李振华, 姜雨菲, 杜锋,等.基于LSTM-Transformer模型的突水条件下矿井涌水量预测[J].河南理工大学学报（自然科学版）,2026,45(1):77-85.

LI Z H, JIANG YU F, DU F,et al.Prediction of mine water inflow under water-inrush conditions based on a LSTM-Transformer model[J].Journal of Henan Polytechnic University(Natural Science) ,2026,45(1):77-85.

基于LSTM-Transformer模型的突水条件下矿井涌水量预测

李振华1,2,3,4, 姜雨菲1, 杜锋1,2,3,4, 王文强1

1.河南理工大学 能源科学与工程学院，河南 焦作  454000；2.煤炭安全生产与清洁高效利用省部共建协同创新中心，河南 焦作  454000；3.河南理工大学 河南省矿井水害防控及水资源利用工程技术研究中心，河南 焦作  454000；4.河南理工大学 郑州高等研究院，河南 郑州  451464

摘要: 目的 矿井涌水量精准预测对预防矿井水害和保障矿井安全生产具有重要意义，为精准预测矿井涌水量，构建适用于华北型煤田受底板L_1-4灰岩含水层和奥陶系灰岩含水层水害威胁的矿井涌水量预测模型。 方法 以河南某典型矿井的水文监测数据为基础，提出LSTM-Transformer模型。利用LSTM捕捉矿井涌水量的动态时序特征，通过Transformer的多头注意力机制分析含水层水位变化和矿井涌水量之间的复杂时序关联，构建水位动态变化驱动下的矿井涌水量精准预测框架。 结果 结果表明，LSTM-Transformer模型预测精度显著优于LSTM，CNN，Transformer和CNN-LSTM模型的，其均方根误差为20.91 m³/h，平均绝对误差为16.08 m³/h，平均绝对百分比误差为1.12%，且和单因素涌水量预测模型相比，水位-涌水量双因素预测模型预测结果更加稳定。 结论 LSTM-Transformer模型成功克服传统方法在捕捉复杂水文地质系统中水位-涌水量动态关联上的局限，为矿井涌水量动态预测提供可解释性强、鲁棒性好的解决方案，也为类似地质条件下矿井涌水量预测提供了新方法。

关键词:涌水量预测;水位动态响应;LSTM-Transformer耦合模型;时间序列预测;注意力机制;矿井安全生产

doi:10.16186/j.cnki.1673-9787.2025080014

基金项目:国家自然科学基金资助项目（U24B2041，52174073）；河南省高校科技创新团队支持计划项目（23IRTSTHN005）

收稿日期:2025/08/09

修回日期:2025/10/15

出版日期:2025-12-03

Prediction of mine water inflow under water-inrush conditions based on a LSTM-Transformer model

Li Zhenhua1,2,3,4, Jiang Yufei1, Du Feng1,2,3,4, Wang Wenqiang1

1.School of Energy Science and Engineering， Henan Polytechnic University， Jiaozuo  454000， Henan， China；2.Collaborative Innovation Center of Coal Work Safety and Clean High Efficiency Utilization， Jiaozuo  454000， Henan，  China；3.Henan Mine Water Disaster Prevention and Control and Water Resources Utilization Engineering Technology Research Center， Henan Polytechnic University， Jiaozuo  454000， Henan， China；4.Zhengzhou Institute for Advanced Research of Henan Polytechnic University， Zhengzhou  451464， Henan， China

Abstract: Objectives Accurate prediction of mine water inflow is crucial for preventing water hazard accidents and ensuring safe production. This study aims to construct a water inflow prediction model suitable for mines in North China-type coalfields affected by water hazards from the underlying L_1-4 limestone aquifer and Ordovician limestone aquifer. Methods Based on hydrogeological monitoring data from a typical coal mine in Henan Province， a coupled LSTM-Transformer model was proposed. The LSTM component captures the dynamic temporal characteristics of mine water inflow， while the multi-head attention mechanism of the Transformer analyzes the complex temporal correlation between aquifer water level variations and mine water inflow. This framework enables accurate prediction of mine water inflow driven by dynamic water level changes. Results The coupled LSTM-Transformer model significantly outperformed LSTM， CNN， Transformer， and CNN-LSTM models in prediction accuracy， with a root mean square error （RMSE） of 20.91 m³/h， mean absolute error （MAE） of 16.08 m³/h， and mean absolute percentage error （MAPE） of 1.12%. Furthermore， compared to the single-factor water inflow prediction model， the two-factor （water level and water inflow） prediction model showed greater stability. Conclusions The LSTM-Transformer coupled model successfully overcomes the limitations of traditional methods in capturing the dynamic water level-discharge relationship within complex hydrogeological systems. It provides a solution for dynamic mine water inflow prediction with strong interpretability and robustness， offering a novel methodology for predicting water inflow under similar geological conditions.

Key words:mine inflow prediction;dynamic response of water level;LSTM-Transformer coupled model;time series forecasting;attention mechanism;mine safety production