Multi-step trend aware graph neural network for traffic flow forecasting

Abstract

Traffic flow prediction plays an important role in smart cities. Although many neural network models already existed that can predict traffic flow, in the face of complex spatio-temporal data, these models still have some shortcomings. Firstly, they although take into account local spatio-temporal relations, ignore global information, leading to inability to capture global trend. Secondly, most models although construct spatio-temporal graphs for convolution, ignore the dynamic characteristics of spatio-temporal graphs, leading to the inability to capture local fluctuation. Finally, the current popular models need to take a lot of training time to obtain better prediction results, resulting in higher computing cost. To this end, we propose a new model: Multi-Step Trend Aware Graph Neural Network (MSTAGNN), which considers the influence of global spatio-temporal information and captures the dynamic characteristics of spatio-temporal graph. It can not only accurately capture local fluctuation, but also extract global trend and dramatically reduce computing cost. The experimental results showed that our proposed model achieved optimal results compared to baseline. Among them, mean absolute error (MAE) was reduced by 6.25% and the total training time was reduced by 79% on the PEMSD8 dataset. The source codes are available at: https://github.com/Vitalitypi/MSTAGNN.