Journal Article

Model-Data Hybrid-Driven Real-Time Optimal Power Flow: A Physics-Informed Reinforcement Learning Approach

Xu, Huan, Lin, Zhida, Yang, Duotong, Ma, Xiyuan, Zhang, Ximing, Zhou, Changcheng, Li, Zhuohuan, Yu, Yun

2025

Overview

With the rapid development of artificial intelligence technology, DRL has shown great potential in solving complex real-time optimal power flow problems of modern power systems. Nevertheless, traditional DRL methodologies confront dual bottlenecks: (a) suboptimal coordination between exploratory behavior policies and experience-based data exploitation in practical applications, compounded by (b) users’ distrust from the opacity of model decision mechanics. To address these, a model–data hybrid-driven physics-informed reinforcement learning (PIRL) algorithm is proposed in this paper. Specifically, the proposed methodology uses the proximal policy optimization (PPO) algorithm as the agent’s foundational framework and constructs a PI-actor network embedded with prior model knowledge derived from power flow sensitivity into the agent’s actor network via the PINN method, which achieves dual optimization objectives: (a) enhanced environmental perceptibility to improve experience utilization efficiency via gradient-awareness from model knowledge during actor network updates, and (b) improved user trustworthiness through mathematically constrained action gradient information derived from explicit model knowledge, ensuring actor updates adhere to safety boundaries. The simulation and validation results show that the PIRL algorithm outperforms the baseline PPO algorithm in terms of training stability, exploration efficiency, economy, and security.

• Publisher: MDPI AG
• Subject: Algorithms / Alternative energy sources / Artificial intelligence / deep reinforcement learning (DRL) / Efficiency / Electric power systems / Energy resources / Learning / Neural networks / Optimization / Parameter estimation / Physics / physics-informed neural network (PINN) / physics-informed reinforcement learning (PIRL) / real-time optimal power flow (RT-OPF) / Renewable resources