基于BP神经网络的咸水层CO_(2)封存注入能力预测方法

Prediction method of CO_(2)injectivity in saline aquifer based on BP neural network

本文采用数值模拟软件CMG-GEM,构建咸水层CO_(2)封存模型,并基于BP神经网络对CO_(2)注入能力进行预测。首先,在数值模拟中考虑了不同地质参数和操作参数,使用拉丁超立方抽样形成220组数值模拟方案,建立机器学习样本库;然后,使用数据归一化、十倍交叉验证等技术,形成CO_(2)注入能力预测的代理模型;最后,利用该模型分别测试Sleipner,Quest和Illinois碳封存项目的注入能力。研究结果表明:最优的BP神经网络的激活函数为satlin,最优的隐藏层神经元数量为10,最优的学习率为0.4;对于Sleipner,Quest和Illinois碳封存项目,训练集、测试集与全部数据的皮尔逊相关系数均超过0.95;使用该代理模型预测的Sleipner,Quest和Illinois碳封存项目年注气量与实际情况年注气量相对误差分别为6.85%,5.57%和3.26%。与传统数值模拟技术相比,该技术具备计算耗时短、准确率高的优点。

In this work,numerical simulation software CMG-GEM was used to build the model of CO_(2)storage in a saline aquifer,and CO_(2)injectivity was predicted based on BP neural network.Firstly,the effects of different geological parameters and operational parameters on CO_(2)injectivity were considered in the numerical simulation.Latin hypercube sampling(LHS)was used to form 220 groups of numerical simulation schemes and establish the machine learning database.Secondly,data normalization and tenfold cross-validation techniques were employed and the proxy model of prediction of CO_(2)injectivity formed.Finally,the proxy model was used to test Sleipner,Quest,and Illinois projects of CO_(2)storage.The results show that the optimal activation function is satlin,the optimal number of neurons in the hidden layer is 10,and the optimal learning rate of the BP neural network is 0.4.For the projects of Sleipner,Quest,and Illinois,the Pearson coefficients are all over 0.95 in the training set,testing set,and all the data set.The relative errors of the projects of Sleipner,Quest,and Illinois between the predicted annual gas injectivity and the actual annual gas injectivity are 6.85%,5.57%,and 3.26%,respectively.Compared with the traditional numerical simulation techniques,this technique has the advantages of short computation time and high accuracy.