor the PACF plot shows a tail, the current time series is suitable for the VARMA model. The VARMA type is determined by setting the values ​​of p, q, and d. If the VAR or VMA model is selected, the model order d is 0.

[0028] Furthermore, when it is determined that the hyperparameters p and q are not both 0, d = [0, 1, 2, 3] is set, an ACF image is plotted, and the standard deviation in the sample sequence is calculated. The optimal value of d is selected as the ACF value with a lag of 1 that is positive, has a small variance and noise.

[0029] Furthermore, in S3, the hyperparameters of the model are estimated and statistically tested. The specific steps are as follows:

[0030] S3-1: Selecting the Akaike Information Criterion V AIC The formula for calculating the model quality evaluation index is as follows:

[0031] V AIC = -2ln(L) + 2k

[0032] In the formula, L is the maximum likelihood estimate; k is the parameter being estimated; and V AIC The smaller the value, the better the model's predictive performance.

[0033] S3-2: Set p = [1,2,3], q = [1,2,3], construct the VARMA(p,d,q) model, exhaustively enumerate the combinations of (p,q) and calculate the corresponding V. AIC Values, select the smallest V in sequence AIC The values ​​(p,q) are used as estimated values ​​of the model parameters.

[0034] S3-3: Perform coefficient variable tests, residual correlation tests, residual variance tests, and residual normality tests on the residual sequence of the selected VARMA(p,d,q) model. If the coefficient variable test value is less than 0.05, and the residual correlation test, residual variance test, and residual normality test values ​​are all greater than 0.05, then the residual distribution of the final output model is close to white noise, indicating that the model can predict the dissolved gas content in transformer oil well. Otherwise, return to step S3-2 to consider changing the hyperparameters of the model.

[0035] Furthermore, S4 provides early warning of transformer faults based on predicted dissolved gases in the oil. The specific steps are as follows:

[0036] S4.1: A transformer fault diagnosis model based on gradient boosting tree is constructed using the internationally recognized IEC T 10 transformer sample dataset;

[0037] S4.2: Optimize the gradient boosting tree model using the Bayesian algorithm to obtain the optimal diagnostic model;

[0038] S4.3: Use the dissolved gas in transformer oil predicted by the VARMA model as input to the trained diagnostic model to achieve early fault warning of the transformer.

[0039] The beneficial effects of this invention are:

[0040] The present invention provides a transformer fault early warning method based on multivariate time series, which can not only reflect the temporal characteristics of dissolved gases in oil, but also fully consider the relationship between different characteristic gases and the influence of external factors, improve the prediction accuracy of online monitoring data of dissolved gases in transformer oil, and provide a reliable guarantee for early fault warning of transformers. Attached Figure Description

[0041] Figure 1 This is a flowchart of the transformer fault early warning method based on multivariate time series in this invention.

[0042] Figure 2 This is a flowchart of parameter estimation for the VARMA model in an embodiment of the present invention.

[0043] Figure 3 This is the diagnostic result based on the gradient boosting tree model in this embodiment of the invention.

[0044] Figure 4 This is the prediction result of the CH4 gas time series in the embodiments of the present invention. Detailed Implementation

[0045] The present invention will be further described below with reference to the embodiments. It should be noted that the following embodiments are only used to further illustrate the present invention and should not be construed as limiting the scope of protection of the present invention. Some non-essential improvements and adjustments made by those skilled in the art based on the above-described invention are still within the scope of protection of the present invention.

[0046] like Figure 1 As shown, a specific embodiment of the present invention provides a transformer fault early warning method based on multivariate time series, including the following steps:

[0047] S1: Convert online monitoring data such as dissolved gas in transformer oil, short-circuit current, and load into time series samples and perform data preprocessing.

[0048] In this embodiment, synchronous historical data from transformer online monitoring, including dissolved gases in transformer oil, short-circuit current, and load, are collected and processed over at least one month to obtain a time series sample. The dissolved gases in transformer oil mainly include seven characteristic gases: H2, CH4, C2H6, C2H4, C2H2, CO, and CO2. Box plots are used to identify and process outliers in the time series sample. If a characteristic value is greater than Q... U +1.5Q C The eigenvalue is Q. U If the eigenvalue is less than Q L +1.5Q C The eigenvalue is Q. L Among them, Q U Q is the upper quartile. L Q is the lower quartile. C For Q U and Q L The numerical difference, i.e., the quartiles. Based on the distribution of missing data, a Lagrange interpolation polynomial is used to piecewise fit the sample data to estimate the missing values. The calculation formula is as follows:

[0049]

[0050] In the formula, x represents missing data in the time series sample; (x0,y0), (x1,y1), ..., (x n ,y n (n+1) data points selected for adjacent sample points of missing values.

[0051] S2: Construct a multivariate time-series VARMA model based on dissolved gases in oil, using short-circuit current and load as exogenous variables. The VARMA model combines the characteristics of vector autoregression (VAR) and vector moving average (VMA), and the calculation formula is as follows:

[0052] Y t =v+φ1Y t-1 +…+φ p Y t-p +Θ1ε t-1 +…+Θ q ε q-1 +ε t +BX t

[0053] In the formula, Y t ε represents the dissolved gas in the oil at time t; v is a constant vector; ε t The white noise error at time t; φ1, φ2, ..., φp For autoregressive parameters; Θ1,Θ2,…,Θ q q represents the moving average parameters; p is the autoregression order, q is the moving average order; X t B represents the exogenous variable at time t; B represents the effect of the exogenous variable on the gas.

[0054] S3: Identify the VARMA model to which the sample time series belongs, estimate and statistically test the hyperparameters of the model, and predict the content of dissolved gas in transformer oil at future times.

[0055] In this embodiment, the hyperparameter estimation verification process based on the VARMA model is as follows: Figure 2 As shown. The VARMA(p,d,q) model was determined based on time series samples from online monitoring of dissolved gases in transformer oil. The specific steps are as follows:

[0056] S3.1: The autocorrelation coefficient measures the correlation between the dissolved gas in the oil at the current time point and the observed values ​​at any historical time point. The calculation formula is as follows:

[0057]

[0058] S3.2: The partial correlation coefficient measures the magnitude of the direct correlation between dissolved gas in oil at the current time point and the observed values ​​at any historical time point. The calculation formula is as follows:

[0059]

[0060]

[0061] In the formula, It is the variance function. Let φ be the autocovariance function. kk This is the partial correlation coefficient.

[0062] S3.3: Input the dissolved gas time series in transformer oil into the VARMA model and plot the autocorrelation (ACF) and partial correlation (PACF) plots. When the ACF plot shows a tail and the PACF plot shows a truncated state, the current time series is suitable for the VAR model, and the lag order of the truncated PACF plot is the ideal value of the hyperparameter p; when the PACF plot shows a tail and the ACF plot shows a truncated state, the current time series is suitable for the VMA model, and the lag order of the truncated ACF plot is the ideal value of the hyperparameter q; when neither the ACF nor the PACF plot shows a tail, the current time series is suitable for the VARMA model. The VARMA type is determined by setting the values ​​of p, q, and d. If the VAR or VMA model is selected, the model order d is 0.

[0063] VARMA(p,0,0) VAR model VARMA(0,0,q) VMA model VARMA(p,d,q) VARMA model

[0064] When the hyperparameters p and q are both non-zero, d = [0, 1, 2, 3] is set, an ACF graph is plotted, and the standard deviation in the sample sequence is calculated. The optimal value of d is selected as the ACF value with a lag of 1 that is positive, has low variance, and low noise. The specific steps for determining the parameters p and q and performing statistical tests on the VARMA model are as follows:

[0065] S3-1: Selecting the Akaike Information Criterion V AIC The formula for calculating the model quality evaluation index is as follows:

[0066] V AIC = -2ln(L) + 2k

[0067] In the formula, L is the maximum likelihood estimate; k is the parameter being estimated; and V AIC The smaller the value, the better the model's predictive performance.

[0068] S3-2: Set p = [1,2,3], q = [1,2,3], construct the VARMA(p,d,q) model, exhaustively enumerate the combinations of (p,q) and calculate the corresponding V. AIC Values, select the smallest V in sequence AIC The values ​​(p, q) corresponding to the values ​​are used as the estimated values ​​of the model parameters.

[0069] S3-3: Perform coefficient variable tests, residual correlation tests, residual variance tests, and residual normality tests on the residual sequence of the selected VARMA(p,d,q) model. If the coefficient variable test value is less than 0.05, and the residual correlation test, residual variance test, and residual normality test values ​​are all greater than 0.05, then the residual distribution of the final output model is close to white noise, indicating that the model can predict the dissolved gas content in transformer oil well. Otherwise, return to step S3-2 to consider changing the hyperparameters of the model.

[0070] S4: Establish a transformer fault diagnosis model based on gradient boosting tree, provide early fault warning based on the predicted dissolved gas content in the oil, and strengthen corresponding maintenance measures based on the diagnosis results.

[0071] In this embodiment, the transformer operating status is divided into normal and abnormal, with abnormal status including five fault types: medium-low temperature overheating (T1), high temperature overheating (T2), partial discharge (PD), low-energy discharge (D1), and high-energy discharge (D2). Using the internationally recognized IEC T 10 transformer sample dataset and field chromatographic data, a transformer fault diagnosis model based on gradient boosting trees is constructed. The gradient boosting tree model is optimized using a Bayesian algorithm to obtain the optimal diagnostic model. The dissolved gas in the transformer oil predicted by the VARMA model is used as input to the trained diagnostic model to achieve early fault warning for the transformer. The transformer fault diagnosis results based on gradient boosting trees are as follows: Figure 3 As shown, the diagnostic results of the optimized gradient boosting tree are distributed diagonally along the confusion matrix, with both normal and medium-low temperature overheating diagnostic accuracies reaching 100%. The gradient boosting tree model achieved a diagnostic accuracy of 93.3% on 150 validation samples, indicating that the proposed diagnostic model can meet the accuracy requirements for predicting transformer faults in the field.

[0072] Verification Example: Using historical data from 100 days of online monitoring of dissolved gases in the oil of a transformer as an example, predict the dissolved gas content in the oil within the next month, and perform transformer fault diagnosis analysis to verify the effectiveness of the method.

[0073] Outlier handling and missing data imputation were performed on the time series data of dissolved gas in transformer oil, load, and short-circuit current withstand capability. The time series samples of dissolved gas in transformer oil were used as endogenous variables, and the time series samples of transformer load and short-circuit current withstand capability were used as exogenous variables, input into a VARMA model based on multivariate time series. Based on these sample series, ACF and PACF plots were plotted, and the type of prediction model was analyzed. The results show that both the ACF and PACF plots exhibit neither tailing nor truncation, and the ACF value with a lag of 1 for d>1 is negative, with low variance and noise. Therefore, the final confirmed prediction model is VARMA(p,0,q). Setting p=[1,2,3] and q=[1,2,3], the V values ​​corresponding to all (p,q) combinations were calculated. AIC Value. When p=1 and q=2, V AIC The value is the smallest, and it passes the coefficient variable test, residual correlation test, residual variance test, and residual normality test. The model residuals follow white noise analysis, and the modeling is complete. The dissolved gas values ​​predicted by the VATMA(1,0,2) model for the seven types of oil are in high agreement with the actual values, such as... Figure 4 This is the prediction result for the CH4 gas time series. The dissolved gas in transformer oil predicted by the VARMA model is input into a transformer fault diagnosis model based on a gradient boosting tree, ultimately predicting medium-low temperature overheating. Based on the diagnosis results, power maintenance personnel should closely monitor the recent trend of dissolved gas changes in the oil and perform pre-maintenance procedures on the transformer.

[0074] The above description merely illustrates preferred embodiments of the present invention, and while the description is relatively specific and detailed, it should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications, improvements, and substitutions without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this patent should be determined by the appended claims.

Claims

1. A transformer fault early warning method based on multivariate time series, characterized in that: Includes the following steps: S1: Convert online monitoring data of dissolved gas in transformer oil, short-circuit current, and load into time series samples and perform data preprocessing; S2: Construct a multivariate time series VARMA model based on dissolved gases in oil, using short-circuit current and load as exogenous variables. The VARMA model combines the characteristics of vector autoregression (VAR) and vector moving average (VMA), and the calculation formula is as follows: In the formula, Y t ε represents the dissolved gas in the oil at time t; v is a constant vector; ε t The white noise error at time t; φ1, φ2, ..., φ p These are autoregressive parameters; q represents the moving average parameters; p is the autoregression order, q is the moving average order; X t B represents the exogenous variable at time t; B is the effect of the exogenous variable on the gas. S3: Identify the VARMA model to which the sample time series belongs, estimate and statistically test the model's hyperparameters, and predict the dissolved gas content in transformer oil at future times. The specific steps for estimating and statistically testing the model's hyperparameters are as follows: S3-1: Select the Akaike Information Criterion V AIC The calculation formula for this model quality evaluation indicator is as follows: In the formula, L is the maximum likelihood estimate; k is the parameter being estimated; and V is the maximum likelihood estimate. AIC The smaller the value, the better the model's prediction performance; S3-2: Set p=[1,2,3], q=[1,2,3], construct the VARMA(p, d, q) model, exhaustively enumerate the combinations of (p,q) and calculate the corresponding V AIC Values, select the smallest V in sequence AIC The values ​​(p, q) are used as model parameter estimates; S3-3: Perform coefficient variable tests, residual correlation tests, residual variance tests, and residual normality tests on the residual sequences of the selected VARMA(p, d, q) model respectively; If the coefficient variable test value is less than 0.05, and the residual correlation test, residual variance test, and residual normality test values ​​are all greater than 0.05, then the residual distribution of the final output model is close to white noise, indicating that the model can predict the content of dissolved gas in transformer oil well. Otherwise, return to step S3-2 to consider changing the hyperparameters of the model. S4: Establish a transformer fault diagnosis model based on gradient boosting tree, provide early fault warning based on the predicted dissolved gas content in the oil, and strengthen corresponding maintenance measures based on the diagnosis results.

2. The transformer fault early warning method based on multivariate time series as described in claim 1, characterized in that: In S1, the time series samples undergo data preprocessing, and the specific steps are as follows: S1.1: Collect and organize synchronous historical data of transformer online monitoring of dissolved gases in oil, short-circuit current and load for at least one month to obtain time series samples, wherein the dissolved gases in transformer oil include 7 characteristic gases: H2, CH4, C2H6, C2H4, C2H2, CO and CO2. S1.2: Use box plots to identify and process outliers in time series samples. If the feature value is greater than Q... U +1.5Q C The eigenvalue is Q. U If the eigenvalue is less than Q L -1.5Q C The eigenvalue is Q. L ; Among them, Q U Q is the upper quartile. L Q is the lower quartile. C For Q U and Q L The numerical difference, i.e., the quartiles; S1.3: Based on the distribution of missing data, the sample data is piecewise fitted using a Lagrange interpolation polynomial to estimate the missing values. The calculation formula is as follows: ， In the formula, x represents missing data in the time series sample; (x0, y0), (x1, y1), ..., (x n , y n (n+1) data points selected for adjacent sample points of missing values.

3. The transformer fault early warning method based on multivariate time series as described in claim 1, characterized in that: The specific steps for identifying the VARMA model to which a sample time series belongs in S3 are as follows: S3.1: The autocorrelation coefficient measures the correlation between the dissolved gas in the oil at the current time point and the observed values ​​at any historical time point. The calculation formula is as follows: ， S3.2: The partial correlation coefficient measures the magnitude of the direct correlation between dissolved gas in oil at the current time point and the observed values ​​at any historical time point. The calculation formula is as follows: ， ， In the formula, It is the variance function. Let φ be the autocovariance function. kk This is the partial correlation coefficient; S3.3: Input the dissolved gas time series in transformer oil into the VARMA model and plot the ACF and PACF plots respectively; when the ACF plot shows a tail and the PACF plot shows a truncated state, the current time series is suitable for the VAR model, and the lag order of the truncated PACF plot is the ideal value of the hyperparameter p; when the PACF plot shows a tail and the ACF plot shows a truncated state, the current time series is suitable for the VMA model, and the lag order of the truncated ACF plot is the ideal value of the hyperparameter q; when neither the ACF nor the PACF plot shows a tail, the current time series is suitable for the VARMA model; determine the type of VARMA by setting the values ​​of p, q, and d; if it is determined to use the VAR or VMA model, the model order d is 0.

4. The transformer fault early warning method based on multivariate time series as described in claim 3, characterized in that: When the hyperparameters p and q are both non-zero, set d=[0, 1, 2, 3], plot the ACF image, calculate the standard deviation in the sample sequence, and select the ACF value with a lag of 1 as the order with the smallest positive variance and noise as the optimal value of d.

5. The transformer fault early warning method based on multivariate time series as described in claim 1, characterized in that: The specific steps for S4 are as follows: S4.1: A transformer fault diagnosis model based on gradient boosting tree is constructed using the internationally recognized IEC T 10 transformer sample dataset; S4.2: The gradient boosting tree model is optimized using the Bayesian optimization algorithm to obtain the optimal transformer fault diagnosis model; S4.3: Use the dissolved gas in transformer oil predicted by the VARMA model as input to the trained diagnostic model to achieve early fault warning of the transformer.

Images