Training Method and System for Transmission Line Fault Rate Prediction Model Based on Machine Learning

Abstract

This invention relates to the field of model training technology, specifically to a training method and system for a transmission line fault rate prediction model based on machine learning. This invention employs an improved inverse distance weighted interpolation algorithm, introducing timeliness weights, reliability weights, and terrain constraint weights to map regional meteorological grid data to tower coordinates and align the time series. It statistically analyzes the frequency of meteorological factors in fault and non-fault samples, calculating single-factor correlation weights and multi-factor collaborative correlation weights. The meteorological factor feature vectors, correlation weights, and fault labels are combined to form a training dataset. A neural network model including an attention layer is constructed, and attention weights are calculated based on single-factor correlation weights to train the model. This invention guides the attention mechanism through correlation weights, enabling the model to prioritize highly correlated meteorological factors, effectively improving prediction accuracy.

Description

Technical Field

[0001] This invention relates to the field of model training technology, specifically to a method and system for training a transmission line fault rate prediction model based on machine learning. Background Technology

[0002] Transmission lines play a vital role in the power system by transmitting electrical energy. Their operating environment is complex and variable, and they are often threatened by faults caused by various meteorological factors such as lightning, strong winds, and flashover. Accurately predicting the fault rate is of great significance to ensuring the safe and stable operation of the power grid.

[0003] Currently, transmission line fault rate prediction methods mainly rely on training neural network models using meteorological data, which improves prediction accuracy compared to methods without meteorological data. However, existing technologies have significant shortcomings: First, directly mapping meteorological data to transmission line locations using traditional interpolation methods ignores the impact of timeliness, site reliability, and terrain factors, resulting in insufficient mapping accuracy. Second, directly inputting meteorological factor features into the neural network fails to quantify the differences in the correlation strength between different meteorological factors and faults, and does not consider the synergistic effect of multiple factors, leading to insufficient ability of the model to capture key meteorological factors and limiting further improvement in prediction accuracy. Summary of the Invention

[0004] This invention provides a training method and system for a transmission line fault rate prediction model based on machine learning, in order to improve the accuracy of fault prediction.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] This invention relates to a machine learning-based method for training a transmission line fault rate prediction model, comprising:

[0007] S100: Acquire regional meteorological grid data and tower coordinate data, and use an improved inverse distance weighted interpolation algorithm to map the regional meteorological grid data to the tower coordinate positions. Align the time series according to the fault occurrence time, extract meteorological factor feature vectors and label them with fault tags to form sample units. The improved inverse distance weighted interpolation algorithm introduces meteorological element timeliness weight, meteorological station reliability weight and terrain constraint weight to determine the comprehensive weight, and calculates the interpolation result based on the comprehensive weight.

[0008] S200: Statistically analyze the frequency of occurrence of each meteorological factor in the fault samples and non-fault samples, calculate the single-factor association weight based on the probability difference between the two, and calculate the multi-factor collaborative association weight based on the covariance of different meteorological factors in the fault samples and the single-factor association weight.

[0009] S300: Combine the meteorological factor feature vectors, single-factor association weights, multi-factor collaborative association weights, and fault labels to form a training dataset;

[0010] S400: Construct a neural network model containing an attention layer, calculate attention weights in the attention layer based on the single-factor association weights, train the neural network model using the training dataset, and obtain a transmission line fault rate prediction model.

[0011] As a preferred embodiment of the present invention, the step of determining the comprehensive weight includes:

[0012] Calculate the timeliness weight of meteorological elements, and correct the station weights by using a time decay factor based on the time difference between the interpolation time of the target point and the actual measurement time of each meteorological station.

[0013] Calculate the reliability weight of meteorological stations, identify abnormal stations whose measured values ​​deviate from the mean of neighboring stations by more than three standard deviations, and assign low reliability weights to abnormal stations.

[0014] Calculate the terrain constraint weights and determine the terrain similarity weights based on the elevation and slope differences between the target point and each meteorological station.

[0015] The timeliness weight of the meteorological elements, the reliability weight of the meteorological stations, and the terrain constraint weight are combined to obtain a comprehensive weight.

[0016] As a preferred embodiment of the present invention, the step of calculating the interpolation result includes:

[0017] Using the tower coordinates as the target point, calculate the number of stations per unit area around the target point, and determine the attenuation coefficient based on the station density. The lower the station density, the greater the attenuation coefficient.

[0018] The number of neighboring stations to be included in the interpolation calculation is determined based on the distance between the target point and the nearest station; the greater the distance, the more neighboring stations are selected.

[0019] The overall weights of each neighboring site are normalized.

[0020] The interpolation result of the target point is obtained by multiplying the measured meteorological data of each neighboring station with the corresponding normalized comprehensive weight and then summing the results.

[0021] As a preferred embodiment of the present invention, the meteorological factors include lightning meteorological factors, strong wind meteorological factors, and short-term heavy precipitation meteorological factors:

[0022] The lightning meteorological factors are characterized by lightning density, lightning intensity, lightning distance attenuation coefficient, and lightning polarity ratio.

[0023] The wind meteorological factors are characterized by the maximum instantaneous wind speed, the rate of change of wind speed over time, the duration of wind speed exceeding the threshold, and the change of wind speed in the vertical direction.

[0024] The short-duration heavy precipitation meteorological factor is characterized by the amount of precipitation per unit time.

[0025] As a preferred embodiment of the present invention, the single-factor association weight is calculated in the following manner:

[0026] The probability of occurrence of each meteorological factor in the fault sample and the probability of occurrence of each meteorological factor in the non-fault sample are statistically analyzed.

[0027] Calculate the ratio of the probability of occurrence of the same meteorological factor in faulty samples to that in non-faulty samples and take the logarithm;

[0028] The positive and negative directions of the weights are determined by the sign function based on the relationship between the probability of occurrence of meteorological factors in faulty samples and non-faulty samples.

[0029] The formula for calculating the single-factor association weight is as follows:

[0030] ;

[0031] in, For the first Single-factor association weights of individual meteorological factors The first fault sample The probability of occurrence of each meteorological factor The first non-fault sample The probability of occurrence of each meteorological factor It is a symbolic function.

[0032] As a preferred embodiment of the present invention, the calculation of the multi-factor synergistic association weights includes:

[0033] Select two different meteorological factors from the fault sample and calculate the covariance between the two meteorological factors.

[0034] Obtain the single-factor correlation weights corresponding to each of the two meteorological factors;

[0035] The covariance is multiplied by the two single-factor association weights, and a synergy coefficient is introduced for adjustment to obtain the multi-factor synergistic association weight of the meteorological factor pair.

[0036] Multi-factor synergistic association weights were calculated for all meteorological factors to obtain a complete set of multi-factor synergistic association weights.

[0037] As a preferred embodiment of the present invention, the step of calculating attention weights includes:

[0038] Obtain the feature vectors of each meteorological factor and their corresponding single-factor correlation weights;

[0039] Calculate the product of the single-factor association weight and the corresponding meteorological factor feature vector;

[0040] The product of all meteorological factor feature vectors is normalized to obtain the attention weights of each meteorological factor feature vector.

[0041] The attention weights are multiplied by the corresponding meteorological factor feature vectors to obtain the attention-weighted features.

[0042] This invention also proposes a machine learning-based transmission line fault rate prediction model training system, comprising:

[0043] The data mapping module is used to acquire regional meteorological grid data and tower coordinate data. An improved inverse distance weighted interpolation algorithm is used to map the regional meteorological grid data to the tower coordinate positions. The time series is aligned according to the fault occurrence time, meteorological factor feature vectors are extracted and fault labels are marked to form sample units. The improved inverse distance weighted interpolation algorithm introduces meteorological element timeliness weight, meteorological station reliability weight and terrain constraint weight to determine the comprehensive weight, and the interpolation result is calculated based on the comprehensive weight.

[0044] The weight quantization module is used to count the frequency of occurrence of each meteorological factor in fault samples and non-fault samples, calculate the single-factor association weight based on the probability difference between the two, and calculate the multi-factor collaborative association weight based on the covariance of different meteorological factors in fault samples and the single-factor association weight.

[0045] The dataset construction module is used to combine the meteorological factor feature vectors, single-factor association weights, multi-factor collaborative association weights, and fault labels to form a training dataset.

[0046] The model training module is used to construct a neural network model containing an attention layer, calculate attention weights based on the single-factor association weights in the attention layer, and train the neural network model using the training dataset to obtain a transmission line fault rate prediction model.

[0047] The present invention also proposes a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described machine learning-based transmission line fault rate prediction model training method.

[0048] The present invention also proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the above-described machine learning-based transmission line fault rate prediction model training method.

[0049] The beneficial effects of this invention are:

[0050] 1. This invention introduces the timeliness weight of meteorological elements, the reliability weight of meteorological stations, and the terrain constraint weight through an improved inverse distance weighted interpolation algorithm. It also adopts a spatial density adaptive strategy to dynamically adjust the attenuation coefficient and the number of neighboring stations, thus solving the problem of insufficient mapping accuracy caused by traditional interpolation methods ignoring time attenuation, station anomalies, and terrain differences. This invention achieves high-precision spatiotemporal mapping of meteorological data to tower locations.

[0051] 2. This invention quantifies the correlation strength between meteorological factors and faults by calculating single-factor correlation weights, captures the synergistic effect between meteorological factors through multi-factor collaborative correlation weights, integrates statistical prior knowledge into the training data, and calculates attention weights based on single-factor correlation weights in the attention layer to form a weight-guided attention mechanism. This enables the model to adaptively focus on highly correlated meteorological factors, significantly improving the accuracy of fault prediction. Attached Figure Description

[0052] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0053] Figure 1 This is a flowchart illustrating the training method for the transmission line fault rate prediction model based on machine learning according to the present invention.

[0054] Figure 2 This is a schematic diagram of the structure of the transmission line fault rate prediction model training system based on machine learning of the present invention;

[0055] Figure 3 This is a structural block diagram of an electronic device provided by the present invention. Detailed Implementation

[0056] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0057] Example 1: As Figure 1 As shown, the present invention provides a machine learning-based method for training a transmission line fault rate prediction model, comprising:

[0058] S100: Acquire regional meteorological grid data and tower coordinate data, and use an improved inverse distance weighted interpolation algorithm to map the regional meteorological grid data to the tower coordinate positions. Align the time series according to the fault occurrence time, extract meteorological factor feature vectors and label them with fault tags to form sample units. The improved inverse distance weighted interpolation algorithm introduces meteorological element timeliness weight, meteorological station reliability weight and terrain constraint weight to determine the comprehensive weight, and calculates the interpolation result based on the comprehensive weight.

[0059] Furthermore, the step of calculating the interpolation result includes:

[0060] Using the tower coordinates as the target point, calculate the number of stations per unit area around the target point, and determine the attenuation coefficient based on the station density. The lower the station density, the greater the attenuation coefficient.

[0061] The number of neighboring stations to be included in the interpolation calculation is determined based on the distance between the target point and the nearest station; the greater the distance, the more neighboring stations are selected.

[0062] The overall weights of each neighboring site are normalized.

[0063] The interpolation result of the target point is obtained by multiplying the measured meteorological data of each neighboring station with the corresponding normalized comprehensive weight and then summing the results.

[0064] Specifically, the tower coordinates of the transmission line GIS data were first refined. The latitude, longitude, and altitude of each tower were re-collected using a BeiDou positioning system to correct the drift errors of traditional GIS data and obtain accurate tower coordinate data. Simultaneously, regional meteorological grid data was acquired. Each regional meteorological grid contains meteorological stations for actual meteorological data measurement.

[0065] In the process of mapping regional meteorological grid data to tower coordinates, an improved inverse distance weighted interpolation algorithm is adopted. This algorithm introduces the timeliness weight of meteorological elements, the reliability weight of meteorological stations, and the terrain constraint weight on the basis of the traditional inverse distance weighted interpolation algorithm to determine the comprehensive weight. Then, the interpolation result is calculated based on the comprehensive weight and the measured regional meteorological grid data.

[0066] Furthermore, the step of determining the comprehensive weights includes:

[0067] Calculate the timeliness weight of meteorological elements, and correct the station weights by using a time decay factor based on the time difference between the interpolation time of the target point and the actual measurement time of each meteorological station.

[0068] Calculate the reliability weight of meteorological stations, identify abnormal stations whose measured values ​​deviate from the mean of neighboring stations by more than three standard deviations, and assign low reliability weights to abnormal stations.

[0069] Calculate the terrain constraint weights and determine the terrain similarity weights based on the elevation and slope differences between the target point and each meteorological station.

[0070] The timeliness weight of the meteorological elements, the reliability weight of the meteorological stations, and the terrain constraint weight are combined to obtain a comprehensive weight.

[0071] Specifically, the timeliness weights of meteorological elements are calculated. Because the correlation between transmission line faults and meteorology is highly time-sensitive—for example, the impact of short-duration heavy rainfall on a fault is concentrated within one hour after the rainfall occurs—it is necessary to correct the weights using a time decay factor. Timeliness weights of meteorological elements at individual meteorological stations The calculation formula is:

[0072] ;

[0073] in, Interpolation time for target point and the first The time difference between the actual measured times at each meteorological station The time decay coefficient is dynamically adjusted according to the weather type, and the target point is the pole point.

[0074] The spatiotemporal joint weight is calculated by combining spatial distance. Spatiotemporal joint weight of individual meteorological stations Represented as:

[0075] ;

[0076] in, For the target point and the first Euclidean distances between weather stations This is the dynamic attenuation coefficient.

[0077] Secondly, the reliability weights of meteorological stations are calculated. This is achieved through... The criterion for identifying outlier sites is to calculate the mean of the measured values ​​from all neighboring sites. and standard deviation If the measured value of a certain meteorological station satisfy If so, it is marked as a low-reliability site.

[0078] In this embodiment, reliability weights are defined. for:

[0079] ;

[0080] Combining the spatiotemporal joint weight with the reliability weight yields the spatiotemporal-reliability joint weight. .

[0081] Next, calculate the terrain constraint weights. Extract the terrain features of the target point and each meteorological station, including elevation differences. and slope difference Define terrain similarity weights for:

[0082] ;

[0083] in, The altitude influence coefficient. This is the slope influence coefficient.

[0084] Finally, the weights of meteorological element timeliness, meteorological station reliability, and terrain constraints are integrated to form a comprehensive weight. Represented as:

[0085] ;

[0086] The composite weights are normalized to obtain the normalized composite weights. :

[0087] ;

[0088] in, This represents the number of dynamically nearby sites.

[0089] The interpolation results are calculated based on the comprehensive weights and measured regional meteorological grid data. The calculation formula is:

[0090] ;

[0091] in, This represents the number of dynamic neighboring sites. For the first Actual measured data from several meteorological stations For the first Normalized comprehensive weights for each meteorological station.

[0092] During the calculation process, the dynamic attenuation coefficient is dynamically adjusted according to the spatial density adaptive strategy. and the number of dynamic neighboring sites The dynamic attenuation coefficient is used to calculate the spatiotemporal joint weight in the timeliness weight of meteorological elements, and the dynamic neighboring station number is used to calculate the interpolation result.

[0093] Specifically, dynamic attenuation coefficient Based on the spatial density of stations surrounding the target point Calculation of (number of stations per unit area):

[0094] ;

[0095] in, Based on the coefficient, To adjust the coefficient, This represents the maximum site density in the region. The sparser the sites (…), the higher the density. Small), The larger the value, the greater the weight of nearby sites, avoiding interpolation smooth transition.

[0096] Dynamic neighboring site count Based on the distance from the target point to the nearest station calculate:

[0097] ;

[0098] in, Number of base sites This is the distance coefficient. To round up, more neighboring stations are selected as the target point is farther away from the station, ensuring interpolation stability.

[0099] After completing the interpolation calculation, meteorological data slices with preset time intervals before the fault are extracted based on the fault trip time, and a time series correspondence between "fault time - precursor meteorological conditions" is constructed to achieve time axis alignment. Subsequently, meteorological factor feature vectors are extracted, and fault labels are marked according to the actual situation (1 is marked when a fault occurs, and 0 is marked when no fault occurs) to form sample units.

[0100] S200: Statistically analyze the frequency of occurrence of each meteorological factor in the fault samples and non-fault samples, calculate the single-factor association weight based on the probability difference between the two, and calculate the multi-factor collaborative association weight based on the covariance of different meteorological factors in the fault samples and the single-factor association weight.

[0101] Furthermore, the meteorological factors include lightning meteorological factors, strong wind meteorological factors, and short-duration heavy precipitation meteorological factors:

[0102] The lightning meteorological factors are characterized by lightning density, lightning intensity, lightning distance attenuation coefficient, and lightning polarity ratio.

[0103] The wind meteorological factors are characterized by the maximum instantaneous wind speed, the rate of change of wind speed over time, the duration of wind speed exceeding the threshold, and the change of wind speed in the vertical direction.

[0104] The short-duration heavy precipitation meteorological factor is characterized by the amount of precipitation per unit time.

[0105] Specifically, the lightning meteorological factors include a lightning strike density feature vector, a lightning strike intensity feature vector, a lightning strike distance attenuation coefficient feature vector, and a lightning polarity proportion feature vector. The lightning strike density feature vector represents the number of lightning strikes per unit time and space; the lightning strike intensity feature vector represents the current amplitude of a single lightning strike; the lightning strike distance attenuation coefficient feature vector represents the degree of attenuation of the impact of lightning distance on the line; and the lightning polarity proportion feature vector represents the proportion of positive lightning strikes. These feature vectors are combined to form a lightning meteorological factor feature vector with a dimension of 4.

[0106] The strong wind meteorological factors include a maximum instantaneous wind speed feature vector, a wind speed change rate feature vector, a duration exceeding the threshold feature vector, and a wind shear index feature vector. Specifically, the maximum instantaneous wind speed feature vector represents the maximum wind speed within the spatiotemporal window; the wind speed change rate feature vector represents the rate of change of wind speed over time; the duration exceeding the threshold feature vector represents the duration for which the wind speed exceeds the line's tolerance threshold; and the wind shear index feature vector represents the vertical wind speed variation. These feature vectors are combined to form a four-dimensional strong wind meteorological factor feature vector.

[0107] The short-duration heavy precipitation meteorological factor includes a short-duration heavy precipitation feature vector, which represents the precipitation amount per unit time, i.e., a short-duration heavy precipitation meteorological factor feature vector with dimension 1.

[0108] Furthermore, the single-factor association weights are calculated in the following manner:

[0109] The probability of occurrence of each meteorological factor in the fault sample and the probability of occurrence of each meteorological factor in the non-fault sample are statistically analyzed.

[0110] Calculate the ratio of the probability of occurrence of the same meteorological factor in faulty samples to that in non-faulty samples and take the logarithm;

[0111] The positive and negative directions of the weights are determined by the sign function based on the relationship between the probability of occurrence of meteorological factors in faulty samples and non-faulty samples.

[0112] The formula for calculating the single-factor association weight is as follows:

[0113] ;

[0114] in, For the first Single-factor association weights of individual meteorological factors The first fault sample The probability of occurrence of each meteorological factor The first non-fault sample The probability of occurrence of each meteorological factor It is a symbolic function.

[0115] Specifically, firstly, the frequency of occurrence of each meteorological factor is statistically analyzed in both faulty and non-faulty samples. When the data of one or more feature vectors of a meteorological factor exceeds a threshold, that meteorological factor is considered to have occurred. Then, the probability of occurrence of each meteorological factor in both faulty and non-faulty samples is statistically analyzed separately.

[0116] Secondly, the ratio of the probability of the same meteorological factor occurring in faulty samples to that in non-faulty samples is calculated and its logarithm is taken to quantify the probability difference between the two. The larger this ratio, the stronger the correlation between the meteorological factor and the fault.

[0117] Next, based on the relationship between the probabilities of meteorological factors occurring in faulty and non-faulty samples, the sign function is used to determine the direction of the weights. If the probability of occurrence in faulty samples is greater than that in non-faulty samples, the weight is positive; otherwise, it is negative.

[0118] Single-factor association weights The calculation formula is:

[0119] ;

[0120] in, The single-factor association weight of the k-th meteorological factor. For the k-th meteorological factor, exemplarily... For lightning meteorological factors, As a meteorological factor of strong winds, This refers to meteorological factors related to short-duration heavy rainfall. For the malfunction to occur, No malfunction occurred; The first fault sample The probability of occurrence of each meteorological factor The first non-fault sample The probability of occurrence of each meteorological factor; It is a symbolic function.

[0121] Furthermore, the calculation of the multi-factor synergistic association weights includes:

[0122] Select two different meteorological factors from the fault sample and calculate the covariance between the two meteorological factors.

[0123] Obtain the single-factor correlation weights corresponding to each of the two meteorological factors;

[0124] The covariance is multiplied by the two single-factor association weights, and a synergy coefficient is introduced for adjustment to obtain the multi-factor synergistic association weight of the meteorological factor pair.

[0125] Multi-factor synergistic association weights were calculated for all meteorological factors to obtain a complete set of multi-factor synergistic association weights.

[0126] Specifically, in actual power transmission line fault scenarios, multiple meteorological factors often work synergistically to cause faults. For example, when lightning and heavy precipitation occur simultaneously, the fault rate is significantly higher than when a single factor occurs alone. Therefore, it is necessary to calculate the synergistic correlation weights of multiple factors to quantify the strength of the synergistic correlation between different meteorological factors.

[0127] First, select two different meteorological factors from the fault sample and calculate the covariance between them. The covariance reflects the co-occurrence pattern of the two meteorological factors in the fault sample; the larger the covariance, the more likely the two meteorological factors are to occur simultaneously.

[0128] Secondly, the individual factor association weights for each of the two meteorological factors are obtained. The individual factor association weights quantify the correlation strength between a single meteorological factor and the fault.

[0129] Next, the covariance is multiplied by the two single-factor association weights, and a synergy coefficient is introduced for adjustment to obtain the multi-factor synergistic association weights for this meteorological factor pair. The synergy coefficient is used to balance the contribution ratio of the covariance and the single-factor weights in the multi-factor weights.

[0130] Multi-factor synergistic association weights The calculation formula is:

[0131] ;

[0132] in, For the first The meteorological factors and the first Multi-factor collaborative correlation weights of meteorological factors The first fault sample , Covariance of each meteorological factor This is the synergy coefficient.

[0133] Finally, multi-factor synergistic association weights were calculated for all meteorological factors to obtain a complete set of multi-factor synergistic association weights. This set contains the synergistic association weights of all pairwise combinations of meteorological factors, which are used to construct the association weight vector in subsequent steps.

[0134] S300: Combine the meteorological factor feature vectors, single-factor association weights, multi-factor collaborative association weights, and fault labels to form a training dataset;

[0135] Specifically, for each sample unit, its meteorological factor feature vector consists of lightning meteorological factor feature vectors, strong wind meteorological factor feature vectors, and short-duration heavy precipitation meteorological factor feature vectors. The lightning meteorological factor feature vector has a dimension of 4, including lightning strike density, lightning strike intensity, lightning distance attenuation coefficient, and lightning polarity ratio; the strong wind meteorological factor feature vector has a dimension of 4, including maximum instantaneous wind speed, wind speed change rate, duration of exceeding threshold, and wind shear index; the short-duration heavy precipitation meteorological factor feature vector has a dimension of 1, representing the precipitation per unit time. Therefore, the total dimension of the meteorological factor feature vectors for each sample unit is 9.

[0136] The single-factor association weights and multi-factor collaborative association weights calculated in step S200 are combined to form an association weight vector. The single-factor association weight vector contains three elements, corresponding to the single-factor association weights of the lightning meteorological factor, the strong wind meteorological factor, and the short-duration heavy precipitation meteorological factor, respectively. The multi-factor collaborative association weight vector contains three elements, corresponding to the multi-factor collaborative association weights of the three meteorological factor pairs: lightning-strong wind, lightning-short-duration heavy precipitation, and strong wind-short-duration heavy precipitation, respectively. The total dimension of the association weight vector is 6.

[0137] The fault label is marked according to whether the sample unit has a fault. The sample with a fault is marked as 1, and the sample without a fault is marked as 0.

[0138] The meteorological factor feature vector (9 dimensions), the association weight vector (6 dimensions), and the fault label (1 dimension) of each sample unit are combined to form a single training sample. This training sample contains the original feature information of the meteorological factors, the association strength information between the meteorological factors and the fault, and the label information of the fault occurrence.

[0139] Repeat the above combination operation for all sample units to form a complete training dataset. This training dataset contains both the feature information of meteorological factors and the correlation weight information obtained through statistical analysis, providing a data foundation for the subsequent construction of a weight-guided attention neural network model.

[0140] S400: Construct a neural network model containing an attention layer, calculate attention weights in the attention layer based on the single-factor association weights, train the neural network model using the training dataset, and obtain a transmission line fault rate prediction model.

[0141] Specifically, in this embodiment of the invention, the neural network model includes an input layer, an attention layer, a hidden layer, and an output layer. The input layer contains the meteorological factor feature vector and the associated weight vector. The meteorological factor feature vector has a dimension of 9, and the associated weight vector has a dimension of 6, therefore the total dimension of the input layer is 15.

[0142] Specifically, association weights are integrated into the attention mechanism to construct a weight-guided attention neural network model. This attention neural network model includes an input layer, an attention layer, a hidden layer, and an output layer. The input layer contains the meteorological factor feature vectors and the association weight vectors. The meteorological factor feature vectors consist of a lightning meteorological factor feature vector (dimension 4), a strong wind meteorological factor feature vector (dimension 4), and a short-duration heavy precipitation meteorological factor feature vector (dimension 1), for a total dimension of 9. The association weight vectors include 3 single-factor association weights and 3 multi-factor collaborative association weights, with a dimension of 6. Therefore, the total dimension of the input layer is 15.

[0143] Furthermore, the step of calculating the attention weights includes:

[0144] Obtain the feature vectors of each meteorological factor and their corresponding single-factor correlation weights;

[0145] Calculate the product of the single-factor association weight and the corresponding meteorological factor feature vector;

[0146] The product of all meteorological factor feature vectors is normalized to obtain the attention weights of each meteorological factor feature vector.

[0147] The attention weights are multiplied by the corresponding meteorological factor feature vectors to obtain the attention-weighted features.

[0148] Specifically, the feature vectors of each meteorological factor and their corresponding single-factor correlation weights are obtained. The meteorological factor feature vectors are organized according to the meteorological factor type, denoted as... ,in This is the feature vector of lightning meteorological factors (dimension 4). This is the feature vector of the gale meteorological factor (dimension 4). This is the feature vector of short-duration heavy precipitation meteorological factors (dimension 1). The corresponding single-factor association weights are denoted as follows: These weights have been calculated in step S200.

[0149] Next, calculate the product of the single-factor association weight and the corresponding meteorological factor feature vector. For the first... Calculate a number of meteorological factors. This product reflects the importance score of the meteorological factor after taking into account its correlation with the fault.

[0150] Next, the product of all meteorological factor feature vectors is normalized to obtain the attention weights for each meteorological factor feature vector. Normalization is then performed using the softmax function to ensure that the sum of all attention weights is 1, enabling the model to adaptively allocate attention resources.

[0151] No. Attention weights of the feature vectors of each meteorological factor The calculation formula is:

[0152] ;

[0153] in, For the first Attention weights for each feature vector. For the first The single-factor association weights of meteorological factors corresponding to each feature vector. For the first There are 3 eigenvectors. Since there are a total of 3 meteorological factors, the summation range is from 1 to 3.

[0154] Finally, the attention weights are multiplied by the corresponding meteorological factor feature vectors to obtain the attention-weighted features. For the first... Each meteorological factor has an attention-weighted characteristic. The attention-weighted features of the three meteorological factors are concatenated to form a complete attention-weighted feature vector. The vector dimension remains 9.

[0155] Through the above calculation process, the attention mechanism can adaptively adjust the weights of each meteorological factor feature based on the single-factor association weights, enabling the model to prioritize meteorological factors highly correlated with the fault. Simultaneously, the multi-factor synergistic association weights, as part of the input vector, work together with the attention-weighted features in subsequent hidden layer processing to capture the synergistic effects between meteorological factors, thereby improving prediction accuracy.

[0156] The hidden layer employs two fully connected layers with Leaky ReLU activation. The first layer has 16 neurons, and the second layer has 8 neurons. Leaky ReLU effectively avoids neuron death and maintains gradient flow. The first fully connected layer receives attention-weighted feature vectors and multi-factor collaborative association weight vectors as input, mapping them to a 16-dimensional feature space. After Leaky ReLU activation, the second fully connected layer compresses the features from 16 dimensions to 8 dimensions, and then performs another Leaky ReLU activation to extract higher-level abstract features.

[0157] The output layer uses the Sigmoid activation function to output the failure probability. The output layer maps the 8-dimensional features of the hidden layer to 1 dimension, representing the probability of a failure occurring, with a value ranging from 0 to 1. Failure Probability The calculation formula is:

[0158] ;

[0159] in, The label indicates that a fault has occurred. For attention-weighted features, This is the output layer weight matrix. This is a bias term.

[0160] The attention neural network model is trained using the training dataset. A cross-entropy loss function is used to measure the difference between the predicted probability and the true label. Model parameters, including the weights and biases of the attention layer, hidden layer, and output layer, are updated using backpropagation. Appropriate learning rates, batch sizes, and training epochs are set during training. Training stops when the model's performance on the validation set reaches its optimum or the loss function converges, yielding the final transmission line fault rate prediction model.

[0161] Example 2: A power company manages a 220kV transmission line network covering various terrains including mountainous and plain areas, with a total line length of approximately 1200 kilometers and over 1800 towers. This region experiences approximately 30 transmission line faults annually due to meteorological factors such as lightning, strong winds, and short-duration heavy rainfall. To improve fault prediction accuracy and reduce the fault rate, the power company adopted the machine learning-based transmission line fault rate prediction model training system of this invention. This system includes:

[0162] The data mapping module is used to acquire regional meteorological grid data and tower coordinate data. An improved inverse distance weighted interpolation algorithm is used to map the regional meteorological grid data to the tower coordinate positions. The time series is aligned according to the fault occurrence time, meteorological factor feature vectors are extracted and fault labels are marked to form sample units. The improved inverse distance weighted interpolation algorithm introduces meteorological element timeliness weight, meteorological station reliability weight and terrain constraint weight to determine the comprehensive weight, and the interpolation result is calculated based on the comprehensive weight.

[0163] The weight quantization module is used to count the frequency of occurrence of each meteorological factor in fault samples and non-fault samples, calculate the single-factor association weight based on the probability difference between the two, and calculate the multi-factor collaborative association weight based on the covariance of different meteorological factors in fault samples and the single-factor association weight.

[0164] The dataset construction module is used to combine the meteorological factor feature vectors, single-factor association weights, multi-factor collaborative association weights, and fault labels to form a training dataset.

[0165] The model training module is used to construct a neural network model containing an attention layer, calculate attention weights based on the single-factor association weights in the attention layer, and train the neural network model using the training dataset to obtain a transmission line fault rate prediction model.

[0166] Specifically, the data mapping module acquired regional meteorological grid data and pole coordinate data from 2022 to 2024. The system employs an improved inverse distance weighted interpolation algorithm, introducing timeliness weights for meteorological elements, reliability weights for meteorological stations, and terrain constraint weights to accurately map regional meteorological data to the location of each pole. The spatial resolution is improved from 1km×1km to 10m×10m, and the temporal resolution is improved from hourly to 5-minute levels, effectively solving the interpolation bias problem caused by the sparseness of stations in mountainous areas.

[0167] The weighting quantification module analyzed fault and non-fault samples over three years, totaling 7200 sample units, including 90 fault samples. By calculating single-factor association weights, the system found that the weight of the lightning meteorological factor was 2.18, the strong wind meteorological factor was 1.76, and the short-duration heavy precipitation meteorological factor was 1.42, indicating that lightning had the most significant impact on faults. Multi-factor synergistic association weight analysis revealed that the synergistic weight between lightning and short-duration heavy precipitation reached 2.89, confirming the amplifying effect of multi-factor synergy on faults.

[0168] The dataset construction module combines meteorological factor feature vectors, association weights, and fault labels to form a training dataset containing 7,200 samples.

[0169] The attention neural network model constructed in the model training module calculates the attention weights of each meteorological factor based on single-factor association weights in the attention layer. Training results show that lightning received an attention weight of 0.49, strong winds 0.33, and short-duration heavy precipitation 0.18, consistent with the relative relationships of the association weights. The model achieved a prediction accuracy of 91.2% on the validation set, a 6.7 percentage point improvement compared to the traditional neural network model without association weights (84.5% accuracy).

[0170] After the model was deployed to the operation and maintenance system, it successfully predicted 13 potential faults during the first half of 2025, of which 12 actually occurred. By taking proactive measures, the fault rate decreased by 32% compared to the same period last year, reducing direct economic losses by approximately 1.8 million yuan and significantly improving the level of power grid safety operation.

[0171] This application case demonstrates the effectiveness of the invention. A high-precision mapping of meteorological data was achieved through an improved inverse distance weighted interpolation algorithm. The correlation strength between meteorological factors and faults was quantified by introducing correlation weights. An attention mechanism enabled the model to adaptively focus on highly correlated factors, effectively improving fault prediction accuracy and demonstrating significant engineering application value.

[0172] Example 3: Another aspect of the present invention provides an electronic device, please refer to... Figure 3The image shows an electronic device according to Embodiment 3 of the present invention, including a memory 20, a processor 10, and a computer program 30 stored in the memory and executable on the processor. When the processor 10 executes the computer program 30, it implements the above-mentioned training method for the transmission line fault rate prediction model based on machine learning.

[0173] In some embodiments, the processor 10 may be a central processing unit (CPU), controller, microcontroller, microprocessor or other data processing chip, used to run program code stored in memory 20 or process data, such as executing access restriction programs.

[0174] The memory 20 includes at least one type of readable storage medium, such as flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 20 can be an internal storage unit of an electronic device, such as the hard disk of the electronic device. In other embodiments, the memory 20 can also be an external storage device of the electronic device, such as a plug-in hard disk, SmartMediaCard (SMC), SecureDigital (SD) card, FlashCard, etc., equipped on the electronic device. Furthermore, the memory 20 can include both internal and external storage units of the electronic device. The memory 20 can be used not only to store application software and various types of data of the electronic device, but also to temporarily store data that has been output or will be output.

[0175] It should be pointed out that, Figure 3 The structure shown does not constitute a limitation on the electronic device. In other embodiments, the electronic device may include fewer or more components than shown, or combine certain components, or have different component arrangements.

[0176] This invention also proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described machine learning-based transmission line fault rate prediction model training method.

[0177] Those skilled in the art will understand that the logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as a ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can mean any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0178] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0179] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0180] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A training method for a transmission line fault rate prediction model based on machine learning, characterized in that, include: S100: Acquire regional meteorological grid data and tower coordinate data, and use an improved inverse distance weighted interpolation algorithm to map the regional meteorological grid data to the tower coordinate positions. Align the time series according to the fault occurrence time, extract meteorological factor feature vectors and label them with fault tags to form sample units. The improved inverse distance weighted interpolation algorithm introduces meteorological element timeliness weight, meteorological station reliability weight and terrain constraint weight to determine the comprehensive weight, and calculates the interpolation result based on the comprehensive weight. S200: Statistically analyze the frequency of occurrence of each meteorological factor in the fault samples and non-fault samples, calculate the single-factor association weight based on the probability difference between the two, and calculate the multi-factor collaborative association weight based on the covariance of different meteorological factors in the fault samples and the single-factor association weight. S300: Combine the meteorological factor feature vectors, single-factor association weights, multi-factor collaborative association weights, and fault labels to form a training dataset; S400: Construct a neural network model containing an attention layer, calculate attention weights in the attention layer based on the single-factor association weights, train the neural network model using the training dataset, and obtain a transmission line fault rate prediction model. The steps for determining the overall weight include: Calculate the timeliness weight of meteorological elements, and correct the station weights by using a time decay factor based on the time difference between the interpolation time of the target point and the actual measurement time of each meteorological station. Calculate the reliability weight of meteorological stations, identify abnormal stations whose measured values ​​deviate from the mean of neighboring stations by more than three standard deviations, and assign low reliability weights to abnormal stations. Calculate the terrain constraint weights and determine the terrain similarity weights based on the elevation and slope differences between the target point and each meteorological station. The timeliness weight of the meteorological elements, the reliability weight of the meteorological stations, and the terrain constraint weight are combined to obtain a comprehensive weight. The steps for calculating the interpolation result include: Using the tower coordinates as the target point, calculate the number of stations per unit area around the target point, and determine the attenuation coefficient based on the station density. The lower the station density, the greater the attenuation coefficient. The number of neighboring stations to be included in the interpolation calculation is determined based on the distance between the target point and the nearest station; the greater the distance, the more neighboring stations are selected. The overall weights of each neighboring site are normalized. The interpolation result of the target point is obtained by multiplying the measured meteorological data of each neighboring station with the corresponding normalized comprehensive weight and then summing the results.

2. The training method for a transmission line fault rate prediction model based on machine learning according to claim 1, characterized in that, The meteorological factors include lightning meteorological factors, strong wind meteorological factors, and short-duration heavy precipitation meteorological factors: The lightning meteorological factors are characterized by lightning density, lightning intensity, lightning distance attenuation coefficient, and lightning polarity ratio. The wind meteorological factors are characterized by the maximum instantaneous wind speed, the rate of change of wind speed over time, the duration of wind speed exceeding the threshold, and the change of wind speed in the vertical direction. The short-duration heavy precipitation meteorological factor is characterized by the amount of precipitation per unit time.

3. The method for training a transmission line fault rate prediction model based on machine learning according to claim 1, characterized in that, The single-factor association weights are calculated in the following manner: The probability of occurrence of each meteorological factor in the fault sample and the probability of occurrence of each meteorological factor in the non-fault sample are statistically analyzed. Calculate the ratio of the probability of occurrence of the same meteorological factor in faulty samples to that in non-faulty samples and take the logarithm; The positive and negative directions of the weights are determined by the sign function based on the probability relationship between the occurrence of meteorological factors in faulty samples and non-faulty samples. The formula for calculating the single-factor association weight is as follows: ； in, For the first Single-factor association weights of individual meteorological factors The first fault sample The probability of occurrence of each meteorological factor The first non-fault sample The probability of occurrence of each meteorological factor It is a symbolic function.

4. The training method for a transmission line fault rate prediction model based on machine learning according to claim 1, characterized in that, The calculation of the multi-factor synergistic association weights includes: Select two different meteorological factors from the fault sample and calculate the covariance between the two meteorological factors. Obtain the single-factor correlation weights corresponding to each of the two meteorological factors; The covariance is multiplied by the two single-factor association weights, and a synergy coefficient is introduced for adjustment to obtain the multi-factor synergistic association weight of the meteorological factor pair. Multi-factor synergistic association weights were calculated for all meteorological factors to obtain a complete set of multi-factor synergistic association weights.

5. The method for training a transmission line fault rate prediction model based on machine learning according to claim 1, characterized in that, The steps for calculating attention weights include: Obtain the feature vectors of each meteorological factor and their corresponding single-factor correlation weights; Calculate the product of the single-factor association weight and the corresponding meteorological factor feature vector; The product of all meteorological factor feature vectors is normalized to obtain the attention weights of each meteorological factor feature vector. The attention weights are multiplied by the corresponding meteorological factor feature vectors to obtain the attention-weighted features.

6. A training system for a transmission line fault rate prediction model based on machine learning, characterized in that, The data mapping module is used to acquire regional meteorological grid data and tower coordinate data. An improved inverse distance weighted interpolation algorithm is used to map the regional meteorological grid data to the tower coordinate positions. The time series is aligned according to the fault occurrence time, meteorological factor feature vectors are extracted and fault labels are marked to form sample units. The improved inverse distance weighted interpolation algorithm introduces meteorological element timeliness weight, meteorological station reliability weight and terrain constraint weight to determine the comprehensive weight, and the interpolation result is calculated based on the comprehensive weight. The weight quantization module is used to count the frequency of occurrence of each meteorological factor in fault samples and non-fault samples, calculate the single-factor association weight based on the probability difference between the two, and calculate the multi-factor collaborative association weight based on the covariance of different meteorological factors in fault samples and the single-factor association weight. The dataset construction module is used to combine the meteorological factor feature vectors, single-factor association weights, multi-factor collaborative association weights, and fault labels to form a training dataset. The model training module is used to construct a neural network model containing an attention layer, calculate attention weights based on the single-factor association weights in the attention layer, train the neural network model using the training dataset, and obtain a transmission line fault rate prediction model. The steps for determining the overall weight include: Calculate the timeliness weight of meteorological elements, and correct the station weights by using a time decay factor based on the time difference between the interpolation time of the target point and the actual measurement time of each meteorological station. Calculate the reliability weight of meteorological stations, identify abnormal stations whose measured values ​​deviate from the mean of neighboring stations by more than three standard deviations, and assign low reliability weights to abnormal stations. Calculate the terrain constraint weights and determine the terrain similarity weights based on the elevation and slope differences between the target point and each meteorological station. The timeliness weight of the meteorological elements, the reliability weight of the meteorological stations, and the terrain constraint weight are combined to obtain a comprehensive weight. The steps for calculating the interpolation result include: Using the tower coordinates as the target point, calculate the number of stations per unit area around the target point, and determine the attenuation coefficient based on the station density. The lower the station density, the greater the attenuation coefficient. The number of neighboring stations to be included in the interpolation calculation is determined based on the distance between the target point and the nearest station; the greater the distance, the more neighboring stations are selected. The overall weights of each neighboring site are normalized. The interpolation result of the target point is obtained by multiplying the measured meteorological data of each neighboring station with the corresponding normalized comprehensive weight and then summing the results.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the machine learning-based transmission line fault rate prediction model training method as described in any one of claims 1-5.

8. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the machine learning-based transmission line fault rate prediction model training method as described in any one of claims 1-5.

Images