In order to achieve a fast and accurate prediction for remaining useful life (RUL) of proton exchange membrane fuel cells (PEMFCs), an improved degradation model named RCLMA is developed which integrates the residual convolutional blocks, the long short-term memory (LSTM) units and the multi-head self-attention layers. Firstly, the stack voltages of two PEMFCs are collected, reconstructed and smoothed. Secondly, the RCLMA model is constructed and validated based on the experimental data, and the effects of input size and output size on the prediction performance are studied. Finally, ablation experiments are conducted to investigate the role of each part of the RCLMA model in the prediction. The results show that the RCLMA model reaches the best prediction performance with a root mean square error (RMSE) of 0.01785 and a ScoreRUL of 0.994589 when the input size is 200 and the output size is 30, and the prediction time does not exceed 5 s. All three parts of the RCLMA model make important contributions to the prediction performance, where the residual part is partial to finding the overall trend of the data, the recurrent part is inclined to obtain the time dependence and the self-attention part plays a good capture of both.