The application of neural networks to fuel processors for fuel-cell vehicles

Passenger vehicles fueled by hydrocarbons or alcohols and powered by proton exchange membrane (PEM) fuel cells address world air quality and fuel supply concerns while avoiding hydrogen infractructure and on-board storage problems. Reduction of the carbon monoxide concentration in the on-board fuel processor's hydrogen-rich gas by the preferential oxidizer (PrOx) under dynamic conditions is crucial to avoid poisoning of the PEM fuel cell's anode catalyst and thus malfunction of the fuel-cell vehicle. A dynamic control scheme is proposed for a single-stage tubular cooled PrOx that performs better than, but retains the reliability and ease of use of, conventional industrial controllers. The proposed hybrid control system contains a cerebellar model articulation controller artificial neural network in parallel with a conventional proportional-integral-derivative (PID) controller. A computer simulation of the preferential oxidation reactor was used to assess the abilities of the proposed controller and compare its performance to the performance of conventional controllers. Realistic input patterns were generated for the PrOx by using models of vehicle power demand and upstream fuel-processor components to convert the speed sequences in the Federal Urban Driving Schedule to PrOx inlet temperatures, concentrations, and flow rates. The proposed hybrid controller generalizes well to novel driving sequences after being trained on other driving sequences with similar or slower transients. Although it is similar to the PID in terms of software requirements and design effort, the hybrid controller performs significantly better than the PID in terms of hydrogen conversion setpoint regulation and PrOx outlet carbon monoxide reduction.