Method for predicting electrical parameters of lightning arrester

Abstract

The invention belongs to the technical field of electric parameter prediction, and particularly relates to a lightning arrester electric parameter prediction method. Collecting electric parameters of the lightning arrester, and storing the electric parameters of the lightning arrester according to a timestamp; constructing a hierarchical hybrid prediction model, and predicting the electrical parameters of the lightning arrester, including decomposing the electrical parameters into a horizontal component, a trend component and a seasonal component, constructing an electrical parameter prediction equation, and performing preliminary prediction on the electrical parameters to obtain an electrical parameter prediction value; calculating a difference value between an electric parameter true value and an electric parameter predicted value, inputting the difference value as a residual sequence into the lightweight time sequence convolution network, extracting a residual nonlinear compensation feature, calculating a residual variance, dynamically adjusting a convolution kernel weight, compressing the residual nonlinear compensation feature through a full connection layer, and outputting an electric parameter correction value. And finally correcting the electric parameter predicted value.

Description

Technical Field

[0001] The present invention belongs to the technical field of electrical parameter prediction, and particularly relates to a method for predicting electrical parameters of a lightning arrester. Background Art

[0002] In a power system, as a key protection device, a lightning arrester is crucial for protecting grid equipment from lightning overvoltage and switching overvoltage, and its performance directly affects the safe and stable operation of the power grid. However, since lightning arresters are mostly placed in the open air, during long-term operation, they will inevitably be affected by the combined effects of power frequency voltage and complex and harsh natural environments, resulting in problems such as aging and moisture absorption of the lightning arrester valve discs, thereby reducing the protection performance of the lightning arrester. In severe cases, it may even cause a large-scale power outage accident in the power grid.

[0003] In order to effectively evaluate the health status of a lightning arrester, timely detect potential fault hazards, and provide a scientific basis for subsequent condition-based maintenance, currently, it is mainly achieved by real-time monitoring of key electrical parameters of the lightning arrester, such as total current, resistive current, lightning strike times, and lightning strike moments. Through the collection and analysis of these electrical parameters, the current health status of the lightning arrester can be evaluated, usually presented in the form of a health status index.

[0004] The electrical parameters of a lightning arrester are affected by various environmental factors such as seasonal changes and load cycles, showing complex non-linear trends and seasonal fluctuation characteristics. However, due to the lack of the ability to analyze these key factors, traditional prediction methods are difficult to accurately capture these change laws of electrical parameters and cannot issue early warnings for abnormal changes in key electrical parameters. Summary of the Invention

[0005] In order to overcome the problems in the prior art, the present invention proposes a method for predicting electrical parameters of a lightning arrester.

[0006] The technical solution of the present invention to solve the above technical problems is as follows: The present invention provides a method for predicting electrical parameters of a lightning arrester, including the following steps: Collect the electrical parameters of the lightning arrester and store the electrical parameters of the lightning arrester according to the time stamp; Construct a hierarchical hybrid prediction model to predict the electrical parameters of the lightning arrester, including: decomposing the electrical parameters into a horizontal component, a trend component, and a seasonal component, constructing an electrical parameter prediction equation to perform a preliminary prediction on the electrical parameters to obtain an electrical parameter prediction value; calculating the difference between the true value of the electrical parameter and the electrical parameter prediction value, inputting the difference as a residual sequence into a lightweight temporal convolutional network to extract residual non-linear compensation features, calculating the residual variance and dynamically adjusting the convolutional kernel weights, and outputting an electrical parameter correction value after compressing the residual non-linear compensation features through a fully connected layer, and finally correcting the electrical parameter prediction value.

[0007] Further, the decomposition of the electrical parameter into a horizontal component, a trend component, and a seasonal component, and the construction of the electrical parameter prediction equation include: Horizontal component: ; Trend component: ; Seasonal component: ; Total current prediction equation: ; In the formula, represents the electrical parameter at time t; represents the horizontal component of the electrical parameter at time t; represents the horizontal component of the electrical parameter at time t - 1; represents the trend component of the electrical parameter at time t; represents the trend component of the electrical parameter at time t - 1; represents the seasonal component of the electrical parameter at time t; represents the seasonal component of the electrical parameter at time represents the seasonal cycle length; represents the horizontal parameter, represents the damping factor, represents the trend parameter, represents the seasonal parameter; represents the prediction time step; represents the predicted value of the electrical parameter at time represents the seasonal component of the electrical parameter at time

[0008] Further, the optimal solution combination of the horizontal parameter, the damping factor, the trend parameter, and the seasonal parameter is iteratively solved by the gradient descent method.

[0009] Further, the difference is input into the lightweight temporal convolutional network as a residual sequence to extract the residual non-linear compensation feature, including: ; In the formula, represents the residual non-linear compensation feature output by the l th layer of the lightweight temporal convolutional network at time ; represents the weight of the l th layer and the kth convolutional kernel in the convolutional network; K represents the convolutional kernel size; represents the bias term; represents the difference between the true value and the predicted value of the electrical parameter at time t.

[0010] Further, the calculation of the residual variance and the dynamic adjustment of the convolutional kernel weight include: Calculate the residual variance: ; In the above formula, represents the current residual variance; represents the size of the sliding window; represents the average value of recent residuals; represents t - i the difference between the true value and the predicted value of the electrical parameter at If the residual variance is greater than the preset residual variance threshold, dynamically adjust the weights of the convolutional kernels in the convolutional network: ; In the above formula, represents the weights of the convolutional kernels in the adjusted lightweight temporal convolutional network; represents the historical residual variance; is the sensitivity coefficient.

[0011] Furthermore, after compressing the residual non-linear compensation features through the fully connected layer, output the corrected value of the electrical parameter: ; In the above formula, l is the number of convolutional layers; is the weight of the fully connected layer; is the bias term of the fully connected layer; represents the corrected value of the electrical parameter at j represents the feature nodes in the fully connected layer; represents the l th j feature node in the th layer of the lightweight temporal convolutional network at time

[0012] Furthermore, finally correct the predicted value of the electrical parameter, including: correcting the predicted value of the electrical parameter based on the corrected value of the electrical parameter: ; In the above formula, represents the corrected predicted value of the electrical parameter at represents the predicted value of the electrical parameter at

[0013] Furthermore, it also includes: real-time identification of mutation events through sliding window outlier detection. If a mutation event is detected, reset the trend component, shorten the seasonal cycle length, reduce the level smoothing parameter, and dynamically adjust the parameters of the full-current hierarchical hybrid prediction model.

[0014] Further, the electrical parameters include total current and resistive current.

[0015] Compared with the prior art, the present invention has the following technical effects: Through the collaborative mechanism of adaptive decomposition-depth residual compensation-dynamic parameter linkage, the present invention systematically solves the triangular contradiction of accuracy, computing power and robustness in the prediction of arrester electrical parameters: a lightweight hierarchical hybrid architecture is proposed - an adaptive damping model is constructed with the triple exponential smoothing algorithm as the core, and the parameters are dynamically optimized by introducing a damping factor and a rolling time window, effectively suppressing trend divergence and enhancing the ability to fit time series periodicity, while avoiding parameter solidification; on this basis, a lightweight temporal convolutional network (L-TCN) is designed as residual compensation. For the non-linear spike residuals caused by transient events such as lightning strikes, low-computing-power and high-precision compensation is achieved through local temporal feature extraction in the convolutional layer, and the prediction error is reduced compared with traditional residual correction methods; furthermore, a real-time detection-parameter linkage rule for mutation events driven by statistics is proposed: based on the outlier detection in the sliding window, abnormal conditions such as lightning strikes and short circuits are quickly identified, the historical trend interference is eliminated by resetting the trend component, the seasonal cycle is dynamically compressed to strengthen the short-term response, and the horizontal smoothing parameter is adaptively attenuated to suppress the long-term noise, so that the model improves the local sensitivity to mutation events while maintaining long-term stability, forming a balance mechanism of "global stability-local agility". Description of the Drawings

[0016] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0017] Figure 1 It is a flowchart of the present invention. Detailed Embodiments

[0018] In order to further elaborate on the technical means and effects adopted by the present invention to achieve the intended invention purpose, the following, in combination with the drawings and preferred embodiments, details the specific embodiments, structures, features and effects of the technical solutions proposed according to the present invention. The specific features, structures or characteristics in one or more embodiments can be combined in any suitable form. Unless otherwise defined, all technical and scientific terms used in the present invention have the same meaning as commonly understood by those skilled in the technical field to which the present invention belongs.

[0019] The present invention proposes a hierarchical hybrid prediction model, which separates the trend, seasonality, and non-linear residual characteristics of time series data through a hierarchical structure, combines traditional statistical models with lightweight deep learning to improve prediction accuracy and robustness, and realizes the adaptability to mutation events through a dynamic residual feedback mechanism and enhanced feature engineering. Finally, it realizes the high-precision and low-latency prediction of the electrical parameters of lightning arresters, solves the problems of the above-mentioned existing technologies, and provides a reliable basis for the state evaluation and fault warning of lightning arresters.

[0020] Referring to Figure 1 , a method for predicting the electrical parameters of a lightning arrester, comprising the following steps: collecting the electrical parameters of the lightning arrester and storing the electrical parameters of the lightning arrester according to the time stamp, wherein the electrical parameters include the total current and the resistive current; constructing a hierarchical hybrid prediction model to predict the electrical parameters of the lightning arrester, including: decomposing the electrical parameters into a horizontal component, a trend component, and a seasonal component, constructing an electrical parameter prediction equation, and performing a preliminary prediction on the electrical parameters to obtain an electrical parameter prediction value; calculating the difference between the actual value of the electrical parameter and the electrical parameter prediction value, inputting the difference as a residual sequence into a lightweight time series convolutional network, extracting the residual non-linear compensation feature, calculating the residual variance and dynamically adjusting the convolutional kernel weight, outputting an electrical parameter correction value after compressing the residual non-linear compensation feature through a fully connected layer, and finally correcting the electrical parameter prediction value.

[0021] The following is a detailed expansion of each of the above steps: Step 100: Collect the electrical parameters of the lightning arrester and store the electrical parameters of the lightning arrester according to the time stamp.

[0022] The on-line monitoring system of the lightning arrester collects the actual value of the total current of the lightning arrester and stores the historical data set of the total current according to the time stamp.

[0023] Step 200: Construct a hierarchical hybrid prediction model to predict the electrical parameters of the lightning arrester.

[0024] Construct a hierarchical hybrid prediction model, where the hierarchical hybrid prediction model includes three layers, the first layer is adaptive damping triple exponential smoothing, the second layer is a lightweight time series convolutional network, and the third layer is dynamic perception of mutation events.

[0025] Introduce horizontal parameters, damping factors, trend parameters, and seasonal parameters, decompose the electrical parameters into a horizontal component, a trend component, and a seasonal component; and construct a prediction equation based on the horizontal component, trend component, and seasonal component to obtain an electrical parameter prediction value, including: Horizontal component: ; Trend component: ; Seasonal component: ; Total current prediction equation: ; In the formula, Denote the electrical parameter at time t; Denote the horizontal component of the electrical parameter level at time t; Denote the horizontal component of the electrical parameter at time t - 1; Denote the trend component of the electrical parameter at time t; Denote the trend component of the electrical parameter at time t - 1; Denote the seasonal component of the electrical parameter at time t; Denote the seasonal component of the electrical parameter at time Denote the seasonal cycle length; Denote the horizontal parameter, Denote the damping factor, Denote the trend parameter, Denote the seasonal parameter; Denote the prediction time step; Denote the predicted value of the electrical parameter at time Denote the seasonal component of the electrical parameter at time. Solve the optimal solution combination of the horizontal parameter, damping factor, trend parameter, and seasonal parameter by iterative gradient descent method.

[0026] Input the difference as a residual sequence into a lightweight temporal convolutional network to extract residual non - linear compensation features, including: ; In the formula, Denote the residual non - linear compensation feature output by convolution at time l in the th layer of the lightweight temporal convolutional network; Denote the weight of the l th layer and the k th convolutional kernel in the convolutional network; K denotes the convolutional kernel size; Denote the bias term; Denote the difference between the true value and the predicted value of the electrical parameter at time t.

[0027] The calculation of the residual variance and the dynamic adjustment of the convolutional kernel weights include: Calculate the residual variance: ; In the above formula, Denote the current residual variance; Denote the size of the sliding window; Denote the average value of recent residuals; Denote t - i the difference between the true value and the predicted value of the electrical parameter at time If the residual variance is greater than a preset residual variance threshold, the convolution kernel weights in the convolutional network are dynamically adjusted: ; In the above formula, represents the convolution kernel weights in the adjusted lightweight temporal convolutional network; represents the historical residual variance; is the sensitivity coefficient.

[0028] The corrected value of the electrical parameter is output after compressing the residual non-linear compensation feature through the fully connected layer: ; In the above formula, l is the number of convolutional layers; is the weight of the fully connected layer; is the bias term of the fully connected layer; represents the corrected value of the electrical parameter at j represents the feature nodes in the fully connected layer; represents the l th j layer and the th feature node in the lightweight temporal convolutional network, and the residual non-linear compensation feature output by convolution at time

[0029] Finally, correct the predicted value of the electrical parameter, including: correcting the predicted value of the electrical parameter based on the corrected value of the electrical parameter: ; In the above formula, represents the corrected predicted value of the electrical parameter at represents the predicted value of the electrical parameter at

[0030] The mutation event is identified in real time through the sliding window outlier detection. If a mutation event is detected, the trend component is reset, the seasonal cycle length is shortened, and the horizontal smoothing parameter is reduced, and the parameters of the full-current hierarchical hybrid prediction model are dynamically adjusted.

[0031] Taking the full current as an example, this step 200 includes: Construct a full-current hierarchical hybrid prediction model, where the full-current hierarchical hybrid prediction model includes three layers. The first layer is the adaptive damping triple exponential smoothing, the second layer is the lightweight temporal convolutional network, and the third layer is the dynamic perception of mutation events.

[0032] Step 210: Decompose the full current into a horizontal component, a trend component, and a seasonal component, construct a full-current prediction equation, make a preliminary prediction of the full current, and obtain the full-current predicted value; calculate the difference between the full-current true value and the full-current predicted value.

[0033] Step 210 may include the following sub-steps: Step 2101: Introduce the horizontal parameter, damping factor, trend parameter, and seasonal parameter, decompose the total current into the total current horizontal component, total current trend component, and total current seasonal component, and construct a prediction equation based on the total current horizontal component, total current trend component, and total current seasonal component to obtain the total current prediction value.

[0034] Total current horizontal component: ; Total current trend component: ; Total current seasonal component: ; Total current prediction equation: ; In the formula, represents the total current at time t; represents the total current horizontal component at time t; represents the total current horizontal component at time t - 1; represents the total current trend component at time t; represents the total current trend component at time t - 1; represents the total current seasonal component at time t; represents the total current seasonal component at time represents the seasonal cycle length; represents the horizontal parameter, represents the damping factor, represents the trend parameter, represents the seasonal parameter; where , , take values in (0, 1), takes values in (0.9, 1); represents the prediction step size; represents the total current prediction value at time represents the seasonal component at time

[0035] Damping factor restricts the exponential growth of the trend, avoiding overfitting. It restricts the exponential growth of the trend component, making the prediction value tend to be reasonable in the long term and solving the "trend explosion" problem of the traditional exponential smoothing model.

[0036] Step 2102: Optimize the horizontal parameter, damping factor, trend parameter, and seasonal parameter to solve the optimal solution parameter combination.

[0037] Adopt rolling time window cross-validation to minimize the mean square error (MSE): ; Among them, represents the predicted value of the total current at time t.

[0038] Iteratively solve the optimal solution through the gradient descent method , , , parameter combination.

[0039] Online update parameters through a rolling time window to adapt to model drift caused by equipment aging and environmental changes. The model can adapt to data distribution changes under different working conditions (such as humid environment, high-temperature environment). Compared with traditional fixed-parameter models, the prediction can reduce errors.

[0040] Step 220: Take the difference between the true value of the total current and the predicted value of the total current as the residual sequence and input it into the lightweight temporal convolutional network to extract residual non-linear compensation features, calculate the residual variance, and dynamically adjust the convolutional kernel weights. After compressing the features through the fully connected layer, output the corrected value of the total current, and finally correct the predicted value of the total current.

[0041] The said step 220 includes the following steps: Step 2201: Take the difference between the true value of the total current and the predicted value of the total current as the residual sequence and input it into the lightweight temporal convolutional network (L-TCN) to perform residual feature extraction and obtain the convolutional output.

[0042] Capture the residual non-linear features after the first layer of prediction through the convolutional operation of the lightweight temporal convolutional network to correct the baseline prediction value.

[0043] Take the predicted residual sequence output from the first layer as the input: ; Among them, represents the predicted value at ; Design causal convolution relying on historical data to ensure temporal causality: ; In the formula, represents the residual non-linear compensation feature of the convolutional output of the l th layer in the convolutional network at time ; is the weight of the l th layer and the kth convolutional kernel in the convolutional network; K is the convolutional kernel size; is the bias term.

[0044] Step 2202: Calculate the residual variance and dynamically adjust the convolutional kernel weights in the lightweight temporal convolutional network according to the statistical characteristics of the residual variance.

[0045] Calculate the residual variance: ; In the above formula, represents the current residual variance; represents the size of the sliding window; represents the average value of the recent residuals.

[0046] If the residual variance is greater than the preset residual variance threshold, dynamically adjust the weights of the convolutional kernels in the convolutional network: ; In the above formula, represents the weights of the convolutional kernels in the adjusted convolutional network; represents the historical residual variance; is the sensitivity coefficient.

[0047] At this time, ; In the formula, represents the residual non - linear compensation feature of the convolutional output of the l th layer in the adjusted convolutional network at time .

[0048] Step 2203: Compress the residual non - linear compensation feature through the fully - connected layer and output the full - current correction value: ; In the above formula, l is the number of convolutional layers; is the weight of the fully - connected layer; is the bias term of the fully - connected layer; represents the full - current correction value; j represents the feature nodes in the fully - connected layer.

[0049] Step 2204: Correct the full - current prediction value based on the full - current correction value: ; In the above formula, represents the full - current prediction correction value at

[0050] Step 230: Real - time identify mutation events through sliding - window outlier detection. If a mutation event is detected, reset the trend component, shorten the seasonal cycle length, and reduce the level smoothing parameter, and dynamically adjust the parameters of the full - current hierarchical hybrid prediction model to enhance the short - term response ability.

[0051] Real - time detect mutation events such as lightning strikes based on statistical methods and dynamically adjust the parameters of the full - current hierarchical hybrid prediction model.

[0052] Step 2301: Based on the residual sequence, real-time identification of mutation events is carried out through sliding window outlier detection: For the predicted residual sequence within the sliding window , , …, calculate the median ; Median absolute deviation ; If , then mark as a mutation event, where is the threshold coefficient, usually taken as 3.

[0053] Step 2302: If a mutation event is detected, reset the trend component to avoid historical trend interference: ; In the above formula, represents the reset total current trend component.

[0054] Shorten the seasonal cycle length , lasting time steps to enhance the short-term response ability: ; In the above formula, represents the reset seasonal cycle length.

[0055] If the residuals exceed the threshold for n consecutive times, reduce the horizontal smoothing parameter : ; The updated horizontal smoothing parameter is passed to the first layer for prediction at the next moment.

[0056] Step 300: Electrical parameter warning and health status index.

[0057] Accurate predicted values of each electrical parameter can be obtained through Steps 100 - 200. The program sets the warning value J as needed. When the predicted value of the electrical parameter > J, the device sends a warning message and displays it to the user.

[0058] Calculate the arrester health status index (actual value and predicted value) through the comprehensive evaluation model: ; Among them, is the dynamic weight, which is calculated in real time by the entropy weight method and reflects the influence degree of each parameter on the health status; It is a membership function, which maps a score from 0 to 1 according to the degree of deviation of the parameter from the reference value.

[0059] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included in the protection scope of the present invention.

Claims

1. A lightning arrester electrical parameter prediction method, characterized in that: The following steps are involved: Collect the electrical parameters of the arrester and store them by timestamp; Constructing a hierarchical hybrid prediction model to predict the electrical parameters of the arrester, including: decomposing the electrical parameters into horizontal components, trend components and seasonal components, constructing electrical parameter prediction equations, making preliminary predictions on the electrical parameters, and obtaining the predicted values ​​of the electrical parameters; The difference between the true value of the electrical parameter and the predicted value of the electrical parameter is calculated, and the difference is input into the lightweight temporal convolutional network as a residual sequence. The residual nonlinear compensation feature is extracted, the residual variance is calculated and the convolution kernel weight is dynamically adjusted. After the residual nonlinear compensation feature is compressed by the fully connected layer, the electrical parameter correction value is output, and finally the electrical parameter prediction value is corrected.

2. A lightning arrester electrical parameter prediction method according to claim 1, characterized in that: The method of decomposing the electric parameters into a horizontal component, a trend component and a seasonal component and constructing an electric parameter prediction equation includes: Horizontal component: ; Trend Component: ; Seasonal Components: ; Full current prediction equation: ; In the formula, represents the electrical parameters at time t; represents the horizontal component of the electrical parameter at time t; Represents the horizontal component of the electrical parameter at t-1; Represents the trend component of the electrical parameter at time t; Represents the trend component of the electrical parameter at t-1; represents the seasonal component of the electrical parameter at time t; express Seasonal component of electrical parameters at ; Indicates the length of the seasonal cycle; represents the level parameter, represents the damping factor, represents the trend parameter, represents seasonal parameters; represents the prediction time step; express The predicted value of electrical parameters at ; express Seasonal component of electrical parameters at .

3. A lightning arrester electrical parameter prediction method according to claim 2, characterized in that: The optimal solution combination of level parameter, damping factor, trend parameter and seasonal parameter is iteratively solved through the gradient descent method.

4. A lightning arrester electrical parameter prediction method according to claim 1, characterized in that: The difference is input as a residual sequence into a lightweight temporal convolutional network to extract residual nonlinear compensation features, including: ； In the formula, Represents the first in the lightweight temporal convolutional network l Layer in time Residual nonlinear compensation characteristics of convolution output; Represents the convolutional network l Tier k convolution kernel weights; K represents the convolution kernel size; represents the bias term; Represents the difference between the true value of the electrical parameter and the predicted value of the electrical parameter at time t.

5. A lightning arrester electrical parameter prediction method according to claim 4, characterized in that: The calculation of residual variance and dynamic adjustment of convolution kernel weights include: Calculate the residual variance: ； In the above formula, represents the current residual variance; Indicates the size of the sliding window; represents the average of recent residuals; express t - i The difference between the actual value of the electrical parameter and the predicted value of the electrical parameter; If the residual variance is greater than the preset residual variance threshold, the convolution kernel weights in the convolution network are dynamically adjusted: ； In the above formula, Represents the adjusted convolution kernel weights in the lightweight temporal convolutional network; represents the historical residual variance; is the sensitivity coefficient.

6. A lightning arrester electrical parameter prediction method according to claim 5, characterized in that: After the residual nonlinear compensation features are compressed by the fully connected layer, the corrected value of the electrical parameters is output: ； In the above formula, l is the number of convolutional layers; is the weight of the fully connected layer; is the bias term of the fully connected layer; express Correction value of electrical parameters; j Represents the feature nodes in the fully connected layer; Represents the first l Tier j Feature nodes at time Residual nonlinear compensation features of convolution output.

7. A lightning arrester electrical parameter prediction method according to claim 6, characterized in that: The electric parameter prediction value is finally corrected, including: correcting the electric parameter prediction value based on the electric parameter correction value: ； In the above formula, express Time-to-hour power parameter prediction correction value; express Prediction of electrical parameters.

8. A lightning arrester electrical parameter prediction method according to claim 2, characterized in that: Also includes: The sliding window outlier detection is used to identify mutation events in real time. If a mutation event is detected, the trend component is reset, the seasonal cycle length is shortened, and the horizontal smoothing parameter is reduced to dynamically adjust the parameters of the full current stratified hybrid prediction model.

9. A lightning arrester electrical parameter prediction method according to claim 1, characterized in that: The electrical parameters include total current and resistive current.

Images