Short-Term Traffic Flow Prediction with Regime Switching Models

Abstract

Accurate short-term prediction of traffic parameters is a critical component for many intelligent transportation system applications. Traffic flow is subject to abrupt disturbances because of various unexpected events (e.g., accidents, weather-induced disruption) that may change the underlying dynamics and the stability of the data generation process. Short-term prediction models that do not account for these changes produce biased and less accurate predictions. This paper proposes a new adaptive approach to short-term prediction that explicitly accounts for occasional regime changes by using statistical change-point detection algorithms. In this context, the expectation maximization and the CUSUM (cumulative sum) algorithms are implemented to detect shifts in the mean level of the process in real time. Autoregressive integrated moving average models are used for developing the forecasting models while the process mean is monitored by the two detection algorithms. The intercept of the forecasting models is updated on the basis of the detected shifts in the mean level to adapt to any potential new regimes. The proposed approach is tested on real-world loop data sets. The results show significant improvements in prediction accuracy compared with traditional autoregressive integrated moving average models with fixed parameters.