A Data-Driven Deep Learning Framework for Pressure Transient Analysis in Large-Scale Underground Gas Storage Systems

Abstract Underground natural gas storage (UNGS) at large scales faces challenges such as drastic changes in injection-withdrawal rates as well as periodic cycling, making it difficult for traditional machine learning methods to achieve accurate transient pressure prediction. To address this issue, this paper proposes a hybrid PiCrossformer framework that combines physics-informed neural networks with Crossformer time series modeling. This framework is used to achieve high-accuracy prediction of transient pressure behavior in UNGS during the pressure build-up period without requiring additional well shut-in operations. The model utilizes the preceding pressure time series and gas injection–withdrawal rates as dynamic input features, and reservoir pressure as the output feature. The cubic spline interpolation is utilized to deal with the data sparsity in the reservoir, and the moving window method is used to calculate pressure derivatives from the predicted pressure series. By explicitly embedding the residuals of the reservoir flow governing equations into the loss function, the model is guided to learn pressure evolution processes that conform to subsurface reservoir physical laws. Additionally, a weighted abrupt change-point loss mechanism is introduced to enhance the model’s ability to detect abrupt changes, and the whale optimization algorithm is utilized to automatically tune network architecture and hyperparameters. The model achieves an early-stage mean squared error (MSE) of 6.49×10-9 and a prediction accuracy of 99.98%, ensuring the reliability of pressure derivative data under low-MSE conditions. The application in the real case of the Hutubi UNGS in China shows that the initial formation pressure is 31.7877 MPa and the permeability is 11.6487 md.