摘要
为预测油田增油控水措施效果中月产油量与含水量,提出一种基于混合空洞卷积神经网络(HDCNN:Hybrid Dilated Convolutional Neural Network)-BIGRU-Attention的措施效果预测模型。模型通过HDCNN,提取生产数据多尺度全局特征;针对措施生产数据时序性较强与波动性较大的特点,利用双向门控循环单元(BIGRU:Bidirectional Gated Recurrent Unit)充分挖掘数据间长期依赖关系,提高时序信息利用率与学习效果;引入缩放点积注意力模块(Attention),为重要信息赋予较高权重并不断调整参数使模型始终关注与预测目标相关性较大的特征。为验证模型的有效性,将LSTM(Long Short-Term Memory)、CNN(Convolutional Neural Network)-LSTM以及LSTM-Attention作为实验对比,结果表明该模型具有更低的预测误差与更好的泛化能力。
Measure planning is the main method to increase oil and control water in oilfield.In order to accurately predict the effect of various measures to increase oil production,a measure effect prediction model based on HDCNN(Hybrid Dilated Convolutional Neural Network)-BIGRU-Attention is proposed with monthly oil production and water content as the prediction targets.The model extracts multi-scale global features of production data through HDCNN.Aiming at the characteristics of strong timing and large volatility of measure production data,the BIGRU(Bidirectional Gated Recurrent Unit)is used to fully mine the long-term dependence between data to improve the utilization rate of time series information and the learning effect.The scaled dot-product attention mechanism(Attention)is introduced,and the weight adjustment strategy is used to make the network focus on the feature dimension with large correlation with the prediction target.In order to verify the effectiveness of the proposed model,LSTM(Long Short-Term Memory),CNN(Convolutional Neural Network)-LSTM and LSTM-attention are taken as experimental comparisons.The results show that the proposed model has lower prediction error and better generalization ability.
作者
张强
李志溢
邓彬
ZHANG Qiang;LI Zhiyi;DENG Bin(School of Computer and Information Technology,Northeast Petroleum University,Daqing 163318,China)
出处
《吉林大学学报（信息科学版）》
CAS
2023年第4期631-638,共8页
Journal of Jilin University（Information Science Edition）
基金
国家自然科学基金资助项目(61702093)
黑龙江省自然科学基金资助项目(F2018003)
黑龙江省博士后专项基金资助项目(LBH-Q20077)。
关键词
油田措施效果预测
双向门控循环单元
混合空洞卷积神经网络
缩放点积注意力机制
prediction of oilfield measures effect
bidirectional gated recurrent unit(BIGRU)
hybrid dilated convolutional neural network(HDCNN)
scaled dot-product attention mechanism
