基于BP神经网络的多相管流模型优选及应用分析  

BP neural network-based selection of the optimum model of multiphase flow in pipes and its application

作　　者：钟海全[1] 李颖川[1,2] 刘永辉[1] 李成见[3] 李伟[3] 

机构地区：[1]西南石油大学石油工程学院,成都610500 [2]西南石油大学油气藏地质及开发工程国家重点实验室,成都610500 [3]中海石油(中国)有限公司北京研究中心,北京100027

出　　处：《中国科技论文在线》2008年第11期873-878,共6页

基　　金：高等学校博士学科点专项科研基金项目(20060615002);四川省自然科学基金项目(2006ZD042)

摘　　要：应用91井次的油气井测试数据对Ansari,Hagedorn & Brown Revised,Duns&Ros,Gray以及Beggs&Brill Revised模型进行了评价分析,结果表明:最适合的模型主要有Hagedorn&Brown Revised模型和Ansari模型,其最佳井数分别占50.5%和35.2%。为了对特定的油井条件优选出最佳的计算模型,提出利用神经网络优选多相管流模型的方法,建立了BP神经网络模型,给出了完整的学习算法程序框图,提出采用动量算法修改隐含层权值,采用基于指数平滑的方法修改输出层权值,从而有效地减少了抖动。应用油气井测试数据及管流模型评价结果,研究了神经网络输入层、输出层及隐含层神经元个数对预测准确率的影响,结果表明:合理的BP神经网络模型能很好的预测最佳管流模型,其预测准确率可高达100%.A total of 91 wells were studied to evaluate several multiphase flow models, including Ansari, Hagedorn Brown Revised, Duns Ros, Gray and Beggs Brill Revised models. The results show that the optimal models are Hagedorn Brown Revised and Ansari models, which give the best description of multiphase flow for 50.5% and 35.2% of the wells, respectively. This study presents an approach to the selection of the optimum multiphase flow model for a given well by artificial neural networks. A BP neural network model is built, and the flowchart for training algorithm is provided. The momentum algorithm is adopted to modify the weight of the hidden layer, and an exponential smoothing method is used to adjust the weight of the output layer, which effectively reduces buffeting. The effects of the number of neurons in the input layer, hidden layer and output layer on the predicted accuracy are investigated through using the test data in wells and the evaluation results. The result shows that a good BP neural network model can predict the optimum multiphase flow model with the possible accuracy of 100%.

关 键 词：石油、天然气能 BP神经网络 管流模型 BP算法 模型优选 误差 

分 类 号：TE319[石油与天然气工程—油气田开发工程]