Uncertainty quantification in Bayesian physics-informed deep learning-based traffic state prediction

Abstract

Accurate and reliable traffic state prediction (TSP) is an essential task for intelligent transportation systems. However, achieving this goal is challenging due to the high-dimensional and coupled nature of traffic feature evolution patterns, which are deeply recessive and make it difficult to effectively characterize and model TSP using purely data-driven methods. Furthermore, a significant limitation of existing TSP methods is their inability to estimate data and model uncertainty, which is crucial for understanding inherent data variations and model limitations. To address these challenges, this study proposes a novel TSP model that combines the diffusion convolutional recurrent neural network (DCRNN) with physical prior knowledge within a Bayesian framework. Specifically, DCRNN captures the spatiotemporal correlation among various sensors. Furthermore, this approach leverages Monte Carlo dropout and heteroskedasticity modeling to quantify epistemic and aleatoric uncertainties. The model's efficacy is evaluated using the Xuancheng China urban dataset and the PeMS04 US highway dataset. Empirical results show that the proposed method outperforms state-of-the-art methods in both prediction accuracy and uncertainty quantification. These findings highlight the advantages of a data-model hybrid-driven approach to achieve accurate and reliable TSP. This study effectively quantifies and mitigates both aleatoric and epistemic uncertainties, holding significant implications for the control and management of real traffic flow.