Efficient Power Grid Topology Control via Two-Stage Action Search

Topology control in large-scale power grid system is a challenging problem because of the enormous state and action space. This paper proposes an efficient two-stage topology control framework to make a trade off between speed and accuracy. The framework consists of two components: a fast nominator for candidate actions generation and a slower but more accurate ranker for final action selection. Differing from previous works, this paper formulates candidates generation as a sequential decision making process, so as to take full advantage of information feedback from the ranker to guide the subsequent candidates generation process. To achieve this, the nominator is built as a GRU-based agent and is trained via reinforcement learning (RL). Experiment results show that the nominator is able to capture valuable information from historical feedback from ranker, and our method outperforms traditional methods on L2RPN NeurIPS 2020 Adaptability Track benchmark.