CN-121980375-A - Power transmission line fault identification method based on improved attention weight

Abstract

The invention discloses a power transmission line fault identification method based on improved attention weight, which comprises the following steps of collecting historical fault data of a power transmission line, extracting a plurality of fault characteristics from the historical fault data, carrying out coding pretreatment to obtain a numerical characteristic vector, learning the characteristic weight of the characteristic vector through an attention mechanism, introducing regularized constraint characteristic weight distribution, identifying and selecting characteristic interaction combinations of the characteristic weights to construct a classification module capable of integrating the characteristic interaction combinations and the interaction weights, establishing a joint optimization target, carrying out end-to-end training on a fault identification model by adopting a gradient optimization algorithm based on the joint optimization target, carrying out training and verification on the integral fault identification model by using the historical fault data, optimizing model parameters until convergence, inputting the real-time collected power transmission line fault characteristics into the trained fault identification model to output fault probability distribution, and determining a fault type according to a preset decision rule for the fault probability distribution.

Inventors

• WANG ZHENG
• ZHANG HANJIE
• WANG HONGQI
• DAN HANYONG
• HUANG QIAN
• Chen Yuemu
• CHEN LIANG
• CHENG MENG
• YANG QIJIA
• XU GANGYI
• XU WANGSHENG
• HU JINGBO
• MENG XIANGLONG
• XU SIBO
• LI HAIFENG
• YANG XI
• ZHANG ZHIRAN

Assignees

• 中国南方电网有限责任公司超高压输电公司贵阳局

Dates

Publication Date

20260505

Application Date

20251209

Claims (10)

1. 1. The power transmission line fault identification method based on the improved attention weight is characterized by comprising the following steps of: Collecting historical fault data of the power transmission line, extracting a plurality of fault characteristics from the historical fault data, and performing coding pretreatment to obtain a numeric characteristic vector; Learning the feature weights of the feature vectors through an attention mechanism, and introducing regularization constraint feature weight distribution; identifying and selecting feature interaction combinations of the feature weights, distributing interaction combination weights for each feature interaction combination, and constructing a classification module capable of integrating the feature interaction combinations and the interaction weights; Establishing a joint optimization target covering classification task loss, characteristic weight distribution constraint and interaction weight distribution constraint, and performing end-to-end training on the fault identification model by adopting a gradient optimization algorithm based on the joint optimization target; training and verifying the integral fault recognition model by using the historical fault data, and optimizing model parameters until convergence; and inputting the real-time collected fault characteristics of the power transmission line into a trained fault recognition model to output fault probability distribution, and determining the fault type according to a preset decision rule of the fault probability distribution.
2. 2. The transmission line fault identification method based on improved attention weight according to claim 1, wherein the fault characteristics comprise weather characteristics, time period characteristics, season characteristics, reclosing characteristics, transition resistance value size characteristics, transition resistance volt-ampere characteristics, zero-sequence current direct current content characteristics and zero-sequence current harmonic content characteristics; The key feature interaction combination comprises an interaction pair of weather features and seasonal features, an interaction pair of time period features and reclosing characteristic features, an interaction pair of transition resistance value magnitude features and transition resistance volt-ampere characteristic features, an interaction pair of zero-sequence current direct-current content features and zero-sequence current harmonic content features, an interaction pair of weather features and reclosing characteristic features and an interaction pair of seasonal features and reclosing characteristic features.
3. 3. The method for identifying a power transmission line fault based on improved attention weight according to claim 2, wherein the obtaining process of the feature vector comprises: performing unified feature coding processing on the acquired fault features, and converting different types of features into unified numeric feature vectors; The feature encoding process includes: Adopting a three-classification coding mode for weather features, wherein the code is 0 in sunny days, 1 in rainy days and 2 in thunder rain; The time slot characteristics adopt a four-classification coding mode, wherein midnight codes are 0, early morning codes are 1, daytime codes are 2, and evening codes are 3; the method comprises the steps of adopting a four-classification coding mode for seasonal features, wherein the spring coding is 0, the summer coding is 1, the autumn coding is 2, and the winter coding is 3; The reclosing characteristic adopts a two-class coding mode, wherein the successful coding is 1, and the unsuccessful coding is 0; And a binary coding mode is adopted for the magnitude of the transition resistance, the volt-ampere characteristic of the transition resistance, the zero sequence current direct current content and the zero sequence current harmonic content, wherein the low-resistance, linear and low-content characteristic codes are 0, and the high-resistance, nonlinear and high-content characteristic codes are 1.
4. 4. The method for identifying a power transmission line fault based on improved attention weight as recited in claim 1, wherein said learning feature weights of said feature vectors by an attention mechanism comprises: For input feature vectors The importance score of each feature is calculated through a multi-layer perceptron network, and the calculation formula is as follows: ; Wherein, the Is the first The weight parameter corresponding to each feature is used for the method, Representing input to The operation of the multi-layer perceptron is applied, Is the first The importance scores corresponding to the individual features are provided, For the bias term, n is the critical fault signature number, To activate the function, the range of the tanh function is ; Normalizing the importance scores of all the features through a softmax function to obtain feature weights of each feature, wherein the calculation formula is as follows: ; The characteristic weight satisfies And is also provided with The probability distribution characteristics of the feature weights are ensured.
5. 5. The method for identifying a power transmission line fault based on improved attention weighting as recited in claim 4, wherein said introducing a distribution of regularization constraint feature weights comprises: The entropy regularization term is introduced to restrict the weight distribution, and the regularization term is defined as: ; Wherein the method comprises the steps of Is a regularization coefficient.
6. 6. The method for identifying the power transmission line fault based on the improved attention weight according to claim 2, wherein the learning of the feature weights of the feature vectors through the attention mechanism comprises the steps of calculating the weights of each feature interaction pair by adopting the attention mechanism, wherein a calculation formula is as follows: ; Wherein, the The feature stitching is represented and is performed, A multi-layer perceptron for processing interactive features.
7. 7. The power transmission line fault identification method based on improved attention weight as claimed in claim 6, wherein the classification module adopts an improved bayesian classifier, and an improved bayesian classifier probability calculation formula is: ; Wherein, the The category of the fault is indicated and, As a result of the single feature weights, The feature interaction weight; The joint conditional probability is estimated by adopting a Laplace smoothing method, and the calculation formula is as follows: ; where α is the smoothing parameter and β is the adjustment factor.
8. 8. The method for identifying a power transmission line fault based on improved attention weighting as recited in claim 6, wherein a total objective function of said joint optimization objective is defined as: ; Wherein, the Classifying the loss function for cross entropy is defined as: ; For the feature weight entropy regularization, For the regularization of the interaction weight entropy, it is defined as: ; ; And Is a regularization coefficient.
9. 9. The power transmission line fault identification method based on improved attention weight as claimed in claim 1, wherein the model training process adopts a small batch gradient descent method for parameter optimization, and a parameter updating formula is as follows: ; Wherein, the In order for the rate of learning to be high, A dynamic decay strategy is adopted to ensure that the dynamic decay time, Including parameters of all MLP networks.
10. 10. The power transmission line fault identification method based on improved attention weight as claimed in claim 1, wherein the fault identification model outputs probability distribution of four fault types including lightning strike fault, wind deflection fault, forest fire fault and tree obstacle fault, and the decision rule adopts a maximum probability principle and is defined as: ; and simultaneously outputting a characteristic weight vector and a characteristic interaction weight matrix for generating an explanatory analysis report.

Description

Power transmission line fault identification method based on improved attention weight Technical Field The invention relates to the technical field of power transmission line fault identification, in particular to a power transmission line fault identification method based on improved attention weight. Background With the continuous expansion of the scale of the power system and the deep promotion of the construction of the smart grid, the power transmission line is used as an important carrier for power transmission, and the running state of the power transmission line is directly related to the safety and stability of the power grid. The transmission line is spread over complex and diverse geographic environments and climatic conditions and faces various fault threats such as lightning stroke, windage yaw, tree barriers, forest fires and the like. The power transmission line fault cause can be accurately and rapidly identified, scientific rush repair guidance can be provided for operation and maintenance personnel, the power failure time is shortened, data support can be provided for line operation and maintenance strategy optimization and risk early warning through fault mode analysis, and the method has important significance in guaranteeing energy safety and improving intelligent level of a power grid. The existing transmission line fault identification technology is mainly based on expert experience rules and a traditional machine learning method, but still faces a plurality of technical challenges. Firstly, the unreasonable problem of characteristic weight distribution is highlighted, namely the traditional method has the defects of high calculation complexity, slow convergence speed, easy sinking into local optimum and the like when the weight is optimized by a complex shape algorithm, and the real contribution degree of different fault characteristics to classification results is difficult to accurately reflect, so that the classification accuracy is limited. Secondly, the characteristic independence assumption is too strong to limit the classification performance, i.e. classical methods such as naive Bayes and the like assume that all characteristic conditions are independent, important correlations among the characteristics such as weather, seasons, time periods, reclosing success rate and the like in an actual fault scene are ignored, and the synergies among the characteristics cannot be fully mined. The existing power transmission line fault identification technology has the problems that the characteristic weight distribution is unreasonable and the characteristic independence is assumed to be too strong, and the accuracy, the reliability and the practicability of the existing power transmission line fault identification technology are difficult to meet the higher requirements of a modern intelligent power grid on fault diagnosis. Disclosure of Invention In order to overcome the defects in the prior art, the invention aims to provide a power transmission line fault identification method based on improved attention weight, which can adaptively learn the feature importance and effectively capture the interaction relationship among features so as to solve the problems in the prior art. The technical scheme adopted for solving the technical problems is that the power transmission line fault identification method based on the improved attention weight comprises the following steps: Collecting historical fault data of the power transmission line, extracting a plurality of fault characteristics from the historical fault data, and performing coding pretreatment to obtain a numeric characteristic vector; Learning the feature weights of the feature vectors through an attention mechanism, and introducing regularization constraint feature weight distribution; identifying and selecting feature interaction combinations of the feature weights, distributing interaction combination weights for each feature interaction combination, and constructing a classification module capable of integrating the feature interaction combinations and the interaction weights; Establishing a joint optimization target covering classification task loss, characteristic weight distribution constraint and interaction weight distribution constraint, and performing end-to-end training on the fault identification model by adopting a gradient optimization algorithm based on the joint optimization target; training and verifying the integral fault recognition model by using the historical fault data, and optimizing model parameters until convergence; and inputting the real-time collected fault characteristics of the power transmission line into a trained fault recognition model to output fault probability distribution, and determining the fault type according to a preset decision rule of the fault probability distribution. The fault characteristics comprise weather characteristics, time period characteristics, seasonal characteristics, rec