Near-Future Traffic Forecasting for Planning-based Traffic Signal Optimisation

Abstract

Accurately forecasting traffic evolution is critical for effective urban traffic management, especially when integrated with operational control. While high-fidelity traffic simulators offer valuable insights, their high computational cost prevents daily use and real-time deployment. Data-driven approaches offer a computationally feasible alternative, at the cost of lower fidelity. This work bridges this gap by exploring the capabilities of near-future traffic forecasting tailored for traffic signal optimisation. We train data-driven surrogate models to predict the impact of various traffic signal configurations on network traffic over short horizons (up to 360 seconds). Our extensive evaluation on real-world data from a UK urban corridor demonstrates remarkable accuracy for multi-horizon predictions. Further, we demonstrate how these forecasts can be integrated into a planning-based traffic signal optimisation framework. We develop two novel, dedicated heuristics that leverage these predictions to guide the search. Our empirical results on real-world data demonstrate substantial improvements: the forecast-informed heuristics consistently improve solution quality while reducing the number of expanded nodes compared to state-of-the-art domain-independent heuristics.