摘要
经典含水饱和度参数预测模型是通过岩石物理实验来确定的.由于获取的岩心样本量在实际工程中非常有限,导致由实验确定的预测模型参数不可靠,最终合理的预测结果难以给出.含水饱和度预测本质属于拟合问题,而机器学习在处理拟合问题方面能力出众,因此是一理想应用手段.高效梯度提升模型(LightGBM)在集成学习理论中是最强模型之一,具有巨大的实际应用潜力,为此被采用进行预测研究.为降低原始数据集异常点及无效特征对模型预测能力和泛化能力带来的负面影响,本文提出利用Tukey算法和主成分分析(PCA)算法进行数据预处理,由此建立了一基于LightGBM的预测策略.本文选用鄂尔多斯盆地长4+5段致密砂岩储层岩心样本数据集对提出的预测策略进行验证.为加强验证效果,本文引入K邻近(KNeighbors)、支持向量拟合(SVR)和随机森林(Random Forest)等3个模型在两个验证集上进行对比.实验结果显示,提出策略在两个实验中均能给出最小均方根误差(RMSE)和平均绝对误差(MAE).验证结果显示提出的基于LightGBM的预测策略能够处理实际含水饱和度参数预测问题,且鲁棒性好,在测井评价研究方向上具有应用和参考价值.
Parametric determination of classic water saturation predictor generally is realized under petrophysical experiment.Since the costly and technical limitations are imposed on the acquisition of cores in the actual petroleum-based engineering,the parameters of water saturation predictors figured out by fewer cores routinely are unreliable.The predicting results,thus,are also questionable.The nature of prediction of water saturation is akin to solving a regression issue,while machine learning currently is acknowledged as the most extraordinary solver in this field.Then,such kind technology becomes fairly potential on the predicting of water saturation.Light Gradient Boosting Machine(Light GBM)has been proved as one of the most powerful models in the ensemble learning,and simultaneously has demonstrated as a promising candidate in the soft computing application,therefore selected to forecast water saturation.To possibly eliminate the negative influence brought by outliers and interference features of the raw dataset,in the preprocessing Tukey and Principal Component Analysis(PCA),two classic and functional algorithms,are introduced,and thereby a Light GBM-based predicting strategy for water saturation is proposed.The data for validation is collected from the tight sandstone reservoirs located in the member of Chang 4+5,Ordos Basin.To highlight the validating effect,three contrastive predicting models,including KNeighbors,Support Vector Regression(SVR),and RF(random forest),are employed to finalize two tests.The experimental results manifest that the proposed strategy is capable of providing the smallest Root Mean Squared Errors(RMSE)and Mean Absolute Error(MAE)in two tests.The findings reveal that the Light GBM-based strategy is effective and robust enough to deal with the real water saturation prediction,and then it is believable that it has a remarkable application value and a reference value in the research of logging interpretation.
作者
丁圣
杨尚锋
路巍
罗仁敏
朱立华
谷宇峰
陈小宏
DING Sheng;YANG ShangFeng;LU Wei;LUO RenMin;ZHU LiHua;GU YuFeng;CHEN XiaoHong(Sinopec Geophysical Research Institute,Nanjing 211103,China;China University of Petroleum(Beijing),Beijing 102249,China;The Fifth Oil Production Plant of PetroChina Changqing Oilfield Company,Xi'an 710200,China;Strategic Research Center of Oil and Gas Resources,Ministry of Natural Resources,Beijing 100034,China)
出处
《地球物理学进展》
CSCD
北大核心
2023年第1期185-200,共16页
Progress in Geophysics
