A hybrid physical-data driven model on proton exchange membrane fuel cells for cooling, heating and power supply with gas heat recovery

The increasing development of hydrogen energy in buildings accelerates the applications of proton exchange membrane fuel cell (PEMFC) combined cooling, heating and power (CCHP) system. However, the PEMFC output thermal properties have not been characterized comprehensively, and the PEMFC model complexity and computational efficiency have not been balanced suitably in long-time scenario operation. In this study, in order to guarantee both model complexity and computational efficiency, three surrogate models of PEMFC are developed to characterize the output thermal properties comprehensively, i.e., semi-empirically physical (SP) model, neural network (NN) model, and hybrid physical-neural network (HPNN) model. Three surrogate models are identified by database training with total 1225 population, and the generalization accuracy of the surrogate models is comparatively analyzed based on in-field and CFD-based simulation experiment results. A CCHP system integrating with PEMFC HPNN model are investigated based on co-simulation with exhaust burning gas heat recovery. The results indicate that, the HPNN model exhibits the prioritized generalization accuracy compared to the SP model and NN model. The coolant net thermal power output of PEMFC stack occupies 98.149% of the total thermal power output. The CCHP system could recover 1.634 × 1010 J thermal energy from the exhaust gas flow during the whole year, leading to a 5.913% energy saving of balance of plant components and 0.720% improvement of system total energy efficiency. The results could provide a guidance for PEMFC modeling method, i.e., the HPNN model structure that could guarantee both model generalization accuracy and physical mechanism interpretability. The system analysis results will give a CCHP system design for high complexity and efficiency that could accelerate the hydrogen energy development.