The principal aim of this study was to develop a method for making a short-term prediction model of traffic flow status (i.e. travel time and a five-step travel-speed-based classification) and test its performance in the real world environment. Specifically, the objective was to find a method that can predict the traffic flow status on a satisfactory level, can be implemented without long delays and is practical for real-time use also in the long term. A sequence of studies shows the development process from offline models with perfect data to online models with field data. Models were based on MLP neural networks and self-organising maps. The purpose of the online model was to produce real-time information of the traffic flow status that can be given to drivers. The models were tested in practice. In conclusion, the results of online use of the prediction models in practice were promising and even a simple prediction model was shown to improve the accuracy of travel time information especially in congested conditions. The results also indicated that the self-adapting principle improved the performance of the model and made it possible to implement the model quite quickly. The model was practical for real-time use also in the long term in terms of the number of carry bits that it requires to restore the history of samples of traffic situations. As self-adapting this model performed better than as a static version i.e. without the self-adapting feature, as the proportion of correctly predicted traffic flow status increased considerably for the self-adapting model during the online trial. This report may be found at http://lib.tkk.fi/Diss/2009/isbn9789513873417/
Abstract