Exploiting dynamic spatio-temporal correlations for citywide traffic flow prediction using attention based neural networks

Abstract For intelligent transportation systems (ITS), predicting urban traffic crowd flows is of great importance. However, it is challenging to represent various complex spatial relationships across distinct regions, as well as dynamic temporal relations among various time periods. To overcome this challenge, we propose DHSTNet, a novel deep Spatio-temporal neural network for predicting traffic crowd flows. Our proposed approach consists of four components: (i) the closeness component considers spontaneous changes of the traffic crowd flows; (ii) the period influence component frequently characterizes variations of daily flows; (iii) the weekly influence component characterizes the weekly arrangements of crowd flows; and (iv) the external branch component identifies various external influences. Our model applies diverse weights to individual branches. Then, it fuses the outcomes of the four features. Extensive experiments on two real-world datasets demonstrate the advantage of the proposed model over the compared baseline methods. Moreover, to verify the generalization of the proposed model, we apply the proposed attention-based mechanism with a previously proposed model, resulting in a hybrid approach known as Att-DHSTNet, to forecast short-term crowd flows. Experimental results also confirm improved performance.