【摘要】 为了提升致密气储层流体识别的可靠性，解决地质因素导致测井参数贡献降低的问题，为油气勘探开发提供支撑，以四川盆地某区块致密储层为研究对象，采用皮尔逊相关系数与SHAP价值图（Shapley Additive exPlanations Plot）结合的方法优选测井参数。建立误差反向传播（Back Propagation,BP）神经网络、支持向量机、贝叶斯优化、优化的分布式梯度提升库（eXtreme Gradient Boosting,XGBoost）这4种机器学习预测模型，通过受试者工作特征曲线（ROC）的平均值I_AUC及实际样本来检验对比模型性能。研究结果表明：（1）综合两种方法优选出弹性模量差比/纵横波速度比（D_R/Z）、体积压缩系数-泊松比（C-P）、三孔隙度差值（A₁）、拉梅常数/泊松比（L_a/P_L）、峰基比(Q_FGB)、斯通利波时差(I_DTST)共6个对流体识别贡献显著的敏感测井参数；（2）4种机器学习预测模型验证ROC曲线的平均值I_AUC依次为0.955、0.994、0.954、0.995；（3）18个检验样本中，XGBoost模型符合率达88.9%，支持向量机、BP神经网络、贝叶斯优化模型符合率分别为83.3%、72.2%、66.6%；（4）D4井盲井段应用显示XGBoost模型预测结果与试气结论吻合度最高，研究区各井总体预测符合率达84.52%。结论认为，XGBoost模型在致密气储层流体识别中性能最优，结合皮尔逊相关系数与SHAP价值图的参数优选方法及多模型对比策略，有效规避了传统方法的主观性，提升了流体识别的客观性与可靠性，为致密气储层流体识别提供了有效思路。更多还原

【Abstract】 To improve the reliability of fluid identification in tight gas reservoirs, address the reduced contribution of logging parameters caused by geological factors, and provide support for oil and gas exploration and development, taking a tight reservoir block in the Sichuan basin as the research object, the method combining Pearson correlation coefficient and SHAP value map is adopted to optimize logging parameters. Four machine learning prediction models（BP neural network, support vector machine, Bayesian optimization, and XGBoost） are established. Model performance is compared via the average I_AUC value of ROC curves and actual sample verification. The research results show that:（1）A total of 6 sensitive logging parameters with significant contributions to fluid identification are optimized by integrating the two methods, including D_R/Z（elastic modulus difference ratio）, C-P（bulk compressibility-Poisson’s ratio）, A₁（three-porosity difference）, L_a/P_L（Lame constant/Poisson’s ratio）, Q_FGB（peak-base ratio）, and I_DTST（Stoneley wave slowness）.（2）The average I_AUC values of the validation ROC curves for the four models are 0.955, 0.994, 0.954, and 0.995 in sequence.（3）Among 18 verification samples, the accuracy rate of the XGBoost model reached 88.9%, while those of the support vector machine, BP neural network, and Bayesian optimization models are 83.3%, 72.2%, and 66.6% respectively.（4）Application in the blind well section of well D4 showed that the XGBoost model’s prediction results had the highest consistency with gas test conclusions, and the overall prediction accuracy rate of all wells in the study area reached 84.52%. It is concluded that the XGBoost model has the optimal performance in fluid identification of tight gas reservoirs. The parameter optimization method combining Pearson correlation coefficient and SHAP value map, along with the multi-model comparison strategy, effectively avoids the subjectivity of traditional methods, improves the objectivity and reliability of fluid identification, and provides an effective idea for fluid identification in tight gas reservoirs.更多还原