The automobile industry is switching from fossil-fuel-based energy sources to green energy sources. In particular, because hydrogen has a high energy density and efficiency compared to other energy sources, it is a major research field for commercial vehicles that require large amounts of energy and travel long distances on average, such as buses and trucks. In fuel-cell electric vehicles (FCEVs), maintaining an adequate energy production intensity is necessary to improve energy production efficiency and prevent falling into a state of inoperability owing to a lack of energy. In this case, energy prediction becomes a significant factor. In this study, a deep-learning-based prediction method for FCEV powertrain energy is proposed. The proposed method uses only internal data, which can be obtained from the vehicle, and does not require external information regarding future routes. Additionally, we designed a model considering the vehicle data time-series characteristics and proposed a shift mixup, which is a data augmentation method that does not compromise the vehicle’s dynamic characteristics, to address the data shortage problem. Furthermore, a pretext task-learning method that can improve model performance without external data during inference is introduced. This method includes pretext tasks designed specifically for the vehicle domain. Finally, a distance-based loss mask (DBLM) and contrastive learning that the representation can learn semantic information are proposed. Experimental results for the actual driving dataset show that our method improved by 28.05% and by 30.04% compared to the exponential moving average (EMA) in terms of root mean squared error (MSE) and mean absolute error (MAE), respectively. We demonstrate with energy management strategy (EMS) that an effective energy prediction algorithm helps sustain an optimal state of charge (SOC).