Epilepsy is one of the most common neurological disorders and affects 1% of the world's population. Early prediction of the onset time of the seizure can enable responsive treatments and greatly improve the quality of life of epileptic patients. Although seizure detection has been extensively studied, reliable prediction of the onset of seizures remains a challenge. Despite recent advances in deep learning, most existing methods require a large amount of data to train the models, which are often not available for privacy, proprietary, or cost reasons. In this work, we show that the performance of seizure prediction can be improved by using data augmentation from only a limited amount of EEG recordings. We developed a generative adversarial network (GAN) and a refiner based on a publicly available CHB-MIT dataset. We trained a classifier using only synthesized data, but tested the classifier using real data from all 23 patients in the dataset. With a seizure prediction horizon (SPH) of 5 min and a seizure occurrence period (SOP) of 30 min, the classifier achieved an average AUC score of 61%. Improvement in the use of the refiner is 2.1% on average in all patients. The proposed method holds great promise for alleviating the problem of data scarcity and imbalance in seizure prediction and many other applications.