基于注意力机制的TCN-LSTM的页岩气产量递减动态预测研究

Research on Dynamic Prediction of Shale Gas Production Decline Based on an Attention-Based TCN-LSTM

摘要: 针对传统递减曲线模型在复杂生产动态刻画方面的局限，以及常规机器学习方法对样本规模依赖较强、中长期外推能力不足的问题，本文构建了融合时间卷积网络(TCN)、长短时记忆网络(LSTM)与注意力机制的ATT-TCN-LSTM页岩气产量预测模型。首先对油嘴调整、气举等生产制度造成的异常产量进行校正，并采用小样本微量扰动方式扩充数据集；随后结合随机森林与主成分分析(PCA)识别并降维地质、工程主控参数；在此基础上，采用AdamW优化器和余弦退火学习率完成模型训练，并利用MSE、MAE和MAPE评价预测效果。结果表明，ATT-TCN-LSTM能够较好刻画页岩气井全生产周期的产量变化，在拟合精度和中长期预测能力方面均优于基础LSTM和TCN-LSTM模型。研究形成的主控因素分析与时序预测方法可为页岩气单井动态配产和压裂参数优化提供参考。

Abstract: In response to the limitations of traditional decline curve models in characterizing complex production dynamics, as well as the strong dependence of conventional machine learning methods on sample size and insufficient ability to extrapolate in the medium to long term, this paper constructs an ATT-TCN-LSTM shale gas production prediction model that integrates temporal convolutional networks (TCN), long short-term memory networks (LSTM), and attention mechanisms. Firstly, the abnormal production caused by production regimes such as nozzle adjustment and gas lift is corrected, and the dataset is expanded using small sample micro disturbance method; Subsequently, random forest and principal component analysis (PCA) are combined to identify and reduce the dimensionality of geological and engineering control parameters; On this basis, the model was trained using the AdamW optimizer and cosine annealing learning rate, and the prediction performance is evaluated using MSE, MAE, and MAPE. The results show that ATT-TCN-LSTM can better characterize the production variations of shale gas wells throughout the entire production cycle, and is superior to the basic LSTM and TCN-LSTM models in terms of fitting accuracy and medium- to long-term prediction ability. The main controlling factor analysis and time series prediction methods developed in this study can provide reference for the dynamic production allocation and fracturing parameter optimization of single shale gas wells.