A Reinforcement Learning Approach for Branch Overload Relief in Power Systems

This work presents a reinforcement learning (RL) approach to the problem of branch overload relief. An agent is trained to re-dispatch generators' real power output in order to adjust the power flow through the network so that none of its branches is overloaded. The generation re-dispatch agent is trained using the deep deterministic policy gradient (DDPG) algorithm. The proposed approach is tested on both the IEEE 14-bus and 39-bus systems. Once trained, the performance of the re-dispatch agents was compared against that of the classical optimal power flow (OPF) approach. The trained agents demonstrated better performance with both having: (a) close to a 100% success rate for cases which had an OPF solution; and (b) 70.14% and 46.48% success rates for cases which had no feasible OPF solution. Such results confirm the potential of the proposed DDPG-based RL re-dispatch approach as a reliable method for managing branch overloading.