Dynamic Jacobi graph and trend-aware flow attention convolutional network for traffic forecasting

Abstract

Traffic forecasting which has attracted wide attention is the critical issue of the intelligent transportation systems. However, despite its prominent achievements, there are still some challenges due to the complexity of spatial and temporal correlations, which exist in plenty of traffic data. In this paper, a dynamic Jacobi graph and trend-aware flow attention convolutional network (JGFACN) is proposed for traffic forecasting. More specially, a dynamic Jacobi graph convolution network with self-attention is developed for acquiring the spatial correlations among different traffic sensors, in which dynamic adjacency matrix can be derived from the data by dynamic adjacency matrix generator and the spatial correlation among traffic sensors can be minded by Jacobi graph convolution network and self-attention. Secondly, a temporal trend-aware multi-head flow attention mechanism following with a causal dilated convolution is introduced for leveraging the temporal correlation of traffic data effectively. With the help of trend-aware multi-head flow attention, the nonlinear and long-term traffic data can be handled effectively via considering comprehensive information from different representation subspaces and reducing the quadratic complexity. Meanwhile, the participation of a causal dilated convolution can exponentially expand the receptive field. Thirdly, because the nearby spatial-temporal observations are often more correlated, the spatial-temporal order information of traffic data plays an important role in traffic forecasting. Consequently, the rotary temporal position embedding (RTPE) and the learnable spatial positional embedding (LSPE) with graph convolution network are equipped for acquiring the spatial-temporal positional information effectively. Meanwhile, the LSPE with graph convolution network can solve the problem of spatial heterogeneity. At last, the results of extensive experiments on vast amounts of traffic datasets demonstrate the effectiveness of our proposed JGFACN.