A Custom Loss Function Approach for Data-Driven Economic Model Predictive Control

This paper proposes a new controller that takes advantage of the properties of the recurrent neural network of the long short-term memory (RNN-LSTM) to imitate the behavior of an economic model predictive control (economic MPC). The approach introduces specific knowledge of the benchmark controller optimization problem and modifies the loss function during the training stage of the RNN-LSTM. This method is easy to implement, and its effectiveness demonstrates itself in a practical application: the energy management system of a microgrid on a university campus. The results show that the modification of the loss function improves the accuracy of the proposed controller compared to the traditional training method.