Application of SARSA learning algorithm for reactive power control in power system

This paper aims to make use of a reinforcement learning method in order to compute an approximation of the optimal control strategy for a set of reactive compensators in power system to maintain acceptable voltage profile and some elements of power system state vector within operating limits under load variations and system contingencies with the minimum usage of reactive power resources. Optimal control settings are learnt by experience and by means of reward values which indicate effectiveness of actions for satisfying state variable constraints. The SARSA learning algorithm which is an on-policy algorithm in RL concept is applied to the IEEE 39-buses New England power system. Results show that SARSA learning algorithm is able to provide optimal or near optimal control settings for power system under varying system conditions.