Deep Reinforcement Learning for Online Vulnerability Identification in Smart Grid under Load Uncertainty

Abstract

The outages of transmission lines can trigger cascading failure in power grid, resulting in serious negative impacts. Identifying those critical lines can help to take precautionary measures and build robust power system. A group of existing works have developed approaches to identify those critical lines. However, the approaches haven’t consider the load uncertainty, which will lead to the changes of critical lines. In this paper, we investigate a critical line identification problem in smart grid considering load uncertainty. Specifically, we first formulate an optimal virtual attacking problem and its objective is to maximize the expected generation loss under the given attacking resources. Due to the existence of large solution space, unknown system dynamics model, and load uncertainty, solving the formulated problem is challenging. Thus, we design an online algorithm based on prioritized multi-agent-attention-actor-critic (PMA3C). Compared with optimization-based methods and evolutionary methods, the designed algorithm can identify critical lines instantly under load uncertainty without searching the whole or partial solution space. Simulation results indicate that the designed algorithm can identify critical lines efficiently.