数据-知识混合驱动的压裂球座坐封有效性智能诊断方法  

Intelligent Diagnosis for Effectiveness of Data-Knowledge Mixed-Driven Fracturing Ball Seat Setting

作　　者：盛茂[1] 范龙昂 张帅[1] 张彦军[2,3] 李川[2,3] 田守嶒[1] SHENG Mao;FAN Longang;ZHANG Shuai;ZHANG Yanjun;LI Chuan;TIAN Shouceng(National Key Laboratory of Petroleum Resources and Engineering(China University of Petroleum(Beijing)),Beijing,102249,China;Oil and Gas Technology Research Institute,PetroChina Changqing Oilfield Company,Xi'an,Shaanxi,710018,China;National Engineering Laboratory for Exploration and Development of Low-Permeability Oil&Gas Fields,Xi'an,Shaanxi,710018,China)

机构地区：[1]油气资源与工程全国重点实验室(中国石油大学(北京)),北京102249 [2]中国石油长庆油田分公司油气工艺研究院,陕西西安710018 [3]低渗透油气田勘探开发国家工程实验室,陕西西安710018

出　　处：《石油钻探技术》2024年第5期76-81,共6页Petroleum Drilling Techniques

基　　金：国家自然科学基金优秀青年科学基金项目“油气井流体力学与工程”(编号:52122401)资助。

摘　　要：水平井桥塞分段压裂时的桥塞球座坐封有效性实时诊断是其关键环节,若球座坐封失效,将无法正常进行后续作业,目前主要依靠人工观察井口压力变化特征,难以快速准确判识。为此,融合专家经验定性判识和坐封数据特征挖掘定量标注,滑动窗口数据切片形成5792组标签数据,优选井口压力–排量二维输入的长短期记忆神经网络,建立了压裂球座坐封有效性智能诊断模型,并采用欠采样平衡数据集方式提升模型预测精度。结果表明,井口压力呈现显著的陡升—陡降—平缓上升的三阶段特征,若井口压力缺失某个阶段特征,则为坐封失效;井口压力斜率统计值分布范围较大,无法形成明确规则实现准确诊断。采用人工智能技术学习不同井口压力形态的有效/无效坐封数据特征,实现了每秒输出诊断结果,测试集准确率96.8%,验证集准确率84.3%。研究结果为桥塞球座坐封有效性实时自动诊断提供了方法。Real-time diagnosis of the effectiveness of the bridge plug ball seat setting is a key step in the staged fracturing of horizontal wells.If the ball seat setting fails,follow-up operations cannot proceed normally.Currently,manual observation of wellhead pressure changes is primarily relied upon,making it difficult to quickly and accurately identify key characteristics.To address this,a combination of expert qualitative judgment and quantitative feature mining of setting data was implemented.Sliding window data was segmented to form 5792 sets of labeled data.A long short-term memory(LSTM)neural network,using a two-dimensional input of wellhead pressure and displacement,was selected.An intelligent diagnosis model for evaluating the effectiveness of the fracturing ball seat setting was established,utilizing an under-sampling balanced dataset to improve the model’s prediction accuracy.The results show that the setting data exhibits a clear three-stage characteristic:a steep rise,a steep drop,and a gentle rise in wellhead pressure.If the wellhead pressure lacks any of these stage characteristics,it indicates an invalid setting.The wellhead pressure slope exhibits a wide distribution range,making it difficult to form explicit rules for accurate diagnosis.Artificial intelligence technology is used to learn the valid/invalid setting data characteristics from various wellhead pressure forms,producing diagnosis results per second with an accuracy of 96.8% for the test set and 84.3% for the validation set.The findings are expected to provide a method for real-time and automatic diagnosis of the effectiveness of the bridge plug ball seat setting.

关 键 词：水力压裂 球座坐封 智能诊断 数据挖掘 深度学习 

分 类 号：TE357.1[石油与天然气工程—油气田开发工程]