Hybrid physics–data model for real-time compressor aerodynamic performance prediction in fuel cells

Abstract

In the actual operation of fuel-cell systems, providing the most accurate centrifugal compressor aerodynamic performance prediction in real-time is crucial for system efficiency and control. One key factor influencing the prediction is choke, which causes blade vibrations, limits the system’s maximum power and leads to a sharp decline in efficiency. However, the embedded multi-dimensional look-up tables typically necessitates a substantial amount of bench tests and makes extrapolation difficult. To address this, the study innovatively introduces a hybrid modeling framework that uses a small-sample physics-based model as a prior constraint and a data-driven method to compensate for physics-based prediction errors. The residual learning serves as a bridge trained with Gaussian process regression (GPR) by minimizing the root mean square error (RMSE). The hybrid model is implemented and validated on a fuel cell centrifugal compressor. Results show that, the physics-based prediction effectively constrains the shape of the characteristic curves, while the data-driven method significantly improves accuracy, especially in the choke region, with a 5% enhancement. In real-time prediction, as training samples expand, the accuracy improves further with the validation data almost all falling within the 95% confidence interval. The maximum training time is 20.96 s, indicating suitability for real-time prediction. This study highlights the potential of hybrid model of fuel-cell centrifugal compressors in engineering applications.