Machine learning aided methods for reducing the dimensionality of the comprehensive energy economy optimization of the fuel cell powertrains

In pursuit of the energy economy, complex high dimensional optimization of design schemes of the fuel cell (FC) powertrain has been developed and it usually requires total reconstruction of the powertrain. Does each design variable equivalently contribute to the energy economy of the powertrain? Could the original complex optimization problem be simplified while maximizing the improvement in energy economy at the minimum cost of modifications? In this study, based on the comprehensive energy dissipation model and aided by the machine learning method for removing computational loads, the importance of design variables involved in the design schemes of FC powertrain to its energy economy is quantified through global sensitivity analysis. According to the quantified importance pattern of design variables, the partial optimization with respect to five most important design variables offers a 17.78% (compared to the benchmark design) improvement in energy saving, which performs almost as well as the complete optimization of all design variables (19.62%). And the partial optimization retains nearly half of the original settings in benchmark design scheme. Also it is verified that there is no significant improvement in energy economy if more factors are considered. It should be noted that the true available design space is shrunk, due to the fact that the optimization is conditioned on a given benchmark design scheme. Therefore, in this case, the comparative validations demonstrate that generic sensitivity analysis considering entire design space provides misleading information for resolving the critical design variables.