Contingency Planning in Pre-Incident based on Reinforcement Learning Assuming Multiple Faults in Transmission Lines by Typhoons

Abstract

Efforts to minimize power outage caused by natural disasters receive much attention. In that context, contingency plan in pre-incident is expected to be effective for reducing potential disaster damage and realizing quick recovery. Existing contingency planning schemes are usually designed as model-based approach to prevent either the worst-case scenario or a few typical scenarios. However, the effectiveness and adaptability of these schemes can be reduced in the long-term disaster such as a typhoon; there are an enormous number of risk scenarios caused by cascading accidents over time. In this paper, we propose a policy-based planning approach based on reinforcement learning for pre-allocation of power-supply cars and re-scheduling generator operation. By learning effective policies via simulations of various disaster scenarios in advance, the model can instantly output plans tailored to the disaster situation even for an enormous number of disaster scenarios. Therefore, the derived plan can follow the progress of the disaster dynamically. The practicality of proposed method is evaluated via numerical experiments based on the original grid simulating Eastern Japan and Typhoon No. 15 in 2019. The result shows the reduction of power outage by the proposed method compared to the existing method under multiple faults in transmission lines.