In this paper, we propose TSception, a multi-scale convolutional neural network, to learn temporal dynamics and spatial asymmetry from affective electroencephalogram (EEG). TSception consists of dynamic temporal, asymmetric spatial, and high-level fusion Layers, which learn discriminative representations in the time and channel dimensions simultaneously. The dynamic temporal layer consists of multi-scale 1D convolutional kernels whose lengths are related to the sampling rate of the EEG signal, which learns its dynamic temporal and frequency representations. The asymmetric spatial layer takes advantage of the asymmetric neural activations underlying emotional responses, learning the discriminative global and hemisphere representations. The learned spatial representations will be fused by a high-level fusion layer. With robust nested cross-validation settings, the proposed method is evaluated on two publicly available datasets DEAP and AMIGOS. And the performance is compared with prior reported methods such as FBFgMDM, FBTSC, Unsupervised learning, DeepConvNet, ShallowConvNet, and EEGNet. The results indicate that the proposed method significantly (p<0.05) outperforms others in terms of classification accuracy. The proposed methods can be utilized in emotion regulation therapy for emotion recognition in the future. The source code can be found at: https://github.com/deepBrains/TSception-New