Robust Self-Attention ConvLSTM-based Traffic Flow Prediction Model

Traffic flow forecasting has been receiving a lot of attention because of its important role in traffic control and management. Accurate traffic forecasting is critical to improving the performance of intelligent transportation systems. However, accurate traffic forecasting still faces the following challenges, including modeling the dynamics of traffic data along the temporal and spatial dimensions, significant differences in peak hour/peak hour traffic, and traffic flow data affected by partial noise. In this paper, we propose a hybrid and robust model with Self-Attention ConvLSTM networks and localized stochastic sensitive (LSS) for traffic flow prediction. The proposed model extracts features with long-range spatiotemporal dependencies with Self-Attention ConvLSTM. To further explore the long-term temporal features, we utilize LSTM module to extract daily and weekly periodic features as assistive features. The LSS reduces sensitivity to unseen samples around training samples and avoids large output fluctuations due to the noise or change of the data. Experiments on real traffic flow datasets show that the proposed method yields better prediction performance compared to other contrast methods.