Accurate prediction of dynamic travel times can assist commuters in making better travel decisions. In this paper, a new algorithm is proposed to accurately predict the expected and confidence levels of dynamic travel times. The algorithm pre-processes the available historical data to identify recurring bottlenecks along the road. Subsequently, the algorithm builds a spatiotemporal congestion probability distribution. This distribution provides the probability of a spatiotemporal section being congested. The proposed algorithm integrates congestion probability and spatiotemporal speed measurements to construct feature vectors that are used as the travel time predictors. A random forest is used to model the relationship between the predictors and the travel time. Consequently, the built random forest can be used to predict the travel time by propagating the new features vector through all trees. The experimental results show that the proposed algorithm achieves more than a 38 percent reduction in the prediction error on congested days compared to the state-of-practice instantaneous algorithm and 28 percent reduction when compared to a genetic programming travel time prediction algorithm. Moreover, the predicted travel time bounds encompass all field observations.
展开▼
