基于LSTM循环神经网络的岩性识别方法

Lithology identification based on LSTM recurrent neural network

针对复杂碳酸盐岩储层岩石组分复杂、岩性多样,常规测井岩性识别方法受限等问题,提出利用长短期记忆神经网络(LSTM)提高岩性识别效果的方法,并结合实际数据进行验证和应用效果分析。考虑到常规机器学习方法在岩性识别中无法充分利用沉积岩石在深度域序列上的潜在信息,从而基于LSTM方法构建了能够提取和学习岩性沉积序列特征的岩性识别手段。以苏里格气田苏东地区下古生界碳酸盐岩储层为例,通过敏感性分析选取自然伽马、光电吸收截面指数、密度、声波时差、补偿中子和电阻率等6种测井参数,构建了基于LSTM的岩性识别模型。结果表明,与朴素贝叶斯,KNN,决策树,SVM和HMM等传统方法相比,LSTM的岩性识别准确率提升幅度介于1.40%~12.25%。高精度的LSTM岩性识别模型为复杂碳酸盐岩储层的表征和评价提供了数据基础。

A lithology recognition method by long-short-term memory neural network(LSTM)was proposed for complex carbonate reservoirs with complex composition and diverse lithology,to overcome obstacles troubling traditional identification,and effective results were showed with a case from gas field.Due to the inadequate ability of general machine learning methods in extracting the characteristics of sedimentary sequence,the LSTM method was introduced into the improvement for lithology identification.Taking the Lower Paleozoic carbonate reservoir in eastern block of Sulige gas field as an example,six sensitive parameters were selected to construct a lithology identification model based on LSTM,such as GR,Pe,DEN,RLLD,AC and CNL.The results show that lithology identification accuracy based on LSTM increases by 1.40%-12.25%above traditional models(Naive Bayes,KNN,Decision Tree,SVM and HMM),and can provide more reliable support for the characterization and evaluation of complex carbonate reservoirs.