电工技术学报2025，Vol.40Issue(5)：1455-1470,16.DOI:10.19595/j.cnki.1000-6753.tces.240326
基于CABFAM-Transformer的输电线路在线测距实测行波预分类方法
Pre-Identification Method of Measured Traveling Wave Data for Online Fault Location of Transmission Line Based on CABFAM-Transformer
摘要
Abstract
This paper presents an innovative approach to fault identification for transmission lines by integrating a convolutional attention-based feature aggregation module(CABFAM)with an adaptive Transformer model.Traditional traveling wave acquisition systems often suffer from frequent triggering,significant noise interference,and difficulty in detecting weak fault signals.To address these limitations,this study proposes a method that combines the strengths of convolutional feature extraction and Transformer-based multi-layer encoding.The objective is to improve the reliability,accuracy,and efficiency of fault identification,providing robust support for the safe operation of power systems. The methodology involves a hybrid architecture that leverages both convolutional modules and Transformer encoders.Specifically,the convolutional block attention module(CBAM)enhances feature extraction by refining channel-wise and spatial features through an attention mechanism.These refined features are then processed by an adaptive Transformer-based encoder,which extracts multi-layered characteristics from complex fault data,ensuring comprehensive analysis and classification. A combination of real-world and simulated data was used to develop and validate the model.The dataset includes 5 076 measured waveforms from 110-220 kV transmission lines in the Yunnan power grid and 15 924 simulated fault scenarios.Key innovations include the use of SiLU(Sigmoid Linear Unit)activation functions to mitigate gradient vanishing problems,improving the stability of the backpropagation process.The model parameters were optimized through iterative training to balance computational efficiency and accuracy,ensuring scalability for large-scale applications. The proposed model demonstrates superior performance in fault detection and classification compared to traditional methods.Key metrics from the experimental evaluation include an mAUC of 0.982,an accuracy of 0.983,a precision of 0.981,and an F1-score of 0.974,indicating the robustness of the approach.Ablation studies further revealed that integrating the CBAM module improved model performance by more than 3 percentage points,while the adaptive Transformer structure contributed an additional 4.5 percentage points increase in key performance indicators. Comparative analysis with existing fault identification techniques highlights the advantages of the proposed model.It not only improves classification accuracy but also maintains high efficiency under noisy conditions and weak fault scenarios.The model's ability to handle diverse fault types and maintain low computational overhead makes it suitable for real-time applications.Additionally,the CABFAM-Transformer framework efficiently processes high-dimensional traveling wave data without compromising on speed,supporting fast decision-making for power grid operators. The results of this study confirm that the CABFAM-Transformer-based fault identification method offers significant improvements over traditional approaches in terms of accuracy,robustness,and efficiency.By combining convolutional attention modules with adaptive Transformer encoders,the framework addresses the challenges posed by noisy data and weak fault signals,ensuring reliable detection across a wide range of scenarios.The use of SiLU activation functions further enhances the model's performance by stabilizing the training process and preventing gradient issues. This method provides a scalable solution for real-world applications in power grid fault management,with the potential to improve operational efficiency and reduce downtime.Future research will explore further optimization of the Transformer architecture to reduce computational costs and enhance real-time capabilities.The application of this model to larger-scale datasets and diverse grid environments will also be investigated,with the aim of developing a comprehensive fault management system for modern power networks.关键词
行波采集装置/基于卷积注意力机制的特征聚合模块/CABFAM/自适应Transformer/实测数据故障辨识Key words
Traveling wave acquisition device/convolutional attention-based feature aggregation module(CABFAM)/adaptive Transformer/measured data samples/fault identification分类
动力与电气工程引用本文复制引用
唐玉涛,束洪春,刘皓铭,苏萱,韩一鸣,代月..基于CABFAM-Transformer的输电线路在线测距实测行波预分类方法[J].电工技术学报,2025,40(5):1455-1470,16.基金项目
国家自然科学基金重点资助项目(52337005). （52337005）
