Coupling TS and ML Models for Short term Traffic Flow Prediction in ITS Supports

TS_ML models were developed by applying the TS model to capture the linear and temporal structures in the traffic flow series, after which the resulting residuals were modelled using ML algorithms to account for the remaining non-linear patterns. Conversely, the ML-TS architecture initially employed ML models to learn the dominant non-linear relationships, while the residuals generated from the ML predictions were subsequently modelled using TS approaches to capture any remaining temporal dependencies and autocorrelation structures.

Across all the road corridors, the most commonly selected architectures were Prophet_GBM (27.5%) and RF_Prophet (17.5%). Thirty-one hybrid-selected series achieved a mean NSE of 0.966 and mean KGE of 0.971. Forecasts from these models is important for the support of real-time ITS applications such as traffic congestion management, and incident response planning.