Predicting Urban Arterial Travel Time with State-Space Neural Networks and Kalman Filters

Abstract

A hybrid model for predicting urban arterial travel time on the basis of so-called state-space neural networks (SSNNs) and the extended Kalman filter (EKF) is presented. Previous research demonstrated that SSNNs can address complex nonlinear spatiotemporal problems. However, SSNN models require off-line training with large sets of input–output data, presenting three main drawbacks: (a) great amounts of time and effort are involved in collecting, preparing, and executing these training sessions; (b) as the input–output mapping changes over time, the model requires complete retraining; and (c) if a different input set becomes available (e.g., from inductive loops) and the input–output mapping has to be changed, then retraining the model is impossible until enough time has passed to compose a representative training data set. To improve SSNN effectiveness, the EKF is proposed to train the SSNN instead of conventional approaches. Moreover, this network topology is derived from the urban travel time prediction problem. Instead of treating the neural network as a “black-box” model, the design explicitly reflects the relationships that exist in physical traffic systems. It allows the interpretation of neuron weights and structure in terms of the inherent mechanism of the network process with clear physical meaning. Model performance was tested on a densely used urban arterial in the Netherlands. Performance of this proposed model is compared with that of two existing models. Results of the comparisons indicate that the proposed model predicts complex nonlinear urban arterial travel times with satisfying effectiveness, robustness, and reliability.