Shale gas production prediction based on TCN-LSTM model

Journal Information

J. Hefei Univ. Tech. (Nat. Sci.), Sept.2025, Vol.48 No.9: 1259-1264, 1275

DOI: 10.3969/j.issn.1003-5060.2025.09.016

Authors

SHI Zhengzheng $ ^{1} $, LI Daolun $ ^{1} $, FU Ning $ ^{2} $, ZHANG Kang $ ^{2} $

(1. School of Mathematics, Hefei University of Technology, Hefei 230601, China; 2. Changqing Downhole Technology Company, CNPC Chuanqing Drilling Engineering Co., Ltd., Xi'an 710016, China)

Abstract and Keywords

Abstract: Accurate prediction of shale gas production helps to plan production schedules in advance and optimize production schemes. Current research on output prediction often requires long-term production data or a large amount of feature data. When the production time is short or the feature data is scarce, it is difficult to accurately predict the output. To this end, this paper proposes a temporal convolutional network-long short-term memory network (TCN-LSTM) model with an attention mechanism. This model is jointly trained using the production data of three wells. The TCN and LSTM modules extract local and global features respectively, which are then fused through the fully connected network. The attention mechanism is used to focus on key information, learn the flow laws from the existing well production data, and improve the prediction accuracy in scenarios where new wells have limited initial data. The results show that the multi-well joint prediction model outperforms the single-well prediction model in terms of accuracy and trend prediction. Based on the mean absolute error (MAE) evaluation index, the prediction accuracy increases by approximately four times. Additionally, the model reduces the reliance on long-term data and multiple features. It is of great significance in reservoir development.

Keywords: temporal convolutional network(TCN); long short-term memory network(LSTM); attention mechanism; production prediction; multi-well

作者信息

史峥峥 $ ^{1} $，李道伦 $ ^{1} $，付宁 $ ^{2} $，张康 $ ^{2} $

(1. 合肥工业大学数学学院, 安徽合肥 230601; 2. 中国石油川庆钻探长庆井下技术作业公司, 陕西西安 710016)

摘要和关键词

摘要: 准确预测页岩气产量有助于提前规划生产计划、优化生产方案。现有研究在进行产量预测时，往往需要长时间的生产数据或大量特征数据，当生产时间较短或特征数据较少时，难以准确预测产量。为此，文章提出一种具有注意力机制的时间卷积网络-长短期记忆网络（temporal convolutional network-long short-term memory network，TCN-LSTM）模型。该模型使用3口井生产数据联合训练，其中TCN和LSTM模块分别提取局部和全局特征，然后用全连接网络融合；并使用注意力机制聚焦关键信息，从已有井生产数据中学习流动规律，提高了对初期数据匮乏的新井的预测精度。结果表明，多井联合预测模型在精度和趋势预测方面均优于单井预测模型，基于平均绝对误差（mean absolute error，MAE）评估指标的预测精度提高了约4倍，并且减少了对长周期数据和多特征的依赖，在油藏开发中具有重要意义。

关键词: 时间卷积网络（TCN）；长短期记忆网络（LSTM）；注意力机制；产量预测；多井

Funding

国家自然科学基金资助项目（12172115；12372244）

合肥工业大学学报(自然科学版)

Navigation