Development of travel time estimation models: consideration of link geometry for Korean motorways

Traffic assignment in transport appraisal is an important procedure that allocates future origindestination trip demand to every route. A link cost function, which is also called a volume-delay function (VDF), has been used to predict travel time consumed by traffic demand passing through each link in traffic assignment. In order to cover various road characteristics such as the number of lanes, road types, different parameters in VDF have been proposed commonly with road capacity and free-flow travel time. VDF has the advantage in that it is simple enough to analyse the entire network, explaining the relationship mainly with traffic flow. However, VDF has two drawbacks: uncertainty in road capacity and difficulty in spatial transferability. The former includes the conceptual uncertainty that there is little consensus of the definition and the measurement uncertainty of which traffic flow is chosen for road capacity (e.g. the highest or 95 percentile one, etc.). The other drawback is that current VDFs cannot account for change in link geometry and as such are not spatially transferable. Because of the two drawbacks, current VDFs could result in inaccurate traffic assignment; and hence cause inappropriate provision of road space. In order to overcome the limitations, this study implements the empirical analysis of 72 Korean motorway links by quantifying the dependent variable of link travel time and the independent variables of traffic flow and link geometric features. The dataset was collected from intelligent transport systems and road design drawings. Fixed effects modelling by least squares dummy variables identified influential factors on travel time. In order to develop feasible travel time estimation models, three statistical methods were introduced as follows: firstly, this study introduces linear statistical estimations, which are ordinary least squares (OLS) and generalised least squares (GLS) estimated by likelihood maximisation. In the modelling process, strict statistical assumptions of the OLS estimation are tested and different variance-covariance structures in the GLS estimation are scrutinised to deal with statistical violations such as heteroscedasticity and serial correlation. Secondly, nonlinear least squares (NLS) estimation, which is widely used for VDF customisation, is applied by combining link geometric variables with an existing model. In order to clarify the uncertainty of road capacity, sensitivity analysis using different road capacity values shows the impact on NLS estimated models. Lastly, the most appropriate model is selected by the comparison with statistical accuracy measures after a 10-fold cross-validation with the application to practical traffic assignment and transport appraisal. In conclusion, this study develops new types of travel time estimation models that include link geometric variables by testing many statistical approaches. The results suggest that not only traffic flow in existing models but also many influential factors such as weather, brightness and link attributes can affect travel time. In particular, some link geometric variables of upgrade, downgrade and tunnel ratios are statistically significant as explanatory variables in the models. In addition, it is worth noting that it is possible to develop a new type of VDF with link geometry instead of road capacity. The statistical significance of the developed models and their application to transport planning demonstrate that the selected model can replace existing VDFs in traffic assignment