Abstract: To cope with the forecast errors caused by non-normal distributions of data in the process of forecasting vehicle speed on highways, this paper proposes a probabilistic forecasting method, which is a hybrid of deep temporal convolutional networks (DeepTCN) and Copula theory. DeepTCN is established first to give a deterministic forecast considering multiple features. Then, an appropriate Copula function is fitted based on the conditional probability distribution of prediction errors according to the obtained forecasts by DeepTCN. Finally, probabilistic results are provided by error compensation. Real traffic data collected on a highway road in Guangzhou, China are utilized to verify the effectiveness of the proposed hybrid method. Data analysis and experimental results show that real-world data reflects significant randomness in traffic flows, with this randomness being relatively mild at both low and high vehicle densities, but more evident at medium densities; DeepTCN demonstrates superior performance in handling long-term time series information compared with various existing methods, making it well-suited for highway vehicle speed forecasting; by incorporating Copula functions, it can compensate for forecast errors caused by data randomness to some extent, further improving forecast accuracy.