基于梯度提升回归树的气井油管积液高度预测  

Prediction of liquid accumulation height in oil pipelines of gas wells using gradient boosting regression trees

作　　者：向华[1] 夏文龙 刘波涛[1,2] 孔梦婷 张玉祥 杨浩波 XIANG Hua;XIA Wenlong;LIU Botao;KONG Mengting;ZHANG Yuxiang;YANG Haobo(School of Computer Science,Yangtze University,Jingzhou 434023,Hubei;Hubei Key Laboratory of Oil and Gas Drilling and Production Engineering(Yangtze University),Wuhan 430101,Hubei)

机构地区：[1]长江大学计算机科学学院,湖北荆州434023 [2]油气钻采工程湖北省重点实验室(长江大学),湖北武汉430101

出　　处：《长江大学学报（自然科学版）》2024年第5期94-101,共8页Journal of Yangtze University(Natural Science Edition)

基　　金：国家自然科学基金项目“页岩气藏水平井压裂缝网流动表征及反演优化方法研究”(52004033);油气钻采工程湖北省重点实验室开放基金项目“机器学习在页岩气井井筒积液预测及泡沫排水适用性诊断中的应用研究”(YQZC202402)。

摘　　要：气井油管积液高度预测是气藏开发的重要环节,更是排水采气不可或缺的一部分。气井开采后期,气井底部会出现积液聚集现象,积液过多会造成气井停产,为了避免停产问题,必须对气井油管积液高度进行预测,但传统石油工程模型预测气井油管积液高度,存在着具体计算需要大量经验参数等问题。提出一个基于梯度提升回归树模型预测气井油管积液高度的方法,以气井的套压、油压、油管下深、油层中深、日产气、日产水、井口温度7种生产数据为特征,采用集成学习方法,结合多个决策树的预测结果,以迭代逐步改进的方式来提高模型的整体性能,从而精确预测气井油管积液高度。通过与32口井仪器探测实测值、回归决策树和随机森林对比分析,梯度提升回归树模型预测值与实测值相符,预测效果也最好,平均相对误差仅3.87%,调整后的相关系数R2为0.85。梯度提升回归树模型与现有的油管内积液量和环空积液量预测模型相比较,平均相对误差降低了1.9%。Prediction of liquid accumulation height in gas well tubing is a crucial aspect of gas reservoir development and an indispensable part of drainage gas recovery.In the late stages of gas well exploitation,liquid accumulation occurs at the bottom of the well,and excessive liquid accumulation can lead to well shutdown.To mitigate this issue,it is essential to predict the liquid accumulation height in the gas well tubing.However,traditional petroleum engineering models for predicting gas well tubing liquid accumulation height face challenges,including the need for a substantial amount of empirical parameters in specific calculations.This paper proposes a method based on the Gradient Boosting Regression Tree(GBRT)model to predict the liquid accumulation height in gas well tubing.The approach utilizes seven production data features,including annular pressure,oil pressure,tubing depth,reservoir depth,daily gas production,daily water production,and wellhead temperature.Employing ensemble learning,the method combines predictions from multiple decision trees in an iterative stepwise manner to enhance the overall performance of the model,thereby accurately predicting the liquid accumulation height in gas well tubing.Through comparative analysis with measurements from 32 instrument-detected wells,regression decision trees,and random forests,the GBRT model exhibits good agreement between predicted and actual values,demonstrating superior predictive performance.The average relative error is only 3.87%,and the adjusted R 2 is 0.85.Compared to existing models for predicting liquid accumulation in the tubing and annulus,the model proposed in this paper reduces the average relative error by 1.9%.

关 键 词：气井积液 预测模型 机器学习 梯度提升回归树 

分 类 号：TE32[石油与天然气工程—油气田开发工程]