Abstract
A hierarchical control framework that combines model predictive control (MPC) and reinforcement learning (RL) to separate decisions on two timescales: long-horizon setpoints for the slow dynamics, and a pretrained RL agent that tracks them in real time. Applied to a battery energy storage system in frequency regulation markets, it balances profit opportunities that last seconds against battery degradation that develops over weeks to months. Compared with MPC baselines, it extended battery lifetime by 84% and increased operational profit by 34%.
Type
Publication
arXiv preprint
Dr Paulina Quintanilla
Assistant Professor in Process Systems Engineering
Assistant Professor in Process Systems Engineering at UCL, leading the LOOPS research group. Research interests include machine learning, optimisation and control for processes and physical systems.