Calibration of fundamental diagrams for travel time predictions based on the cell transmission model

Abstract: Road traffic increases constantly and the negative consequences in the form of traffic jams can be realized especially in urban areas. In order to provide real time traffic information to road users and traffic managers, accurate computer models gain relevance. A software called Mobile Millennium Stockholm (MMS) was developed to estimate and predict travel times and has been implemented on a 7km test stretch in the north of Stockholm. The core of the software is the cell transmission model (CTM) which is a macroscopic traffic flow model based on aggregated speed observations. This thesis focuses on different calibration techniques of the so called fundamental diagram as an important input factor to the CTM. The diagrams illustrate the mathematical function which defines the relation between traffic flow, density and speed. The calibration is performed in different scenarios based on the least square (LS) and total least square (TLS) error minimization. Furthermore, sources, representing the traffic demand, and sinks, representing the surrounding of the modeled network, are implemented as dynamic parameters to model the change in traffic behavior throughout the day. Split ratios, as a representation of the drivers‘ route choice in the CTM are estimated and implemented as well. For the framework of this work, the MMS software is run in a pure prediction mode. The CTM is based on the source, sink, split and fundamental diagram parameters only and run forward in time. For each fundamental diagram calibration scenario an independent model run is performed. The evaluation of the scenarios is based on the output of the model. The results are compared to existing Bluetooth travel time measurements for the test stretch, which are used as ground truth observations, and a mean average percentage error (MAPE) is calculated. This leads to a most reasonable technique for the fundamental diagram calibration – the total least square error minimization.