径向基神经网络预测天然气凝析油爆炸极限  

Prediction of explosion limit of natural gas condensate using radial basis function artificial neural networks

作　　者：范峥[1] 田润芝 景根辉 林亮 田磊[3] CHEN Shengchieh FAN Zheng;TIAN Run-zhi;JING Gen-hui;LIN Liang;TIAN Lei;CHEN Shengchieh(College of Chemistry and Chemical Engineering,Xi’an Shiyou University,Xi’an 710065,China;Xi’an Changqing Technology Engineering Co.,Ltd.,Xi’an 710018,China;Shaanxi Provincial Research and Design Institute of Petroleum and Chemical Industry,Xi’an 710054,China;College of Engineering,Virginia Commonwealth University,Richmond 23284,USA)

机构地区：[1]西安石油大学化学化工学院,西安710065 [2]西安长庆科技工程有限责任公司,西安710018 [3]陕西省石油化工研究设计院,西安710054 [4]弗吉尼亚联邦大学工学院,里士满23284

出　　处：《天然气化工—C1化学与化工》2020年第5期91-95,102,共6页Natural Gas Chemical Industry

基　　金：中国国家留学基金(201908610135);陕西省科学技术研究与发展计划项目(2016GY-150);西安石油大学研究生创新与实践能力培养项目(YCS19212063)。

摘　　要：为了对天然气凝析油爆炸极限进行准确预测,首先在不同组分含量和现场工况条件下测定了它的爆炸上、下限,并将此实验结果作为径向基人工神经网络训练、验证和测试的样本数据库,然后以天然气凝析油中的C5、C6、C7、C8、C9+质量分数,气相中的O2物质的量分数以及操作温度为输入信号,以爆炸上、下限为输出信号,建立了天然气凝析油爆炸极限预测模型。结果表明,当隐含层节点数为34时,7-34-2型径向基人工神经网络结构合理且准确度良好,经过2190次反复迭代后,该模型的均方误差0.0099小于允许收敛误差限0.0100,预测值和期望值近似呈线性,其在训练阶段、验证阶段与测试阶段的决定系数分别为0.9997、0.9998、0.9999,具有较高的相关性,同时,除了C9+质量分数外,天然气凝析油中的C5、C6、C7、C8质量分数,气相中的O2物质的量分数和操作温度对爆炸上、下限的影响非常显著,建议给予重点关注。本文可为天然气凝析油爆炸风险地有效识别、合理控制与及时消除提供科学、可靠的理论支撑和数据来源。For the sake of predicting the explosion limit of natural gas condensate accurately, firstly, the upper and lower explosion limits were detected under different component contents and working conditions. These obtained experimental results were collected as a sample database for training, verifying and testing of radial basis artificial neural networks. Afterward, with the mass contents of C5, C6, C7, C8, C9+ in natural gas condensate, O2 molar content in the gas phase, and the operating temperature as input signals, and the upper and lower explosion limits as output signals, a prediction model of explosion limit for natural gas condensate was established. The results showed that when the number of nodes in the hidden layer is 34, the radial basis artificial neural network with 7-34-2 type has a reasonable structure and good accuracy. The mean square error becomes 0.0099 after 2190 iterations,lower than a specific convergence tolerance of 0.0100. The predicted value was approximately linear with the desired value. The determination coefficients of training, verifying and testing stages were 0.9997, 0.9998 and 0.9999 respectively, demonstrating high relevance. Meanwhile, the mass contents of C5, C6, C7, C8 in the natural gas condensate, O2 molar content in the gas phase and the operating temperature except for C9+ mass content have significant effects on the upper and lower explosion limits, which deserves to be focused on. Therefore, this paper could provide scientific and reliable theoretical support and data source for effective identification, reasonable control and timely elimination of the explosion risk of natural gas condensate.

关 键 词：天然气凝析油 爆炸极限 径向基人工神经网络 预测 多因素方差分析 

分 类 号：TE38[石油与天然气工程—油气田开发工程]