机理约束下钻井机械钻速智能预测泛化方法  

A generalization method of intelligent ROP prediction under mechanism constraints

作　　者：祝兆鹏 朱林 宋先知[1] 李永钊 张仕民[1,2] 柯迪丽娅·帕力哈提 张诚恺 王超尘 ZHU Zhaopeng;ZHU Lin;SONG Xianzhi;LI Yongzhao;ZHANG Shimin;Kediliya Palihati;ZHANG Chengkai;WANG Chaochen(China University of Petroleum-Beijing,Beijing 102249,China;Center of Advanced Oil and Gas Equipment,China University of Petroleum-Beijing,Beijing 102249,China;GWDC Drilling Engineering and Technology Research Institute,Panjin,Liaoning 124010,China;Engineering Technology Research Institute of PetroChina Xinjiang Oilfield Company,Karamay,Xinjiang 834000,China)

机构地区：[1]中国石油大学(北京),102249 [2]中国石油大学(北京)高端油气装备智能设计与制造研究中心 [3]中国石油长城钻探工程公司工程技术研究院,102249 [4]中国石油新疆油田公司工程技术研究院

出　　处：《天然气工业》2024年第9期179-189,共11页Natural Gas Industry

基　　金：国家重点研发计划项目“复杂油气智能钻井理论与方法”(编号:2019YFA0708300);国家自然科学基金委员会杰出青年科学基金项目“油气井流体力学与工程”(编号:52125401);中国石油科技创新基金项目“油气钻井破岩智能监测与优化调控技术”(编号:2022DQ02-0308)。

摘　　要：钻井机械钻速的准确预测可辅助油气井钻井前科学配置资源,对制订更加合理的钻井作业方案以及钻井提效、降本增效具有重要现实意义。智能化预测钻井机械钻速已成为行业研究热点,为解决常规智能模型在不同井间迁移能力较差的问题,在对综合录井数据进行降噪、补全等预处理的基础上,利用钻井专业知识构造约束条件,引入了域对抗神经网络(DANN),建立了机械钻速模型在不同井间的迁移机制,结合滑动窗口、增量更新与实时录井数据,形成了机械钻速模型随井下工况的实时更新方法。研究结果表明:(1)数据层约束和网络层约束均可提高智能模型的精度与稳定性,且双机理约束下的BP模型相比于普通BP模型预测精度明显提高;(2)基于域对抗神经网络的机械钻速预测模型可有效地将邻井(源域)数据知识迁移到测试井(目标域);(3)基于增量学习算法建立的双滑动窗口数据更新机制,使模型实时适应地下钻进环境变化,预测精度和泛化能力进一步提升;(4)机理约束、迁移训练与实时更新对模型泛化性能的强化作用具有叠加效应,新井机械钻速预测平均相对误差降低至20.2%。结论认为,建立的机械钻速预测模型及迁移方法相较于传统钻速预测模型,具有更好的迁移性和更高的准确度,减少了迁移过程中重复训练时间,为机械钻速智能预测提供了新的思路和方向。Accurate prediction of the rate of penetration(ROP)is supportive to rational resource allocation before well drilling,and has important practical significance in making proper drilling plan and increasing drilling efficiency at lower cost.Intelligent ROP prediction has become a hot spot in research.However,conventional intelligent models show poor transferability between wells.In this paper,comprehensive logging data are preprocessed by denoising and completion techniques.Then,constraints are constructed depending on drilling knowledge.By introducing the domain adversarial neural network(DANN),the transfer mechanism of ROP models between wells is established.Finally,based on sliding window,incremental updating and real-time logging data,a method for real-time update of the ROP model with downhole conditions is formed.The following results are obtained.First,both data layer constraint and network layer constrain can improve the accuracy and stability of the intelligent model,and the BP model under two mechanism constraints yields much more accurate results than ordinary BP model.Second,the DANN-based ROP model can effectively transfer the data knowledge from the offset well(source domain)to the test well(target domain).Third,the double-sliding-window data updating mechanism based on the incremental learning algorithm allows the model to adapt to the subsurface drilling conditions in a real-time manner,with further enhanced predication accuracy and generalization ability.Fourth,mechanism constraints,transfer learning,and real-time updates play a combined role in strengthening the generalization performance of the model.The average relative error of ROP prediction for new wells is reduced to 20.2%.It is concluded that the proposed ROP model yields better transferability and higher accuracy than traditional models,providing a new approach and direction for intelligent prediction of ROP.

关 键 词：机械钻速 机理约束 域对抗神经网络 迁移学习 增量更新 模型泛化 

分 类 号：TE242[石油与天然气工程—油气井工程]