Physics-Constrained Vulnerability Assessment of Deep Reinforcement Learning-Based SCOPF

The decarbonization of energy systems has posed unprecedented challenges in system complexity and operational uncertainty that render it imperative to exploit cutting-edge artificial intelligence (AI) technologies to realize real-time, autonomous power system operation and control. In particular, deep reinforcement learning (DRL)-based approaches in power systems are extensively studied and implemented in several trials worldwide. Nevertheless, the vulnerability of DRL brings new security threats to power systems that have not been well identified and investigated in the literature. To this end, this paper proposes a physics-constrained vulnerability assessment methodological framework for the DRL-based power system operation and control, with a special focus on the problem of security-constrained optimal power flow (SCOPF). In particular, we develop a novel adversarial example generation method, defined as a false data injection attack against the DRL-based SCOPF (FDIAI), to realize a targeted adversarial attack considering the nonlinear physical constraints in power systems via two main stages of constructor function design and unconstrained optimization problem transformation. In this way, the proposed FDIAI can significantly influence the decision-making procedure of DRL while successfully evading the bad data detection mechanism in power systems. Case studies are conducted to explore the stealthiness and effectiveness of FDIAI and then show its severe impacts on system operation and control on the winners' models of the Learning to Run a Power Network (L2RPN) competitions, including L2RPN IJCNN 2019 (IJCNN), L2RPN WCCI 2020 (WCCI), and L2RPN NeurIPS 2020 (NeurIPS).

Automated summary

This paper proposes a physics-constrained vulnerability assessment framework for DRL-based power system operation and control, addressing the vulnerabilities and security threats. A novel adversarial example generation method is developed to conduct targeted adversarial attacks and evade bad data detection mechanisms. Case studies on the winners' models of the L2RPN competitions demonstrate the severe impacts on system operation and control.