Traffic Flow Prediction Model Based on Gated Temporal Attention and Down-Sampling Convolutional Network

Effective road congestion solutions and accurate prediction of traffic flow play a key role in intelligent transportation systems. Due to the complexity of traffic data, the task of traffic flow prediction becomes very important. Existing temporal feature extraction models may not be able to effectively capture long-term dependencies when dealing with temporal data, and also lack the ability to capture temporal correlations at different temporal resolutions. To address the above issues, we present a traffic flow forecasting method called GTA-DSCN, which incorporates gated temporal attention blocks and down-sampled convolutional networks. The gated temporal attention block captures the local context information of the time series through the normalized causal trend attention mechanism and the gating mechanism, while the down-sampled convolutional network captures the temporal correlation at different temporal resolutions through a special structure and an interactive learning strategy. Experimental results on two public transportation network datasets, PEMS04 and PEMS08, show that the GTA-DSCN model outperforms multiple baseline methods in forecasting traffic flow during the following hour. The model shows better performance in evaluation indicators such as MAE, MAPE and RMSE.