大庆石油地质与开发2025，Vol.44Issue(4)：90-100,11.DOI:10.19597/J.ISSN.1000-3754.202404030
基于物理模型约束与注意力机制的卷积长短期记忆网络在火驱产量预测中的应用
Application of CNN-LSTM based on physical model constraint and attention mechanism in in-situ combustion production prediction
摘要
Abstract
In-situ combustion is a major strategic replacement technique for the development of heavy oil in China,with significant advantages of in-situ heat generation,high recovery rate and green low-carbon.The extremely com-plex subsurface physical and chemical processes cause difficulty to unify physical model and data mining method,resulting in much challenge in predicting in-situ combustion production.A physical model for predicting in-situ combustion oil production is derived based on material balance theory by combustion tube experiment.A convolu-tional long short-term memory network(CNN-LSTM)prediction model based on physical model constraint and atten-tion mechanism is established by combining deep learning algorithm and incorporating physical model constraint in-to loss function.The results show that absolute error,mean square error(MSE),root mean square error(RMSE)and mean absolute percentage error of CNN-LSTM method based on physical model constraint and attention mecha-nism are 0.094 8,0.011 3,0.106 4 and 15.28%,respectively,with lower errors and better fitting effect than other methods.The introduction of physical constraint effectively overcomes the lack of interpretability of complex neural networks as"black boxes,"while significantly improving accuracy and robustness of the model.This study can be used in production prediction of producers of in-situ combustion oilfields,providing basis for decision-making of in-situ combustion development of heavy oil.关键词
稠油油藏/火驱技术/卷积神经网络/物理实验模型/产量预测Key words
heavy oil reservoir/in-situ combustion technique/convolutional neural network/physical experiment model/production prediction分类
能源科技引用本文复制引用
袁元,赵仁保,徐浩天,赵邑镇,孙梓齐,杨凤祥,展宏洋,李海波,吕世瑶..基于物理模型约束与注意力机制的卷积长短期记忆网络在火驱产量预测中的应用[J].大庆石油地质与开发,2025,44(4):90-100,11.基金项目
国家自然科学基金项目"火线传播速度的预测和火腔在三维空间中演化的控制机制探索"(52274052) （52274052）
新疆维吾尔族自治区高层次人才引进项目"基于垂向火驱井网条件下超稠油的原位改质辅助开发机理研究"(JXDF0221). （JXDF0221）
