基于Transformer模型的光伏系统故障分类与预测方法  

Transformer Model-Based Fault Classification and Prediction Method for Photovoltaic Systems

作　　者：林江[1] 蔡晓龙 周剑桥 LIN Jiang;CAI Xiaolong;ZHOU Jianqiao(School of Electrical Engineering,Shanghai Jiaotong University,Shanghai 200240,China)

机构地区：[1]上海交通大学电气工程学院,上海200240

出　　处：《智慧电力》2025年第4期96-102,共7页Smart Power

基　　金：国家自然科学基金资助项目(52107200)。

摘　　要：光伏系统故障分类对保障电力系统稳定运行至关重要。针对传统算法在相似场景下易混淆故障类型的问题,提出一种基于Transformer模型的光伏系统故障分类与预测方法。通过光伏仿真系统构建包含无故障、轻微遮蔽、中等遮蔽及重度遮蔽下短路故障的多故障类型数据集,基于Transformer混合架构模型利用自注意力机制捕捉故障特征全局关联性,结合U-Net解码器优化时序特征提取,利用ReLU激活函数与Adam优化器动态调整参数,并采用SoftMax函数实现多级故障分类。仿真分析表明,所提方法可有效捕捉故障特征的时空交互规律,在复杂故障场景下展现出更高分类精度与鲁棒性,为光伏系统故障诊断提供可靠方案。Fault classification in photovoltaic(PV)systems is crucial for ensuring the stable operation of power systems.To address the limitations of traditional algorithms,which are prone to misclassifying fault types under similar scenarios,this paper proposes a Transformer model-based method for PV system fault classification and prediction.A multi-fault dataset is constructed through PV system simulations,encompassing no-fault conditions,partial shading(mild,moderate,and severe),and short-circuit faults.The proposed Transformer hybrid architecture leverages self-attention mechanisms to capture global dependencies among fault features,integrates a U-Net decoder to optimize temporal feature extraction,dynamically adjusts parameters via ReLU activation functions and the Adam optimizer,and employs SoftMax functions for multi-level fault classification.Simulation results demonstrate that the proposed method effectively captures the spatiotemporal interaction patterns of fault characteristics,exhibiting superior classification accuracy and robustness in complex fault scenarios.This approach provides a reliable solution for PV system fault diagnosis.

关 键 词：光伏系统 神经网络 故障分类 故障预测 

分 类 号：TM615[电气工程—电力系统及自动化]