Short-term inter-urban traffic forecasts using neural networks

Abstract

Back-propagation neural networks were trained to make short-term forecasts of traffic flow, speed and occupancy in the Utrecht/Rotterdam/Hague region of The Netherlands. A problem which had to be faced when designing the system was the vast number of possible input parameters. Whilst neural networks which utilised all available inputs performed well, their size made them impractical for implementation. A technique of stepwise reduction of network size was developed by elasticity testing the large neural networks, showing a way of overcoming this difficulty. Results for occupancy and flow forecasts by this method show some promise, but do not out-perform naive predictors. Forecasts of vehicle speed were much less successful, perhaps because of the distorting effect of slow moving vehicles, particularly in low flow conditions. The elasticity tests were found to be useful, not only as a means of enabling network size reduction, but as a means of interpreting the neural network model.