A Resilience-enhanced Multi-agent Framework for Power Systems Based on Proximal Policy Optimization

Existing power system resilience enhancement methods, such as mobile power scheduling, active generation rescheduling, and network topology reconfiguration, have not fully explored the potential of the reactive power compensator to maintain voltage stability during and after an outage event and flexibility. Additionally, they require a high degree of accuracy in system modeling, which poses challenges for them to scale up their application to large-scale integrated power grids. To this end, in the paper we propose a multi-agent framework based on deep reinforcement learning for proximal policy optimization methods, aiming to address the computational and scalability issues associated with accurate system modeling. Moreover, we develop a reactive compensator deployment plan for power system resilience enhancement. This plan incorporates a hybrid soft action critic algorithm based on agent for offline localization, classification, and on-line control of reactive compensators to improve their voltage recovery capability. The multi-agent framework learns from previous experiences and is trained to determine the appropriate location and sizing of reactive power compensators, thereby preventing bus voltage overshoot during multi-line faults. As a case study, the voltage overshoot problem caused by a multi-line outage during a storm is validated on an IEEE 39 bus system. The results indicate that the multi-agent framework can effectively coordinate and control the reactive power compensators to eliminate or minimize the bus voltage overruns during faulted line off-grid conditions, which improves the resilience of the power system.