In this paper, an innovative robust prediction algorithm for performance degradation of proton exchange membrane fuel cell (PEMFC) is proposed based on a combination of model-based and data-driven prognostic method. A novel approach using the moving window method is applied, in order to 1) train the developed models; 2) update the weight factors of each method and 3) further fuse the predicted results iteratively. In the proposed approach, both model-based and data-driven methods are simultaneously used to achieve a better accuracy. During the prediction process, each dataset in the proposed moving window are divided into three sections respectively: training, evaluation and prediction. The training data are used first to identify the models parameters. The evaluation data are then used to measure the weight of each method, which represents the degree of confidence of each method in the actual state. Based on these dynamically adjusting weight factors, the prediction results from different methods are then fused using weighted average methodology to calculate the final prediction results. In order to verify the proposed method, three experimental validations with different aging testing profiles have been performed. The results demonstrate that the proposed hybrid prognostic approach can achieve a higher accuracy than conventional prediction methods. In addition, in order to find the satisfactory trade-off between the prediction accuracy and forecast time for optimizing on-line prognostic, the performance variation of proposed approach with different moving window length is further showed and discussed. (C) 2017 Elsevier Ltd. All rights reserved.