Uncertainty-Aware Coordinated Multi-Ramp Metering Based on an Action-Conditioned Traffic Prediction Operator Under Sparse Observations

Abstract

Coordinated multi-ramp metering (CRM) is an important approach to mitigating freeway congestion, but sparse observations in practice complicate the state estimation and model calibration required by predictive control. This study develops an Action-Conditioned Traffic Prediction Operator (ATPO) driven by sparse, heterogeneous observations and candidate multi-ramp actions, and embeds it in an uncertainty-aware CRM framework (ATPOUCRM). ATPO forms a four-channel history from the density and speed measurements of fixed detectors and probe vehicles and their separate observation masks. Queries at future spatiotemporal points attend to all valid historical observations before incorporating upstream mainline demand and on-ramp queue information. Candidate metering actions condition the latent field through feature-wise linear modulation, a spatial Fourier operator, and a causal temporal kernel. The operator produces action-conditioned predictive distributions of future density and speed throughout the corridor; a bootstrap ensemble additionally characterizes disagreement among models. Using these predictions, ATPOUCRM primarily minimizes expected total time spent, including mainline travel and on-ramp waiting, while penalizing predictive uncertainty near merge areas. It uses the cross-entropy method to search candidate actions and applies a safety filter before execution. For five-minute global traffic state prediction on a simulated 7-km freeway corridor, ATPO achieves root mean square errors of 2.54 veh/km/lane, 4.56 km/h, and 0.76 veh*h for density, speed, and total time spent over the forecast horizon, respectively, outperforming the comparison methods. Across 100 control scenarios, ATPOUCRM reduces mean total time spent by 26.17% relative to no control. ATPO also generates predictions on finer spatiotemporal grids without retraining.