Managing Power Flows in Microgrids Using Multi-Agent Reinforcement Learning

Smart Microgrids bring numerous challenges, including how to leverage the potential benefits of renewable energy sources while maintaining acceptable levels of reliability in the power infrastructure. One way to tackle this challenging problem is to use intelligent storage systems (batteries and supercapacitors). Charging and discharging them at the proper time by exploiting the variablity of the renewable energy sources guarantees to balance supply and demand at any time. Reinforcement Learning (RL) is a branch of artificial intelligence encompassing techniques that allow agents (in our case electrical devices) to learn to behave rationnally, that is to perform sequences of decisions in order to optimize a given performance criteria. The theoretically sound framework of Reinforcement Learning makes these techniques to be increasingly used for solving difficult control problems. In this paper, a multi-agent reinforcement learning technique is proposed as an exploratory approach for controling a gridtied microgrid in a fully distributed manner, using multiple energy storage units and the grid. Preliminary simulation results using different scenarios show the feasibility and validity of the approach on a test microgrid, and open the way for future work in the field of agent-based learning control strategies in Smart Microgrids.