A sparrow search algorithm was proposes for optimizing gated recurrent units for the remaining useful life of PEMFC (proton exchange membrane fuel cell) with a maximum rated power of 110 kW. More than 600 hours of dynamic cycle endurance tests were conducted to simulate the operation of on-board PEMFC under different road conditions. To accurately predict the remaining useful life of high-power PEMFC, it is necessary to consider its output voltage under different operating states, and classify and predict the output voltage according to different power points. Firstly, the sampling array was filtered to reduce peak values and smooth noise reduction. Then, based on data-driven methods, voltage data at different operating states and different training set partitions were used as inputs. The accuracy of this model was confirmed by the selected evaluation indicators and different common time series regression algorithms. The experiment took 60% of the data as the training set as an example, the prediction results of SSA-GRU compared to TCN showed a decrease of 0.110 5%, 0.525 7%, 0.308 4%, 0.402 1%, and 0.831 9% in MAPE at 30, 50, 70, 90, and 110 kW, respectively. At the specified deadline for useful life, the minimum prediction error for useful life is only 0.733%, and the prediction error under different working conditions is superior to other prediction algorithms.