Three-Pressure Prediction Method for Formation Based on Xgboost-gnn Hybrid Model

Abstract

Abstract Accurate prediction of Three-Pressure data in geological formations can assist in determining drilling fluid design, wellbore stability assessment, and optimization of drilling parameters, thereby reducing the probability of drilling risks. Conventional methods for predicting triplet pressure in geological formations often involve complex calculations, numerous empirical parameters, low prediction accuracy, limited universality, and a certain degree of lag. Therefore, there is an urgent need for new methods that are efficient, simple, and accurate in predicting triplet pressure in geological formations. To address the aforementioned issues, this study focuses on the Penglai gas area in the Sichuan Basin. By employing the XGBoost algorithm, three well logging parameters, namely acoustic time difference, compensating density, and natural gamma, are selected to classify the strata into two types: clastic rocks and carbonate rocks. Additionally, using 11 well logging and drilling parameters, including well depth, acoustic time difference, compensating density, natural gamma, drilling time, drilling pressure, and torque, a graph neural network (GNN) is applied to capture the spatial geological features of the strata. Separate GNN prediction models are established for both clastic rocks and carbonate rocks, and the predicted results are compared and validated against field-measured data. The results indicate that the XGBoost algorithm achieves a classification accuracy of 94.31% and an AUC of 0.99. The GNN prediction models exhibit good accuracy and stability. When compared with the field-measured data, the clastic rock model shows an average MAPE of 3.963% and an average R2 value of 0.869 for the testing set, while the carbonate rock model shows an average MAPE of 1.681% and an average R2 value of 0.885 for the testing set. Compared with conventional rock mechanics three-layer pressure prediction methods such as the Eaton method, the XGBoost-GNN algorithm demonstrates higher accuracy, precision, stability, and more accurate data for predicting layer positions. By utilizing the XGBoost-GNN algorithm, this study proposes a classification-first, prediction-second methodology, which effectively captures the spatial and geological features of the strata by modeling the graph structure. This approach provides more accurate prediction results and supports drilling engineering design and safe and efficient drilling.