Short-term traffic flow prediction using Markov models

Sammanfattning

With a growing urban population, the need for proper traffic management becomes ever greater. This study explores how well Markov-based models can predict short-term transitions between traffic speed states. First-order and second-order Markov-based models, both with and without time-dependency were studied in this paper. The transition matrices for each model were estimated from three weeks of traffic sensor data in Stockholm using maximum likelihood estimation and then evaluated using log-loss, perplexity, Brier score, and accuracy. The best performing model was the second-order time-of-day-dependent model. The results from this study show that the Markov-based models outperform naive baselines, indicating that a Markov-based approach to short-term traffic flow prediction is viable. The models in this paper can be built upon and improved, for example, by adding spatial dependency from neighboring sensors.