A Machine Learning Framework for Clustering and Calibration of Roadway Performance Models with Application in the Large-Scale Traffic Assignment

Roadway performance models are one component of travel demand model systems, whose primary purpose is to replicate the congestion effect in traffic assignment. In this sub-model, accurate estimates of roadway travel time (delay), which is sensitive to road traffic volumes, have a principal role in properly assigning trips from origin zones to destination zones to various paths through the road network. One of the main challenges of developing these volume-delay models is the availability of data, which historically was largely lacking. Thanks to emerging data sources, the speed and volume for a wide range of roadway segments in the network are available in excellent temporal and spatial resolution. This study proposes a multi-stage machine-learning-based framework to clean, classify and calibrate roadway performance models of various roadway functional classes in the Greater Toronto and Hamilton Area (GTHA). Recurrent spatial and temporal trends of roadway performance are further investigated, and distinctive patterns are observed for road segments with specific physical attributes.