Population synthesis for travel demand modelling in Australian capital cities

Abstract

Microsimulation analysis in travel demand modelling provides an important basis to investigate travel behaviour in a spatial context. The disaggregated nature of the model design is well suited to represent complex travel behaviour and simulate spatial interactions. An integral part of building a microsimulation travel demand model is to obtain a comprehensive set of spatial microdata for the entire model region in small geographies. Data at this level of detail have been collected in Australian censuses. However, detailed geocoded microdata is generally restricted in access due to privacy reasons. Population synthesis techniques have been developed as viable alternatives to supplement the lack of completeness in spatial microdata for microsimulation analysis. These techniques are used to generate synthetic microdata that are statistically sound enough for microsimulation while preserving the privacy of the actual population. Several synthetic population generators have been developed for use in microsimulation travel demand models in recent decades. In Australia, the use of population synthesis techniques to create synthetic populations remains a challenging and time-consuming task that often hinders the progress of further development in microsimulation travel demand models. The primary focus of this thesis is to establish a reproducible population synthesis routine for creating synthetic microdata that can be readily fed into an activity-based model to simulate travel activity schedules at household and person level in Australian capital cities. The aim is to ease the process of preparing the necessary microdata for microsimulation and incentivise further development of microsimulation modelling in travel demand. The main approach adopted for the synthesis routine is based on the Iterative Proportional Updates (IPU) algorithm, a modified Iterative Proportional Fitting (IPF) procedure. IPU differs from the standard IPF procedure in that distributions at the household and person level are iteratively fitted simultaneously. The IPU procedure is solely based on a mathematical algorithm. Therefore, the accuracy of generated synthetic population relies on the quality and integrity of input data. In this thesis, two new heuristic procedures were formulated for data treatments before and after IPU using Australian census data. The procedure proposed for data treatment before the synthesis routine ensures the consistency of the input data, whereas the procedure proposed for data treatment after the synthesis routine extends under-synthesised estimates to a near complete synthetic population. In this research study, complete sets of synthetic populations at the household and person level were generated for Greater Sydney, Greater Melbourne and Greater Brisbane in the smallest geographical units available. These generated datasets have been extensively validated and benchmarked against actual aggregate census data to evaluate their representativeness in small geographies. The performance statistics for all three cities have consistently displayed excellent fit with high level of confidence in matching the synthesised to actual data. Multiple experiments have been conducted to test the efficacy and robustness of the IPU algorithm. The treated post-synthesised estimates have also been validated and proven to further increase the accuracy of the synthesised estimates. A case analysis is presented to illustrate the application of the generated synthetic population in policy analysis. In a broader context, this thesis contributes in setting up an efficient and replicable population synthesis routine that can be included into a standard methodological toolbox for transport researchers and for mainstream social scientists to produce synthetic populations using Australian census data. The lack of implementation details or transparency in validations of existing population synthesis procedures often imposes the need to redevelop a new synthesis routine whenever a synthetic population is required for microsimulation analysis. This research intends to alleviate the cumbersome and costly process of building synthetic microdata from scratch by presenting a practical pathway to building synthetic populations for microsimulation analysis.