吴家坪组深层页岩井壁稳定性与三压力预测研究

Study on Wellbore Stability and Three-Pressure Prediction in Deep Shale of the Wujiaping Formation

摘要: 为解决吴家坪组深层页岩气钻井中井壁坍塌频发、三压力预测精度不足的技术难题，通过岩心力学实验、已钻井测井数据分析及多模型融合深度学习方法，构建“地层特征–压力系统–智能预测”一体化研究体系。结果表明：1) 吴家坪组页岩非均质性强，同一露头岩心单轴抗压强度差异达51.0 MPa (86.9~137.9 MPa)，钻井液浸泡1天后强度降幅超50%，强脆性破坏导致微裂隙发育加剧坍塌风险；2) 已钻井(大页1井、大201井、大页1H2-4井)吴家坪组上覆岩层压力当量密度2.60~2.65 g/cm³，孔隙压力峰值1.80~2.24 g/cm³，水平地应力方位集中于N90˚E~N100˚E，大页1井吴一段~吴二段存在“漏垮共存”风险(坍塌压力上限 > 2.50 g/cm³、漏失压力下限 ≈ 2.0 g/cm³)；3) 基于18项测井特征构建的深度神经网络(DNN)模型，在大页1H1-1井训练集上孔隙压力预测R²达0.996 (RMSE = 0.31 MPa)，引入梯度提升树、XGBoost、LightGBM等多模型融合(加权平均 + 堆叠集成)后，大页1H2-1井预测误差显著降低，孔隙压力误差0.27%、坍塌压力误差8.19%、破裂压力误差3.62%，较单纯神经网络分别降低0.27%、4.03%、1.32%。研究成果为川中地区吴家坪组特定井(大页1井、大201井、大页1H2-4井等)深层页岩气安全钻井液密度窗口设计与高效钻井提供关键技术支撑，其模型适用范围限于研究区域内目标层段。

Abstract: In order to solve the technical problems of frequent borehole collapse and insufficient prediction accuracy of three pressures in deep shale gas drilling in Wujiaping formation, an integrated research system of “formation characteristics pressure system intelligent prediction” was established through core mechanics experiment, well logging data analysis and multi model fusion in-depth learning method. The results show that: 1) the heterogeneity of Wujiaping formation shale is strong, the difference of uniaxial compressive strength of the same outcrop core is 51.0 MPa (86.9~137.9 MPa), and the strength decreases by more than 50% after soaking in drilling fluid for 1 day. The strong brittle failure leads to the development of microcracks and exacerbates the risk of collapse; 2) The pressure equivalent density of the overlying strata of the Wujiaping formation in the drilled wells (Daye 1 well, Daye 201 well and Daye 1h2-4 well) is 2.60~2.65 g/cm³, the peak pore pressure is 1.80~2.24 g/cm³, the horizontal in-situ stress is concentrated in N90˚E-N100˚E, and there is a risk of “leakage and collapse” in the wuyi-wuii section of Daye 1 well (the upper limit of collapse pressure > 2.50 g/cm³, and the lower limit of leakage pressure ≈ 2.0 g/cm³); 3) The depth neural network (DNN) model based on 18 logging characteristics predicted the pore pressure of 0.996 (RMSE = 0.31 MPa) in the training set of Daye 1h1-1 well. After introducing the gradient lifting tree, XGBoost, LightGBM and other multi model fusion (weighted average + stack integration), the prediction error of Daye 1h2-1 well was significantly reduced, with the pore pressure error of 0.27%, collapse pressure error of 8.19%, and fracture pressure error of 3.62%, which were 0.27%, 4.03%, and 1.32% lower than that of the simple neural network, respectively. The research results provide key technical support for the design of safe drilling fluid density window and efficient drilling of deep shale gas in specific wells of Wujiaping formation in Central Sichuan (Daye 1 well, Daye 201 well, Daye 1h2-4 well, etc.), and the scope of application of the model is limited to the target interval in the study area.