Proton exchange membrane fuel cells (PEMFCs) offer power generation capabilities for diverse applications including commercial enterprises, industrial sectors, and residential technologies. Nevertheless, the comprehensive integration of PEMFC applications could be improved by challenges related to degradation and durability. The imperative development of efficient performance prognostic models assumes a pivotal role in the prognosis of remaining useful life (RUL), health monitoring, and effective utilization of PEMFCs. This paper centers on the prognostication of critical components within PEMFCs and introduces a transfer learning approach based on variational autoencoder and bi-directional long short-term memory with an attention mechanism (Bi-LSTM-AM) model. This approach combines feature fusion, knee-point detection, and a sophisticated deep-learning-based predictive model. Notably, incorporating the variational autoencoder as the framework for feature fusion introduces a novel perspective previously unexplored. Identifying the knee point and knowing the start point on the training data, facilitates optimized parameter computation. The application of transfer learning facilitates the transfer of optimal model parameters and weights from a source to a target dataset. Conclusively, the estimation of stack voltage degradation and real-time RUL prediction based on the test dataset is executed by implementing our proposed method. The stack voltage prediction findings showcase the Bi-LSTM-AM model’s superior performance relative to comparison models. The proposed online rolling prediction model, utilizing a sliding window technique for RUL prediction, yields significantly enhanced accuracy, culminating in a relative error margin ranging from approximately 1.69% to 5.04%.