<p>Vehicle speed prediction is of great significance for the vehicle safety system and energy management system. Due to the influences of various factors on vehicle behaviors, vehicle speed usually has the characteristics of nonlinearity, time variability, and uncertainty, which makes its prediction a challenging problem and gain plenty of research attentions. The speed of a target vehicle is easily directly affected by traffic conditions and the speed of surrounding vehicles. However, existing studies on the vehicle speed prediction failed to consider and utilize the impacts of these two factors fully. And there are also deficiencies in extracting temporal features from the data. In order to solve this problem, this paper proposes a novel hybrid deep learning based vehicle speed prediction model through assembling Temporal Convolution Network (TCN) and Graph Attention Network (GAT). Specifically, in the GAT module, we construct a graph adjacency matrix according to the spatial positional connectivity between the target vehicle and surrounding vehicles to model their interactions, and define a grid rule to extract and represent the traffic status information around the target vehicle. TCN is used to extract features of historical data. Experimental results based on the public datasets, Next Generation Simulation (NGSIM) and the Highway Drone Dataset (HighD), verify that our model outperforms the classical interactive deep learning model in terms of accuracy, and it is proved that the information of surrounding vehicles and surrounding traffic information can promote the prediction accuracy.</p>