Spatiotemporal Attention Networks for Traffic Demand Prediction

Travel demand, which include taxi, bus, and bike demand, forecasting the travel demand is an important part of intelligent city intelligent transportation system. Accurate prediction models can help cities pre allocate resources to meet travel demand, reduce energy waste. Travel demand prediction can be summarized as spatiotemporal sequence prediction. For a long time, in the field of spatiotemporal sequence prediction, most of them will emphasize the effective capture and modeling of nonlinear and complex spatiotemporal dependency, which can effectively improve the accuracy of spatiotemporal sequence prediction. At the same time, the demand for multi-step prediction is also increasing. Long-time high-precision prediction can effectively improve the auxiliary role of the model for decision-making. To address these issues, we propose a Spatiotemporal Attention Network (STATTN) with a novel spatiotemporal attention mechanism that capture dependency in time-dimension and spatial-dimension at the same time which is spatiotemporal dependency. In order to learn high-quality representation of spatial points in spatiotemporal sequence units, we adopt dilated temporal 1d convolutional neural networks which has ability to learn representation from data through back propagation. To alleviate the error propagation, we use the generate-style decoder which can generate the output without iteration steps. Through extensive experiments on two prediction tasks, we demonstrate the advantages of STATTN in short-term and long-term prediction scenarios.