This report presents the development of a travel-time prediction model for non-recurrent situations of road traffic. Furthermore, the developed model applies for urban road networks and freeways. The purpose of such a travel time predictor is to support the drivers’ decision-making in non-regular situations such as incidents, road works or social events. An extensive literature review is conducted aiming to detect insights about the requirements of a reliable travel time prediction model, the relationship between congestion, traffic parameters and travel time and the already available methods used for travel time prediction. The literature review concludes that most of the available prediction models are used on freeways and not being specifically designed for congested traffic situations. Thus, there is a true need for models that precisely aim to provide travel time information when it is urgently needed, namely in case of non-recurrent situations. In the first instance, a modelling framework is defined. Various traffic conditions along with the defined non-recurrent situations are simulated to create sufficient input/output pattern for testing two different neural network classes, the popular Feed Forward neural network and the recurrent ELMAN neural network. Analysing the neural network properties and their training parameters enabled us to detect the optimal neural network parameter set. Making use of the microscopic simulation tool PARAMICS for the road network design and to generate the traffic flow data (flows, speeds and occupancies from detector stations), the model employs recurrent ELMAN neural networks for the prediction task. The prediction models core competency is to incorporate changes of traffic states in time and space by sectoring the road network into clusters of common attributes. In case of the urban road model, upstream and downstream detector stations of the adjacent links and of the link of interest are used as input (same driving direction). The freeway is divided into sections with detector stations bordering the sections to predict the travel time for the entire freeway stretch. The three major impediments emerging during the research were: I. the difficulty to simulate non-recurrent situations with realistic driver behaviour in PARAMICS. II. the difficulty to generate, store and arrange the necessary input/output pattern with PARAMICS (travel times, aggregated data) III. the problematic nature of travel time prediction for urban road networks because of the complexity caused by (signalized) intersections. For the freeway sector, the sectored model shows a promising performance with a MAPE of 15.43%, while for the urban link sector the MAPE is at 35.8%. Obviously, the complexity of the controlled character of urban road traffic influences the performance of the urban road sector model. In conclusion, this novel approach of developing different neural network predictors that sector the road network and by this means incorporate the spatial attributes of the target site can be a successful model definition to deal with the difficulty to predict travel times for challenging traffic situations such as non-recurrent situations. (Author/publisher)
Abstract