A reinforcement learning solution for the unit commitment problem

Abstract

This work proposes a solution method for the unit commitment (UC) problem using the reinforcement learning (RL) technique. The UC problem is an optimization problem whose objective is to minimize the electrical power system operational cost. The solution of an UC problem yields an operational schedule for a set of generation units while satisfying unit operational constraints and system demand. This paper considers a four-state definition (maximum, minimum, banking and off) for the thermal generation units. In this scenario, it is verified that the modelling of the UC problem has the markovian property which allows the use of the RL method and thus giving the possibility to solve the UC problem by the means of rewards. These rewards are derived from unit operational costs, the system demand and the electrical system constraints. A two-step algorithm is proposed for searching a solution (considering the subsequent and sub subsequent states) which allows the learning agent to be able to efficiently evaluate every alternative and choose the best available one. Finally, the method performance is measured considering a 10-unit system proving the effectiveness of the proposed method.