A Physics-Informed Deep learning framework for traffic state estimation in signalized arterial roads

• Introduces an enhanced Physics-Informed Deep Learning framework tailored for traffic state estimation on signalized arterials. • Incorporates spatial, signal-phase, and congestion-aware loss function for better learning near queues. • Demonstrates consistent outperformance over existing methods in estimating traffic states and queue lengths. Efficient traffic management and effective signal planning can decrease intersection congestion and minimize delays. While Physics-Informed Deep Learning (PIDL) has shown promise in highway contexts, its application to urban arterial roads remains underexplored. This paper presents a structurally enhanced PIDL framework designed for traffic state estimation (TSE) on signalized arterials using partially observed high-resolution traffic data. The proposed model integrates a diffusively corrected PDE to capture driver behavior and incorporates spatial, signal-phase, and congestion-aware reweighting mechanisms into the loss function to improve learning around queue-prone regions. Three FD models–3-parameter LWR, Greenberg, and Newell–are evaluated across two realistic traffic conditions using simulated loop detector and probe vehicle data. A hybrid collocation sampling strategy and FD parameter estimation are employed within the learning architecture. Results demonstrate the model’s robustness across varying probe penetration rates, data sparsity, and congestion regimes. Notably, the proposed PIDL consistently outperforms its state-of-the-art counterpart in both estimation accuracy and queue length prediction. These findings highlight the proposed framework’s suitability for real-time arterial traffic operations and signal optimization.