基于残差神经网络模型的高压气井水合物生成预测方法

Prediction of Natural Gas Hydrate Formation in High-pressure Gas Wells Based on Residual Network

综合天然气多组分混合比例、临界压力和温度等因素影响,构建跳过连接的卷积神经网络模型,提出利用残差神经网络(ResNet)预测高压气井多元体系天然气水合物生成的方法。井场试验结果表明,相较于传统全连接神经网络(FCNN)与小波神经网络(WNN),基于大数据集参数自动调优和动态微调残差神经网络模型的多元体系水合物生成预测效果较优且吻合程度较高,预测均方误差由WNN的0.006~3.417和FCNN的0.008~3.722降至ResNet的0.001~1.020,并可量化分析单组分体系以及二元和多组分体系水合物相平衡实际压力随温度动态变化关系。

Taking into account factors such as component mixing ratio,critical pressure,and temperature in multi-component system of high-pressure gas wells,a connection-skipping convolutional neural network model is constructed,and the generation conditions of natural gas hydrate are predicted using residual neural network(ResNet).The well site test results show that compared to traditional fully connected neural network(FCNN)and wavelet neural network(WNN),the prediction result of hydrate generation in the multi-component system using the residual neural network(ResNet)which is based on the automatic tuning and dynamic fine-tuning of network parameters is more accurate.It makes the mean square error of prediction results decreased to 0.001~1.020 from 0.006~3.417 of WNN and 0.008~3.722 of FCNN.Residual neural networks can also be used to quantitatively analyze the dynamic relationship between the actual phase equilibrium pressure and temperature of single component and multi-component hydrate systems.