Dynamic PEMFC Model as a Base for a State Classifier and Controller

Abstract

Online state classification of PEM fuel cell systems is a challenging task. While electrochemical impedance spectroscopy (EIS), the measurement of the 1kHz impedance, cyclic voltammetry, and the polarisation curve are the state-of-the-art in laboratory testing, it is hardly possible to install these analytic methods in application-oriented fuel cell systems. The necessary technical equipment is expensive, and hybridisation would be essential to create artificial operating conditions necessary during the measurement. A more economic solution for online state classification is the integration of a real-time model-based state classifier. A model, operating in parallel to the fuel cell, is able to provide extra information about the fuel cell such as the humidity of the membrane, which is not measurable in situ. On the other hand, online comparison of measurement data with the theoretical calculation allows us to detect malfunction of components. Based on a detailed model and a complex classification database, including error states of the system, it is possible to classify the state of a fuel cell or detect system errors. The voltage drift between classical physical models and a real fuel cell caused by irreversible degradation (platinum diversion, membrane thinning, carbon corrosion) and reversible degradation (contamination) is not yet completely understood, and therefore modelling so far remains imprecise. To avoid the misinterpretation caused by these constantly growing differences between the model and the reality, the described modelling approach has been focused on dynamic fuel cell reactions. Dynamic step responses of the fuel cell for abruptly changing operation conditions are being simulated in real time. The changing operation conditions may result from a real load cycle or, as for theses analyses, by forced stoichiometric steps. The comparison of the real and the simulated step responses is the basis for the state classifier and the controller to react. A conclusion of the series of measurement is shown, the model and its parameters are presented, and the opportunities of the state classifier and controller are discussed in this paper.