Freight Choice Model for Mode and Crossing: A Forecast Model for the ├śresund Region
J Rich, P M Holmblad, C O Hansen, DTU Transport, Technical University of Denmark, DK
The paper describes the estimation and application of a freight forecast model for choice of mode and crossing. The model is validated statistically by reviewing the elasticity structure and the value-of-time across commodity groups and modes.
Intro: The paper describes the estimation and application of a freight forecast model for choice of mode and crossing. The model is estimated as an aggregate weighted discrete choice model, in which OD matrices and Level-Of-Service (LOS) matrices are applied directly to specify utility functions for the different choice alternatives. The model is estimated and applied for 13 commodity groups with very different mode-choice patterns, spanning from strict bulk commodities to high-value commodities. In the model, utility functions are estimated only from the mode-choice, which consist of five modes; Truck, Combi-Rail, Combi-Ship, Rail, and Ship. The traffic loads on crossings are calibrated subsequently so that the model replicates base matrices according to modal split and the crossing pattern. The zone structure of the model covers most of Europe, although the geographical focal point of the model is the ├śresund region. As a result, the spatial resolution level is intentionally fine-grained around the ├śresund region and less so in the periphery of the model area.
The model is validated statistically by reviewing the elasticity structure and the value-of-time across commodity groups and modes. Generally, the result of the model are sound and in good accordance with reference literature. Moreover, the model is used to simulate a sequence of reference scenarios in order to test the applicability of the model in more general.
Model formulation: Generally, all of the models has been specified in a log-space formulation, e.g. LOS variables related to time as well as cost are transformed on a logarithmic scale. There have been two major reasons for this. Firstly, the log-space has generally performed better in terms of goodness-of-fit. Secondly, by the log-space formulation, we avoid scale-dependency in the derived demand responses. In the log-space, we obtain a direct proportional relationship between demand and supply. It should be remembered that, although the linear-in-variable utility specification, cause the mapping between variables and utility to be linear, it causes the mapping between variables and demand to be exponentially scaled. This form may be fully valid for passenger models and especially for short-range modes, however, for freight models less so. Another reason to avoid scale-dependence is that the longer travel time, the more uncertain the supply data. Of course, this may be true for passenger models as well; however, it is especially true for freight. The problem is that the underlying assignment model does not include a ?logistic module?. In other words we use relative rigid standards in the calculation of waiting time, terminal time, and time for re-loading. These time components are generally speaking ?fixed constants?, which on longer trips may accumulate to days and even weeks in some extreme cases.
Estimation: The model estimation applies a two-stage estimation strategy. The problem is that for most commodity groups, the bulk-modes (rail and ship) are very different from the other modes: The cost is generally much lower, the travel time much higher, and so are the volumes on particular links. Consequently, it has been difficult to estimate the effect of different time components (e.g. short vs. long trips and waiting time vs. in-vehicle-time) in a full joint estimation with all modes included. The problem is that for non-bulk modes, time is decomposed into vehicle time and waiting time, whereas for bulk-modes there are only a single time and cost component. It means that estimating time and cost for all modes, will seriously affect the scaling between, e.g. waiting time and in-vehicle time. A solution to this problem, would be to estimate an individual scaling of time for bulk modes, however, this is not possible due to strong correlated between LOS variables. Another way to overcome a) the scaling and b) the correlation problem is to implement a two-stage estimation procedure. In the first stage only non-bulk modes are considered. It enables the estimation of the different time components for truck and combi modes. In the second stage, the scaling of the different time components from stage 1 is fixed. The purpose is to estimate an overall scaling of time and cost.
Results: Due to the two-stage estimation, it has been possible to overcome correlation problems and to estimate different time components, e.g. waiting time, in-vehicle time and divided into different distance bands. Generally, the results of the model considering demand sensitivity and the value-of-time, conforms well to the reference literature, and suggest that aggregate choice estimation is an option when disaggregate data is not available.
Association for European Transport