Mitigating Online Microgrid Controller Deployment Issues with Learning Techniques: An Approximate MPC Approach

Due to the increasing complexity of energy systems, standard rule-based controllers may become inadequate for Energy Management System (EMS) supervision. Advanced optimal techniques as Model Predictive Control (MPC) can be too intensive in computational resources, especially for small agents with low-performance environment. Therefore, control issues about reducing implementation complexity gains interest in literature. In particular, surrogate models integrating learning techniques come within this scope by mitigating online implementation issues with offline data-driven approximations. However, there is a lack of works dealing with the analysis of performance robustness of such approximate controllers, when any uncertainties occur in online control. In this paper, we propose a freestanding approximate explicit MPC method, mobilizing learning techniques that approximate an analytical control law derived from a formal multiparametric programming. We propose to characterize the performance of the so-called reduced controller on simulation of a PV-storage microgrid, with historical data and forecast mismatch. We found numerically that the reduced controller has very similar performances compared to the baseline MPC when the forecast error is approximatively limited to 20%, and globally presents a relative change of economical loss less than 4% in average on simulation tests.