基于半监督自学习的光伏系统故障检测方法

A Semi-self-supervised Fault Detection Method for Photovoltaic Systems

对光伏系统的主要设备进行及时的故障检测能有效提高系统发电量、减少安全隐患。传统的基于机器学习和深度学习的方法通常需要大量标签数据来建立数据模型,然而在光伏系统的运行过程中可用的标签数据非常有限,造成无法建立一个可靠的数据模型,导致模型的泛化性不高。此外,人工标记数据既成本高昂又易出错。为了解决这个挑战,设计了基于自学习预训练和半监督模型调优的半监督自学习的故障检测方法。该方法通过最大程度地利用少量的有标签的数据(即有限的监督信息)和大量的无标签数据,实现高性能的故障检测。该方法已经部署到国内一个2.5 MW的光伏电站上,数月的运维结果表明,相对于传统的机器学习和深度学习的方法,本文所提出的方法针对光伏组串的故障检测的F1分数提升了4.49%或更多,从而有效地提升了现场运维的效率。

Fault detection for photovoltaic(PV)system components has gained increasing attention from operations and maintenance due to its ability to improve system electricity generation and reduce safety hazards.Conventional machine learning and deep learning methods typically rely on a considerable amount of labeled data to establish effective models.However,the labeling is limited to support effective models in PV systems.Moreover,manul labelling is both costly and prone to errors.To address these challenges,this work proposes a semi-self-supervised learning fault detection method.The proposed method first pre-trains a model based on self-learning using a large amount of unlabeled data.It then fine-tunes the pre-trained model using limited supervised information to achieve effective fault detection for the main components of PV systems,such as PV strings and PV modules.This approach has been deployed at a 2.5 MW PV system located in North China.Months of operation and maintenance demonstrate that the proposed fault detection method improves fault detection performance by 4.49% or more in terms of F1-score compared with conventional machine learning based method,thereby effectively enhancing on-site operation and maintenance.