A Reinforcement Learning Approach for Sequential Decision-Making Process of Attacks in Smart Grid

摘要

An attacker can very possibly make significant damage for the power grid with a proper sequence of timing and attacks. Existing approaches neglect the power system generation loss and also identification of critical attack sequences. In this paper, we investigate a reinforcement learning approach to identify the minimum number of attacks/actions to reach blackout threshold. The attacker will only have limited topological information of the power systems. Proper state vectors, action vectors and also reward are designed in this smart grid security environment. The proposed method is evaluated on a W & W 6 bus system and an IEEE 30 bus system. The attack performance is tested for different percentages of line outage. The amount of load shedding is also considered as an attack objective and demonstrated on W & W 6 bus system. The optimal attack sequence is identified through a trial-and-error learning process and is then validated on a power system simulator.