Objective Simulation constitutes a core technology for intelligent pipeline network dispatching. Data-driven approaches leverage large-scale datasets and pre-trained models to significantly reduce real-time simulation latency.

Method This study introduces a steady-state simulation framework based on graph neural networks (GNN). First, a graph representation of the pipeline system was constructed using cross-sections and adjacency matrices. Subsequently, inspired by the Newton-Raphson method widely employed in mechanistic simulations, the original GNN was reconstructed by designing a linear incremental update function, leading to the development of the Newton-GNN model. Finally, pipeline system simulation was achieved by mimicking the inverse Jacobian matrix and linear iteration steps through learned approximations.

Result Validation demonstrates that: First, across 6000 diverse cases, the proposed model enhances accuracy by 72.3% compared to the baseline GNN, with a relative deviation as low as 0.06%. Second, in generalization tests, the model adapts effectively to various scenarios with parameter variations. Third, in two case studies, including the XYX and FR-SH sections, the proposed model yields relative deviations of 2.17% and 1.70% for mainline pressure predictions, respectively.

Conclusion The Newton-GNN model exhibits high accuracy and strong generalization ability, making it promising for large-scale natural gas pipeline network simulations.