Neural operator-based modeling of gas dynamics in energy sector pipeline systems

This paper presents a study on the application of neural operators, specifically the Fourier Neural Operator (FNO), for modeling gas flow regimes in gas transportation systems (GTS). The research aims to accelerate the calculation of transient gas flow in extensive pipelines while maintaining accuracy and spatial-temporal invariance. The proposed method integrates classical numerical solutions of gas dynamics equations into the training of the neural operator. The first part of the study validates the accuracy of the FNO-based model in predicting gas pressure and flow rates, achieving errors below 0.05% in Mean Absolute Percentage Error (MAPE) and 0.5% in Maximum Absolute Percentage Error (max APE). The second part addresses scenarios involving flow direction reversals in pipelines by extending the training dataset to include cases with reverse pressure gradients and adjusted boundary conditions. Results demonstrate that the model effectively replicates numerical solutions for such cases. Further, the study explores the modeling of GTS with complex topologies, representing the system as a directed graph. The optimization-based approach minimizes flow imbalances at pipeline junctions using a gradient descent algorithm (Adam), leveraging the differentiable nature of the FNO. Computational experiments reveal acceptable computation times and scalability, with GPU acceleration significantly enhancing individual pipeline predictions. The proposed hybrid method combines neural network modeling, optimization, and graph theory to enable the simulation of large-scale GTS. The findings highlight the potential of neural operators in automating and optimizing real-world pipeline network operations, offering a robust framework for future advancements in engineering system modeling.