A novel hybrid deep learning methodology for real-time wellhead pressure forecasting and risk warning during shale gas hydraulic fracturing

Abstract

Accurate real-time forecasting of wellhead pressure significantly impacts risk warning and optimization of fracturing parameters. However, the complexity and non-stationary of data limit the accuracy of traditional deep learning (DL). We propose a novel hybrid DL method to enhance risk warning capabilities. The proposed method integrates the complex forecasting process into four modules. Firstly, the VMD-Fuzzy entropy module classifies intrinsic mode functions (IMFs) obtained from variational mode decomposition to significantly reduce feature redundancy. Then the Attention-GNN automatically learns latent features between multiple variables to automatically update the graph structure and incorporate controllable future input features. Additionally, the temporal–spatial feature extraction module captures spatial and temporal correlations to improve accuracy. The uncertainty quantification module employs a backtrack loss function and multi-head attention to enhance the capturing capability for critical data features. The method is verified using fracturing data from a shale gas block in Sichuan, China. The average root mean square error (RMSE), average maximum allowable error (MAE) and average R${}^2$ of the target area are 1.31 (MPa), 1.27 (MPa) and 0.94, respectively, which are significantly better than the traditional DL. In addition, the data of 4 overpressure well stages were used for example verification, and the corresponding traffic light risk warning system was developed. The verification results prove that the proposed method can effectively improve the warning timeliness, and provide an effective technical way to achieve intelligent and efficient hydraulic fracturing.