Abstract

Thesis (PhD)--University of South Australia, 2012. Includes bibliographical references. Trip distribution forecasting has been undertaken by using the gravity model for more than four decades. Literature review suggests that traffic forecasts can be overestimated up to 60 per cent above the actual numbers. Underestimated traffic forecasts are also found, with an average of 21.3 per cent below the actual numbers. Thus, a question emerges: Has traffic forecasting become more accurate over the last four decades? In fact the literature indicates that traffic forecasts have not improved over the last 30-years. Thus, a new, transparent and accountable forecasting method is recommended in order to prevent more severe financial risk and perhaps more damage to social and economic welfare. This research proposes a trip distribution modelling framework using the neural network approach. Although neural networks have been used in trip distribution modelling for more than two decades, the forecasting capability is not yet optimally explored and understood. This is due to its performance being dependent on multiple properties. It requires working in an integrated system in solving various tasks, including the forecasting task. This research focuses on the development of that integrated system by considering the internal and external factors related to the neural network approach and the characteristics of fully constrained spatial distribution models. Therefore, the capability of neural network approach in forecasting spatial distribution can be optimized. The gravity model forecasts the trip distribution numbers by using a specific deterrence function and also, on occasions, by incorporating socioeconomic adjustment factors (K-factor). It theoretically assumes that the trip number increases when the deterrence variable decreases, and vice versa. Then, the K-factor is incorporated in the gravity model so that it generates more accurate results. Reliance on these functions and factors is considered as one of the weaknesses of the gravity model. Firstly, the assumption related to the deterrence variable is not always true. Secondly, the mechanism of K-factor is unclear and it is difficult to estimate its future values. Thirdly, collecting the data for estimating the K-factor is difficult, time consuming, and costly. Meanwhile, the neural network approach for spatial distribution forecasting works by capturing the relationship between independent and dependent variables without reliance on any function representing the relationship between these variables. It also works without requiring additional data representing the socioeconomic adjustment factor. It forecasts the trip distribution numbers through a learning process and generalizes the trip numbers based on that relationship. Therefore, more accurate estimation of trip numbers is expected. This is demonstrated by application of the proposed framework. The calibration and generalization performance of the neural model for passenger travel and commodity flow are found to be superior to the doubly constrained gravity model. About 88 per cent of the neural models have lower error and higher goodness-of-fit than the gravity model. About 75 per cent of the neural models significantly outperform the gravity model for passenger trip distribution forecasting, while this is about 63 per cent for commodity flow distribution. Another benefit of the proposed framework is it can be used for any number of traffic analysis zones.