Insulator surface pollution prediction method based on meteorological environmental factors

Abstract

The invention relates to the technical field of overhead transmission lines, and discloses a meteorological environment factor-based insulator surface contamination prediction method, which comprises the following steps of: firstly, acquiring multi-source heterogeneous data of insulator body attributes, a micro-meteorological environment and atmospheric chemical components, and constructing a historical sample database; obtaining a normalized time sequence matrix through data cleaning and feature screening; constructing a hybrid neural network model, extracting local features through a convolutional neural network, generating attention masks based on relative humidity by using a dynamic weighted chemical component feature fusion network, performing dynamic weighting on feature vectors, and inputting the feature vectors into a long-short-term memory network to perform time sequence evolution prediction; and finally, outputting predicted values of equivalent salt density and insoluble deposition density of the surface of the insulator by utilizing the model for training convergence. By introducing a humidity attention mechanism, deep fusion of the meteorological environment and the chemical component characteristics is realized, and the prediction precision of insulator dirt retention in a complex environment is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of overhead transmission line technology, specifically to a method for predicting surface contamination of insulators based on meteorological environmental factors. Background Technology

[0002] Insulators, as critical external insulation components in transmission and distribution systems, directly affect the safety and stability of the power system. During long-term operation, insulator surfaces gradually accumulate a layer of contaminants composed of industrial dust, salt spray, and other pollutants. When exposed to humid weather, this contaminant layer dissolves, forming a conductive water film. This leads to a sharp increase in leakage current on the insulator surface, a decrease in flashover voltage, and ultimately, serious power accidents such as flashover and tripping. With the continuous expansion of the power grid and the extension of power lines to complex environments such as coastal areas, heavy industrial zones, and windy sandy regions, the problem of insulator surface contamination is becoming increasingly prominent, posing a severe challenge to the operation, maintenance, and repair of power equipment.

[0003] The accumulation of contaminants in insulators is a complex dynamic process modulated by multiple factors. It depends not only on the concentration of suspended particulate matter and chemical gases in the atmosphere, but also on micrometeorological conditions such as ambient temperature, relative humidity, rainfall, wind speed, and wind direction. There are complex nonlinear coupling effects among these factors. For example, relative humidity not only affects the degree of moisture absorption of contaminant particles, but also changes the deliquescence characteristics and adhesion of chemical components, thus significantly affecting the deposition rate.

[0004] However, traditional insulator pollution level assessment mainly relies on manual periodic tower climbing sampling or obtaining local data based on simple point monitoring equipment. These methods often have significant time lags, making it difficult to reflect the dynamic changes in pollution over time in real time. Furthermore, manual testing is costly and time-consuming, failing to meet the needs of intelligent power grid operation and maintenance. Although the development of IoT and online monitoring technologies in recent years has made it possible to acquire massive amounts of meteorological and environmental data, existing pollution prediction technologies still have significant limitations. Current prediction models are mostly based on traditional mathematical statistical regression or shallow machine learning algorithms. These methods often treat micrometeorological parameters and atmospheric chemical components as independent parallel feature inputs, ignoring the deep-seated interaction mechanisms between different physical fields, particularly failing to effectively characterize the nonlinear modulation effect of relative humidity on the deposition efficiency of specific chemical components. Furthermore, the process of pollution accumulation has a long-term time-dependent nature, and conventional models cannot simultaneously take into account the spatial feature extraction of multi-source heterogeneous data and the capture of evolutionary patterns over long periods of time. As a result, under complex and variable meteorological conditions, the prediction accuracy of the equivalent salt density (ESDD) and non-soluble deposition density (NSDD) on the insulator surface is difficult to meet the actual requirements of refined and differentiated operation and maintenance. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method for predicting insulator surface pollution based on meteorological environmental factors. This method solves the problem that existing transmission line pollution prediction technologies fail to fully consider the nonlinear strong coupling relationship between atmospheric chemical components and the micro-meteorological environment, resulting in insufficient prediction accuracy under complex meteorological conditions.

[0006] To achieve the above objectives, this invention provides a method for predicting insulator surface pollution based on meteorological environmental factors. The method in this embodiment first constructs a multi-source heterogeneous database containing insulator properties, micro-meteorological environment, and atmospheric chemical components, and aligns the data sources along the time dimension to form a historical sample set for supervised learning. Addressing the noise interference and redundancy in the original monitoring data, this method generates a high-quality normalized multivariate time series matrix through data cleaning and feature selection strategies, providing a standardized data foundation for model input.

[0007] In constructing the core prediction model, this invention proposes a hybrid neural network architecture based on environmental and meteorological mutual attention and convolutional neural network long short-term memory network. The design of this architecture follows a logical path from local feature extraction to multi-factor fusion, and then to temporal evolution deduction. Specifically, the model uses a convolutional neural network to process the input multivariate time series and extract high-dimensional local feature vectors implicit in the data.

[0008] The key innovation of this invention lies in the introduction of a dynamically weighted chemical component feature fusion network. Considering that relative humidity is a crucial catalyst affecting the efficiency of particulate matter absorption, deliquescence, and adhesion on insulator surfaces in actual physical processes, this method does not treat humidity data merely as a general feature, but rather as a control variable. By mapping relative humidity to an attention mask, the high-dimensional feature vector output by the convolutional neural network is dynamically weighted. This mechanism simulates the physical regulation effect of meteorological environment on pollutant deposition and accumulation at the feature level, enabling the model to adaptively focus on key chemical component features based on the current humidity level, thereby achieving deep feature fusion of environmental and chemical factors.

[0009] The fused feature vectors are then fed into a Long Short-Term Memory (LSTM) network. Utilizing the network's unique gating mechanism, the model captures long-term dependencies and temporal evolution patterns in the contamination accumulation process. During model training, a combination of backpropagation, adaptive moment estimation (IME), and random deactivation regularization ensures that the model parameters converge quickly to the optimal solution and exhibit good generalization ability.

[0010] Finally, this method restores the normalized values ​​of the model output to equivalent salt density and insoluble deposition density with physical dimensions through linear inverse transformation, providing an intuitive and quantitative basis for the insulation status assessment of transmission lines.

[0011] In a preferred embodiment, to comprehensively characterize the physical fields influencing pollution accumulation, the multi-source heterogeneous data covers three dimensions: intrinsic property data describing the geometry and material of the insulators, micrometeorological environment data determining atmospheric diffusion and humidity conditions, and atmospheric chemical composition data constituting the sources of pollutants. This comprehensive input feature construction ensures that the prediction model has sufficient information completeness.

[0012] In a preferred embodiment, this invention employs a decision tree-based anomaly detection mechanism to address potential outliers in the monitoring data. By constructing a decision tree and analyzing the depth and sample density of the leaf nodes, abnormal samples generated by sensor malfunctions or extreme emergencies can be effectively identified and removed, thereby improving the purity and distribution stability of the training data.

[0013] In a preferred embodiment, to reduce the dimensionality of the model input and eliminate the interference of redundant information, this invention utilizes grey relational analysis to quantify the correlation strength between each input variable and the surface contamination of the insulator. By setting a correlation threshold, key feature indicators that significantly influence contamination accumulation are selected, achieving dimensionality reduction and optimization of the feature space.

[0014] In a preferred embodiment, the specific computational logic of the dynamically weighted chemical component feature fusion network projects scalar humidity data into a high-dimensional feature space through a fully connected layer, and uses a Sigmoid activation function to transform it into attention weights with values ​​constrained between 0 and 1. Through element-wise multiplication, this weight vector recalibrates the feature channels, enhancing the model's sensitivity to specific pollutant features under high humidity conditions.

[0015] In a preferred embodiment, to accurately evaluate prediction performance and guide model optimization, the training process uses mean squared error as the loss function and introduces inverse normalization before the output layer. This ensures that the model not only achieves optimal mathematical convergence but also that the output directly corresponds to the pollution level indicators of the power industry standard.

[0016] The technical solution provided by this invention, by embedding a meteorological environment mutual attention mechanism in a deep learning network, effectively analyzes the nonlinear modulation mechanism of humidity field on particulate matter deposition process, and significantly improves the prediction accuracy and robustness of equivalent salt density and insoluble deposition density on insulator surface.

[0017] This invention provides a method for predicting surface contamination of insulators based on meteorological environmental factors. It has the following beneficial effects:

[0018] 1. This invention constructs a dynamically weighted chemical component feature fusion network and uses relative humidity to generate an attention mask to dynamically weight the high-dimensional features extracted by the convolutional neural network. This mechanism effectively simulates the physical regulation effect of atmospheric humidity on the deliquescence and adsorption efficiency of particulate matter, enabling the model to adaptively adjust feature weights according to changes in the meteorological environment. This significantly improves the prediction accuracy of the equivalent salt density and insoluble deposition density on the insulator surface under complex and variable meteorological conditions.

[0019] 2. This invention employs a data cleaning strategy based on decision trees and a feature selection strategy based on grey relational analysis. Abnormal data is logically removed by using the leaf node depth and sample ratio of the decision tree, and key influencing factors are quantitatively selected using grey relational analysis. This effectively removes noise interference and redundant information from the monitoring data. This not only reduces the input dimension and computational load of the neural network, but also improves the convergence stability and generalization performance of the model when facing non-stationary monitoring data in the field.

[0020] 3. This invention constructs a hybrid architecture that connects convolutional neural networks and long short-term memory networks. It utilizes convolutional neural networks to mine high-dimensional coupling features between insulator body, micrometeorological, and chemical composition data, and uses long short-term memory networks to capture long-term temporal dependencies in the pollution accumulation process. This spatiotemporal feature joint modeling method overcomes the shortcomings of traditional single models in terms of insufficient feature extraction capability when processing multivariate long-sequence data, and achieves accurate prediction of the pollution accumulation evolution trend of insulators. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the method flow in Example 1;

[0022] Figure 2 This is a schematic diagram of the data cleaning decision tree logic of the present invention. Detailed Implementation

[0023] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] See attached document Figure 1 This invention provides a method for predicting surface contamination of insulators based on meteorological environmental factors. The method includes:

[0025] Step 1: Acquisition of multi-source heterogeneous data and construction of sample library.

[0026] This step involves acquiring multi-source heterogeneous data related to contamination accumulation on insulator surfaces through online monitoring devices, micro-meteorological stations, and regional environmental monitoring platforms deployed on transmission line towers. The data includes insulator intrinsic property data, micro-meteorological environment data, and atmospheric chemical composition data.

[0027] Insulator body attribute data is used to characterize the physical structure and material properties of the insulator, specifically including the insulator's structural height, nominal diameter, umbrella structure parameters, and insulator material type. Among these, the umbrella structure parameters include umbrella extension and umbrella spacing, and the insulator material type includes glass, ceramic, or composite materials.

[0028] Micrometeorological environmental data is used to characterize the physical environment in which the insulators are located, specifically including ambient temperature, relative humidity, wind speed, wind direction, and rainfall at the sampling time. Among these, relative humidity data serves as the basic variable for generating query vectors in the subsequent feature fusion network, representing the saturation level of water molecules in the air.

[0029] Atmospheric chemical composition data are used to characterize the chemical composition of pollutants in the air surrounding the insulators, specifically including atmospheric particulate matter concentration data and gaseous pollutant concentration data. The atmospheric particulate matter concentration data includes PM2.5 concentrations (aerodynamic diameter ≤ 2.5 micrometers) and PM10 concentrations (aerodynamic diameter ≤ 10 micrometers). The gaseous pollutant concentration data includes sulfur dioxide (SO2), nitrogen dioxide (NO2), carbon monoxide (CO), and ozone (O3).

[0030] In the above atmospheric chemical composition data, PM2.5 and PM10 mainly serve as sources of insoluble deposition density on the insulator surface; SO2 and NO2, as acidic precursors, are converted into sulfates and nitrates under specific humidity conditions, and mainly serve as sources of equivalent salt density on the insulator surface.

[0031] While acquiring the above input variables, the corresponding measured values ​​of insulator surface contamination, including equivalent salt density (ESDD) and insoluble deposition density (NSDD), are obtained through periodic insulator surface sampling or equivalent sensor monitoring, and used as target labels for model training and validation.

[0032] The collected insulator property data, micro-meteorological environment data, atmospheric chemical composition data, and corresponding measured pollution levels are aligned according to a unified timestamp.

[0033] The aligned data is divided into training and test sets and stored in a historical sample database. The training set is used for subsequent parameter iteration and optimization of convolutional neural networks and long short-term memory networks, while the test set is used to evaluate the predictive performance of the trained model.

[0034] Step 2: Data cleaning and preprocessing.

[0035] See attached document Figure 2 Step two involves cleaning, serializing, and feature filtering the multi-source heterogeneous data in the historical sample database obtained in step one, in order to eliminate sensor noise, unify the measurement scale, and reduce data redundancy.

[0036] First, for abnormal observation data caused by occasional malfunctions or communication errors in the online monitoring device, a decision tree algorithm is used for identification and elimination. The measured value of insulator surface contamination is used as the target variable of the decision tree, and the corresponding micrometeorological environment data and atmospheric chemical composition data are used as feature variables to construct the decision tree model.

[0037] Decision tree models learn data distribution patterns by recursively segmenting the feature space. After training, the distribution of training samples in each leaf node of the decision tree is statistically analyzed. Data conforming to normal physical laws is usually distributed in leaf nodes deeper in the tree structure, and these nodes contain a larger number of samples. Conversely, anomalous data that does not conform to the conventional coupling relationship between meteorology and pollution tends to fall into leaf nodes with shallower tree depths or very few samples.

[0038] Traverse all leaf nodes of the decision tree, calculating the depth of each leaf node and the proportion of samples contained within it. Leaf nodes with a depth less than a preset depth threshold or a sample proportion less than a preset proportion threshold are marked as anomalous nodes. All sample data falling into anomalous nodes are removed from the historical sample database, completing data cleaning.

[0039] Next, the cleaned discrete observation data are organized into a multivariate time series matrix in chronological order. Let the total number of samples be... Each sample contains Using 10 characteristic indicators, construct the original data sequence matrix Its expression is as follows:

[0040] ;

[0041] in, Indicates the first The first characteristic index in the th Observations at each time step The value range is 1 to , The value range is 1 to .

[0042] To eliminate the differences in dimensions and orders of magnitude among insulator structural parameters, meteorological parameters, and chemical composition parameters, the original data sequence matrix was processed. Each row vector in the dataset is normalized. The min-max normalization method is used to linearly map the data to the [0,1] interval. The normalized data... The calculation formula is as follows:

[0043] ;

[0044] in, Indicates the first The minimum value among a sequence of characteristic indicators Indicates the first The maximum value among the sequence of feature indicators.

[0045] Meanwhile, in order to unify the statistical distribution benchmark in subsequent error analysis, the deviation standardization method was used to calculate the mean and variance of each characteristic index. The mean of each characteristic indicator and variance The calculation is as follows:

[0046] ;

[0047] ;

[0048] Based on mean and variance To obtain the standardized value for:

[0049] ;

[0050] Finally, grey relational analysis was used to assess and screen the importance of multi-source input features. The measured sequence of insulator surface contamination (ESDD or NSDD) was defined as the reference sequence. The normalized sequences of insulator intrinsic property data, micrometeorological environment data, and atmospheric chemical composition data are defined as comparison sequences. .

[0051] Calculate the comparison sequence With reference sequence Grey relational degree between The calculation formula is as follows:

[0052] ;

[0053] in, The resolution coefficient has a value range of (0,1) and is set to 0.5 in this embodiment. Indicates the reference sequence and the first The comparison sequence of the first feature index is in the th... The absolute difference at each point in time. This represents the minimum difference between two levels. This represents the maximum difference between the two levels.

[0054] Calculate the grey relational degree of all input features. Afterwards, The values ​​are arranged in descending order. A correlation threshold is set, and feature indicators with a gray correlation degree greater than the threshold are retained, while feature indicators with a gray correlation degree less than or equal to the threshold are removed. The normalized data sequences corresponding to the retained feature indicators are used as the input vectors for subsequent neural network models.

[0055] Step 3: Construction of CNN-LSTM model based on environmental and meteorological mutual attention.

[0056] This step constructs a hybrid neural network model that connects a local feature extraction network, a dynamically weighted feature fusion network, and a temporal evolution prediction network. (See attached diagram.) Figure 2 , Figure 2 This is a schematic diagram of the local feature extraction network structure of a convolutional neural network (CNN) according to an embodiment of the present invention.

[0057] First, the normalized multivariate time series matrix, filtered and retained from step two, is input into the CNN network. The dimension of this input matrix is... ,in Indicates the input time step. This represents the number of feature indicators after screening by grey relational analysis. The CNN network slides along the time dimension through one-dimensional convolution operations to capture the local coupling patterns between different meteorological environmental parameters and atmospheric chemical component parameters within a short period. The CNN network mainly consists of a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, and a flattening layer connected sequentially.

[0058] The first convolutional layer contains 64 convolutional kernels (filters), each with a size of 3. The kernels slide along the time axis of the input matrix with a stride of 1. At each position, the weight matrix of the kernel is multiplied by the corresponding local window data of the input matrix using the Hadamard Product, and then summed to generate a feature map. This operation extracts the correlation between meteorological abrupt changes and pollutant concentration fluctuations across three time steps. The output of the first convolutional layer is processed by a Modified Linear Unit (ReLU) activation function, mapping all negative values ​​to 0 to introduce non-linearity and remove redundant low-response features.

[0059] The next step is the first pooling layer, which employs max pooling. The pooling window size is set to 2. The pooling layer slides across the feature map, selecting the maximum value within the window's coverage area as the output at each step, discarding the rest. This operation halves the temporal dimension of the feature map while preserving the most salient feature responses, thereby reducing data dimensionality and improving the model's invariance to small temporal shifts.

[0060] The feature data output from the first pooling layer is then fed into the second convolutional layer. This second convolutional layer also has 64 kernels, each with a kernel size of 3, and uses the ReLU activation function. The second convolutional layer performs deep abstraction on the dimensionality-reduced features, extracting higher-level combined features. Subsequently, the data undergoes a second dimensionality reduction through a second max-pooling layer with a window size of 2.

[0061] After two layers of convolution and pooling operations, the original multivariate time series is transformed into a set of abstract feature matrices containing multiple channels. This feature matrix is ​​then fed into a flattening layer. The flattening layer unfolds the multidimensional feature matrix into a one-dimensional high-dimensional feature vector. This feature vector It includes all static correlation features and short-term dynamic features extracted from the input window, which serve as the input basis for the subsequent dynamic weighted network.

[0062] The dynamically weighted chemical component feature fusion network is configured to adaptively adjust the weights of each component in the feature vector extracted by the convolutional neural network based on the real-time ambient humidity, in order to generate a fused feature vector. This network receives two input signals: the first input is the high-dimensional feature vector output from the CNN flattening layer in step three. The second input is the current time step. normalized relative humidity values .

[0063] The network first includes a humidity attention branch network. This branch network consists of a fully connected (DenseLayer) layer and a non-linear activation layer connected in series. The normalized relative humidity value... The input is fed into this fully connected layer. The fully connected layer is configured with high-dimensional feature vectors. Using the same number of neurons of the same dimension, scalar humidity data is mapped to a dimensionless form through matrix multiplication and bias addition. The humidity feature vector, where equal The length.

[0064] Subsequently, the humidity feature vector is passed through a non-linear activation layer using the Sigmoid function. The Sigmoid function compresses each element of the vector into the open interval (0,1), generating a humidity attention mask vector. Each value in this mask vector represents the importance coefficient of the corresponding feature channel under the current humidity conditions.

[0065] Next, feature fusion is performed. The high-dimensional feature vector output by the CNN is then fused. With humidity attention mask vector Perform element-wise multiplication. This operation is equivalent to multiplying... Each feature component is assigned a gain coefficient determined by the current humidity, resulting in a dynamically weighted fused feature vector. .

[0066] This treatment process is based on the physicochemical mechanism of insulator contamination: the impact of contaminant components on insulator surface on insulation performance is primarily determined by ambient humidity. In environments with high relative humidity, sulfur dioxide in the air... and nitrogen dioxide Gaseous pollutants are more easily converted into ionic compounds such as sulfates and nitrates. In this case, the characteristic component associated with soluble salt density (ESDD) plays a dominant role in pollution level prediction. However, in dry environments with low relative humidity, atmospheric particulate matter... and Physical sedimentation dominates, at which point the feature component associated with non-soluble sediment density (NSDD) plays a dominant role. Through iterative training of the model on a large amount of historical data, the humidity attention branch network can adaptively learn this nonlinear mapping relationship, that is, automatically outputting a high-weight mask for soluble component features when the input is high humidity, and automatically outputting a high-weight mask for insoluble particle features when the input is low humidity.

[0067] The final generated fusion feature vector This vector not only includes the local spatiotemporal coupling features of multi-source data extracted by CNN, but also explicitly embeds the feature importance distribution information under current meteorological conditions. It is then transmitted to the input layer of the Long Short-Term Memory (LSTM) network.

[0068] This step will generate the dynamically weighted fused feature vector output in step three. The input is fed into an LSTM time-series evolution prediction network. This network consists of several LSTM memory cells expanded according to time series, and uses a gating mechanism to model the long-term dependence and nonlinear evolution process of insulator pollution. Each LSTM memory cell maintains a cell state. This state persists throughout the entire timeline, serving as a channel for information transmission.

[0069] at any time step The LSTM unit receives the fused feature vector at the current time step. and the hidden layer state at the previous time step As input, the computation process within the unit consists of three stages: the forget gate, the input gate, and the output gate.

[0070] The first stage involves the calculation of the forget gate. The forget gate determines the cell's state from the previous time step. How much information is discarded? This process is implemented through a Sigmoid activation function layer, whose output value is between 0 and 1, where 0 represents complete forgetting and 1 represents complete retention. Physically, this process simulates the loss of contaminants from the insulator surface under the action of natural rain or strong winds. (Forget gate output) The calculation formula is as follows:

[0071] ;

[0072] in, This represents the Sigmoid activation function; The weight matrix representing the forget gate; This indicates that the hidden state from the previous time step is concatenated with the current input vector. This represents the bias term for the forget gate.

[0073] The second stage involves input gate computation and cell state updating. This stage determines how much of the new information at the current moment will be stored in the cell state. First, the input gate layer determines the extent of the update value. And a new candidate cell state vector is constructed through the Tanh layer. Physically, this process simulates the accumulation of new pollutants on the surface of insulators under current meteorological and chemical conditions. The calculation formula is as follows:

[0074] ;

[0075] ;

[0076] in, and These are the weight matrices for the input gate and the candidate state, respectively; and For the corresponding bias term; It is the hyperbolic tangent activation function.

[0077] Subsequently, by combining the outputs of the forget gate and the input gate, the cell state is updated to obtain the current cell state. :

[0078] ;

[0079] Among them, This represents the Hadamard product of a matrix, which is the product of corresponding elements. The formula indicates that the current total amount of contamination equals the sum of the amount of residual contamination after cleaning and the amount of newly deposited contamination.

[0080] The third stage is the calculation of the output gate. This stage is based on the updated cell state. This determines the output of the hidden layer at the current moment. The output gate first determines the output portion of the cell state through a Sigmoid layer. Then the cell state was treated with Tanh and then... Multiplication. The calculation formula is as follows:

[0081] ;

[0082] ;

[0083] Among them, This is the weight matrix of the output gate; This is the bias term for the output gate.

[0084] After the time-series recursive computation of the LSTM layer, the hidden state at the last time step. The input is fed into a fully connected regression layer (DenseLayer). The fully connected regression layer maps the high-dimensional hidden state to a one-dimensional scalar output, which is the predicted value of the insulator surface contamination. .

[0085] Predicted value This corresponds to the equivalent salt density (ESDD) or non-soluble deposition density (NSDD) on the insulator surface.

[0086] Step 4: Model training strategy and parameter optimization.

[0087] After constructing the aforementioned CNN-LSTM hybrid neural network model, this embodiment of the invention enters the model training and parameter optimization stage. This stage minimizes the error between the predicted and actual values ​​using the backpropagation algorithm, thereby determining all weight matrices in the network. and bias vector The optimal solution.

[0088] First, configure the model's training hyperparameters. The Adam (Adaptive Moment Estimation) optimization algorithm is chosen as the parameter update strategy. The Adam algorithm combines the advantages of momentum gradient descent and RMSProp, designing independent adaptive learning rates for different parameters by calculating the first and second moment estimates of the gradient. In this embodiment, the initial learning rate of the Adam optimizer is set to 0.001, and the exponential decay rate of the first moment estimate is... The exponential decay rate is set to 0.9 for the second moment estimation. The numerical stability constant is set to 0.999. Set to 10 -8 .

[0089] Define the model's loss function. Since insulator surface contamination prediction is a regression problem, this embodiment uses Mean Squared Error (MSE) as the loss function. This loss function is used to quantify the model's predictions in the current training batch. Compared with the actual label value The differences between them. For those containing A training batch of 100 samples, loss function The calculation is as follows:

[0090] ;

[0091] in, For the first The measured soil sludge level (ESDD or NSDD) of the sample for each characteristic indicator. This is the predicted value output by the model for this sample. The number of samples used to calculate the loss in this case.

[0092] The training process employs mini-batch gradient descent. The training set data constructed in step one is divided into several batches, with the batch size set to 32 or 64. In each training epoch, a complete batch of data is input into the network for forward propagation. The data sequentially passes through a CNN feature extraction layer, a dynamically weighted feature fusion layer, and an LSTM temporal prediction layer, finally outputting the prediction result at a fully connected layer.

[0093] Calculate the loss function value for the current batch. Then, backpropagation is performed. The loss function is calculated relative to the weights of each layer in the network. and bias The gradient is calculated. Based on the calculated gradient and the learning rate set by the Adam optimizer, the parameters of each layer are updated. This process is repeated until a preset number of iterations is reached (e.g., 100 to 200 iterations), or the loss value on the validation set no longer decreases within a certain number of consecutive iterations (i.e., early stopping is triggered) to prevent the model from overfitting.

[0094] To further improve the model's generalization ability and prevent overfitting, a Dropout (random deactivation) regularization strategy is introduced between the LSTM layer and the final fully connected layer. The Dropout ratio is set to 0.2, meaning that in each training iteration, a portion of the neuron activation values ​​in the LSTM hidden layer output vector are randomly set to 0 with a 20% probability. This operation forces the network to learn more robust distribution features rather than relying on certain specific local features.

[0095] After training, the weights of the convolutional kernels, the weights of the fully connected layers of the attention mechanism, the weights of each gate unit of the LSTM, and the final values ​​of all bias terms are saved to obtain the trained insulator surface contamination prediction model.

[0096] See attached document Figure 1 The optimization algorithm used in this embodiment of the invention mainly refers to the Adaptive Moment Estimation (Adam) algorithm for updating neural network model parameters and its specific execution strategy in hybrid networks.

[0097] In each iteration of model training, the core task of the optimization algorithm is to fine-tune the weight matrix and bias vector in the network based on the gradient information of the loss function relative to each parameter. Since the model constructed in this invention includes a convolutional neural network (CNN), an environmental meteorological mutual attention network, and a long short-term memory network (LSTM), the parameters at different levels have different sensitivities and gradient scales. Therefore, the Adam algorithm is used to achieve adaptive learning rate adjustment for each parameter.

[0098] The specific execution of the optimization process includes the following technical aspects:

[0099] First, backpropagation of the gradient is performed. After completing forward propagation and calculating the mean squared error loss for the current batch, the algorithm first calculates the gradient of the output layer (fully connected regression layer) relative to the prediction error. This gradient information is then propagated backward along the network structure. For the LSTM layer, the Backpropagation Through Time (BPTT) algorithm is used to expand the error gradient along the time axis, propagating it forward step by step to calculate the forget gate, input gate, output gate, and weight matrix in cell state updates. The gradients of the bias term and the multiplication term are calculated. For a dynamically weighted feature fusion network, the gradients are split by the multiplication operation at the feature fusion node, one path leading to the fully connected layer of the attention mechanism, and the other path leading to the flattening layer of the CNN. Finally, the gradients are calculated by inverse operations of the pooling layer and the convolutional layer to obtain the gradients of all convolutional kernel weights in the first and second convolutional layers.

[0100] Secondly, the Adam algorithm is used to maintain the first and second moment estimates of the gradient. For each parameter to be optimized in the network, the algorithm calculates the exponential moving average of its gradient (i.e., the first moment, representing the average direction of the gradient) and the exponential moving average of the squared gradient (i.e., the second moment, representing the dispersion or variance of the gradient). These two moment estimates are updated in each iteration based on the current gradient value and the moment estimates from the previous time step, with the update rate determined by the aforementioned exponential decay rate. and control.

[0101] Next, bias correction and parameter updates are performed. Since the initial values ​​of the first and second moments are typically initialized as zero vectors, the estimates will be biased towards zero in the early stages of training. Therefore, the algorithm performs bias correction on the two moment estimates to eliminate the influence of initialization bias. After correction, the algorithm divides the first moment estimate by the square root of the second moment estimate (adding the numerical stability constant to the denominator). The update step size is calculated using a step size calculation method. Finally, the current parameter value is subtracted from the product of the update step size and the global learning rate to complete one parameter update.

[0102] Through the aforementioned mechanism, the Adam algorithm can automatically allocate larger update step sizes to parameters with lower update frequencies (such as weights corresponding to sparse meteorological features), while allocating smaller update step sizes to parameters with higher update frequencies. This mechanism effectively solves the convergence difficulty problem caused by the difference in feature sparsity in multi-source data feature fusion networks, ensuring that the CNN local feature extractor and the LSTM temporal predictor can converge collaboratively to the optimal solution.

[0103] In this embodiment of the invention, the loss function network is configured as a core evaluation metric to quantify the model's predictive performance and guide the direction of parameter updates.

[0104] Since the prediction targets of this invention—the equivalent salt density (ESDD) and insoluble deposition density (NSDD) on the insulator surface—are both continuously varying real values, this is a typical regression prediction problem. Therefore, this embodiment constructs an objective function with mean squared error (MSE) as its core. This loss function is not merely a mathematical formula; at a physical level, it characterizes the degree of deviation between the pollution accumulation trajectory predicted by the model and the actual physicochemical deposition process occurring on the insulator surface.

[0105] The specific loss calculation process is as follows: In each iteration of the training phase, the model receives a batch of input data and outputs the predicted value vector corresponding to each sample in the batch through forward propagation. Simultaneously, it retrieves the time-synchronized measured label values ​​that strictly correspond to the input of that batch from the historical sample database. The loss function network calculates the squared Euclidean distance between the predicted value and the measured label value for each sample.

[0106] The significance of using squaring operations instead of absolute value operations lies in the fact that the squaring term nonlinearly amplifies the error. For insulator pollution prediction, small prediction deviations (such as fluctuations within the measurement error range) are compressed, while larger prediction deviations (such as failure to accurately predict sudden pollution events) are significantly amplified. This mechanism forces the model to prioritize and correct those serious prediction errors during training, thereby more sensitively capturing the drastic fluctuations in pollution caused by meteorological changes (such as sandstorms or high-humidity dense fog).

[0107] The final loss value for the current batch is obtained by summing the squared errors of all samples within the batch and taking the average. The loss value As the starting signal for backpropagation, it directly determines the direction and magnitude of gradient descent.

[0108] The minimization of this loss function directly drives the parameter learning of the aforementioned dynamically weighted chemical component feature fusion network. Specifically, if the model fails to correctly adjust the weights of chemical components according to relative humidity (e.g., ignoring the conversion of SO2 in high humidity environments), it will lead to a significant deviation between the predicted ESDD value and the measured value, resulting in a substantial loss. The strong gradient signal generated by this loss value is backpropagated to the weight matrix of the attention network, forcing the network to adjust its response mode to humidity features until the model can simulate the weighted logic of humidity components that conforms to the physicochemical mechanism. In this way, the loss function achieves a closed-loop fusion of physical mechanism constraints and data-driven training.

[0109] Step 5: Output the prediction results.

[0110] In this embodiment of the invention, after the model outputs numerical values ​​through a fully connected regression layer, it does not directly use them as the final result. Instead, it performs data post-processing and physical dimension restoration operations to generate insulator surface contamination prediction indicators with practical engineering guidance significance.

[0111] First, the raw predicted values ​​output by the model at the fully connected layer. It lies within the normalized numerical range of [0,1]. To convert it to physical dimensions conforming to power industry standards, the system calls the statistical parameters of the training set label data recorded in the data preprocessing stage of step two, i.e., the maximum value. and minimum value The inverse normalization algorithm is used to perform a linear inverse transformation on the original predicted values. The calculation process involves normalizing the predicted values... Multiply by the range plus the minimum value This allows us to reconstruct the true predicted value. After the above inverse transformation process, the present invention finally outputs two independent physical quantities: equivalent salt density (ESDD) and insoluble sediment density (NSDD).

[0112] The unit of equivalent salt density (ESDD) is milligrams per square centimeter. This index quantifies the total amount of conductive soluble salts deposited on the surface of the insulator. Physically, the ESDD value directly corresponds to the conductivity level of the electrolyte solution formed after the insulator surface becomes damp and dissolves, and is a core parameter determining the magnitude of the leakage current and the flashover voltage threshold of the insulator under humid conditions. The unit for non-soluble deposition density (NSDD) is also milligrams per square centimeter (mg / cm²). This index quantifies the total amount of insoluble particulate matter such as dust and particulate matter deposited on the surface of the insulator. Physically, while NSDD does not directly participate in conductivity, it constitutes a porous dielectric layer with a sponge-like ability to absorb and retain moisture. This layer structure can significantly prolong the wetting time of the insulator surface and, to some extent, hinder the movement of the discharge arc.

[0113] The system outputs the calculated ESDD and NSDD predicted values ​​to a visual monitoring terminal or a downstream insulation safety assessment network. Maintenance personnel or automated systems use these two indicators, combined with the insulator's specific creepage distance parameters, and refer to power industry pollution level standards (such as GB / T26218) to determine the current pollution level (Level I to IV) of the insulator. When the predicted value exceeds a preset safety threshold, the system generates a cleaning and maintenance command, thereby achieving condition-based transmission line insulation maintenance.

[0114] To objectively and quantitatively verify the model's performance in predicting insulator surface contamination, this invention constructs an evaluation index system encompassing multiple statistical dimensions. This system is configured to measure the degree of agreement between the model's predicted value sequence and the actual monitored value sequence from different perspectives, specifically employing the root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (COP). () is used as the core evaluation indicator.

[0115] First, the root mean square error (RMSE) is used to measure the absolute error level of the predicted values. RMSE squares the error, making it extremely sensitive to very large errors. In insulator pollution monitoring, sudden changes in pollution levels due to severe weather (such as sandstorms) are often key triggers for flashover accidents. The RMSE index effectively penalizes the model's prediction bias at such extreme points, ensuring the model's ability to capture sudden pollution events. The formula for calculating RMSE is as follows:

[0116] ;

[0117] in, This represents the total number of test samples. Indicates the first Measured pollution levels (ESDD or NSDD) of insulators for each sample. This represents the model's predicted value for this sample. The smaller the RMSE value, the lower the dispersion between the model's predicted value and the true value, and the higher the prediction accuracy.

[0118] Secondly, the Mean Absolute Percentage Error (MAPE) is used to measure the relative accuracy of the prediction results. This is because the surface contamination levels of insulators vary significantly across different regions and voltage levels (e.g., ESDD in lightly polluted areas may be as low as 0.03 mg / cm³). 2， In heavily polluted areas, the concentration can reach 0.3 mg / cm³. 2， Absolute error alone is insufficient for comparing model performance across different scenarios. MAPE, by calculating relative bias, eliminates the influence of the base numerical magnitude of contamination, intuitively reflecting the percentage deviation of the prediction from the true value. The formula for calculating MAPE is as follows:

[0119] ;

[0120] A MAPE value close to 0% indicates that the model's relative error is extremely small, demonstrating good generalization adaptability. Finally, the coefficient of determination (...) This is used to evaluate the model's ability to explain data variability, i.e., goodness of fit. The index quantifies the extent to which the model's predictions can accurately reflect the fluctuations in insulator pollution levels caused by meteorological and chemical factors. Its calculation formula involves the ratio of the sum of squared predicted residuals to the sum of squared total deviations, as follows:

[0121] ;

[0122] in, This represents the arithmetic mean of the measured values ​​of all test samples. The value range of is usually between 0 and 1. When The closer the value is to 1, the more successfully the model has extracted the complex nonlinear mapping relationship between the input features and the output target, and the more closely the prediction curve follows the evolution trajectory of the true filth level; conversely, if... A lower value indicates that the model has failed to effectively capture the key patterns in the data.

[0123] In practical applications, the above three indicators are used in combination. Only when both RMSE and MAPE are below the preset tolerance threshold, and Only when the model's performance exceeds the preset fitting threshold is it deemed qualified and deployed to the online monitoring system for transmission lines to perform actual prediction tasks.

Claims

1. A method for predicting surface contamination of insulators based on meteorological environmental factors, characterized in that, Includes the following steps: Step 1: Obtain multi-source heterogeneous data related to the accumulation of contaminants on the insulator surface. The multi-source heterogeneous data includes insulator body property data, micro-meteorological environment data, and atmospheric chemical composition data, and obtain the corresponding measured values ​​of insulator surface contamination. Then, align the multi-source heterogeneous data with the measured values ​​of insulator surface contamination over time to construct a historical sample database. Step 2: Perform data cleaning, normalization, and feature filtering on the multi-source heterogeneous data in the historical sample database to obtain a normalized multivariate time series matrix; Step 3: Construct a hybrid neural network model based on environmental meteorological mutual attention and convolutional neural network long short-term memory network; the hybrid neural network model includes a convolutional neural network local feature extraction network, a dynamically weighted chemical component feature fusion network, and a long short-term memory network temporal evolution prediction network connected in sequence; Step 4: Based on the data in the historical sample database, train the hybrid neural network model using the backpropagation algorithm until the model converges; Step 5: Use the trained hybrid neural network model to predict the surface contamination of the insulator during the prediction period and output the prediction results.

2. The insulator surface pollution prediction method based on meteorological environmental factors according to claim 1, characterized in that, In step one, the insulator body attribute data includes the insulator structural height, nominal diameter, umbrella extension amount, umbrella spacing, and insulator material type; The micrometeorological environmental data includes ambient temperature, relative humidity, wind speed, wind direction, and rainfall. The atmospheric chemical composition data includes the concentration of particulate matter with an aerodynamic diameter of less than or equal to 2.5 micrometers, the concentration of particulate matter with an aerodynamic diameter of less than or equal to 10 micrometers, the concentration of sulfur dioxide, the concentration of nitrogen dioxide, the concentration of carbon monoxide, and the concentration of ozone. The measured values ​​of surface contamination of the insulator include equivalent salt density and insoluble deposition density.

3. The insulator surface pollution prediction method based on meteorological environmental factors according to claim 1, characterized in that, In step two, the data cleaning steps include: The measured values ​​of insulator surface contamination were used as the target variable, and micro-meteorological environment data and atmospheric chemical composition data were used as feature variables to construct a decision tree model. The depth of each leaf node and the proportion of samples contained in the node are calculated by traversing all leaf nodes of the decision tree model. Leaf nodes whose depth is less than a preset depth threshold or whose sample ratio is less than a preset ratio threshold are marked as abnormal nodes. All sample data falling into the abnormal node will be removed from the historical sample database.

4. The insulator surface pollution prediction method based on meteorological environmental factors according to claim 1, characterized in that, In step two, the feature filtering steps include: Grey relational analysis was used to define the sequence of measured values ​​of insulator surface contamination as the reference sequence, and the sequences of normalized insulator body property data, micrometeorological environment data, and atmospheric chemical composition data as the comparison sequence. The grey relational analysis algorithm was used to calculate the grey relational degree between each comparison sequence and the reference sequence. Feature indicators with a gray correlation degree greater than a preset correlation degree threshold are retained, while feature indicators with a gray correlation degree less than or equal to the preset correlation degree threshold are removed.

5. The insulator surface pollution prediction method based on meteorological environmental factors according to claim 1, characterized in that, In step three, the convolutional neural network local feature extraction network includes a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, and a flattening layer connected in sequence. By inputting the normalized multivariate time series matrix into the first convolutional layer, performing two convolutional operations and max pooling operations, and then expanding it using the flattening layer, a one-dimensional high-dimensional feature vector is calculated. The dynamically weighted chemical component feature fusion network performs the following operations: The system receives the high-dimensional feature vector output by the local feature extraction network of the convolutional neural network and the relative humidity data processed in step two. It generates an attention mask using the relative humidity data and uses the attention mask to weight the high-dimensional feature vector to obtain a fused feature vector. The fused feature vector is then input into the long short-term memory network temporal evolution prediction network.

6. The insulator surface pollution prediction method based on meteorological environmental factors according to claim 5, characterized in that, In step three, the specific operation process of the dynamically weighted chemical component feature fusion network includes: Obtain the normalized relative humidity value at the current time step; The normalized relative humidity value is input into a fully connected layer, and the scalar humidity data is mapped into a humidity feature vector through matrix multiplication. The dimension of the humidity feature vector is the same as the dimension of the high-dimensional feature vector. The humidity feature vector is processed by the Sigmoid activation function to generate a humidity attention mask vector with values ​​between 0 and 1. The fused feature vector is obtained by multiplying the high-dimensional feature vector with the humidity attention mask vector element by element.

7. The insulator surface pollution prediction method based on meteorological environmental factors according to claim 6, characterized in that, In step three, the Long Short-Term Memory Network Temporal Evolution Prediction Network is composed of several Long Short-Term Memory Network storage units expanded in a time sequence. Each Long Short-Term Memory (LSTM) network storage unit receives the fused feature vector at the current time step and the hidden layer state at the previous time step as input; The long short-term memory network storage unit calculates forgotten information through a forgetting gate, calculates cell state update information through an input gate, and calculates the hidden layer output at the current moment through an output gate. The hidden layer output of the last time step is connected to a fully connected regression layer, which outputs a normalized predicted value of the insulator surface contamination.

8. The insulator surface pollution prediction method based on meteorological environmental factors according to claim 1, characterized in that, In step four, the training strategy includes: The mean square error is used as the loss function, and the error between the normalized predicted value output by the model and the normalized value of the measured surface contamination of the insulator is calculated through the loss function. An adaptive moment estimation optimization algorithm is adopted. Based on the gradient of the loss function with respect to the network parameters, the first-order moment estimate and second-order moment estimate of the gradient are calculated by the adaptive moment estimation formula, and the weight matrix and bias vector in the network are updated adaptively. A random deactivation regularization strategy is introduced between the temporal evolution prediction network of the long short-term memory network and the fully connected regression layer, in which the activation values ​​of some neurons are randomly set to zero during training iterations.

9. The insulator surface pollution prediction method based on meteorological environmental factors according to claim 8, characterized in that, The loss function, i.e., the mean squared error, is calculated as follows: Obtain the model prediction value and the corresponding measured label value for each sample in the training batch; The square of the Euclidean distance between the model's predicted value and the measured label value for each sample is calculated using the Euclidean distance formula. The loss value of the current batch is calculated by summing the squared errors of all samples in the training batch and taking the average.

10. The insulator surface pollution prediction method based on meteorological environmental factors according to claim 1, characterized in that, In step five, the step of outputting the prediction result includes: Obtain the maximum and minimum values ​​of the measured surface contamination of insulators in the training set data; The true predicted value with physical dimensions is obtained by performing an inverse linear transformation on the normalized predicted value output by the hybrid neural network model using the maximum and minimum values. The true predicted values ​​include the equivalent salt density prediction value in milligrams per square centimeter and the insoluble sediment density prediction value.

Images