A transfer learning approach for cross-area hydraulic fracturing pressure prediction

The data-driven algorithms provide a powerful tool for predictions of hydraulic fracturing pressures, which is crucial for the design of pumping schedules and safety operations. For unconvetional oil and gas, the rapid declines in productions require continuous exploration of new blocks, during which cross-area pressure prediction becomes essential. However, significant regional variations in geology and limited access to geological data, make data-driven models mainly reliable in the data source region and restrict their generalization across regions. This study builds a transfer learning framework for pressure prediction across areas. In this framework, a deep learning model is first trained by historical data from the basic region, and then fine-tuned using eight fracturing stages of data (geological, wellbore, and pumping records) from the target region. The results show that using data from six fracturing stages is sufficient to achieve desirable generalization performance. This strategy transfers the experiences learned from the basic region to new regions, which may break the data-dependence barrier. Different transfer learning strategies, the core technique for knowledge transfer, are optimized to boost the predicting accuracy and transferring efficiency with minimum tuning dataset. Taking shale gas fracturing for instance, the performance of the new framework is demonstrated by the errors of pressure predictions, with root mean square error (RMSE) 2.26–8.22 MPa, r-square (R²) error 0.50–0.91, and symmetric mean absolute percentage (SMAP) error 2.09–8.55 %. The gradually-unfreezen strategy demonstrates superior performance than the full-unfreezen and full-freezing strategies, and performs better as more tuning data are incorporated. A comparison between the new framework and traditional algorithms (Support Vector Regression and Random Forest) further demonstrates the accuracy of our new method. The SHAP (SHapley Additive exPlanations) value indicates that pump rate is the most influential factor for pressure prediction, followed by perforation friction and well depth. The successful application of the transfer learning framework bridges the gap between different regions, leveraging past experience to improve the efficiency of new developments, particularly beneficial for unconventional shale developments that sustain productions by new explorations. Moreover, the transfer learning strategy may improve the inherent data-dependence weakness of data-driven algorithms, and then promote the generalization of well-trained machine-learning models.