Surge arrester residual service life prediction method based on digital-analog linkage

Abstract

The application discloses a lightning arrester residual service life prediction method based on digital-analog linkage, comprising the following steps: obtaining real-time monitoring data of a lightning arrester to be predicted; preprocessing the real-time monitoring data of the lightning arrester to be predicted to obtain input data; inputting the input data into a lightning arrester residual service life prediction model based on an XGBoost algorithm which is constructed in advance, and outputting the predicted residual service life of the lightning arrester; obtaining the confidence interval of the predicted residual service life of the lightning arrester based on a target Wiener process model; taking the output residual service life of the lightning arrester as a predicted value, and judging whether the predicted value exceeds a third harmonic current setting value or not. The application realizes the precision, intelligence and stabilization of the lightning arrester residual service life prediction by combining a physical model with a data-driven model, effectively makes up for the shortcomings of a traditional feature determination method, and simultaneously improves the adaptability of a data-driven method in a complex power grid environment.

Description

Technical Field

[0001] This invention relates to the field of surge arrester service life prediction, and more specifically, to a method for predicting the remaining service life of surge arresters based on digital-analog linkage. Background Technology

[0002] Metal oxide surge arresters (MOSAs) play a crucial role in protecting power grid equipment from surge voltage damage, especially in modern power systems where MOSA health monitoring and lifetime prediction have become key technologies for ensuring the long-term stable operation of power equipment. However, due to the complexity of the power grid operating environment, the variability of external disturbances, and the nonlinear evolution of equipment aging processes, accurate prediction of MOSA lifetime still faces many challenges.

[0003] Currently, the aging process of MOSA (Modular Equipment for Assay) is typically assessed by monitoring key status parameters such as leakage current and temperature. These indicators reflect the health status and degradation trend of the equipment. Traditional prediction methods mainly rely on empirical models or statistical methods, such as Adaptive Neural Fuzzy Inference Systems (ANFIS) and Support Vector Regression (SVR). However, these methods often exhibit insufficient prediction accuracy and poor adaptability when faced with multi-source heterogeneous data, high-dimensional nonlinear relationships, and complex environmental interference. Furthermore, traditional prediction methods struggle to dynamically adjust the prediction model, making the prediction results susceptible to data noise and affecting the stability and reliability of long-term monitoring.

[0004] In recent years, with the rapid development of data-driven methods and machine learning technologies, using efficient regression models to accurately predict the lifespan of MOSAs has become a research hotspot. The XGBoost algorithm, in particular, with its powerful modeling capabilities and automatic feature selection, can better capture equipment aging characteristics in complex data environments, improving prediction accuracy and generalization ability. However, relying solely on a single data-driven method still has significant shortcomings: 1) As a pure data-driven model, the XGBoost algorithm, while able to learn equipment aging trends from historical data, has its prediction performance heavily dependent on data quality, and may misjudge or lag non-stationary aging processes; 2) A single third harmonic feature cannot fully characterize the physical aging mechanism of surge arresters and is easily affected by environmental interference such as power grid harmonic pollution and temperature fluctuations, leading to a decrease in model stability and robustness. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method for predicting the remaining service life of surge arresters based on digital-analog linkage. By combining a data-driven model with a physical model, a high-precision and highly stable surge arrester service life prediction model is constructed, thereby improving the model's adaptability and robustness in complex power grid environments.

[0006] The technical solution adopted in this invention is:

[0007] A method for predicting the remaining service life of surge arresters based on digital-analog linkage includes:

[0008] S1. Obtain real-time monitoring data of the surge arrester to be predicted;

[0009] S2. Preprocess the real-time monitoring data of the surge arrester to be predicted to obtain the input data;

[0010] S3. Input the input data into the pre-built arrester remaining service life prediction model based on the XGBoost algorithm, and output the predicted value of the arrester remaining service life.

[0011] S4. Based on the target Wiener process model, output the confidence interval of the predicted value of the remaining service life of the surge arrester;

[0012] Specifically, S41, collect the degradation data of the surge arrester and preprocess it to obtain the target degradation data, wherein the degradation data includes the historical remaining service life data, aging rate data and degradation increment data of the surge arrester;

[0013] S42. Construct the Wiener process model and perform parameter estimation to obtain the target Wiener process model;

[0014] S43. Based on the predicted remaining service life of the surge arrester, the aging rate and degradation increment of the surge arrester are calculated.

[0015] S44. Input the predicted remaining service life of the surge arrester, aging rate and degradation increment as target degradation data into the target Wiener process model to obtain the confidence interval of the predicted remaining service life of the surge arrester, which is used to judge the reliability of the predicted remaining service life of the surge arrester.

[0016] The parameter estimation in S42 includes:

[0017] 1) Set initial conditions and iteration termination criteria: Define the drift coefficient and diffusion coefficient Initialize using the average or empirical value of historical data; set the maximum number of iterations. and convergence threshold ;

[0018] 2) Iterative optimization of the parameters of the Wiener process model based on the degradation data of the surge arrester:

[0019] Based on the degradation data of the surge arrester, the current parameters are calculated using the maximum likelihood estimation method. The likelihood function is as follows:

[0020] (10)

[0021] In equation (10), For degradation increment, For time intervals, For aging rate, This represents the random fluctuation term during the aging process; the parameters are updated using the gradient ascent method:

[0022] (11)

[0023] In equation (11), Let be the random fluctuation term in the aging process during the (k+1)th iteration. Let be the random fluctuation term in the aging process during the k-th iteration. For learning rate, As a log-likelihood function, it needs to be dynamically adjusted to achieve a balance between convergence speed and stability;

[0024] 3) Convergence judgment and dynamic adjustment: If the parameter change satisfies and reaches the maximum number of iterations K, then the iteration ends and proceeds to the next step; otherwise, adjust the learning rate according to the current parameter estimation error: use the adaptive learning rate algorithm Adagrad, and repeat step 2) to further optimize the parameters;

[0025] 4) Output the final optimized parameters and validate the model: ① Verify whether the residual sequence conforms to the independent and identically distributed assumption through degradation path residual analysis; ② Calculate the confidence interval coverage through Monte Carlo simulation to ensure that the confidence interval of the parameter estimate can effectively cover the true value;

[0026] S5. Based on the predicted and set values ​​of the remaining service life of the surge arrester, determine the working status of the surge arrester and formulate corresponding maintenance strategies.

[0027] Furthermore, the real-time monitoring data of the surge arrester to be predicted in S1 includes the surge arrester's leakage current, ambient temperature, vibration amplitude, and operating voltage. The specific acquisition process is as follows:

[0028] 1) Leakage current acquisition: The leakage current signal of the surge arrester is acquired using a magnetic current sensor with a resolution of 0.01mA and a portable digital oscilloscope with a sampling rate of 1Msamples / s. The data format is set to time series.

[0029] 2) Collect ambient temperature: The temperature inside the surge arrester is collected as the ambient temperature by a high-precision temperature sensor installed on the surface of the surge arrester, and the sampling frequency is 1 time / hour.

[0030] 3) Vibration amplitude acquisition: The vibration signal of the surge arrester is acquired by a piezoelectric vibration sensor with a sensitivity of 0.01g installed on the base of the surge arrester. The sampling rate is 10kHz. The average vibration amplitude is taken after the data is filtered.

[0031] 4) Collect operating voltage: Measure the actual operating voltage across the surge arrester using a voltage sensor and use the daily average value of the actual operating voltage as the operating voltage.

[0032] Furthermore, S2 includes:

[0033] S21. Based on the Fast Fourier Transform (FFT), feature extraction is performed on the acquired leakage current signal to obtain the third harmonic of the leakage current. The calculation formula is as follows:

[0034] (1)

[0035] In equation (1), The third harmonic of the leakage current. These are the real Fourier coefficients of the third harmonic. The imaginary Fourier coefficients of the third harmonic;

[0036] S22. Normalize the third harmonic of the leakage current, ambient temperature, vibration amplitude, and operating voltage to obtain the input data. The calculation formula for normalization is as follows:

[0037] (2)

[0038] In equation (2), The normalized value is the normalized value. The original data in the dataset, The minimum value in the dataset. The maximum value in the dataset; the dataset is a set of the third harmonic of the leakage current, ambient temperature, vibration amplitude, and operating voltage;

[0039] S23. Define the input data format as follows:

[0040] (3)

[0041] In equation (3), The third harmonic of the leakage current. For ambient temperature, The amplitude of the vibration. This is the operating voltage.

[0042] Furthermore, the pre-built XGBoost algorithm-based arrester remaining service life prediction model in S3 includes the following steps:

[0043] S31. Obtain multi-dimensional surge arrester data:

[0044] 1) Acquire and preprocess historical surge arrester data at different operating times, including leakage current data, ambient temperature data, vibration amplitude data, and operating voltage data;

[0045] 2) Based on Fast Fourier Transform (FFT), feature extraction is performed on the leakage current data to obtain the third harmonic data of the leakage current;

[0046] 3) The target data were obtained by standardizing the third harmonic current, vibration amplitude, operating voltage, and ambient temperature using Pandas.

[0047] 4) Divide the target data into a 70% training set, a 20% validation set, and a 10% test set;

[0048] S32. Construct a surge arrester remaining service life prediction model based on gradient boosting tree, and train it to learn the nonlinear relationship between the third harmonic of leakage current, vibration amplitude, ambient temperature, and operating voltage and the remaining service life of the surge arrester. Specifically, this includes the following:

[0049] 1) Determine the objective function:

[0050] (4)

[0051] in,

[0052] (5)

[0053] (6)

[0054] In equations (4) to (6), The objective function is... The loss function; For regularization terms; The actual value; This is a predicted value; This represents the number of leaf nodes in the current tree. The weight of the leaf node is the square of its weight. , For regularization parameters;

[0055] 2) Construct a decision tree and calculate the gradient. and the split gain of each node The calculation formula is as follows:

[0056] (7)

[0057] (8)

[0058] In equations (7) to (8), , Let be the gradient sums of the left and right child nodes, respectively. , Let be the sum of the second-order gradients of the left and right child nodes, respectively. This is the regularization parameter; if the gain is less than If so, then no split will occur;

[0059] 3) Perform multiple rounds of iterative training, updating the predicted value after each round of training, using the following formula:

[0060] (9)

[0061] In equation (9), For the sample In the Predicted values ​​in rounds of iteration For the sample In the Predicted values ​​in rounds of iteration For learning rate, To input features in the t-th decision tree The corresponding predicted increment value; where the sample For a specific set of data in the training set;

[0062] S33. Input the test set into the trained model and output the preliminary prediction data of the remaining service life of the surge arrester.

[0063] Furthermore, the preprocessing of historical surge arrester data includes:

[0064] 1) Use wavelet denoising or bandpass filtering to remove high-frequency noise in order to eliminate errors caused by power grid interference;

[0065] 2) Remove extreme outlier data through outlier detection methods to improve data reliability;

[0066] 3) All data are organized according to time series to meet the needs of subsequent model training and prediction.

[0067] Furthermore, the multi-round iterative training process of the model is as follows:

[0068] 1) Determine the initial hyperparameters based on the training set and the target error range;

[0069] 2) Define the number of training rounds k=1 and the maximum number of training iterations K;

[0070] 3) Train the model using the training set to obtain preliminary prediction results;

[0071] 4) Dynamically adjust hyperparameters based on the validation set, including learning rate, tree depth, and regularization parameters;

[0072] 5) Determine if the maximum number of training iterations K has been reached: If not, continue to optimize the model's hyperparameters; if training converges or the maximum number of iterations is reached, stop training and save the final optimized model.

[0073] Furthermore, in step 2), the ABC-Gibbs method can also be used: ① construct an approximate posterior distribution and update it alternately through Gibbs sampling; ② use the mean square error between the simulated data and the observed data of the degenerate path as an acceptance criterion to ensure the stability of the parameter estimation.

[0074] Furthermore, S5 includes:

[0075] S51. Based on the predicted remaining service life of the surge arrester obtained from the output and the confidence interval of the predicted remaining service life of the surge arrester, determine whether the predicted remaining service life of the surge arrester obtained from the output is reliable: if reliable, proceed to the next step; otherwise, directly determine the working state as dangerous.

[0076] S52. Determine whether the predicted remaining service life of the surge arrester obtained from the output exceeds the third harmonic setting value of the leakage current: if it exceeds, the working state of the surge arrester is determined to be dangerous and corresponding maintenance strategies are formulated; if it does not exceed, the working state of the surge arrester is determined to be normal and periodic monitoring is sufficient.

[0077] Compared with the prior art, the present invention has the following advantages:

[0078] 1) Use non-invasive measuring devices to ensure the quality of data acquisition;

[0079] 2) Collect the third harmonic, vibration amplitude, ambient temperature and operating voltage of the surge arrester to comprehensively characterize the aging state of the surge arrester from multiple dimensions and improve the accuracy of the prediction model for the remaining service life of the surge arrester.

[0080] 3) The XGBoost algorithm is used to construct a prediction model for the remaining service life of surge arresters. It can not only handle complex nonlinear relationships, but also has the ability to automatically select features and build models efficiently. Compared with traditional feature determination methods, it has higher prediction accuracy and stronger adaptability.

[0081] 4) The Wiener degradation process model is introduced into the constructed surge arrester remaining service life prediction model, which can effectively describe the uncertainty of surge arrester aging and provide the confidence interval for the prediction of the remaining service life of the surge arrester. This makes the prediction of the remaining service life of the surge arrester no longer a single fixed value, but a dynamic evaluation including upper and lower bounds, which can be used for intelligent operation and maintenance decision-making and optimize equipment replacement cycle.

[0082] 5) By combining the physical model with the data-driven model, the physical degradation mechanism data characteristics of the surge arrester are optimized, so that the model can maintain high predictive stability even when the data is incomplete or the operating conditions change greatly.

[0083] 6) This invention achieves more accurate, intelligent and stable prediction of surge arrester lifespan by integrating multi-dimensional data, XGBoost algorithm and Wiener process. It effectively makes up for the shortcomings of traditional feature judgment methods, improves the adaptability of data-driven methods in complex power grid environments, provides a more scientific and reliable lifespan prediction scheme, and provides important technical support for the intelligent operation and maintenance of power grid equipment. Attached Figure Description

[0084] The present invention will be further described below with reference to the accompanying drawings and specific embodiments:

[0085] Figure 1 This is a flowchart illustrating the process in this invention;

[0086] Figure 2 This is a graph showing the trend of the third harmonic variation of the leakage current in this invention;

[0087] Figure 3 This is a graph showing the trend of ambient temperature changes in this invention;

[0088] Figure 4 This is a graph showing the trend of vibration amplitude variation in this invention;

[0089] Figure 5 This is a graph showing the trend of operating voltage variation in this invention;

[0090] Figure 6 This is a flowchart of the training process for the surge arrester remaining service life prediction model based on the XGBoost algorithm in this invention.

[0091] Figure 7 This is the result of optimizing the training accuracy of the surge arrester remaining service life prediction model constructed based on the XGBoost algorithm in this invention;

[0092] Figure 8 This is a schematic diagram of the confidence interval results obtained from the Wiener process model in this invention. Detailed Implementation

[0093] Example

[0094] like Figure 1 As shown, a method for predicting the remaining service life of surge arresters based on digital-analog linkage includes:

[0095] S1. Obtain real-time monitoring data of the surge arrester to be predicted:

[0096] Specifically, the real-time monitoring data of the surge arrester to be predicted includes the surge arrester's leakage current, ambient temperature, vibration amplitude, and operating voltage. The specific acquisition process is as follows:

[0097] 1) Leakage current acquisition: The leakage current signal of the surge arrester is acquired using a magnetic current sensor with a resolution of 0.01mA and a portable digital oscilloscope with a sampling rate of 1Msamples / s. The data format is set to time series.

[0098] 2) Collect ambient temperature: The temperature inside the surge arrester is collected as the ambient temperature by a high-precision temperature sensor installed on the surface of the surge arrester, and the sampling frequency is 1 time / hour.

[0099] 3) Vibration amplitude acquisition: The vibration signal of the surge arrester is acquired by a piezoelectric vibration sensor with a sensitivity of 0.01g installed on the base of the surge arrester. The sampling rate is 10kHz. The average vibration amplitude is taken after the data is filtered.

[0100] 4) Acquiring operating voltage: The actual operating voltage across the surge arrester is measured using a voltage sensor, and the daily average value of the actual operating voltage is used as the operating voltage;

[0101] Table 1 shows some of the real-time monitoring data of the surge arresters to be predicted, collected at different time points:

[0102] Table 1. Real-time monitoring data of some surge arresters to be predicted, collected at different time points.

[0103]

[0104] Note: Leakage current is acquired using a non-invasive measurement method to avoid interference with equipment operation. A magnetic current sensor with a resolution of 0.01 mA is used to measure the time-domain signal of the surge arrester's leakage current, enabling it to capture minute current changes. A portable digital oscilloscope is used for recording to ensure signal acquisition accuracy, with a sampling rate set to 1 Msamples / s to guarantee high data resolution. Vibration amplitude, ambient temperature, and operating voltage are acquired using a direct contact measurement method.

[0105] S2. Normalize the real-time monitoring data of the surge arrester to be predicted to obtain the input data:

[0106] Specifically, S21, based on the Fast Fourier Transform (FFT), the features of the acquired leakage current signal are extracted to obtain the third harmonic of the leakage current, and its calculation formula is as follows:

[0107] (1)

[0108] In equation (1), The third harmonic of the leakage current. These are the real Fourier coefficients of the third harmonic. The imaginary Fourier coefficients of the third harmonic;

[0109] The extracted third harmonic data are shown in Table 2 (retaining two decimal places):

[0110] Table 2. Extracted partial third harmonic data of the leakage current.

[0111]

[0112] S22. Normalize the third harmonic of the leakage current, ambient temperature, vibration amplitude, and operating voltage to obtain the input data. The calculation formula for normalization is as follows:

[0113] (2)

[0114] In equation (2), The normalized value is the normalized value. The original data in the dataset, The minimum value in the dataset. The maximum value in the dataset; the dataset is a set of the third harmonic of the leakage current, ambient temperature, vibration amplitude, and operating voltage;

[0115] The normalized portion of the input data is shown in Table 3:

[0116] Table 3. Partial Input Data After Normalization

[0117]

[0118] S23. Define the input data format as follows:

[0119] (3)

[0120] In equation (3), The third harmonic of the leakage current. For ambient temperature, The amplitude of the vibration. This is the operating voltage;

[0121] Note: After data acquisition, feature extraction is required to obtain key variables characterizing the aging state of surge arresters. The core of feature extraction is calculating the third harmonic of the leakage current, which is an important indicator of the degradation of the surge arrester's insulation performance. After feature extraction, statistical analysis is performed on the third harmonic data under different operating times. For example, for a surge arrester that has been operating for 8 years, its third harmonic is typically distributed between 6μA and 11μA, with a median of 8.5μA. This data can be used to establish a surge arrester aging database and further used as training data for subsequent machine learning models.

[0122] S3. Input the input data into the pre-built arrester remaining service life prediction model based on the XGBoost algorithm, and output the predicted value of the arrester remaining service life.

[0123] Specifically, the pre-built arrester remaining service life prediction model based on the XGBoost algorithm includes the following steps:

[0124] S31. Obtain multi-dimensional surge arrester data:

[0125] 1) Acquire and preprocess historical surge arrester data at different operating times, including leakage current data, ambient temperature data, vibration amplitude data, and operating voltage data;

[0126] 2) Based on Fast Fourier Transform (FFT), feature extraction is performed on the leakage current data to obtain the third harmonic data of the leakage current;

[0127] 3) The target data were obtained by standardizing the third harmonic current, vibration amplitude, operating voltage, and ambient temperature using Pandas.

[0128] 4) Divide the target data into a 70% training set, a 20% validation set, and a 10% test set;

[0129] The historical surge arrester data and preprocessing include:

[0130] ① Use wavelet denoising or bandpass filtering to remove high-frequency noise in order to eliminate errors caused by power grid interference;

[0131] ②Use outlier detection methods to remove extreme outlier data and improve data reliability;

[0132] ③ All data are organized according to time series to meet the needs of subsequent model training and prediction;

[0133] like Figures 2 to 5The figures show the dynamic trends of the third harmonic leakage current, vibration amplitude, ambient temperature, and operating voltage of the surge arrester. Data on these parameters—including third harmonic leakage current, vibration amplitude, ambient temperature, and operating voltage—from the period of operation between 2000 and 2020 were selected as historical surge arrester data. The third harmonic leakage current, vibration amplitude, ambient temperature, and operating voltage interact and jointly determine the aging rate of the surge arrester. The operating voltage fluctuates between 200V and 220V, influenced by factors such as changes in grid load and short-term overvoltages. The ambient temperature generally shows an upward trend, reflecting the heat accumulation effect of long-term operation. The nonlinear growth of the third harmonic leakage current further indicates that the aging process of the surge arrester is not uniform but accelerates during certain periods, possibly due to the amplifying effect of high temperatures or voltage fluctuations. Therefore, relying on a single variable is insufficient to accurately assess the lifespan of a surge arrester; a comprehensive analysis of multiple variables provides a more reliable assessment of its health status.

[0134] S32. Construct a surge arrester remaining service life prediction model based on gradient boosting tree, and train it to learn the nonlinear relationship between the third harmonic of leakage current, vibration amplitude, ambient temperature, and operating voltage and the remaining service life of the surge arrester. Specifically, this includes the following:

[0135] 1) Determine the objective function:

[0136] (4)

[0137] in,

[0138] (5)

[0139] (6)

[0140] In equations (4) to (6), The objective function is... The loss function; For regularization terms; The actual value; This is a predicted value; This represents the number of leaf nodes in the current tree. The weight of the leaf node is the square of its weight. , For regularization parameters;

[0141] 2) Construct a decision tree and calculate the gradient. and the split gain of each node The calculation formula is as follows:

[0142] (7)

[0143] (8)

[0144] In equations (7) to (8), , Let be the gradient sums of the left and right child nodes, respectively. , Let be the sum of the second-order gradients of the left and right child nodes, respectively. This is the regularization parameter; if the gain is less than If so, then no split will occur;

[0145] 3) Perform multiple rounds of iterative training, updating the predicted value after each round of training, using the following formula:

[0146] (9)

[0147] In equation (9), For the sample In the Predicted values ​​in rounds of iteration For the sample In the Predicted values ​​in rounds of iteration For learning rate, To input features in the t-th decision tree The corresponding predicted increment value; where the sample For a specific set of data in the training set;

[0148] like Figure 6 As shown, the multi-round iterative training process of the model is as follows:

[0149] 1) Based on the training set and the target error range, determine the initial hyperparameters: set the initial learning rate to 0.1, the maximum tree depth range to limit the depth of a single tree to prevent overfitting, with a typical value of 6, and the subsample ratio to control the random sampling ratio in each round of training, with a typical value of 0.8.

[0150] 2) Define the number of training rounds k=1 and the maximum number of training iterations K: Training starts from the first round. Set the total number of training rounds, which is determined according to the training set and the target error range. Typical values ​​are 100-500.

[0151] 3) Train the model using the training set to obtain preliminary prediction results: Train the model using the initially set hyperparameters. In each round of training, fit the residual between the current predicted value and the true value to build a new regression tree; optimize the predicted value through gradient update rules to gradually reduce the model error.

[0152] 4) Dynamically adjust hyperparameters based on the validation set, including learning rate, tree depth, and regularization parameters, to improve the model's generalization ability and prevent overfitting or underfitting.

[0153] 5) Determine if the maximum number of training iterations K has been reached: If not, continue to optimize the model's hyperparameters; if training converges or the maximum number of iterations has been reached, stop training and save the final optimized model.

[0154] In addition, after each training round, the training round number k is updated to k+1, and the process returns to the model training step to continue optimizing the model parameters and structure.

[0155] like Figure 7 The figure shows the optimization results of the training accuracy of the surge arrester remaining service life prediction model built based on the XGBoost algorithm. As the number of training rounds increases, the error gradually decreases and stabilizes after 40 rounds, indicating that the model successfully learns the characteristic patterns of surge arrester aging and effectively fits the complex relationship between vibration, ambient temperature, operating voltage, and the remaining service life of the surge arrester. The rapid convergence of the training error demonstrates that the surge arrester remaining service life prediction model built based on the XGBoost algorithm can efficiently capture the nonlinear characteristics of the data and achieve superior prediction results in fewer training rounds, giving it the potential for rapid deployment in actual power grid operation and maintenance.

[0156] S33. Input the test set into the trained model and output the preliminary prediction data of the remaining service life of the surge arrester.

[0157] In addition, after the model training is completed, cross-validation is used to optimize the hyperparameters to ensure that the model has good generalization ability.

[0158] Note: After data preprocessing and feature extraction, a surge arrester remaining service life prediction model based on a combination of data-driven and physical modeling methods is established. This model includes an input layer, a feature extraction layer, an XGBoost modeling layer, a digital-analog linkage optimization layer, and an output layer. A gradient boosting tree-based prediction model is constructed to learn the nonlinear relationship between leakage current, operating voltage, and ambient temperature on the surge arrester's remaining service life. Wiener process modeling is introduced in the digital-analog linkage optimization layer to enhance the stability of the surge arrester remaining service life prediction model.

[0159] S4. Based on the Wiener process model, obtain the confidence interval of the predicted remaining service life of the surge arrester;

[0160] Specifically, S41, collect the degradation data of the surge arrester and preprocess it to obtain the target degradation data, wherein the degradation data includes the historical remaining service life data, aging rate data and degradation increment data of the surge arrester;

[0161] S42. Construct the Wiener process model and perform parameter estimation to obtain the target Wiener process model;

[0162] S43. Based on the predicted remaining service life of the surge arrester, the aging rate and degradation increment of the surge arrester are calculated.

[0163] S44. Input the predicted remaining service life of the surge arrester, aging rate and degradation increment as target degradation data into the target Wiener process model to obtain the confidence interval of the predicted remaining service life of the surge arrester, which is used to judge the reliability of the predicted remaining service life of the surge arrester.

[0164] The parameter estimation in S42 includes:

[0165] 1) Set initial conditions and iteration termination criteria: Define the drift coefficient and diffusion coefficient Initialize using the average or empirical value of historical data; set the maximum number of iterations. and convergence threshold ;

[0166] 2) Iterative optimization of the parameters of the Wiener process model based on the degradation data of the surge arrester:

[0167] Based on the degradation data of the surge arrester, the current parameters are calculated using the maximum likelihood estimation method. The likelihood function is as follows:

[0168] (10)

[0169] In equation (10), For degradation increment, For time intervals, For aging rate, This represents the random fluctuation term during the aging process; the parameters are updated using the gradient ascent method:

[0170] (11)

[0171] In equation (11), Let be the random fluctuation term in the aging process during the (k+1)th iteration. Let be the random fluctuation term in the aging process during the k-th iteration. For learning rate, As a log-likelihood function, it needs to be dynamically adjusted to achieve a balance between convergence speed and stability;

[0172] 3) Convergence judgment and dynamic adjustment: If the parameter change satisfies and reaches the maximum number of iterations K, then the iteration ends and proceeds to the next step; otherwise, adjust the learning rate according to the current parameter estimation error: use the adaptive learning rate algorithm Adagrad, and repeat step 2) to further optimize the parameters;

[0173] 4) Output the final optimized parameters and validate the model: ① Verify whether the residual sequence conforms to the independent and identically distributed assumption through degradation path residual analysis; ② Calculate the confidence interval coverage through Monte Carlo simulation to ensure that the confidence interval of the parameter estimate can effectively cover the true value;

[0174] In addition, in step 2), the ABC-Gibbs method can also be used: ① construct an approximate posterior distribution and update it alternately through Gibbs sampling; ② use the mean square error between the simulated data and the observed data of the degradation path as an acceptance criterion to ensure the stability of the parameter estimation.

[0175] like Figure 8 As shown, the confidence intervals for the predicted remaining service life of surge arresters based on the Wiener process model are represented. The black solid line represents the prediction trend, and the green area represents the 95% confidence interval. The range of the confidence interval gradually expands over time, indicating that the uncertainty of the prediction increases with the extension of future time. This uncertainty modeling method can help power grid maintenance personnel formulate maintenance plans more scientifically, enabling proactive maintenance, avoiding sudden surge arrester failures, and improving the operational reliability of the power grid.

[0176] S5. Based on the predicted and set values ​​of the remaining service life of the surge arrester, determine the working status of the surge arrester and formulate corresponding maintenance strategies.

[0177] Note: The constructed surge arrester remaining service life prediction model is ultimately deployed to the online monitoring system for real-time monitoring of the surge arrester's health status and dynamic updates to the prediction results. Through the intelligent operation and maintenance system, the prediction results are used for equipment maintenance decisions: inputting data into the model for real-time prediction; periodically updating the prediction results and generating operation reports. For surge arresters with predicted service life below the safety threshold, early replacement is recommended to prevent grid failures. Furthermore, by combining historical data, surge arrester replacement cycles can be optimized, and more scientific operation and maintenance plans can be developed, thereby improving the reliability and stability of the power system.

[0178] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the principles and essence of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for predicting the residual service life of a lightning arrester based on digital-analog interaction, characterized by, The method comprises the following steps: S1, obtaining real-time monitoring data of a lightning arrester to be predicted; S2, preprocessing the real-time monitoring data of the lightning arrester to be predicted to obtain input data; S3, inputting the input data into a lightning arrester residual service life prediction model based on an XGBoost algorithm to output a lightning arrester residual service life prediction value; S4, based on a target Wiener process model, outputting a confidence interval of the lightning arrester residual service life prediction value; Specifically, S41, collecting degradation data of the lightning arrester and preprocessing the degradation data to obtain target degradation data, wherein the degradation data comprises historical residual service life data, aging rate data and degradation increment data of the lightning arrester; S42, constructing a Wiener process model and performing parameter estimation to obtain the target Wiener process model; S43, based on the output lightning arrester residual service life prediction value, calculating the aging rate and the degradation increment of the lightning arrester; S44, inputting the lightning arrester residual service life prediction value, the aging rate and the degradation increment into the target Wiener process model as the target degradation data to obtain the confidence interval of the lightning arrester residual service life prediction value, which is used to judge the reliability of the lightning arrester residual service life prediction value; The parameter estimation in S42 comprises: 1) Set initial conditions and iteration termination criteria: define drift coefficient and diffusion coefficient , initialize by mean of historical data or empirical values; set maximum number of iterations and convergence threshold ; 2) iteratively optimizing the parameters of the Wiener process model based on the degradation data of the lightning arrester: Based on the degradation data of the surge arrester, the current parameters are calculated by the maximum likelihood estimation method The likelihood function under the following: （10） In formula (10), is a degradation increment, is a time interval, is an aging rate, is a random fluctuation term in the aging process; wherein the parameters are updated by a gradient ascent method: （11） In formula (11), is the random fluctuation term in the aging process for the k+1th iteration, is the random fluctuation term in the aging process for the kth iteration, is the learning rate, is the log-likelihood function, which needs to be dynamically adjusted to strike a balance between convergence speed and stability; 3) convergence judgment and dynamic adjustment: if the parameter variation satisfies and / or reaches the maximum iteration number K, the iteration is ended and the next step is entered; otherwise, the learning rate is adjusted according to the current parameter estimation error: the adaptive learning rate algorithm Adagrad is used, and step 2) is repeatedly executed to further optimize the parameters; 4) outputting the finally optimized parameters and verifying the model: ① verifying whether the residual error sequence conforms to the independent and identically distributed assumption through residual error analysis; ② ensuring that the confidence interval of the parameter estimation can effectively cover the true value by calculating the confidence interval coverage rate through Monte Carlo simulation; S5, based on the lightning arrester residual service life prediction value and a set value, determining the working state of the lightning arrester and formulating a corresponding maintenance strategy.

2. The method of claim 1, wherein the method is characterized by: The real-time monitoring data of the lightning arrester to be predicted in S1 comprises leakage current, environmental temperature, vibration amplitude and operating voltage of the lightning arrester, and the specific acquisition process is as follows: 1) collecting the leakage current: collecting the leakage current signal of the lightning arrester through a magnetic current sensor with a resolution of 0.01 mA and a portable digital oscilloscope with a sampling rate of 1 Msamples / s, and setting the data format as time series; 2) collecting the environmental temperature: collecting the temperature inside the lightning arrester as the environmental temperature through a high-precision temperature sensor installed on the surface of the lightning arrester, and the sampling frequency is 1 time / hour; 3) collecting the vibration amplitude: collecting the vibration signal of the lightning arrester through a piezoelectric vibration sensor installed on the base of the lightning arrester with a sensitivity of 0.01 g, and the sampling rate is 10 kHz; the data is filtered and the vibration amplitude mean value is taken; 4) collecting the operating voltage: measuring the actual operating voltage between the two ends of the lightning arrester through a voltage sensor and taking the daily average of the actual operating voltage as the operating voltage.

3. The method of claim 2, wherein S2 The method comprises the following steps: S21, feature extraction is performed on the collected leakage current signal based on fast Fourier transform (FFT), and the third harmonic of the leakage current is obtained, and the calculation formula is as follows: （1） In formula (1), is the third harmonic of the leakage current, is the real part Fourier coefficient of the third harmonic, is the imaginary part Fourier coefficient of the third harmonic; S22, the third harmonic of the leakage current, the ambient temperature, the vibration amplitude and the working voltage are normalized to obtain the input data, and the calculation formula of the normalization processing is as follows: （2） In formula (2), is the normalized value, is the original data in the data set, is the minimum value in the data set, is the maximum value in the data set; the data set is a data set formed by the third harmonic of the leakage current, the ambient temperature, the vibration amplitude, and the operating voltage. S23, the input data format is defined as: （3） In formula (3), is the third harmonic of the leakage current, is the ambient temperature, is the vibration amplitude, is the operating voltage.

4. The method of claim 1, wherein the method further comprises: The lightning arrester residual service life prediction model based on XGBoost algorithm pre-constructed in S3 includes the following steps: S31, obtain multi-dimensional lightning arrester data: 1) obtain historical lightning arrester data at different operating times and pre-process, including leakage current data, ambient temperature data, vibration amplitude data and working voltage data; 2) feature extraction is performed on the leakage current data based on fast Fourier transform (FFT) to obtain third harmonic data of the leakage current; 3) standardize the third harmonic current, vibration amplitude, working voltage and ambient temperature based on Pandas to obtain target data; 4) divide the target data into 70% training set, 20% validation set and 10% test set; S32, construct a lightning arrester residual service life prediction model based on gradient boosting tree, and train to learn the nonlinear relationship between the third harmonic of the leakage current, the vibration amplitude, the ambient temperature and the working voltage and the residual service life of the lightning arrester, which includes the following: 1) determine the objective function: （4） wherein, （5） （6） In formulas (4)-(6), is an objective function; is a loss function; is a regularization term; is a true value; is a predicted value; is a number of leaf nodes of a current tree; is a square of a weight of a leaf node, , is a regularization parameter; 2) Construct decision tree, calculate gradient and split gain of each node The calculation formula is as follows: （7） （8） In formulas (7) to (8), , are the gradient sums of the left and right child nodes, respectively, , are the second-order gradient sums of the left and right child nodes, respectively, is a regularization parameter; if the gain is less than , no split is performed; 3) perform multiple rounds of iterative training, and update the prediction value after each round of iterative training, as follows: （9） In formula (9), is a sample In the first iteration, the predicted value is is a sample In the first iteration, the predicted value is is a learning rate, is the input feature in the tth decision tree; and is the corresponding predicted incremental value; wherein the sample is a group of data in the training set; S33, input the test set into the trained model to output the preliminary lightning arrester residual service life prediction value data.

5. The method of claim 4, wherein the method further comprises: The historical lightning arrester data is pre-processed, including: 3) use wavelet denoising method or band-pass filtering method to remove high-frequency noise to eliminate errors caused by power grid interference; 4) eliminate extreme abnormal data through outlier detection method to improve the credibility of the data; 3) all data are arranged in time series to meet the needs of subsequent model training and prediction.

6. The method of claim 4, wherein the method further comprises: The multiple rounds of iterative training process of the model are as follows: 1) determine the initial hyperparameters based on the training set and the target error range; 2) define the training round k=1 and the maximum training iteration number K; 3) train the model using the training set to obtain the preliminary prediction result; 4) dynamically adjust the hyperparameters, including learning rate, tree depth and regularization parameter, based on the validation set; 5) determine whether the maximum training iteration number K is reached: if not, continue to optimize the hyperparameters of the model; if the training converges or the maximum iteration number is reached, stop training and save the final optimized model.

7. The method of claim 1, wherein the method further comprises: In step 2), the ABC-Gibbs method can also be used: ① construct an approximate posterior distribution and update a alternately through Gibbs sampling; ② use the mean square error between the simulated data of the degradation path and the observed data as the acceptance criterion to ensure the stability of the parameter estimation.

Images