A Predictive–Prescriptive Analytics Framework forRoute Planning via Discrete Time–SpaceGraph and ST-GCN-GRU Model

Abstract

Route planning in dynamic traffic environments requires decisions that anticipate future traffic evolution rather than rely on time-invariant costs. This study proposes a predictive–prescriptive analytics framework for time-dependent route planning that integrates short-term traffic speed forecasting with optimization on a first-in–first-out–consistent discrete time–space network. The predictive module employs a spatiotemporal graph convolutional network integrated with gated recurrent units (ST-GCN-GRU) to forecast multihorizon link speeds from historical observations, while the prescriptive module embeds these forecasts into a discrete time–space representation that transforms the time-dependent shortest-path problem into a polynomially solvable form. The framework is evaluated in two complementary ways. First, a real-world end-to-end experiment on an OpenStreetMap-derived Nanjing subnetwork tests whether prediction-informed routing improves downstream decisions under a unified replay protocol. Second, controlled experiments on the Sioux Falls benchmark are used for mechanism exploration, examining when the value of future-aware routing becomes more pronounced. Results show that the proposed framework improves traffic forecasting accuracy, supports near-oracle routing performance in realistic settings, and clarifies how trip length and temporal variability influence routing gains.