Road network free flow speed estimation using microscopic simulation and point-to-point travel times

Abstract

Roads in mesoscopic traffic simulation models are often parametrized by a free speed attribute, which describes the speed at which a vehicle can traverse a road in the absence of congestion. Unlike microscopic models, mesoscopic models do not typically simulate traffic lights, individual driver behavior or acceleration and deceleration of vehicles. In consequence, the free speed parameter of each road must account for all the aforementioned effects. Modeling free speed accurately is important, as it determines the travel time, which in turn affects the decision-making of agents in the simulation.This paper introduces an approach combining microscopic simulation with machine learning models to estimate road segment free speeds. Using the microscopic simulator SUMO, we generate training data for a model search employing Bayesian optimization. Subsequently, these models undergo fine-tuning via gradient-based optimization using real-world point-to-point travel times. A significant advantage is the adaptability of this approach to diverse road networks, including those from OpenStreetMap, and the ability to incorporate routing data from various online providers independently.Our evaluation illustrates a notable decrease in prediction error between 30% and 60% compared to baseline models that assume a uniform free speed reduction for all urban roads. The fine-tuned models are able to generalize well to unseen regions and are therefore applicable to case studies where data for new or altered roads is unavailable.