Prediction of Tailings Dam Displacement Based on RIME Optimization Algorithm and CNN-Attention-BiLSTM Model: Tailings Pond in Nujiang Prefecture, Yunnan Province
1 Faculty of Public Security and Emergency Management, Kunming University of Science and Technology, Kunming 650093, China
2 Yunnan Jinding Zinc Industry Co., Ltd., Nujiang 671400, China
3 Faculty of Land Resource Engineering, Kunming University of Science and Technology, Kunming 650093, China
Show less
文章历史+
Received
Published
Revised
2025-10-18
2026-06-18
2025-12-16
Online Date
2026-08-04
PDF (11606K)
摘要
为实现对尾矿坝位移的精准预测,针对目前预测中存在的数据预处理粗糙、特征选择不准确、模型参数欠优及组合模型预测效果较差的问题,提出一种基于霜冰优化算法(rime optimization algorithm,RIME)和卷积神经网络(convolutional neural network,CNN)-注意力机制(Attention)-双向长短期记忆神经网络(bidirectional long short-term memory,BiLSTM)相结合的尾矿坝位移预测模型。首先运用自适应噪声完备集合经验模态分解算法(complete ensemble empirical mode decomposition with adaptive noise,CEEMDAN)将坝体累计位移分解为趋势项位移和波动项位移,并利用优化高斯过程回归(improved Gaussian process regression,IGPR)模型预测趋势项位移;其次通过随机森林(random forest,RF)-斯皮尔曼(Spearman)融合的特征选择方法,筛选出4个关键特征因子作为使用RIME参数优化的CNN-Attention-BiLSTM模型的输入,波动项位移为输出;最后将IGPR趋势项预测结果与RIME-CNN-Attention-BiLSTM波动项预测结果叠加得到尾矿坝累计位移。以云南省某尾矿坝为例进行实例验证,结果表明:坝体累计位移预测结果R2为0.997,MAE为0.546 mm,RMSE为0.628 mm,与真实值基本一致,误差小。可见所提方法预测精度高,能够显著提高对坝体复杂位移变化情况的预测水平,为尾矿坝安全预警以及决策提供依据。
Abstract
To achieve precise prediction of tailings dam displacement and address existing issues such as rough data preprocessing, inaccurate feature selection, suboptimal model parameters, and poor prediction performance of combined models, a novel prediction model integrating the rime optimization algorithm (RIME) and the convolutional neural network (CNN)-Attention-BiLSTM model was proposed. Firstly, the cumulative displacement of the dam body was decomposed into trend and fluctuation displacements using the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) algorithm. The trend displacement was then predicted using the improved Gaussian process regression (IGPR) model. Subsequently, four key feature factors were selected through a feature selection method combining random forest (RF) and Spearman correlation, and these factors were used as the input for the CNN-Attention-BiLSTM model optimized by RIME parameters, with fluctuation displacement as the output. Finally, the cumulative displacement of the tailings dam was obtained by superimposing the trend prediction results from IGPR and the fluctuation prediction results from the RIME-CNN-Attention-BiLSTM model. The model was validated using a tailings dam in Yunnan Province, China. The prediction results show that the coefficient of determination (R2) is 0.997, the mean absolute error (MAE) is 0.546 mm, and the root mean square error (RMSE) is 0.628 mm, indicating a high degree of consistency with the actual values and minimal error. These results demonstrate that the proposed method has high prediction accuracy and can significantly improve the prediction of complex displacement changes in the dam body, providing early warnings and decision-making support for tailings dam safety.
WANG Ze-ye, YUAN Li-wei, YANG Si-mei, LI Su-min, HE Shu-hua, LUO Pan-wei.
Prediction of Tailings Dam Displacement Based on RIME Optimization Algorithm and CNN-Attention-BiLSTM Model: Tailings Pond in Nujiang Prefecture, Yunnan Province[J].
Science Technology and Engineering, 2026, 26(17): 7534-7544 DOI:10.12404/j.issn.1671-1815.2507296
目前基于多模型组合的深度学习位移预测主要聚焦于滑坡和边坡,曾子健等[3]利用CRITIC(criteria importance though intercrieria correlation)赋权法-粒子群算法(partical swarm optimization,PSO)-支持向量回归(support vector regression,SVR)相结合的模型实现对滑坡位移的有效预测。Song等[4]将经验模态分解(empirical mode decomposition,EMD)、改进PSO和长短期记忆神经网络(long short-term memory,LSTM)模型结合从而提高了滑坡位移的预测精度。许江波等[5]提出了一种基于最大信息系数(maximal information coefficient, MIC)-极端梯度提升(extreme gradient boosting,XGBoost)-LSTM的组合模型,并成功准确预测了边坡的位移。Wang等[6]将细菌觅食优化算法(bacterial foraging optimization algorithm,BFOA)结合最小二乘支持向量机(least squares support vector machine,LSSVM)的多变量时间序列预测模型应用于位移预测,结果表明预测结果准确,能够实现边坡风险的预测和预警。与滑坡和边坡的位移预测研究现状相比,针对尾矿坝的预测模型相对较少,且结构较为简单,难以适应尾矿坝位移预测所面临的复杂情况。近年来,开始有学者探索将多模型组合的深度学习预测模型应用在尾矿坝的位移上。杨玉好等[7]提出结合多头注意力机制(multi-head attention mechanism,MHA)-双向长短期记忆神经网络(bidirectional long short-term memory,BiLSTM)的位移预测模型。Liu等[8]采用主成分分析(principal component analysis,PCA)和移动平均(moving average,MA)分别与LSTM相结合,得到LSTM-PCA和LSTM-MA两种耦合预测模型,研究结果表明多模型组合模型的预测效果较以往模型预测能力明显提高。使用复杂模型进行预测时模型参数的选择非常重要,因此朱明远等[9]建立了粒子群优化算法结合支持向量机的深度学习模型,并应用在大坝的变形预测上。刘迪等[10]将改进麻雀搜索算法(multi strategy improved sparrow search algorithm,MISSA)、卷积神经网络(convolutional neural networks,CNN)和BiLSTM 3种模型结合用来预测坝体位移。两位学者将优化算法结合到模型中提高了模型参数选择和预测的准确性。尾矿坝位移数据的预处理直接影响模型的最终预测效果,所以有学者引入算法对数据进行预处理。Zhu等[11]提出基于EMD-LSTM-注意力机制(attention mechanism,Attention)的长期变形预测方法。唐宇峰等[12]将改进自适应噪声完备集合经验模态分解算法(improved complete ensemble empirical model decomposition adaptive noise,ICEEMDAN)和麻雀搜索优化算法(sparrow search algorithm,SSA)-LSTM-Attention模型相结合,对尾矿坝位移进行预测。Hao等[13]融合自适应噪声完备集合经验模态分解算法(complete ensemble empirical model decomposition adaptive noise,CEEMDAN)-鲸鱼优化算法(whale optimization algorithm,WOA)-Elman神经网络,得到一种混合区间预测模型。三位学者的研究都证明经过对数据预处理后再预测能够充分提高预测精度。上述学者都在尾矿坝位移预测方面取得了很好的进展,但是目前针对尾矿坝位移的预测仍普遍存在数据预处理粗糙、特征选择不准确、所选参数不能充分发挥模型能力、组合模型预测效果较差、没有同时兼顾上述不足等问题。
鉴于此,现首先针对数据预处理问题,使用CEEMDAN将坝体累计位移分解为趋势项位移和波动项位移,降低原始数据的复杂程度,并利用优化高斯过程回归(improved Gaussian process regression,IGPR)模型预测趋势项位移。因为模型的特征输入对预测结果的影响重大,所以为防止特征选择不准确,提出随机森林(random forest,RF)-斯皮尔曼(Spearman)相融合的特征选择方法,通过学习坝体位移和各特征因子之间复杂的非线性关系筛选出4个关键特征因子作为后续波动项位移模型的输入。深度学习模型参数的选取是最关键的环节,所以使用具有多方面优势的霜冰优化算法来选取模型各部分关键参数,以确保组合模型发挥最佳性能,组合模型搭建框架为CNN-Attention-BiLSTM。首先使用CNN模块进行输入特征的信息提取;最后连接Attention模块学习提取信息的重要性,聚焦关键部分并给予高权重后输入BiLSTM模块,该模块具有双向的强时序学习能力;使用RIME-CNN-Attention-BiLSTM组合模型对波动项位移进行预测,并将趋势项预测结果与波动项预测结果叠加得到尾矿坝累计位移。所提出分解-预测-重构的试验思路,兼顾目前位移预测中普遍存在的问题,并通过实例证明该方法的可行性和组合模型的适用性,以期实现对尾矿坝位移的精准预测,为尾矿坝安全提供预警以及决策提供依据。
WangKun, YangXiuzhi, WangDi, et al. Research progress and development trend of tailings dam failure analysis and simulation methods[J]. Nonferrous Metals (Chinese Edition), 2025, 15(8): 1396-1407.
ZengZijian, XiaoHui, XuHaning, et al. Landslide surface displacement prediction based on CRITIC weight method and PSO-SVR model[J]. Science Technology and Engineering, 2024, 24(35): 14940-14946.
[7]
SongK, YangH, LiangD, et al. Step-like displacement prediction and failure mechanism analysis of slow-moving reservoir landslide[J]. Journal of Hydrology, 2024, 628: 16.
XuJiangbo, HouXinmin, WuXiong, et al. Slope displacement prediction based on MIC-XGBoost-LSTM model[J]. China Journal of Highway and Transport, 2024, 37(10): 38-48.
[10]
WangX, BaiZ, LuY, et al. Displacement prediction of tunnel entrance slope based on LSSVM and bacterial foraging optimization algorithm[J]. Scientific Reports, 2024, 14(1): 25819.
YangYuhao, YangBin, HuJun, et al. Prediction of tailings dam displacement based on MHA-BiLSTM[J]. Nonferrous Metals Engineering, 2024, 14(10): 147-157.
[13]
LiuW, PanJ, RenY, et al. Coupling prediction model for long -term displacements of arch dams based on long short -term memory network[J]. Structural Control and Health Monitoring, 2020, 27(7): 15.
ZhuMingyuan, WuYan, RuanXinmin, et al. Deformation prediction model of roller compacted concrete dam based on PSO-SVM[J]. Hydropower Engineering, 2022, 48(3): 64-69.
LiuDi, YangHui, LuCaiwu, et al. Prediction of displacement of tailings dams based on MISSA-CNN-BiLSTM model[J]. China Safety Science Journal, 2024, 34(9): 145-154.
[18]
ZhuY, GaoY, WangZ, et al. A tailings dam long-term deforma tion prediction method based on empirical mode decomposition and LSTM model combined with attention mechanism[J]. Water, 2022, 14(8): 1229.
TangYufeng, ChenXinghong, CaiYu, et al. Prediction of tailings dam displacement based on time series decomposition and SSA-LSTM-Attention model[J]. Science Technology and Engineering, 2023, 23(29): 12753-12759.
[21]
HaoT, XuK, ZhengX, et al. A tailings dam displacement interval prediction model based on time series decomposition and an improved Elman neural network model[J]. Stochastic Environmental Research and Risk Assessment, 2025, 39: 5959-5976.
XuJiayuan, ZouLei, ZhangLiping, et al. Monthly runoff prediction in the upper reaches of the Yangtze River based on the CEEMDAN-FESC-OVMD-Transformer coupling model[J]. Water Resources Protection, 2025, 41(4): 197-209.
[24]
SuH, ZhaoD, AliA H, et al. RIME: a physics-based optimization[J]. Neurocomputing, 2023, 532: 183-214.
ZhuTing, YanQisheng. Wind speed prediction based on modal decomposition and RIME-CNN-BiLSTM-AM[J]. Science Technology and Engineering, 2025, 25(20): 8514-8525.
[27]
HubelD H, WieselT N. Receptive fields of single neurones in the cat’s striate cortex[J]. Journal of Physiology-London, 2009, 587(12): 2722-2732.
LiShiwan, YuanMingdao, XuYunqian, et al. CNN-LSTM seepage quantity prediction model of earth-rock dam based on attention mechanism[J]. Science Technology and Engineering, 2025, 25(21): 9102-9108.
XiangWenqiang, ShaoWei, LuoShaohui, et al. Simulation of soil moisture in permafrost region of Qinghai-Tibet Plateau based on LSTM model[J]. Science Technology and Engineering, 2025, 25(31): 13236-13245.
QiuHongsheng, YangYu, LiDongjian, et al. Subgrade settlement prediction based on equidistance processing and IOWHA operator[J]. Science Technology and Engineering, 2022, 22(20): 8884-8892.
SongChao, ZhaoTengyuan, XuLing. Estimation method of rock uniaxial compressive strength based on Bayesian Gaussian process regression and model selection[J]. Chinese Journal of Geotechnical Engineering, 2023, 45(8): 1664-1673.
ShenFu, ZhangWei, XuXiaoyuan, et al. Research on distribution network topology identification based on random forest and maximum information coefficient key feature selection[J]. Power System Protection and Control, 2024, 52(17): 1-11.
[38]
SpearmanC. The proof and measurement of association between two things[J]. International Journal of Epidemiology, 2010, 39(5): 1137-1150.
LiHao, ZhaoQing, CuiChenzhou, et al. Classification algorithm of stellar spectra based on CNN and LSTM composite deep model[J]. Spectroscopy and Spectral Analysis, 2024, 44(6): 1668-1675.