Deep Learning-Based Predictive Control for Optimal Battery Management in Microgrids

Effective microgrid management necessitates sophisticated strategies to optimally balance grid components and minimize power exchanges with the main grid. Central to this challenge is the energy storage system, typically comprised of lithium-ion batteries, which must operate within specific safety thresholds. Among the different approaches used for battery management in microgrids, model predictive control appears particularly suitable due to its ability to deal with nonlinear systems and constraints. However, the practical deployment of predictive control is often constrained by its substantial computational demands. Notably, achieving high performance typically requires a long prediction horizon, which exacerbates the computational complexity that increases superlinearly with the horizon length. To overcome these limitations, this paper exploits a neural network to approximate the predictive control law, thereby maintaining constant online time complexity regardless of the prediction horizon and facilitating real-time application. This innovative deep learning-based strategy is applied and specifically adapted for the first time to microgrid battery management, incorporating a comparative analysis of several machine learning models to identify the most efficient solution for this application. The results demonstrate that this approach can achieve performance comparable to traditional controllers while ensuring scalability and efficiency. Specifically, the proposed methodology is able to approximate the predictive control action with a mean error of 0.24A and a standard deviation of 2.11A, while reducing the required computational cost by over 200 times when considering a two-day ahead prediction horizon.