BiONet-TSE: A Bidirectional Operator-Guided Neural Network for Traffic State Estimation

Accurate traffic state estimation (TSE) is critical for the deployment of intelligent transportation systems. Physics-informed neural networks represent a significant approach for achieving TSE; however, their applicability is constrained by the availability and accuracy of physical information. To address this limitation, this study proposes a multi-task TSE method guided by operator learning (BiONet). The core architecture of BiONet is a multi-task network. Prior to network training, a bidirectional DeepONet is employed to pre-learn the mapping operators among various physical variables, thereby enabling a generalized representation of physical laws. Building upon this foundation, BiONet incorporates a reliability-gated mechanism to modulate the influence of the operator guidance, effectively preventing performance degradation caused by operator failure. Experimental results on real-world data demonstrate that BiONet can accurately reconstruct traffic states and exhibits enhanced robustness across varying levels of data penetration.