摘要
机器学习的可解释性是其在电力系统领域安全、可靠应用的关键环节与重要基础之一。针对电力系统智能分析的机器学习模型可解释性方法进行初步探讨。首先,阐述了机器学习模型可解释性的基本概念、数学描述与相关评价维度;之后,梳理了实现机器学习可解释的整体思路与技术路线,将可解释方法分为建模前解释、训练后解释与模型自解释3大类,并对其在模型诊断、安全评估、数据纠偏、知识发现等场景的应用进行了分析;最后,对目前电力智能分析的机器学习可解释性研究面临的挑战进行了展望。
The interpretability of machine learning is one of the key points and important foundation for the safe and reliable application in power systems.In this paper,the interpretable machine learning methods applied in power system intelligent analysis are studied preliminarily.First,the basic concept,mathematical description,main theories and evaluative dimension of the interpretable methods are elaborated.Then,the general methodology and technology framework are analyzed,in which the interpretable methods of machine learning are categorized into ante-hoc,post-hoc and self-explained approaches.The application areas of the machine learning interpretability,including model diagnosis,security evaluation,data rectification and knowledge discovery,are studied.Finally,the current research challenges of interpretability are discussed for the machine learning application in power systems intelligent analysis.
作者
蒲天骄
乔骥
赵紫璇
赵鹏
PU Tianjiao;QIAO Ji;ZHAO Zixuan;ZHAO Peng(China Electric Power Research Institute,Haidian District,Beijing 100192,China)
出处
《中国电机工程学报》
EI
CSCD
北大核心
2023年第18期7010-7029,共20页
Proceedings of the CSEE
基金
国家重点研发计划项目(2020YFB0905900)
国家自然科学基金项目(U2066213)。
关键词
机器学习
可解释性
电力系统智能分析
人机交互
人工智能(AI)应用安全
machine learning
interpretability
power systems intelligent analysis
human-AI interaction
artificial intelligence(AI)application security
