Polymer electrolyte membrane fuel cells are emerging as a promising technology for achieving carbon-free mobility, particularly in the automotive sector. In fuel cell-dominant vehicles, highly dynamic operation of the fuel cell system (FCS) is required, while simultaneously maintaining high efficiency and safe operation. This entails avoiding oxygen starvation, compressor surge and choke, as well as ensuring proper membrane hydration, all of which are also critical for system longevity. This thesis presents a model predictive control (MPC) framework designed, implemented, and evaluated for an automotive FCS in a fuel cell-dominant hybrid electric vehicle (FCHEV). The control structure is arranged as a hierarchical MPC scheme, composed of a high-level and a low-level control layer. The high-level control layer includes a power-split nonlinear MPC (NMPC) and a target selector. The former optimizes the power distribution between the FCS and the hybrid battery to maximize FCS efficiency and improve battery charge sustainability. Given this power distribution, the target selector ensures a statically optimal allocation of the compressor operating point, preventing overactuation of the FCS. The high-level control layer, which considers a dynamic model of the battery and a static model of the FCS, operates on a one-second sampling time in line with the power demand dynamics. The low-level control layer comprises a tracking MPC and a state and disturbance estimator. Considering the transient behavior of the FCS, the tracking MPC computes the optimal inputs for FCHEV actuation, with the objective to dynamically track the references provided by the high-level control layer while respecting the system’s operational bounds. The extended Kalman filter estimates internal states of the fuel cell stack and provides a static model error for zero steady-state offset tracking. To capture the transient system behavior, the tracking MPC operates at a faster rate, which poses challenges for real-time feasibility on embedded hardware. In this context, a comparison between linear time-varying MPC and NMPC shows that nonlinear prediction is advantageous under highly dynamic operation, particularly when rapid power changes interact with safety-critical constraints. This motivates a careful design of the NMPC, including the selection of sampling time, numerical integration scheme, and prediction horizon to ensure real-time feasibility. The real-time feasibility of the resulting controller is demonstrated on embedded hardware. The overall MPC framework, including the real-time capable tracking MPC, is tuned using multi-objective Bayesian optimization to balance dynamic capability, system efficiency, and constraint satisfaction. A further contribution of this work is the incorporation of humidity into the tracking MPC, which is essential for accurately capturing the dynamic behavior of the FCS and enabling proper water management.
Thuc Anh Nguyen (Wed,) studied this question.