A Neural Network Based PEMFC Dynamic Model for Hardware-in-the-Loop Application

As the pace of decarbonization continues to accelerate around the world, hydrogen energy as one of the energy solutions of the future has achieved considerable development. Automobiles based on PEMFC are one of the main products for the promotion of hydrogen energy at this stage. The complex operating conditions of fuel cell vehicles put higher requirements on the fuel cell system's dynamic response performance, and the system's dynamic response performance is closely related to the system’s dynamic control strategy. The hardware-in-the-loop testbench is an important role in the toolchain of the fuel cell system's controller development. The fuel cell system model in the real-time controller and signal transceiver board can be used to simulate the characteristics of the system. Facing fuel cell system dynamic control strategy development needs, the hardware-in-the-loop testbench requires a high-precision dynamic model. The data-driven neural network model has long been widely used in the field of nonlinear modeling. The input of the regression neural network structure considers both historical output and current input parameters, which is consistent with the actual situation of fuel cells. In this paper, experiments are carried out based on an 80kW fuel cell system, and the test sequences are designed within the allowable range of load change rules. Then the test result of the whole test sequence is divided into a training dataset and a verification dataset. The fuel cell model based on the regression neural network was trained by using the training dataset, and the influence of network hyperparameters on the loss function was analyzed. Finally, the model simulation accuracy is verified in the hardware-in-the-loop application scenario.