Real-Time Optimization of Fuel Cell Cogeneration Systems with Safety-Aware Self-Learning Algorithms

We propose a self-learning real-time optimization framework with safety constraints for power target allocation of parallel hydrogen fuel cell co-generation stacks, where the performance characteristics of the fuel cells for electrical and thermal power output are initially unknown. For this purpose, Gaussian Processes (GPs) are utilized to model the unknown fuel cell characteristics while safety constraints for total thermal power output and total hydrogen consumption are enforced using upper confidence intervals of the learned GPs. Additionally, the estimation uncertainty of GPs is incorporated into the objective function as a contributing term to enable exploration, where the weight of this term is varied according to the process operating conditions. The expected tracking errors between the desired and achieved electrical and thermal power outputs are incorporated into the objective function as well and are penalized with predetermined weights. Simulation results show that the proposed framework can reduce the overall hydrogen consumption and meet both electrical power and thermal power targets - when they are feasible - while always satisfying the safety constraint. Comparison with an oracle algorithm with access to the true values of the fuel cell performance curves, reveal that the proposed approach attains comparable performance following a small number of external reference changes and independent exploration moves by the algorithm.