A novel algorithm for parametric calibration of speed-density relationships in mesoscopic traffic simulator

Speed-density model is one of the relationships used by mesoscopic traffic simulator to represent traffic dynamics. While the classical speed density relationships provide useful insight into the traffic dynamics problem and have a theoretical value to traffic flow, for such applications they are limited. This paper focuses on calibrating parameters for speed-density relationships with machine learning methods, and introduces a new algorithm called ldquoclustering-locally weighted regressionrdquo. In order to improve the precision of parametric calibration, we also preprocess sensor data, including finding missing sensor data, detecting error data, and repairing both of them. Finally, so as to fuse more information into the process of calibration, this paper utilizes densities and flows as variables to calibrate parameters for the speed-density relationships. The proposed approaches are tested with sensor data from the 3^rd ring road in Beijing. The testing results show that the proposed algorithms have great performance on the parametric calibration and are appropriate for the simulation-based DTA models.