摘要
气井积液是大部分天然气井在开发中后期存在的主要问题,柱塞气举是目前最有效的排水采气工艺之一。精确预测柱塞在井筒中的运动规律有助于柱塞气举井进行故障诊断、关键设计参数优选及工作制度优化。柱塞在井筒中的运动属于复杂的变速运动,其受到多种因素影响,目前所建立的柱塞气举瞬态运动模型存在较多的假设条件,计算需要大量迭代,无法快速、准确地预测其运动规律。传统数据驱动方法缺乏实际物理意义,预测结果可能违背物理规律。为此,通过收集井场大量实际井监测数据,分析不同变量间的相互关系,深入挖掘柱塞瞬态运动过程中的物理规律,提出了一种双阶段物理引导神经网络(PGNN),在预测柱塞速度和井口压力过程中,添加关键物理变量约束作为网络的神经元并加入到损失函数中,使其训练过程受到物理条件约束,再采用现场实际生产数据对建立的DNN-PGNN(物理引导深度神经网络)、LSTM-PGNN(物理引导长短期记忆神经网络)、BiLSTM-PGNN(物理引导双向长短期记忆神经网络)以及未添加物理引导的DNN(深度神经网络)网络进行对比。研究结果表明,所建立的机理与数据双驱动柱塞气举瞬态仿真模型准确性最高,相比于未加入物理约束的深度神经网络均方根误差(RMSE)降低83.78%,平均绝对误差(MAE)降低86.90%。将预测结果与实测数据对比可知,在深度学习网络中加入物理约束可提高模型的准确性、物理一致性以及泛化性。
Liquid loading in gas wells is a main problem existed in the middle-late development stage of most gas wells,and plunger gas lift is one of the most effective water withdrawal and gas recovery technologies.Accurate prediction on the motion law of plunger in wellbore is helpful for fault diagnosis,key design parameter optimization and working system optimization of plunger gas lift wells.The motion of plunger in wellbore is a complex variable motion due to be affected by many factors.The currently established transient motion model for plunger gas lift involves numerous assumptions,requiring extensive iterations for calculations,which hinders rapid and accurate prediction of its motion law.Traditional data-driven methods lack actual physical significance,and their prediction results may violate physical laws.Therefore,by collecting monitoring data from a large number of real wells at well site,analyzing the relationship between different variables,and deeply exploring the physical laws in the transient motion process of the plunger,a two-stage physics-guided neural network(PGNN)was proposed.In the process of predicting plunger speed and wellhead pressure,key physical variable constraints were added as network neurons and incorporated into the loss function,ensuring the training process is constrained by physical conditions.Finally,field actual production data were used to compare the established DNN-PGNN with LSTM-PGNN,BiLSTM-PGNN and the DNN network without physical guidance.The research results show that the established mechanism and data dual-driven plunger gas lift transient simulation model achieves the highest accuracy,with root mean square error(RMSE)reduced by 83.78% and mean absolute error(MAE)reduced by 86.90% compared to deep neural networks without physical constraints.The comparison between predicted results and measured data shows that incorporating physical constraints into deep learning networks enhances the accuracy,physical consistency and generalization of the model.
作者
邢志晟
韩国庆
贾友亮
路鑫
田伟
龚航飞
张田
Xing Zhisheng;Han Guoqing;Jia Youliang;Lu Xin;Tian Wei;Gong Hangfei;Zhang Tian(College of Petroleum Engineering,China University of Petroleum(Beijing);Oil and Gas Technology Institute,PetroChina Changqing Oilfield Company;China Petroleum Technology&Development Corporation)
出处
《石油机械》
北大核心
2025年第12期119-128,共10页
China Petroleum Machinery
基金
国家自然科学基金青年科学基金项目“致密油储层压裂后渗吸动用程度微观控制机理研究”(52204059)。
关键词
气井积液
柱塞气举
物理引导神经网络
瞬态仿真
机理数据双驱动
井口压力
liquid loading in gas well
plunger gas lift
physics-guided neural network
transient simulation
mechanism and data dual-driven
wellhead pressure
