A Fast Spatial-temporal Information Compression algorithm for online real-time forecasting of traffic flow with complex nonlinear patterns

Traffic flow usually contains complex nonlinear patterns. Deep learning can model nonlinear fluctuations through iterative updates of trainable parameters. It generally requires a large computational cost and may not apply to online real-time traffic flow forecasting tasks. Compared with offline forecasting, online real-time forecasting can provide more real-time and accurate traffic information, which is important to help reduce traffic accidents and improve the real-time decision-making ability of traffic management authorities. Current research has not adequately addressed the issue of online real-time traffic flow forecasting. Therefore, it is crucial to discuss the balance between accuracy and computational cost. A Fast Spatial-temporal Information Compression (FSTIC) algorithm is proposed for online real-time traffic flow forecasting. Experimental results show that Time Step Screening and Tucker Decomposition can compress spatial-temporal information. Besides, the Tensor Kernel Ridge Regression in the FSTIC algorithm can model nonlinear small sample data with high accuracy and low computational cost. In comparison to baselines, FSTIC reduces MAE, RMSE, and computational cost by an average of 41.66 %, 35.40 %, and 96.63 %, respectively.