Machine Learning for Optimal Control of a Hydrogen Gas Reformer Used in a PEM Fuel Cell

In this paper we demonstrate how to use the reinforcement learning policy iterations, a machine learning technique, to efficiently learn the optimal feedback gain of a hydrogen gas reformer that supplies hydrogen to a proton exchange membrane (PEM) fuel cell that are used for vehicular applications (electric vehicles). The obtained results show that in the considered problem, the optimal gain can be learned in only a few policy iterations despite the fact that the produced fuel cell current represents an antagonistic disturbance to the state variables of the hydrogen gas reformer (as well as to the PEM fuel cell), whose state space mathematical model is of high dimensions and represented by a system of ten coupled differential equations. The presented approach provides more accurate results than the corresponding optimization problem when the disturbance is neglected.