【作者】 张毅；

【导师】 刘洋；

【作者基本信息】 中国石油大学(北京) ， 地球物理学， 2022， 硕士

【摘要】 油田的经济可行性取决于岩性和流体分布预测的质量和准确性,以及潜在储层的非均质性。准确识别岩性和流体是油气勘探和生产成功的关键。非常规资源勘探的兴起和常规区块的日益复杂,使得准确的岩性和流体预测变得更加关键。但是传统方法的岩性、流体识别具有依赖人工、主观性强、灵活性差等问题。为提高岩性及流体识别的精度和效率,本文研究了基于Conv LSTM神经网络的岩性、流体识别方法。考虑到地层的沉积作用是时序渐变的,而岩性和流体是地层沉积特征的响应,具有一定的时序特征,利用基于Long-Short-Term Memory(LSTM)神经网络改进的Conv LSTM神经网络进行岩性、流体的识别。主要工作流程如下:首先是对测井数据预处理及测井数据重构;然后进行地震弹性参数的反演及敏感性分析;接下来通过测井解释成果与井旁地震道岩性、流体敏感性较高的参数构建网络训练样本;最后将实际地震数据输入到神经网络得到预测剖面并进行结果分析。测试结果表明,使用NVIDIA RTX2070运算,训练处理过程大约需要1小时,岩性及流体的准确率分别达到90.2%和90.3%。说明Conv LSTM神经网络能够实现地震剖面岩性和流体的自动识别,并且预测的精度和效率也达到实际生产的标准。更多还原

【Abstract】 The economic viability of a field depends on the quality and accuracy of predictions of lithology and fluid distribution,as well as the heterogeneity of the underlying reservoir.Accurate identification of lithology and fluids is the key to successful oil and gas exploration and production.The rise of unconventional resource exploration and the increasing complexity of conventional blocks have made accurate lithology and fluid predictions even more critical.However,the traditional methods of lithology and fluid identification have problems such as relying on manual work,strong subjectivity,and poor flexibility.In order to improve the accuracy and efficiency of lithology and fluid identification,this thesis studies a lithologic fluid identification method based on Conv LSTM neural network.Considering that the deposition of the stratum is temporally gradual,and the lithology and fluid are the response of the stratum depositional characteristics,with certain temporal characteristics,the improved Conv LSTM neural network based on LSTM neural network is used to identify the lithologic fluid.The main work flow is as follows: First,preprocess logging data and reconstruct logging data;Then perform seismic elastic parameter inversion and sensitivity analysis;Construct the network training samples with high parameters;Finally,input the actual seismic data into the neural network to obtain the predicted profile and analyze the results.The test results show that using NVIDIA RTX2070 operation,the training process takes about 1 hour,and the accuracy of lithology and fluid reaches 90.2% and 90.3%,respectively.It shows that the Conv LSTM neural network can realize the automatic identification of the lithology and fluid of the seismic section,and the prediction accuracy and efficiency also reach the standard of actual production.更多还原