Day-to-Day Learning Framework for Online Origin–Destination Demand Estimation and Network State Prediction

Origin–destination (O–D) demand is a critical component in both online and offline dynamic traffic assignment (DTA) systems. Recent advances in real-time DTA applications in large networks call for robust and efficient methodologies for online O–D demand estimation and prediction. This study presents a day-to-day learning framework for a priori O–D demand, along with a predictive data-driven O–D correction approach for online consistency between predicted and observed (sensor) values. When deviations between simulation and real world are observed, a consistency-checking module initiates O–D demand correction for the given prediction horizon. Two predictive correction methods are suggested: 1) simple gradient method, and 2) Taylor approximation method. New O–D demand matrices, corrected for 24 simulation hours by the correction module, are used as the updated a priori demand for the next day simulation. The methodology is tested in a real-world network, Kansas City, MO, for a 3-day period. Actual tests in real-world networks of online DTA systems have been very limited in the literature and in actual practice. The test results are analyzed in time and space dimensions. The overall performance of observed links is assessed. To measure the impact of O–D correction and daily O–D updates, traffic prediction performance with the new modules is compared with the base case. Predictive O–D correction improves prediction performance in a long prediction window. Also, daily updated O–D demand provides better initial states for traffic prediction, enhancing prediction in short prediction windows. The two modules collectively improve traffic prediction performance of the real-time DTA system.