A New Energy Management Method Based on Reinforcement Q-learning for Minimizing the Operational Cost of an Energy System

This paper presents a new method of energy management (EM). The proposed method was defined as a reinforcement learning problem for a system consisting of solar panels, wind turbines, gas engines and batteries. The best corresponding energy managements for all the possible system states were found by employing the Q-learning algorithm to solve the EM problem. These results were then used for managing the power supply of four residential buildings in Negin Island Bushehr located in the southwest of Iran. The simulation was performed for 8760 hours of a year. The proposed energy management method was compared with the load following energy management, which resulted in a 2.4% reduction in annual operational costs and a 3.7% reduction in CO2 emissions. With the same quality of the results, the new EM method took about 2.5 times less computational time compared to the optimum energy management performed by genetic algorithm. The effect of iteration number on the convergence of algorithm was also investigated.