Deep reinforcement learning (DRL)-based volt-var control (VVC) in active distribution networks (ADNs) is prone to cyberattacks; malicious attackers can design adversarial attacks to tamper with the real-time measurements and bypass the bad data detection (BDD) mechanism, causing voltage limit violations. Previous work in this area predominantly focuses on designing suitable attack vectors, assuming unlimited attacker resources and neglecting the BDD mechanism. To this end, a new objective as a function of attack timing is designed to maximize the frequency of voltage violation of ADNs with the fewest number of attacks. Furthermore, perturbation set transformation is implemented on the attack vectors to bypass the BDD, ensuring the stealthiness of the attack. For solving the attack optimization with complex constraints, this article reformulates the adversarial attack on DRL-based VVC agents as a hybrid-action state-adversarial Markov decision process (MDP) and proposes a novel knowledge transferred DRL-based adversary to launch strategically timed (when) and stealthy (how) adversarial attacks. Case studies are conducted to verify that the strategically timed adversarial samples generated by the adversary significantly impact the voltage regulation capability of VVC agents while ensuring stealthiness against the BDD mechanism, through benchmarking against existing baseline adversarial attack methods.
Chen et al. (Thu,) studied this question.