Classification of shale gas “sweet spot” based on Random Forest machine learning

NIE Yunli; GAO Guozhong

doi:10.13809/j.cnki.cn32-1825/te.2023.03.011

Views : 0 下载量: 0 CSCD: 1 CNKI被引量: 18 Scopus: 0 更多指标

PDF
Export
Share
Collection
Album

Classification of shale gas “sweet spot” based on Random Forest machine learning
Issue 3, Pages: 358-367(2023)
Published：2023
DOI： 10.13809/j.cnki.cn32-1825/te.2023.03.011
稿件说明：

移动端阅览

DOI：

NIE Yunli, GAO Guozhong. Classification of shale gas “sweet spot” based on Random Forest machine learning[J]. Petroleum Reservoir Evaluation and Development, 2023, (3): 358-367. DOI： 10.13809/j.cnki.cn32-1825/te.2023.03.011.

摘要

为了解决页岩气“甜点”分类识别涉及指标多、需要根据个人经验判别、耗时耗力的问题，提出了一种基于随机森林模型的页岩气“甜点”分类方法。首先，选取长宁区的10口井数据，利用肯德尔相关分析筛选出用于识别的11种特征。然后再分别采用单棵决策树和随机森林方法进行预测，得到页岩气“甜点”识别结果。最后，对预测结果分类并进行算法参数优化。实际应用结果表明，单棵决策树预测精度虽可以达到97.7%，但呈现过拟合趋势，且剪枝之后拟合精度大大降低到只有70.7%；采用的随机森林方法避免了单棵决策树的缺陷，并且预测的精度达到98%，而且，计算代价小，能有效降低时间损耗、节省人力成本。证明随机森林机器学习方法结合多源信息是实现页岩气“甜点”识别预测的一种有效手段。

Abstract

The classification and identification of shale gas “sweet spot” involves a variety of different factors

which requires personal experience

and is usually time and resources consuming. In order to solve this problem

an efficient and effective classification and identification method for shale gas “sweet spot” based on the Random Forest method is proposed. Firstly

data from ten wells in Changning area are selected and eleven features are selected for “sweet spot” classification by the Kendall correlation. Then

the single decision tree and the Random Forest method are used for the “sweet spot” classification and identification. Finally

the results are verified and the Random Forest parameters are optimized. The experimental results show that although the prediction accuracy of a single decision tree can reach 97.7 %

it shows a trend of overfitting

and the fitting accuracy is greatly reduced by only 70.7 % after pruning. The Random Forest method avoids the disadvantage of the single decision tree method

and the prediction accuracy reaches 98 %. Moreover

the computational cost is low

which can effectively reduce the time loss and save the labor cost. As a result

the proposed Random Forest machine learning method with multi-source information is an effective shale gas “sweet spot” classification and identification method.