Accurate seizure prediction is important for design of wearable and implantable devices that can improve the lives of subjects with epilepsy. Such implantable devices can be used for closed-loop neuromodulation. However, there are many challenges that inhibit the performance of prediction models. One challenge in accurately predicting seizures is the nonstationarity of the EEG signals. This paper presents a patient-specific deep learning approach to improve predictive performance by transforming EEG data before extracting features for seizure prediction. In the proposed approach, a Sequence Transformer Network (STN) is first used to learn temporal and magnitude invariances in EEG data. The proposed method further computes the short-time Fourier transform (STFT) of the transformed EEG signals as input features to a convolutional neural network (CNN). A k-out-of-n post-processing method is used to reduce the significance of isolated false positives. The approach is tested using intracranial EEG from the American Epilepsy Society Seizure Prediction Challenge dataset. Leave-one-out cross validation is used to evaluate the model. The proposed model achieves an overall sensitivity of 82%, false prediction rate of 0.38/hour, and average AUC of 0.746.