Activity-based Demand Modelling on a Large Scale: Experience from the New Danish National Model
J Rich, C Giacomo Prato, DTU Transport, DK; A Daly, RAND Europe, UK
This paper presents an activity-based approach to demand modelling within the framework of the new Danish national model.
In 2009, DTU Transport was commissioned by the Danish Ministry of Transport to develop a national transport model for Denmark. An analysis of the specification required by potential users revealed that the objectives for producing forecasts with the model were very diverse, from planning transport infrastructure to designing traffic management measures, from proposing policy interventions to anticipating demand patterns. The wide variety of these objectives calls on the one hand for the model to be flexible and detailed, and introduces on the other hand great challenges in terms of applicability and run-time. This paper outlines an activity-based model framework and, in particular, emphasises the challenges related to the large-scale implementation. As the model will operate on a zone-system with more than 3,500 zones and will allow individuals to be represented at a precision level that notionally corresponds to approximately 250,000 different socio-economic groups in Denmark, these challenges are noteworthy.
Initially, this paper provides a detailed outline of the proposed model framework, which at an upper level involves decomposition between long-term strategic choices and short-term daily choices. The long-term model addresses strategic choices such as residential location, workplace location and car ownership. The short-term model addresses daily travel behaviour in terms of car availability, mode, time-of-day and destination, as well as choice of an activity sequence for the given day. As the choice of activities involves selecting between different sequences of activities and accordingly of trips, the trip chaining issue is also (indirectly) addressed.
Then, the paper enters into the details of the system of discrete choice models to be estimated on the basis of RP data collected within the national travel survey (TU data), combined with SP data to better reflect time-of-day choice. As the TU data only contain information about single individuals, choice model estimation concerns only these individuals. However, also the contribution of other individuals in their households needs to be considered from a demand evaluation perspective. This issue is particularly relevant when addressing car availability in households with more than one adult and only a single car available. A suggestion is to calculate the utility functions for household members whose trips are not directly observed on the basis of parameters estimated for the observed segment and attributes and choice sets for the non-observed segment. The problems and limitations of this approach are discussed.
Last, this paper focuses on issues in the forecasting methodology related to population generation and to the complex trip chains considered in the model. As the model adopts a prototypical sample enumeration approach, the expansion of the micro-data is required to be consistent with future targets. Expansion factors are calculated as the ratio between a population matrix and a survey-sample matrix. The construction of the prototypical population matrix applies the well-known and commonly used iterative proportional fitting (IPF) algorithm. However, as the population matrix extends over a large number of dimensions, non-trivial issues arise. Firstly, an equivalent linear-programming problem is proposed to make the constraints of the IPF internally consistent. Secondly, a flexible aggregation procedure is introduced to eliminate the existence of structural zeros.
As the model proposes an activity-based approach, the representation of complex trip chains is involved. This complexity introduces a challenge in the representation of origin-destination matrices for a large number of zones and multiple destinations. In fact, 3,500 zones and three destinations imply that more than four billion entries are needed to represent a single attribute, causing the trip chain representation to be unmanageable. A ?semi-sparse? approach in which matrices are decomposed into two types is proposed. The first type represents up to two-dimensional trip chains (production-attraction matrices). These matrices are ?complete? in that all entries are represented, and are manageable from the computational perspective because they involve only one origin and destination at a time. The second type represents trip chains with more than a single main destination. These matrices are ?sparse? in the sense that only non-zero elements are represented. As the first type of matrices accounts for approximately 70-80% of the total demand, the problem of structural zeros is largely eliminated. Moreover, as the complex trip chains are consistently represented, these trips will be correctly assigned to the network.
Association for European Transport