基于循环神经网络和数据差分处理的油田产量预测方法  

A method for oilfield production prediction based on recurrent neural network and data differential processing

作　　者：高亚军 唐力辉 王振鹏 谢晓庆 徐海涛[3] GAO Yajun;TANG Lihui;WANG Zhenpeng;XIE Xiaoqing;XU Haitao(CNOOC Research Institute Ltd.,Beijing 100028,China;State Key Laboratory of Offshore Oil Exploitation,Beijing 100028,China;CNOOC EnerTech-Drilling&Production Co.,Tianjin 300459,China)

机构地区：[1]中海油研究总院有限责任公司,北京100028 [2]海洋石油高效开发国家重点实验室,北京100028 [3]中海油能源发展股份有限公司工程技术分公司,天津300459

出　　处：《中国海上油气》2023年第3期126-136,共11页China Offshore Oil and Gas

摘　　要：油田产量预测是油田开发生产过程中关键问题之一,机器学习的数据驱动技术在时间序列的石油产量预测应用中较为广泛。针对早期神经网络对油田时序信息不敏感、预测值容易出现持续偏差的问题,建立了基于长短期记忆(LSTM)和门控递归单元(GRU)两种循环神经网络的产量预测模型,克服了传统方法在产量预测方面的的局限性。引入了回归插补法降噪和差分的数据处理方式,较好地解决了油田生产数据存在异常点、阶段性波动特征的问题,大幅提高了LSTM和GRU神经网络在油田产量预测的精度。利用实际油田数据对建立的模型进行训练和评价,并与传统的支持向量机(SVM)、集成学习回归(AdaBoost)、反向传播(BP)神经网络、循环神经网络(RNN)预测结果进行了对比,发现LSTM和GRU循环神经网络模型具有较高的预测精度,能较好地应用于石油产量的时间序列预测。Oilfield production prediction is one of the key issues in the process of oilfield development.The data-driven technology of machine learning is widely used in the prediction of oil production based on time series.In order to solve the problem that the early neural network is not sensitive to the time series information of the oilfield and the predicted value is prone to continuous deviation,the production prediction model based on two kinds of recurrent neural networks which are Long Short-Term Memory(LSTM)and Gated Recurrent Unit(GRU)is established to overcomes the limitations of traditional methods in production prediction.The methods of noise reduction by regression interpolation and data differential processing are innovatively proposed to commendably solve the problem of abnormal points and periodic fluctuations of production data,greatly improving the accuracy of LSTM and GRU neural networks in oilfield production prediction.The model is trained and evaluated using actual oilfield data,and comparing with the prediction results of support vector machine(SVM),ensemble learning regression(AdaBoost),back propagation(BP)neural network and recurrent neural network(RNN),it is found that LSTM and GRU recurrent neural network models can obtain a higher prediction accuracy,and can be well applied to oil production prediction based on time series.

关 键 词：产量预测 循环神经网络 数据差分处理 回归插补法降噪 

分 类 号：TP328[自动化与计算机技术—计算机系统结构]