Data-Driven Predictive Control With Online Adaption: Application to a Fuel Cell System

Fuel cell systems constitute an electrochemical energy conversion system increasingly used in stationary and mobile applications. Complying with operational limits in transient operation can be achieved by model-based predictive control algorithms. The key challenge arises from the identification of suitable models for embedded real-time optimization. This article presents a data-driven predictive control approach for the air path and power control of a fuel cell system. In particular, we use data-enabled predictive control (DeePC) based on a concise system representation using column subset selection (CSS). The impact of problem formulation, regularization, and different solvers for quadratic programs (QPs) on the turnaround time on embedded hardware is investigated. In addition, we provide an online update algorithm for the system representation to account for the operating regions not contained in the initial dataset. The proposed approach is validated on a high-fidelity fuel cell system simulation and hardware-in-the-loop (HiL) experiments. We demonstrate safe and fast closed-loop control using the column subset algorithms for a comprehensive dataset and reduction in closed-loop cost for unknown operating areas of up to 25%. The control algorithm and the update algorithm are shown to be real-time feasible on a single-core embedded hardware.