A novel learning-based MPC with embedded profiles prediction for microgrid energy management*

Abstract

This paper presents a novel algorithm for microgrid energy management based on a Differentiable learning-based Model Predictive Control (MPC) for jointly optimising profiles prediction and control performance. Specifically, we propose an algorithm for the online training of a Neural Network (NN) that predicts the unknown parameters of the MPC optimisation problem during control operation. Since the training is performed online at each time step the controller adapts to possible changes in the system parameters, while avoiding the offline training phase. Differently to standard methods in the literature, the proposed NN is trained by minimising a performance-based loss, i.e. the total cost of the energy trading with the utility grid. Simulation results show that the proposed approach outperforms the traditional approach minimising an estimation-only MSE loss, both when the model parameters are perfectly known and when they are uncertain.