In this study, a novel multi-scale dilated 3D convolution neural network (CNN) was proposed to improve the performance of seizure prediction algorithm. The electroencephalography (EEG) signals were converted into 3D tensors by short time Fourier transform to show the changes in the frequency domain. 3D CNN was applied to extract features from three dimensions, which are time, frequency and channel respectively. Moreover, dilated convolution kernels enlarged the receptive field of the model and helped to obtain more abstract information. Evaluation results indicate that the proposed model outperforms other state-of-the-art models.