Transformer-enhanced periodic temporal convolution network for long short-term traffic flow forecasting

Recently, Temporal Convolution Networks(TCNs) and Graph Convolution Network(GCN) have been developed for traffic forecasting and obtained promising results as their capability of modeling the spatial and temporal correlations of traffic data. However, few of existing studies are satisfied with both long and short-term prediction tasks. Recent research has shown the superiority of transformer in handling long-range time series forecasting problems. Aimed at the shortcoming of existing solutions, in this paper, we propose a novel Transformer-enhanced Temporal Convolution Network(TE-TCN) to capture spatial, long and short-term periodical dependencies to improve the accuracy of traffic flow forecasting, especially for long-term prediction. TE-TCN integrates transformer multi-head attention mechanism and GRU to discover the long-term periodic patterns. Meanwhile, two paralleled temporal convolution networks are applied to solve the short-term periodic dependencies. The proposed method is evaluated by extensive traffic forecasting experiments on four real-world datasets and the experimental results demonstrate that TE-TCN outperforms the state-of-the-art related methods, especially for long-term traffic flow forecasting.