Data-driven Parameterized Policies for Microgrid Control

If we account for data uncertainty, centralized microgrid control can be decomposed in four tasks: estimating the parameters of the microgrid devices, forecasting the consumption and the renewable production, operational planning to anticipate weather effects and human activities, and real-time control to adapt planned decision to the reality of the moment. As the devices of a microgrid deteriorate over their lifetime and microgrids are by definition small systems, it is of paramount importance to automate the meta-parameterization of these stages to maximize the microgrid efficiency and decrease its maintenance costs. This paper studies how reinforcement learning can be used to address this problem. As reinforcement learning makes use of microgrid operation data (or of simulated data before the microgrid is actually operated) to learn an operation policy, it inherently merges the four stages above, and can in theory adapt to some types of changes without having to perform manual tuning.