基于人工神经网络方法预测油气资源丰度——以渤海湾盆地东濮凹陷文留地区古近系沙河街组三段为例

Prediction of petroleum resource abundance based on artificial neural network method: a case study of third member of Paleogene Shahejie Formation in Wenliu area of Dongpu Sag, Bohai Bay Basin

杨子杰 陈冬霞 王翘楚 王福伟 李莎 田梓葉 陈淑敏 张婉蓉 姚东升 王昱超

YANG Zijie CHEN Dongxia WANG Qiaochu WANG Fuwei LI Sha TIAN Ziye CHEN Shumin ZHANG Wanrong YAO Dongsheng WANG Yuchao

1. 中国石油大学(北京) 地球科学学院, 北京 102249; 2. 油气资源与工程全国重点实验室 中国石油大学(北京), 北京 102249

1. College of Geosciences, China University of Petroleum(Beijing), Beijing 102249, China; 2. National Key Laboratory of Petroleum Resources and Engineering, China University of Petroleum(Beijing), Beijing 102249, China

油气资源丰度通常受多个因素控制，其相关参数信息种类繁杂、数据量庞大，应用传统的地质统计学方法定量预测准确度不高。为了快速预测油气资源量丰度并明确其主控因素，以渤海湾盆地东濮凹陷文留地区古近系沙河街组三段为例，采用基于多层感知器神经网络(MLP)方法对油气资源丰度进行定量预测，同时采用Boosting集成学习算法优化预测模型，分别对66组样本油气资源丰度数据进行预测。结果表明，训练集数据实测值与预测值相关系数分别达0.789和0.989，验证集数据实测值与预测值相关系数分别达0.618和0.825，测试数据中实测值和预测值相关系数分别达0.689和0.845；有效厚度、平均渗透率、有效孔隙度是影响油气资源丰度最主要的3个地质因素，重要性系数分别为33.93%、20.12%和19.53%，圈闭面积、地面原油密度、生烃中心贡献等参数为次要影响因素。采用Boosting集成学习算法优化之后的多层感知器模型预测准确性得到了很大的提升，能为有利目标优选及勘探开发方案调整提供可靠依据，为凹陷内其他区块油气资源评价提供较好的参考和借鉴。

The abundance of petroleum resource is influenced by various factors and involves complex parameters and extensive data. Consequently, traditional geostatistical methods often lack precision in quantitative prediction. To address this issue, this study focuses on the third member of Paleogene Shahejie Formation (member Es3) in the Wenliu area of the Dongpu Sag and utilizes a multi-layer perceptron neural network (MLP) for predicting petroleum resource abundance and employed the Boosting ensemble learning algorithm to optimize the prediction model. The MLP and MLP-Boosting algorithm models were test on 66 sample groups, yielding correlation coefficients of 0.789 and 0.989 for the training set, 0.618 and 0.825 for the validation set and 0.689 and 0.845 for the test set. The analysis identified effective thickness, average permeability and effective porosity are the most significant geological factors influencing petroleum resource abundance, with importance coefficients of 33.93%, 20.12% and 19.53%, respectively. Other factors such as trap area, surface crude oil density and sedimentary facies assignment were found to be less influential. Overall, the Boosting ensemble learning algorithm significantly enhanced the prediction accuracy of the multi-layer perceptron model, offering valuable insights for target optimization, exploration planning and petroleum resource evaluation in other blocks in the sag.

国家自然科学基金面上项目(41972124)资助。

https://doi.org/10.11781/sysydz202402428