Accelerated Phase Equilibrium Predictions for Subsurface Reservoirs Using Deep Learning Methods
Accelerated Phase Equilibrium Predictions for Subsurface Reservoirs Using Deep Learning Methods
复制标题
使用深度学习方法加速地下储层相平衡预测
DOI：
10.1007/978-3-030-22747-0_47
复制
发表时间：
2019
期刊：
影响因子：
--
通讯作者：
Shuyu Sun
中科院分区：
文献类型：
--
作者：
Tao Zhang;Yiteng Li;Shuyu Sun
复杂组成的多相流体流动是一个日益引人注目的研究课题，相关的工程问题，包括全球变暖和温室效应、提高采收率和地下水污染治理等，越来越受到人们的重视。在研究多组分多相流的流动行为和相变之前，首先要准确预测流体混合物中存在的总相数，然后才能确定相平衡状态。本文提出了一种新的基于深度学习的快速预测方法。使用选定的VT动态闪存计算方案生成训练数据，并在激活函数上对网络结构进行深度优化。与文献中提出的加速汽液相平衡计算的机器学习技术相比，首先确定混合物中存在的总相数，然后再估计其他相平衡性质，因此我们不再需要确保混合物处于两相状态。我们的方法可以处理具有复杂组成的流体混合物，在我们的例子中有8种不同的组分，并且原始数据很大。对采用不同激活函数的不同神经网络的不同深度学习模型的预测性能进行分析，可以为以后的研究选择特征来构建类似工程问题的神经网络提供帮助。最后给出了一些结论和评论，以帮助读者了解我们的主要贡献和对未来相关研究的洞察。
Multiphase fluid flow with complex compositions is an increasingly attractive research topic with more and more attentions paid on related engineering problems, including global warming and green house effect, oil recovery enhancement and subsurface water pollution treatment. Prior to study the flow behaviors and phase transitions in multi-component multiphase flow, the first effort should be focused on the accurate prediction of the total phase numbers existing in the fluid mixture, and then the phase equilibrium status can be determined. In this paper, a novel and fast prediction technique is proposed based on deep learning method. The training data is generated using a selected VT dynamic flash calculation scheme and the network constructions are deeply optimized on the activation functions. Compared to previous machine learning techniques proposed in literatures to accelerate vapor liquid phase equilibrium calculation, the total number of phases existing in the mixture is determined first and other phase equilibrium properteis will be estimated then, so that we do not need to ensure that the mixture is in two phase conditions any more. Our method could handle fluid mixtures with complex compositions, with 8 different components in our example and the original data is in a large amount. The analysis on prediction performance of different deep learning models with various neural networks using different activation functions can help future researches selecting the features to construct the neural network for similar engineering problems. Some conclusions and remarks are presented at the end to help readers catch our main contributions and insight the future related researches.