摘要：

页岩岩相的识别与预测对于分析确定页岩油气甜点层段非常重要。在缺乏岩心信息进行单井岩相研究时,测井数据扮演着十分重要的角色,而基于XGBoost算法可以充分挖掘多维测井数据所揭示的页岩岩相信息,从而达到预测单井页岩岩相的目的。本研究应用具有监督学习算法的XGBoost机器学习方法,利用常规测井数据作为变量数据集,建立了可预测页岩岩相类型的计算模型。首先建立适合具体研究区的页岩岩相划分标准,该标准应能体现研究区页岩岩相的辨识差异性,再统计不同矿物含量,确定不同岩相的具体矿物含量和TOC含量界限。在建立计算模型时,相关变量可能会提供相似的信息,导致模型过于依赖这些特征,需注意去除相似信息。XGBoost算法在参数优选方面,其网格搜索具有全面性,在网格搜索过程中应该进行多次优选,不断缩小搜索范围以求取最优值。以松辽盆地松南地区赞字井区块为例,采用矿物组分含量、沉积构造及TOC含量建立页岩岩相划分标准,青山口组可划分出5类主要页岩; 在应用XGBoost算法进行变量优选时,对于具有较高相关性的深侧向电阻率(LLD)和浅侧向电阻率(LLS)曲线,保留一条即可,结果表明模型准确率可提高4%左右; 经过变量选择及参数调优后,最终模型预测岩相的准确率可达90.03%。

关键词: 页岩岩相预测, XGBoost算法, 变量选择, 参数调优, 测井信息, 青山口组, 松辽盆地

Abstract:

The identification and prediction of shale lithofacies are crucial for identifying favorable intervals(“sweet spots”)in shale oil and gas reservoirs. In the absence of core data,logging data plays a key role in lithofacies analysis at the single-well level. By applying the XGBoost algorithm,useful lithofacies information can be extracted from multidimensional logging data,enabling effective prediction of shale lithofacies in individual wells. In this study,the XGBoost machine learning method,a supervised learning algorithm,is used to build a predictive model based on conventional logging datasets. First,a lithofacies classification scheme tailored to the specific study area is established,which captures the variability in shale lithofacies identification. The boundaries of mineral compositions and TOC content for different lithofacies types are then determined using statistical proportion analysis. During model construction,care must be taken to eliminate redundant variables,as highly correlated features may provide overlapping information and cause overfitting. XGBoost's grid search approach allows comprehensive parameter tuning. Multiple rounds of optimization should be conducted,with the search range gradually narrowed to determine the optimal parameter set. Using the Zanzijing block in the Songnan area as a case study,five major shale lithofacies types are defined based on mineral composition,sedimentary structures,and TOC content. During variable selection,for instance,only one of the highly correlated LLD and LLS logs is retained,which results in a model accuracy improvement of approximately 4%. After feature selection and parameter tuning,the final model achieves a lithofacies prediction accuracy of up to 90.03%.

Key words: shale lithofacies prediction, XGboost algorithm, variable selection, parameter tuning, well-logging data, Qingshankou Formation, Songliao Basin

中图分类号: 

• P512.2