This paper describes work carried out to investigate a possible method of forecasting traffic parameters into the short-term future. The work was undertaken as part of the European Communities' DRIVE II GERDIEN project. The overall goal of this project is to specify the requirements of a dynamic traffic monitoring system. Data collected from a test site on the Dutch A12 motorway formed the basis for the work. Back-propagation neural networks were identified as a non-linear forecasting tool which might be useful. A particular problem which had to be faced when designing the forecasting system was the sheer number of possible input parameters. Whilst neural networks utilised all possible inputs possible inputs performed well, their size made them impractical from actual implementation. A technique of stepwise reduction of network size by performing elasticity tests on the large neural networks produced a novel way of overcoming this difficulty. Results for occupancy and flow forecasts by this method are found to be promising. Forecasts of vehicle speed were much less successful. The results from the elasticity tests have a useful function apart from enabling network size reduction, as a great many useful traffic engineering inferences can be drawn from them.
Abstract