基于深度学习的复杂储层流体性质测井识别——以车排子油田某井区为例  

Logging Identification of Complex Reservoir Fluid Properties Based on Deep Learning: A Case Study of One Well Block in Chepaizi Oilfield

作　　者：蓝茜茜 张逸伦 康志宏[1] LAN Xi-xi;ZHANG Yi-lun;KANG Zhi-hong(School of Energy Resources,China University of Geosciences(Beijing),Beijing 100083,China;School of Earth and Space Sciences,Peking University,Beijing 100871,China)

机构地区：[1]中国地质大学(北京)能源学院,北京100083 [2]北京大学地球与空间科学学院,北京100871

出　　处：《科学技术与工程》2020年第29期11923-11930,共8页Science Technology and Engineering

基　　金：“十三五”国家科技重大专项(2017ZX05009-001);中国地质调查局地质调查项目(DD20190085)。

摘　　要：测井资料人工解释是目前主流的储层流体性质识别手段,但其应用于复杂储层时存在识别率低、非智能化的缺陷;而近年来发展起来的深度学习方法可以从海量数据中自动提取数据特征,非线性预测能力强。基于目标区块已有大量测井资料和试油结果数据,在应用常规深度神经网络的基础上,提出一种采用混合采样技术、ReLU-Softmax激活函数和Dropout正则化的组合优化新方法。优化后的网络模型对流体识别问题适应性强,且有效避免了样本不均衡、过拟合等问题。将该方法应用于车排子油田低渗油藏某井区,对12口井的水层、干层、油水同层、油层4种流体进行识别,结果显示总体识别准确率达82.7%,单一流体识别率也均较高。且组合优化方法的识别效果明显优于其他方法,尤其使得小样本类——油层和油水同层的识别率得到显著提高。展现了深度学习在复杂储层流体性质识别中良好的应用效果。Currently,manual interpretation of logging data is the mainstream method for identifying reservoir fluid properties.But this method has a low identification rate when applied to complex reservoirs,as well as a great intelligence defect.Instead,recent developed deep learning technology with strong nonlinear prediction ability can extract data features from massive data automatically.Lots of well logging data and oil test results of target block was taken as input data.Based on the application of conventional DNN method,a new combinatorial optimization method was proposed.The method combined with hybrid sampling strategy,ReLU-Softmax activation function and Dropout regularization trick.The optimized network model had a strong adaptability to the fluid identification problem.Moreover,it effectively avoided sample imbalance problem,over fitting problem,et al.Then,the method was applied to one well block with low-permeability reservoir in Chepaizi Oilfield.The identification results of properties of four fluids in 12 wells show that the overall accuracy rate is 82.7%.And the rates for four fluid as water layer,dry layer,oil-water layer and oil layer are also high.The identification effect of the combination optimization method is obviously better than other methods.Especially,the identification effect of classes with few samples such as oil layers and oil-water layers has improved significantly.The proposed method shows a good application effect of deep learning in the field of complex reservoir fluid property identification.

关 键 词：流体性质识别 深度学习 混合采样 ReLU-Softmax Dropout正则化 车排子油田 

分 类 号：TE348[石油与天然气工程—油气田开发工程]