一种基于一维卷积神经网络的试井模型智能识别方法  

An intelligent method for identifying well testing models based on one-dimensional convolutional neural network

作　　者：齐占奎 张新鹏 刘旭亮 查文舒[4] 李道伦[4] QI Zhankui;ZHANG Xinpeng;LIU Xuliang;ZHA Wenshu;LI Daolun(Daqing Well Logging Technology Service Company,Daqing,Heilongjiang 163453,China;The Third Oil Production Plant of PetroChina Huabei Oilfield Company,Hejian,Hebei 062450,China;School of Artificial Intelligence and Big Data,Hefei University,Hefei,Anhui 230601,China;Department of Applied Mathematics,Hefei Technology University,Hefei,Anhui 230026,China)

机构地区：[1]中国石油大庆油田测试技术服务分公司,黑龙江大庆163453 [2]中国石油华北油田公司第三采油厂,河北河间062450 [3]合肥大学人工智能与大数据学院,安徽合肥230601 [4]合肥工业大学数学学院,安徽合肥230026

出　　处：《油气井测试》2024年第2期72-78,共7页Well Testing

基　　金：国家自然科学基金项目(1217020361);大庆油田项目“水驱常规试井资料智能解释系统研究”(dqp-2020-cs-ky-002)。

摘　　要：为提高试井分析工作效率,实现试井模型的自动识别,提出了基于一维卷积神经网络(1D CNN)的试井模型智能识别方法。根据实测数据的特点,提出基于理论曲线构建样本库的原则与方法,并构建了4种常用油藏模型的训练样本库;建立了一维卷积神经网络模型,将样本库中双对数曲线的压力变化和压力导数数据作为输入,油藏类别作为网络输出训练及优化网络,总识别准确率可达99.16%,敏感度均在98%以上。经4口井实例应用,正确识别试井模型的概率大于0.99,与二维卷积神经网络相比,1D CNN显著降低了计算复杂度和时间成本,加快了训练速度。这表明基于试井理论所构建的样本库是有效的,能满足实测数据模型识别的需求;同时证明了方法的有效性、实用性和普适性。In order to improve the efficiency of well testing analysis and achieve the automatic identification of well testing models,an intelligent identification method based on one-dimensional convolutional neural network(1D CNN)was proposed.According to the characteristics of measured data,principles and methods based on constructing a sample library using theoretical curves were proposed,and training sample libraries for four commonly used reservoir models were constructed.A one-dimensional convolutional neural network model was established.By using the pressure variation data and pressure derivative data of double logarithmic curves in the sample library as inputs,and the reservoir category as the output for training and optimizing the network,the overall identification accuracy can reach 99.16%,with sensitivities all above 98%.Through application cases from four wells,the probability of correctly identifying well testing models is greater than 0.99.Compared with two-dimensional convolutional neural networks,1D CNN significantly reduces computational complexity and time costs,which speeds up the training process.This indicates that the sample library constructed based on well testing theory is effective and can meet the requirements of model identification for measured data,and simultaneously,it also demonstrates the effectiveness,practicality,and universality of this method.

关 键 词：试井模型 一维卷积神经网络 智能识别 深度学习 自动解释 模型识别 样本库 

分 类 号：TE353[石油与天然气工程—油气田开发工程]