基于支持向量机(SVM)的不同咸化程度烃源岩总有机碳含量预测模型

总有机碳含量（TOC）作为评价烃源岩有机质丰度的重要参数，其精确预测对油气勘探开发具有重要意义。目前总有机碳含量预测以ΔlogR方法、多元回归分析等基于统计分析的方法为主，存在泛化能力弱、主观性强等问题。机器学习方法的引入，可有效解决这类非稳定性、非线性、高复杂性的问题，但当下的研究仍停留在方法的比较与选取层面，没有对优良模型进行深入分析并检验其适用性。采用应用效果更好的支持向量机模型进行总有机碳含量预测，选取渤海湾盆地渤中凹陷古近系东营组淡水湖相和柴达木盆地西部狮子沟地区古近系咸化湖相烃源岩作为研究对象，对模型的效果进行检验与对比。通过相关性和XGBoost特征重要性分析，选定声波时差（DT）、体积密度（DEN）、自然电位（SP）、自然伽马（GR）、深度等作为输入层，以总有机碳含量作为输出层，确立SVM烃源岩总有机碳含量预测模型。研究结果表明，模型在应用至差异较大的沉积环境时具有很强的泛化能力以适应不同地区的地质特征；测井曲线对于烃源岩有机质丰度的敏感性由于沉积环境存在差异而有所区别。该模型在渤海湾盆地淡水湖相区域的应用中相关性更高，误差更小。 

Prediction model of TOC contents in source rocks with different salinity degrees based on Support Vector Machine (SVM)

The total organic carbon (TOC) content is an important parameter for the evaluation of abundance of organic matter in source rocks, and its predicting accuracy is of great significance to oil and gas exploration and development. At present, TOC prediction is mainly based on statistical analysis methods such as ΔlogR method and multiple regression analysis, problems such as weak generalization ability and strong subjectivity exist. The introduction of machine learning methods can effectively solve these problems of instability, nonlinearity, and high complexity. However, current research remains at the level of method comparison and selection with no indepth analysis of good models and their applicability. In this paper, a Support Vector Machine (SVM) model with better application effects was used to predict TOC contents of source rocks with different salinity degrees. As source rocks of freshwater lacustrine facies, the Paleogene Dongying Formation in the Bozhong Sag of Bohai Bay Basin and Paleogene source rocks in the Shizigou area of the western Qaidam Basin as saline lacustrine facies source rocks were selected to test and compare the effectiveness of the model. Through correlation analysis and XGBoost feature importance analysis, the logging sonic differential time (DT), volume density (DEN), spontaneous potential (SP), Gamma ray (GR) and depth were selected as the input layer, while the TOC was used as the output layer to establish a TOC prediction model based on SVM. Results show a strong generalization ability when applied to different sedimentary environments. It can adapt to the geological characteristics of different regions. The sensitivity of logging curves to the abundance of organic matter in source rocks varies in different sedimentary environments, which makes the model more relevant when applying to the fresh water lacustrine facies area in the Bohai Bay Basin.