摘要
为解决声波测井曲线缺失或失真问题,使用传统方法重构测井曲线时,往往导致测井曲线精度不够。深度学习具有很强的数据表征能力,但建立模型面临超参数设定的不确定性和时间成本问题。为此,将异步连续减半算法(ASHA)与长短期记忆神经网络(LSTM)模型相结合,设计实现了一种基于超参数优化LSTM的声波测井曲线生成技术,对缺失或失真曲线进行补全。以大庆油田6口井为例,首先通过相关性分析,优选自然伽马、密度、补偿中子曲线作为输入特征量搭建LSTM学习模型,然后采用ASHA对LSTM模型进行超参数调优,并与常见的贝叶斯优化、粒子群优化算法进行时效及精度对比,最后将调优得到的超参数组合应用于LSTM模型,并与多元回归、GRU、BILSTM 3种模型进行对比。该技术的应用结果表明:ASHA算法能更加高效准确地确定模型超参数,节省时间与人力成本,提高建模效率。基于ASHA优化的LSTM模型生成的声波测井曲线精度更高,该技术具有较好的适用性和精确性。
The accuracy of traditional reconstruction methods is often insufficient to solve the problem of data missing or distortion on acoustic log curves.Deep learning has a strong ability of data characterization,but model building suffers from hyperparameter uncertainties and time cost.To solve these problems,the asynchronous successive halving algorithm(ASHA)is combined with the long short-term memory neural network(LSTM)to formulate hyperparameter optimized LSTM for data reconstruction.A case study in Daqing Oilfield involves 6 wells.Through a correlation analysis,natural gamma ray,density and compensated neutron are selected as input characteristic parameters to build the LSTM learning model,for which hyperparameter optimization is performed using the ASHA.In view of efficiency and accuracy,we compare ASHA optimization with Bayesian optimization and particle swarm optimization.The portfolio of optimized hyperparameters is finally applied to the LSTM model.Compared with multiple regression,GRU and BILSTM models,the ASHA can determine model hyperparameters with improved efficiency and accuracy and less time and labor costs.The ASHA optimized LSTM model could reconstruct acoustic log curves with high accuracy.
作者
刘建建
周军
余卫东
陈江浩
樊琦
鄢高韩
LIU Jianjian;ZHOU Jun;YU Weidong;CHEN Jianghao;FAN Qi;YAN Gaohan(Logging Technology Research Institute,China National Logging Corporation,PetroChina,Xi’an 710077,China)
出处
《石油物探》
CSCD
北大核心
2024年第5期1061-1074,共14页
Geophysical Prospecting For Petroleum
基金
中国石油集团测井有限公司课题“CPLog智能远程采集软件系统研发与应用”(CPL2021-A07)资助。
