A Multiple SVR Approach with Time Lags for Traffic Flow Prediction

A multiple support vector regression (SVR) model with time lags was proposed for short term traffic flow prediction. Time lags between current traffic flow and upstream traffic flow were estimated in order to make better use of spatial-temporal correlation between the upstream and the downstream. The time lags could help identify the upstream flow series most similar to that of the current road and to be used as the model input. A global SVR model with a time lag was constructed and we found it performed not so well during some time intervals where the traffic flow was dramatically fluctuant. Local SVR models with time lags were constructed especially for those intervals and improved the performance. Combining both of the global and the local models, the multiple model was applied to 5-minute freeway data observed by loop detectors in the project freeway performance measurement system (PeMS) of California. Comparisons with several other methods showed that the multiple SVR model with time lags was a promising and effective approach for traffic flow prediction.