Data Driven Models for a PEM fuel cell stack prediction

Global energy policies are promoting energy efficiency, distributed generation (DG) and renewable energy resources (RES), increasing the number of "green" installations and their energy production [1-3]. When used in Distributed Generation (DG) applications, Fuel Cells (FC) have the potential to save energy and reduce emissions, however their inherent fuel-flexibility could help address energy shortage issues through energy diversity. In addition, FC have the potential to be quieter, more reliable, and have lower maintenance costs than most technologies used for DG. Polymer Electrolyte Membrane (PEM) technology have attracted a significant amount of research interest in the last years, both in stationary and mobile application, for their ability to be efficient and clean replacements for conventional power generators. Ideally, PEM fuel cells can operate on a H2 rich fuel (i.e.,"re¬formate") generated from fossil fuels such as coal, natural gas, gasoline and landfill gas, or alcohols in conjunction with fuel processing. In a multi-source hybrid plant, FC could thus play two key roles working as a micro-generator fed with H2 rich gas from natural gas (NG) by a reformer, and allowing the energy storage of H2 [4-6]. The operating principles of PEM fuel cells system involve electrochemistry, thermodynamics and hydrodynamics theory for which it is not always easy to establish a mathematical model. In this paper two different methods to model a commercial PEM fuel cell stack are discussed and compared. The models presented are non linear, derived from a black-box approach based on a set of measurable exogenous inputs and are able to predict the output voltage and cathode temperature of a 5 kW module working at the TAE Institute. A FC stack fed with H2 rich gas is employed to experimentally investigate the dynamic behaviour and to reveal the most influential factors. The performance obtained using a Neural Networks (NNs) model are compared with a PLS (Partial Least Square) based model. The results show that both strategy are capable of simulating the effects of different stoichiometric ratio in the output variables under different working conditions.