关键词: 川西坳陷, 须家河组, 裂缝发育特征, 智能化预测方法, 随机森林

Abstract:

The Upper Triassic Xujiahe Formation in the western Sichuan Depression, Sichuan Basin is an important area for the increase in reserves and production of tight sandstone gas (hereinafter referred to as “tight gas”) in the Sichuan Basin. In practical production, high-yield and stable production wells are highly correlated with the dense development of fractures. Fractures provide pathways and spaces for gas migration and storage, and whether fractures develop is a key factor restricting the formation of high-quality reservoirs. To evaluate the “sweet spot” enrichment area of the Xujiahe Formation gas reservoir, fracture development characteristics are identified and an effective fracture recognition method is established based on core observation, logging data, and intelligent algorithms. The research suggests that structural fractures, diagenetic fractures, and overpressure fractures all develop in the study area. The structural fractures are mainly divided into three phases: Phase 1 (NW-SE orientation) predominantly develops low-angle fractures, with occasional high-angle fractures; Phase 2 (NNE-SSW orientation) mainly develops high-angle fractures; Phase 3 (E-W orientation) predominantly develops high-angle fractures. The fracture segments in the tight gas reservoir display characteristics of low density, high neutron density, high sonic time difference, and positive amplitude differences in deep and shallow lateral resistivity. The conventional logging data with fracture and non-fracture labels were normalized, and machine learning algorithms were applied for fracture intelligent prediction. The F1 scores for the K-nearest neighbors (KNN), support vector machine (SVM), extreme gradient boosting (XGBoost), and random forest algorithms were 0.65, 0.83, 0.88, and 0.91, respectively. It was found that the random forest algorithm demonstrated strong robustness and anti-interference capabilities, with higher prediction accuracy and efficiency compared to the other three algorithms. Additionally, to balance computational efficiency and accuracy, the genetic algorithm was selected as the optimization algorithm for hyperparameter tuning, outperforming grid search, Bayesian optimization, and particle swarm optimization algorithms. Shapley Additive Explanations (SHAP) were used to calculate the contribution of different influencing factors to the predictions. It was found that the sonic time difference, neutron density, and compensated density were the main logging curves influencing prediction accuracy. The fracture density exhibited a clear spatial distribution pattern, decreasing from the southwestern part to the northwestern part of the Sichuan Basin. The research results can provide a practical and feasible intelligent prediction model for the fracture “sweet spot” zone in tight gas reservoirs in the western Sichuan Basin, laying the foundation for increasing reserves and production of tight gas.

Key words: western Sichuan Deoression, Xujiahe Formation, fracture development characteristics, intelligent prediction method, random forest