摘要
为了准确高效地表征致密油储层孔隙度的空间分布特征,同时对机器学习模型的可解释性进行评价,采用Z-Score方法对特征属性进行归一化处理,并应用Optuna超参数优化框架对模型的超参数进行调优,建立了一种基于LightGBM算法的孔隙度预测模型,与GBDT和XGBoost算法模型进行了预测效果的综合对比,并利用SHAP算法对LightGBM模型的输出结果进行了可视化解释分析。研究结果表明:LightGBM模型在训练数据集和测试数据集上的预测决定系数分别为0.984和0.855,模型预测准确度高、泛化能力强,综合预测效果好于GBDT和XGBoost模型。应用SHAP算法对LightGBM模型结果的可解释性进行分析,结果表明,影响LightGBM孔隙度预测模型最重要的5项测井参数为密度、阵列感应电阻率、自然伽马、声波时差和光电吸收截面指数。在研究区某单井X致密层段孔隙度的预测实例中,LightGBM模型预测准确度达93.9%,分别高于GBDT和XGBoost模型的预测准确度86.53%和89.08%;训练时长为0.016 s,分别为GBDT和XGBoost模型训练时长的0.096倍和0.025倍;预测时长为0.01 s,分别为GBDT和XGBoost模型预测时长的0.42倍和0.19倍;LightGBM模型的预测效率相对GBDT和XGBoost模型具有明显优势,其在取心井段上对孔隙度的预测误差更小,预测能力更强,且能更好地拟合低值孔隙度。该方法的应用不仅解决了单井致密层段获取完整准确孔隙度分布的难题,而且提高了孔隙度预测的精度和效率,对致密油储层的评价及高效勘探开发具有一定的参考价值。
In order to accurately and efficiently characterize the spatial distribution characteristics of porosity in tight oil reservoirs and evaluate the interpretability of machine learning models,the Z-Score method was used to normalize the feature attributes,and the Optuna hyperparameter optimization framework was applied to optimize the hyperparameters of the model.A porosity prediction model based on the LightGBM algorithm was established,and the prediction performance was comprehensively compared with GBDT and XGBoost algorithm models.The SHAP algorithm was used to visually interpret and analyze the output results of the LightGBM model.The research results indicate that the LightGBM model has prediction determination coefficients of 0.984 and 0.855 on the training and testing datasets,respectively.The model has high prediction accuracy and strong generalization ability,and its overall prediction performance is better than the GBDT and XGBoost models.The interpretability of the LightGBM model results was analyzed by using the SHAP algorithm.The results show that the five most important logging parameters affecting the LightGBM porosity prediction model are density,array induction resistivity,natural gamma,interval transit time,and photoelectric absorption cross-section index.In the porosity prediction example of X tight interval of a single well in the research area,the LightGBM model achieves a prediction accuracy of 93.9%,which is higher than the prediction accuracy of the GBDT and XGBoost models at 86.53%and 89.08%,respectively.The training duration is 0.016 s,which is 0.096 times and 0.025 times the training duration of GBDT and XGBoost models,respectively.The prediction duration is 0.01 s,which is 0.42 times and 0.19 times the prediction duration of GBDT and XGBoost models,respectively.The prediction efficiency of the LightGBM model has significant advantages over the GBDT and XGBoost models.The LightGBM model has smaller prediction errors for porosity in the cored interval,stronger prediction ability,and can better fit low porosity values.This method could solve the problem of obtaining complete and accurate porosity distribution in single well tight intervals and improve the accuracy and efficiency of porosity prediction,which has a certain reference value for the evaluation and efficient exploration and development of tight oil reservoirs.
作者
王伟
党海龙
康胜松
肖前华
丁磊
石立华
WANG Wei;DANG Hailong;KANG Shengsong;XIAO Qianhua;DING Lei;SHI Lihua(Shaanxi Yanchang Petroleum(Group)Co.,Ltd.,Xi’an City,Shaanxi Province,710075,China;College of Petroleum and Natural Gas Engineering,Chongqing University of Science and Technology,Chongqing City,401331,China)
出处
《油气地质与采收率》
北大核心
2025年第5期90-99,共10页
Petroleum Geology and Recovery Efficiency
基金
国家油气重大专项“超低渗油藏力学作用机制及渗流数学模型”(2017ZX05013-001)
陕西省自然科学基础研究计划项目“致密砂岩储层微观孔隙结构及薄膜流动对渗吸规律的影响”(2023-JC-YB-423)。
作者简介
王伟(1983-),男,湖北孝感人,高级工程师,硕士,从事非常规储层评价与有利区预测研究工作。E-mail:wwei0773@qq.com。
