A physics-guided modular deep learning framework for scalable flow state prediction in large-scale gas pipeline networks

Accurate and efficient prediction of flow states in large-scale natural gas pipeline networks (NGPNs) is essential for real-time monitoring, scheduling, and control. Traditional numerical solvers, while physically rigorous, face severe scalability limitations due to the complexity of interconnected topologies and frequently changing boundary conditions. To address these challenges, this study proposes a physics-guided modular deep learning framework that integrates domain knowledge with a divide-and-conquer strategy. The method decomposes the pipeline into subnetworks, employs physics-informed feature pre-allocation for nodal initialisation, and utilises a modified U-Net architecture for steady-state flow prediction. Subnetwork results are then reconstructed through interface consistency enforcement to yield the global solution. Benchmark and real-world validations demonstrate that the framework achieves about 1% mean absolute error in predictions while reducing computation time by more than three orders of magnitude compared to traditional monolithic solvers. This balance of accuracy and efficiency highlights the framework's potential for deployment in digital twins and real-time decision support systems for energy infrastructure. HighlightsA physics-guided deep learning framework is proposed for scalable flow state prediction in large-scale natural gas pipeline networks.The method decomposes complex pipeline systems into modular subnetworks and employs a U-Net model with physics-informed feature pre-allocation.The proposed framework achieves comparable accuracy to traditional solvers while reducing computation time.