An Attention-Based Spatial-Temporal Traffic Flow Prediction Method with Pattern Similarity Analysis

Road traffic prediction plays a remarkable role in intelligent transportation system managements. Most current research usually considers the two dimensions of time and space dependencies. However, merely considering the upstream and downstream information in the spatial dimension is incapable to capture sufficient spatial features. In this paper, a traffic prediction model using Spatial-Temporal information and Traffic Pattern Similarity information (STTPS) is proposed. It selects observation points with similar traffic pattern to supplement spatial information through similarity analysis and attention mechanism. Simultaneously, many influences of seasonality and super-recent factors are fully considered in the time dimension. Furthermore, a time convolution network and a gated recurrent unit are used for feature extraction. Finally, the time characteristics of variant scales is fused to obtain the prediction result. The proposed prediction is compared with the state-of-art models to demonstrate its capability in both single-step (5 min) and multi-step (1 hour) traffic forecasting.