Prognostics of proton exchange membrane fuel cell stack in a particle filtering framework including characterization disturbances and voltage recovery

In the perspective of decreasing polluting emissions and developing alternative energies, fuel cells, and more precisely Proton Exchange Membrane Fuel Cells (PEMFC), represent a promising solution. Even if this technology is close to being competitive, it still suffers from too short life duration. As a consequence, prognostic seems to be a great solution to anticipate PEMFC stacks degradation. However, a PEMFC implies multiphysics and multiscale phenomena making the construction of an aging model only based on physics very complex. One solution consists in using a hybrid approach for prognostics combining the use of models and available data. Among these hybrid approaches, particle filtering methods seem to be really appropriate as they offer the possibility to compute models with time varying parameters and to update them all along the prognostics process. But to be efficient, not only should the prognostics system take into account the aging of the stack but also external events influencing this aging. Indeed, some acquisition techniques introduce disturbances in the fuel cell behavior and a voltage recovery can be observed at the end of the characterization process. This paper proposes to tackle this problem. First, PEMFC fuel cells and their complexities are introduced. Then, the impact of characterization of the fuel cell behavior is described. Empirical models are built and introduced in both learning and prediction phases of the prognostics model by combining three particle filters. The new prognostic framework is used to perform remaining useful life estimates and the whole proposition is illustrated with a long term experiment data set of a PEMFC in constant load solicitation and stable operating conditions. Estimates can be given with an error less than 5% for life durations of more than 1000 hours. Finally, the results are compared to a previous work to show that introducing a disturbance modeling can dramatically reduce the uncertainty coming with the predictions.