PipeFormer: A physics-informed tokenized sparse transformer for natural gas pipeline networks

Real time simulation is vital for the safety and optimal scheduling of natural gas pipeline networks. However, traditional numerical solvers are often too slow for rapid decision making, while previous deep learning models frequently violate physical laws or struggle with the complex connectivity of pipeline equipment. To address these challenges, we propose PipeFormer, a physics-informed framework explicitly designed for pipeline structures. This approach employs a specialized attention mechanism that restricts computation to physically connected neighbors, which efficiently handles the extensive large network without excessive computational cost. We also introduce a discrete tokenization strategy that maps continuous sensor readings into finite categories to better manage the irregular distributions found in real operations. Furthermore, we integrate the conservation laws of mass and momentum directly into the training process to ensure the predictions remain consistent with fluid dynamics. We validated the model on a large industrial dataset. The results demonstrate that our method achieves better accuracy than other deep learning models with a mean absolute percentage error of 0.271% and an inference time of tens of milliseconds per time step. This work provides a promising solution for the intelligent scheduling of large scale energy infrastructure.