Theory-Constrained Neural Network With Modular Interpretability for Fuel Cell Vehicle Modeling

Combining the physical significance of theoretical models with the high fitting accuracy of data-driven models is crucial for precise characterization of fuel cell electric truck (FCET) energy consumption. This paper proposes a theory-constrained neural network (TCNN), integrating physical significance without compromising accuracy. A theory-guided filter is applied to ensure the interpretability of each sub-module. Fuel cell temperature and voltage are modeled separately, each serving as an important input for the other. Sub-networks are trained individually under the constraints of the theory-guided filter to minimize the error between the final output and the data. Additionally, a neural network based on the CNN-BiLSTM-MHSA architecture ensures model accuracy. Results show that the proposed neural network significantly improves fitting accuracy for both fuel cell temperature and voltage. The proposed TCNN successfully integrates the theoretical model, improving accuracy by 33.7% compared to the theoretical model and maintaining accuracy comparable to the purely data-driven model. Hydrogen consumption calculated based on this model shows accuracy improvements of 63.7% and 49.4% compared to traditional methods using efficiency curves and hydrogen consumption curves, respectively. The proposed voltage model and temperature model using TCNN enable accurate hydrogen consumption calculations, which can be broadly applied to energy management, component sizing, and other related areas.