Abstract: Accurately predicting the degradation characteristics of fuel cells can provide a solid basis for control and diagnosis decisions. However, mainstream data-driven methods often do not take into account uncertainty factors such as measurement errors caused by experimental conditions and the model's dependence on the data during the modeling phase. Therefore, only a single point estimate can be provided, resulting in a lack of sufficient credibility in the performance degradation results. This paper proposes a mixed-probability data-driven model (MPDD) that combines the characteristics of multiple data-driven models using Bayesian theory, which could quantify the uncertainty of the model's dependence on the data and provide prediction results for fuel cell performance degradation trends in both point estimates and interval estimates. Based on the full operating condition data in fuel cell dynamic load cycle (FC-DLC), the point estimate results of the MPDD model outperform those of a single data-driven model. Furthermore, based on the steady-state data in FC-DLC, the MPDD model achieves up to a 33% improvement in the concentration rate of interval estimates compared to Gaussian process regression (GPR). The prediction results indicate that this forecasting method possesses excellent uncertainty quantification capabilities and could provide more practical decision recommendations for the operation of electro-hydrogen coupling devices.