Data-driven and model-based hybrid reinforcement learning to reduce stress on power systems branches

This work proposes a reinforcement learning (RL) approach to tackle the control problem of branch overload relief in large power systems. Accordingly, a control agent is trained to change generators' real power output in order to relieve the stressed branches. For large power systems, this control problem becomes one whose decision space (i.e., the action space) is both highly-dimensioned and continuous. This makes it extremely difficult to have successful training for RL-based agents. To improve the effectiveness, a data-driven and model-based hybrid approach is proposed to optimize the control by combining RL-agent actions and generator shifting factor-driven actions. Accordingly, with the proposed approach the RL-agent successfully trains on large power systems. The proposed design is tested on both the IEEE 118-bus testing system and a 2749-bus real system. The obtained results show that the proposed hybrid approach outperforms the data-driven training approach.