<p>Multivariate Time Series Classification (MTSC) is one of the challenging problems in machine learning, which is widely exists in various fields of real life. Existing MTSC methods focus on the global or local features of the time series data while ignoring the correlation and complementary among them. In this paper, we propose an Adaptive Temporal Feature Fusion method (ATFF) for MTSC to explore the intrinsic relationships between the local and global features of time series data. First, the ATFF is proposed to capture not only the short-term dependency features using a Gated Recurrent Units network, but also the global information using a Temporal Convolutional Network at different time scales. Next, we construct an Adaptive Temporal Feature Fusion Module (ATFFM) to effectively fuse the local and global temporal features. Finally, the fusion temporal features are combined with the spatial features extracted by a Residual Network with 10 layers for classification. The experiments on the 26 public datasets from the UEA MTSC archive and the experimental results show that our ATFF achieves the best performance in classification accuracy compared with the state-of-the-art methods, indicating the effectiveness of our proposed ATFFM for MTSC.</p>