Lightning arrester performance degradation index construction method based on improved neural network

Abstract

The invention discloses a lightning arrester performance degradation index construction method based on an improved neural network. According to the method, environment temperature, humidity, accumulated impact times and operation time are collected as input, resistive current and capacitive current are taken as output, and a neural network model is constructed; an attention mechanism is introduced, and model parameters are optimized and updated in combination with a back propagation algorithm and an improved cosine annealing learning rate; performance degradation indexes are constructed, the state of the lightning arrester can be visually displayed through changes of electrical parameters, whether the lightning arrester can operate normally or not is judged, and operation is supervised or shutdown replacement is carried out; according to the invention, the processing capability of the model on complex data can be effectively improved; a reliable basis is provided for an electric power system to timely and accurately judge the lightning arrester performance reduction condition, and the method has important significance and popularization value for guaranteeing safe and stable operation of the electric power system.

Description

Technical Field

[0001] The present invention relates to the technical field of electrical equipment status monitoring and fault prediction, and in particular to a method for constructing arrester performance degradation indicators based on an improved neural network, which is particularly suitable for status assessment and fault warning of offshore wind power zinc oxide arresters. Background Art

[0002] As a crucial overvoltage protection device, the reliability of zinc oxide arresters (MOAs) is directly linked to the safe and stable operation of power systems. Over time, MOA performance is affected by environmental factors (such as temperature and humidity) and electrical shocks (such as lightning strikes and switching overvoltages), leading to a gradual degradation of their internal insulation performance. Traditional MOA performance testing methods rely on periodic preventative testing, which suffers from long inspection cycles and inability to detect potential faults in a timely manner. On the one hand, manual inspections are limited in frequency and are highly subjective, making it difficult to accurately capture subtle changes in MOA performance in real time. On the other hand, relying solely on a single or a few electrical parameters fails to comprehensively and comprehensively assess the actual operating status of the MOA, leading to potential misjudgments and omissions. Resistive and capacitive currents are key parameters that reflect MOA performance. As MOA performance deteriorates, these currents typically increase. Therefore, using advanced algorithms to accurately predict MOA resistive and capacitive currents, thereby enabling early detection of MOA performance changes, has become a research hotspot in the field of power equipment condition monitoring. Currently, several machine learning-based methods have been used to predict the state of power equipment. However, these methods suffer from high model complexity when dealing with arrester performance under the influence of multiple factors. Traditional threshold judgment methods based on a single parameter cannot fully reflect the performance degradation patterns under the coupling of multiple factors, resulting in low prediction accuracy. Therefore, developing a method that can accurately predict arrester performance and construct a degradation index based on performance degradation trends is of great practical significance. Summary of the Invention

[0003] To address the aforementioned technical issues of performance degradation in existing arresters, this paper provides a method for constructing arrester performance degradation indicators based on an improved neural network. By integrating multi-dimensional monitoring data and utilizing a recurrent neural network algorithm to establish an arrester performance prediction model, this method predicts arrester performance degradation trends and constructs a performance degradation indicator to quantitatively assess arrester health, thereby improving the safety and operational efficiency of offshore wind power systems.

[0004] The present invention is achieved by adopting the following technical solutions: The method for constructing a lightning arrester performance degradation index based on an improved neural network includes the following steps: (1) Data collection and preprocessing (1.1) Multi-source monitoring data acquisition. The essence of the performance degradation of metal oxide arresters (hereinafter referred to as MOA) is the degradation of the nonlinear characteristics of the internal valve plate. As the MOA ages, the arrester becomes damp, the ambient temperature rises, and the number of operations increases, the resistive current and capacitive current in the leakage current both tend to increase. Therefore, the magnitude of the resistive current and the capacitive current are used as indicators of MOA performance degradation. By regularly monitoring the changes in electrical parameters, the degree of performance degradation can be directly reflected. The main test item of the arrester electrical test is to test the leakage current. The resistive current and capacitive current are solved by the full current and power angle: ; ; At the same time, the ambient temperature T, humidity H, cumulative number of impacts N, and operating time Y must be monitored simultaneously.

[0005] (1.2) Normalize the collected multi-source monitoring data: ; in, is the original data, and are the minimum and maximum values ​​of the data sequence, respectively, and are normalized data.

[0006] (2) Neural network model construction and training (2.1) Construct a three-layer recurrent neural network (RNN) model, consisting of an input layer, a hidden layer, and an output layer. NNN models can preserve historical information, thereby capturing temporal dependencies in sequential data. Each data point normalized to the interval [0, 1] is used as the input node of the neural network, and the output node is the resistive and capacitive currents.

[0007] The weight matrix of the attention mechanism introduced to the input layer Of , through the formula Calculate the attention score ( is the value after scoring, It is the input layer data. ), which enables the model to automatically focus on the key factors that have a greater impact on the performance of the arrester, such as ambient humidity and number of impacts, and improves the model's ability to handle complex data relationships.

[0008] (2.2) Neural network training: Divide the preprocessed data into a training set and a test set according to a certain ratio. Use the training set data to train the neural network model, and calculate the difference between the predicted value and the true value using the loss function (MSE): ; n is the sample size, 、 are the predicted value and the true value respectively.

[0009] The backpropagation algorithm is used during training. This algorithm is based on the principle of gradient descent. It calculates the gradient of the loss function with respect to the model parameters (weights and biases) and updates the parameters in the opposite direction of the gradient. The weights and biases are optimized and updated according to the cosine annealing learning rate: ; 、 is the current number of iterations and the maximum number of iterations, 、 are the minimum and maximum learning rates.

[0010] The neural network parameters are continuously adjusted until the loss function converges to a certain level or the maximum number of training rounds is reached. In the backpropagation algorithm, the goal is to minimize the loss function by continuously adjusting the parameters (weights and biases) in the neural network, so that the model's predictions are as close to the true values ​​as possible.

[0011] (3) Lightning arrester current prediction: The preprocessed data is input into the trained neural network model to output the predicted values ​​of the lightning arrester's resistive current and capacitive current.

[0012] (4) Performance degradation index construction: The pre-processed real-time data is input into the trained neural network model. The model output is the resistive current and capacitive current of the arrester. The initial current is compared with the model predicted current to obtain the resistive current increment and capacitive current increment. The performance degradation index D is calculated and compared with the set threshold to determine the health status of the arrester.

[0013] The present invention has the following beneficial technical effects: 1. Multi-parameter fusion: The method of the present invention simultaneously considers the coupling effect of resistive current and environmental factors, can capture the nonlinear and time-varying characteristics of arrester degradation, and is suitable for complex working conditions such as humid weather; 2. The method of the present invention constructs a quantitative preventive maintenance index through trend prediction, intuitively judges the trend of arrester performance degradation, provides a scientific basis for the maintenance and repair of the power system, and reduces operation and maintenance costs and failure risks; 3. The method of the present invention has strong scalability and can add input conditions that may affect the performance of the arrester according to the actual project, thereby improving the prediction accuracy; 4. Accurately focus on key influencing factors: After the method of the present invention introduces the attention mechanism, the model can evaluate the importance of input factors such as ambient temperature, humidity, cumulative number of impacts and operating time on the performance of the lightning arrester by calculating the attention weight; for example, in some operating scenarios, the ambient temperature and cumulative number of impacts may have a greater impact on the performance of the lightning arrester. The attention mechanism will give these two factors a higher weight, allowing the model to focus more on capturing the relationship between them and changes in lightning arrester performance, effectively avoiding interference from secondary information, and greatly improving the model's ability to process complex data relationships and the accuracy of capturing key information. DETAILED DESCRIPTION

[0014] In order to describe the present invention more specifically, the technical solution of the present invention is described in detail below.

[0015] The method for constructing a lightning arrester performance degradation index based on an improved neural network includes the following steps: S1. Data collection: Regularly collect the ambient temperature T, humidity H, cumulative number of impacts N, operating time Y, and resistive current I during the operation of the arrester r , capacitive current I c , and normalize it: ; ; ; ; ; ; Among them, the maximum and minimum values ​​of each indicator are generally obtained based on the statistics of the real data range. After normalization, the training data set {(T i ,H i ,N i ,Y i ),(Ir i ,Ic i )}, where i=1,2,...,n, and n is the number of data samples.

[0016] S2. Neural network structure design: The preprocessed data was divided into training and test sets in a 7:3 ratio. A neural network model with one hidden layer was constructed. The input layer had four nodes (x1-x4 corresponded to the four input parameters: ambient temperature, humidity, cumulative shock count, and operating time, respectively), and the output layer had two nodes (one for the resistive current Ir and one for the capacitive current Ic, respectively).

[0017] S3, forward propagation calculation: S3.1 Introducing the attention mechanism: Initializing the weight matrix W of the attention mechanism a , through the formula Calculate the attention score and normalize the score using the softmax function.

[0018] in, is the value after scoring, is the input layer data, yes Normalized values: ; .

[0019] S3.2 Input layer to hidden layer: According to the calculated attention weight , perform weighted summation on the input layer data. Input layer nodes j ( j= 1,2,3,4) to the hidden layer nodes k ( k =1,2,3,4) is , the bias of the hidden layer node is .

[0020] Hidden layer nodes k Input for: ; Hidden layer nodes k Output Through the activation function calculate: ; ; S3.3 Hidden layer to output layer: hidden layer nodes k With the output layer node l ( l =1,2, corresponding to resistive current and capacitive current respectively) is , output layer node l The bias is .

[0021] Output layer nodes l Input for: ; .

[0022] Is the predicted output of the output layer, where the identity output is used.

[0023] S4. Loss function definition: The mean square error (MSE) is used as the loss function L , which measures the difference between the predicted output and the actual capacitive and resistive currents: is the predicted value, is the corresponding true value: .

[0024] S5. Back propagation algorithm: Based on the principle of gradient descent, the model parameters (weights) are calculated by calculating the loss function. 、 and bias 、 ), and update the parameters in the opposite direction of the gradient.

[0025] Calculate the error of the output layer: ; Calculate the error of the hidden layer: ;in is the derivative of the hidden layer activation function.

[0026] Calculate the gradient of the loss function with respect to weights and biases, and update the parameters in such a way that the parameters change in a direction that can reduce the loss function value: ; ; ; .

[0027] S6, parameter update: weights and biases are optimized and updated based on the cosine annealing learning rate. The formula for calculating the cosine annealing learning rate is: ; is the minimum learning rate, is the maximum learning rate, t is the current training round number, is the maximum number of training rounds. The update formula is: ; ; ; .

[0028] S7. Model training and prediction: Repeat steps 3-6 and continuously adjust the parameters of the neural network until the preset maximum number of training rounds is reached.

[0029] S8. Construction of performance degradation indicators: When judging the performance of the arrester, the pre-processed real-time data is input into the trained neural network model to output the resistive current. and capacitive current And compare it with the data at the time of production, and judge according to indicator D: ; ; ; When D<0.1, the arrester performance is considered to be in normal condition and routine inspection is recommended. When 0.1≤D<0.4, the arrester performance is judged to be declining and further inspection and maintenance are required, shortening the monitoring cycle. When D≥0.4, the arrester performance is judged to be deteriorating and shutdown for maintenance is required.

Claims

1. A method for constructing arrester performance degradation indicators based on an improved neural network, characterized by: The following steps are involved: (1) Data collection and preprocessing: The ambient temperature, humidity, cumulative number of shocks, and operating time of the arrester operating environment are collected as input data, and the collected data are preprocessed; (2) Neural network model construction and training: Construct a neural network model with at least one hidden layer, introduce an attention mechanism, combine the backpropagation algorithm with an improved cosine annealing learning rate optimization strategy, and use the preprocessed data to train the neural network model; (3) Arrester current prediction: The pre-processed data is input into the trained neural network model to output the predicted values ​​of the arrester's resistive current and capacitive current; (4) Construction of performance degradation index and determination of health status: The degradation index is calculated based on the difference between the predicted value and the initial value after production to determine its health status.

2. The method for constructing arrester performance degradation index based on improved neural network according to claim 1, characterized in that: The specific process of (2) is as follows: (2.1) Construct a recurrent neural network model, which includes an input layer, a hidden layer, and an output layer; (2.2) Introduce the weight matrix of the attention mechanism into the input layer, calculate the attention score, and assign weights to the factors that affect the performance of the lightning arrester; (2.3) Divide the preprocessed data into a training set and a test set, use the training set data to train the neural network model, and calculate the difference between the predicted value and the true value through the loss function MSE.

3. The method for constructing arrester performance degradation index based on improved neural network according to claim 2, characterized in that: In the model training process of (2.3), the back propagation algorithm is adopted. Based on the principle of gradient descent, the gradient of the loss function with respect to the model parameters is calculated and the parameters are updated in the opposite direction of the gradient. The weights and biases of the factors affecting the performance of the arrester are optimized and updated based on the improved cosine annealing learning rate.

4. The method for constructing a lightning arrester performance degradation index based on an improved neural network according to claim 3, characterized in that: In the improved cosine annealing learning rate optimization strategy, the model parameters are optimized, the cosine annealing learning rate is dynamically adjusted according to the number of training rounds, and the learning rate gradually decreases with the number of iterations.

5. The method for constructing arrester performance degradation index based on improved neural network according to claim 1, characterized in that: The specific process of (4) is as follows: The preprocessed data is input into the trained neural network model, and the initial current is compared with the predicted current output by the model to obtain the resistive current increment and the capacitive current increment. The performance degradation index D is calculated and compared with the set threshold to determine the health status of the lightning arrester.