This paper discusses conceptual and practical advantages and disadvantages of neural networks and time series methods. The paper summarizes the findings from empirical investigations on motorway traffic. The first one was carried out in the context of the European Communities' GERDIEN project.This investigation involved the development of neural networks for forecasting 5, 15 and 30 minutes ahead for a motorway near Utrecht in the Netherlands. The second study arose under a United Kingdom Science and Engineering Research Council grant. The aim was to compare the performance of neural networks with the Athena forecasting model. This model was developed by INRETS for forecasting 1-2 hours ahead for a major north-south ahead for a major north-south corridor in France. In both cases, neural networks were found to be good predictors, especially when using upstream data. However, the Athena model was at an advantage for the longer forecasting periods.
Abstract