基于长短时记忆神经网络LSTM的煤体结构识别方法

煤体结构的定量识别是煤和煤层气等资源勘探开发中的关键性问题之一。本次研究以鄂尔多斯盆地A地区煤岩为研究对象,选取5条常规测井曲线井径CAL、自然伽马GR、声波时差AC、补偿中子CNL和密度DEN曲线,构建基于LSTM神经网络的煤体结构识别模型,实现了煤体结构的定量识别。研究表明：随着煤层被破坏程度的增大,构造煤具有DEN曲线和GR曲线值偏低,AC曲线、CNL曲线和CAL曲线值偏大的特征；LSTM网络模型参数是影响模型识别准确度的重要因素,通过多次试验得出网络训练的迭代次数(Epoch)、LSTM神经网络包含的隐藏单元数(numHiddenUits)和数据最小分块尺寸(MinBatchSize)参数的最优值分别为800、32和28；基于LSTM神经网络的煤体结构识别模型准确度达到85.5%,并且利用未参与模型构建的验证井进行可靠性的验证,认为该方法可有效识别煤体结构,基本满足实际生产的需求,对于相似地区煤体结构的识别也具有一定的借鉴意义。

Coal structure recognition method based on LSTM neural network

Quantitative identification of coal structure is one of the key issues in the exploration and development of coal and coalbed methane resources. In this study, coal rock in A area of Ordos Basin is taken as the research object. Five conventional logging curves, including caliper CAL, natural gamma GR, acoustic time difference AC, compensated neutron CNL and density DEN, are selected to build A coal structure recognition model based on LSTM neural network and achieve quantitative identification of coal structure. The results show that with the increase of the degree of coal seam damage, the structural coal has the characteristics of low DEN curve and GR curve, and high AC curve, CNL curve and CAL curve. LSTM network model parameters are important factors affecting model recognition accuracy. The optimal values of Epoch, numHiddenUits and MinBatchSize of LSTM neural network are 800, 32 and 28 respectively. The accuracy of the coal structure recognition model based on LSTM neural network reached 85.5%, and the reliability of the verification Wells not involved in the model construction was verified. It is considered that the method can effectively identify the coal structure, basically meet the needs of actual production, and has certain reference significance for the identification of coal structure in similar areas.