Modern intelligent transportation systems require the integration of accurate short-term traffic predictions for the purpose of enhanced network management. The present paper proposes a novel conceptual approach to traffic flow prediction based on the joint consideration of time-lagged volume and occupancy data, and information on the statistical characteristics of the short-term traffic flow evolution. For the purpose of prediction, a genetically optimised modular neural network architecture is developed. Findings indicate that the regimes in critical flow near capacity (where traffic flow exhibits unstable nonlinear evolution) are more difficult to predict when compared to the free-flow and jammed conditions. Moreover, the predictability of the critical-flow regimes decreases sharply in multiple-steps-ahead prediction, whereas the predictability of the regimes where traffic flow exhibits a more stable temporal behaviour is less affected by the expansion of the predictive horizon. Finally, a comparative study shows that the proposed modular neural network represents more accurately the anticipated traffic flow regimes when compared to the modular neural networks trained without the statistical information on the traffic flow evolution. (a).
Abstract