生产动态预测对于页岩气资源开发制定有效的气井管理策略、优化生产以及最大化经济回报具有重要作用，传统的递减分析及产能模型在页岩气产能预测方面具有一定局限性，以8口北美页岩气水平井为研究对象，将页岩气井的生产数据处理成时间序列，采用长短期记忆网络（long short-term memory, LSTM）的机器学习和贝叶斯优化方法，较为准确地实现了页岩气井返排动态预测，并筛选出了产量预测的关键参数。研究结果表明：井底流压和日产水量是LSTM模型预测页岩气产量的关键输入参数；随着预测天数的增加，LSTM模型的准确性有所下降；研究结果可为时间序列分析方法在页岩气井动态预测方面的应用提供重要参考。

Production forecasting plays an important role in well management, optimizing production, and maximizing economic returns in shale gas development. Traditional rate-decline analysis and analytical models face limitations in forecasting shale gas production. Eight horizontal shale gas wells from North America were selected. The production data of these wells were processed into time series. A long short-term memory (LSTM) network combined with Bayesian optimization was applied to predict shale gas well flowback performance. Key parameters for production forecasting were identified. The results show that flowing bottomhole pressure and daily water rate are the most important input features. The prediction accuracy of the LSTM model generally decreases as the forecast period increases. These findings provide valuable insights into the application of time series analysis methods in production forecasting of shale gas wells.