Aluminum electrolytic capacitors (AECs) play a crucial role in traction power electronic converters, which are also the most likely to be responsible for breakdowns. Fault prediction for AECs is helpful to realize preventive maintenance and to reduce the cost of the entire system. However, it is restricted by the collected data scale and the period of the deteriorative process of AECs. Thus, this paper takes the advantages of the synthetic minority oversampling technique (SMOTE) that is used to augment the degradation data information and the gate recurrent unit (GRU) that can be suitable for degradation samples to establish a prediction model. An improved particle swarm optimization (IPSO) algorithm is utilized to optimize the hyper parameters of the GRU to promote the feature learning and prediction performance. Thereupon, a prognostics model based on data augmentation and a GRU optimized by IPSO is simulated on a degradation dataset of AECs under an aging test. The results show that the integrated prediction model achieves better accuracy and reliability when compared to some traditional models. Furthermore, the relative error of each prediction point is less than 2.5% for single step and 3.0% for multi-step, respectively.