Environmental issues, especially global warming due to the greenhouse effect, have become more and more critical in recent decades. As one potential candidate among different alternative "green energy" solutions for sustainable development, the proton exchange membrane fuel cell (PEMFC) has received extensive research attention for many years for energy and transportation applications. However, the relatively short lifespan of PEMFCs operating under non-steady-state conditions (for vehicles, for example) impedes its massive use. The accurate prediction of their aging mechanisms can thus help to design proper maintenance patterns of PEMFCs by providing foreseeable performance degradation information. In addition, the prediction could also help to avoid or mitigate the unwanted degradation of PEMFC systems during operation. In this paper, an advanced self-adaptive relevance vector machine (RVM) has been developed and demonstrated to predict the performance degradation of PEMFCs. In order to prove the versatility of proposed RVM method, the predictive results are experimentally validated using two different PEMFC stacks aging data under different operating patterns. Furthermore, the obtained results are compared with results from both classic support vector machine and original RVM methods in order to highlight the effectiveness of the proposed self-adaptive RVM method with a modified design matrix. A comparison between single-step-ahead and multiple-step-ahead predictions of the proposed method is also given and discussed. The results show that the proposed novel RVM method is powerful and effective for PEMFC degradation prediction.