东北大学学报(自然科学版) ›› 2025, Vol. 46 ›› Issue (7): 163-170.DOI: 10.12068/j.issn.1005-3026.2025.20240197

• 智能矿山 • 上一篇    

基于ICEEMDAN与Attention-LSTM的矿山边坡位移预测

李荟(), 韩晓飞, 朱万成, 毛嘉石   

  1. 东北大学 资源与土木工程学院,辽宁 沈阳 110819
  • 收稿日期:2024-10-30 出版日期:2025-07-15 发布日期:2025-09-24
  • 通讯作者: 李荟
  • 作者简介:朱万成(1974—),男,新疆呼图壁人,东北大学教授,博士生导师.
  • 基金资助:
    “十四五”国家重点研发计划项目(2022YFC2903903);国家自然科学基金资助项目(52304167);中央高校基本科研业务费专项资金资助项目(N2301020);辽宁省自然科学基金联合基金资助项目(2023-MSBA-122)

Mine Slope Displacement Prediction Based on ICEEMDAN and Attention-LSTM

Hui LI(), Xiao-fei HAN, Wan-cheng ZHU, Jia-shi MAO   

  1. School of Resources & Civil Engineering,Northeastern University,Shenyang 110819,China.
  • Received:2024-10-30 Online:2025-07-15 Published:2025-09-24
  • Contact: Hui LI

摘要:

为了提高矿山边坡位移预测的精度,提出了1种基于改进自适应噪声完备集合经验模态分解(ICEEMDAN)、最小二乘拟合与融入注意力机制的长短时记忆网络的矿山边坡位移预测模型.首先,针对矿山边坡位移监测数据的时序性、非线性等特点,引入改进的模态分解方法将累积位移进行时频分解,获得趋势项、周期项和随机项,有效降低了数据复杂度.其次,针对趋势项预测,采用最小二乘拟合方法建立三次多项式回归预测模型;针对周期项预测,引入注意力机制区分不同时刻位移数据的重要程度,有效捕捉了长时间位移序列内部的依赖关系.最终,通过整合趋势项和周期项预测结果,获得矿山边坡累积位移的预测结果.以西藏巨龙铜矿边坡为例,测试了该模型的性能.结果表明,提出的边坡位移预测模型的均方根误差和平均绝对百分比误差分别为5.99 mm和5.94%,与传统长短时记忆网络模型相比,分别下降了51.30%和55.17%,预测精度显著提高.

关键词: 位移预测, 露天矿, 滑坡, 长短时记忆网络, 注意力机制

Abstract:

In order to improve the accuracy of mine slope displacement prediction, a mine slope displacement prediction model based on an improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN), least squares fitting method and long short-term memory network integrated into the attention mechanism was proposed. Firstly, considering the temporal and nonlinear characteristics of the mining slope displacement monitoring data, an improved mode decomposition method was employed to perform the time-frequency decomposition of cumulative displacement, resulting in trend, periodic, and random components, thereby effectively reducing the data complexity. Secondly, to predict the trend component, a cubic polynomial regression prediction model was developed by using the least squares fitting method. To predict the periodic component, an attention mechanism was introduced to distinguish the importance of displacement data at different times, which effectively captured the internal dependencies within the long-term displacement sequences. Finally, the predicted results of the trend and periodic components were integrated to obtain the cumulative displacement of the mining slope. Taking the slope of Julong Copper Mine in Xizang as an example, the performance of the proposed method was tested. The results demonstrate that the proposed slope displacement prediction model achieves a root mean square error (RMSE) of 5.99 mm and a mean absolute percentage error (MAPE) of 5.94%. The RMSE and MAPE decrease by 51.30% and 55.17% compared with the traditional LSTM model, respectively. These results highlight the significant improvement in prediction accuracy achieved by the proposed method.

Key words: displacement prediction, open-pit mining, landslide, long short-term memory(LSTM)network, attention mechanism

中图分类号: