Global Explainability for Trustworthy RL in Renewable-Integrated Grids: A Framework for Substation Topology Control
To Tien Hung, Nguyễn Quốc Minh
Trường Đại học Khoa học và Công nghệ Hà Nội Hanoi University of Science and Technology
阅读操作
确认中在文库中上传 PDF 后可生成中文音频讲解。
摘要与影响
The rapid transition toward decarbonization and the large-scale integration of renewable energy sources have significantly increased the operational complexity of modern electricity systems. As grids become more dynamic and decentralized to accommodate green energy uncertainty, human operators face intensifying cognitive demands, requiring intelligent decision support systems for rapid, coordinated control. Reinforcement Learning (RL) has emerged as a powerful paradigm for managing these complexities through autonomous topology optimization. However, the “black-box” nature of RL agents remains a critical barrier to their deployment in safety-critical control rooms. This paper addresses this trust gap by presenting a comprehensive interpretability framework for an RL agent performing substation topology control on the IEEE 14-bus system. We propose a workflow that utilizes an XGBoost surrogate model to replicate the agent's decision-making logic with high fidelity. Through the application of post-hoc explainability techniques, specifically Partial Dependence Plots (PDP) and Accumulated Local Effects (ALE) via the OmniXAI library, the global trends of the agent's behavior are systematically uncovered. The analysis reveals that the agent's actions are primarily governed by line overflow durations, demonstrating physically consistent strategies such as immediate contingency resolution or threshold-based adaptive responses. These findings align with established grid operational standards, proving that the agent learns intuitively reasonable control logic. This study provides a scalable routine for integrating transparency into power system RL, paving the way for trustworthy, human-aligned artificial intelligence in future grid operations.
逐年被引趋势
暂无年度引用数据
关键指标
同类平均 = 1
同领域 · 同年份 · 同类型
Google Scholar 与 OpenAlex 的被引统计范围不同,数值存在差异属正常。
AI 辅助阅读
依据:摘要
可就本文提问;依据不足时会说明。
学术脉络
学科主题
计算机 / AIExplainable Artificial Intelligence (XAI)
Power System Optimization and Stability · Smart Grid and Power Systems