A self-attention dynamic graph convolution network model for traffic flow prediction

Precise and reliable traffic predictions play a vital role in contemporary traffic management, particularly within complex traffic networks. Currently, the approach which utilizes static graph convolution with recurrent neural networks for traffic flow prediction fails in digging the complicated spatial and dynamic temporal correlations deeply. As a solution, we propose a dynamic graph convolution network model, aimed at enhancing traffic forecasting. Within our model, we first utilize the dynamic Laplace matrix, generated by assessing the similarity of time series changes, to facilitate graph convolution, enabling a more in-depth exploration of spatial correlations among nodes. Then, we capture local time correlations through convolutional neural networks and address long-term correlations using the self-attention mechanism. To mitigate the over-smoothing challenge related to graph fusion and deep networks, we introduce the autoregressive module to enhance the predictperformance for aperiodic sequences. Thereby, we contemplate not just the dynamic relationship between multiple time series but also the diverse facets of long-term, short-term, periodic, andnon-periodic changes inherent in time series. Experimental findings corroborate the superior performance of our proposed model compared to all benchmark models