多算法融合因子筛选的LSTM模型在大坝变形预测中的应用

doi:10.16265/j.cnki.issn1003-3033.2025.12.0198

摘要/Abstract

摘要：

针对传统大坝变形预测中影响因子选择方法单一、难以全面捕捉高维数据复杂因子关联性的问题,提出一种基于多算法融合的因子筛选方法,并应用于构建最优因子的长短期记忆(LSTM)网络模型。使用相关系数法、邻域成分分析(NCA)算法、最小绝对值收缩和选择算子法(LASSO)分别筛选影响因子,并合并筛选结果;由于因子间会存在相关性,再根据对称不确定性(SU)剔除高度相关的因子,实现因子优选,利用 LSTM 网络建立大坝变形预测模型;以新疆某混凝土面板堆石坝为例,以均方根误差(RMSE)、平均绝对误差(MAE)、均方误差(MSE)及决定系数(R²)作为评价指标评估模型性能。结果表明:相较于传统因子筛选方法,本文方法融合多种算法优势,能够全面、准确地筛选出对大坝变形影响显著的因子,其中MSE较其他方法降低20.11%~59.09%,RMSE降低10.61%~36.05%,MAE降低9.95%~37.86%,构建出更优的预测模型;部分测点的MSE、RMSE和MAE较其他对比模型最大降幅分别达53.5%、31.9%和34.7%,且决定系数最高达0.986 0。

关键词: 多算法融合, 因子筛选, 最优因子长短期记忆网络模型(LSTM), 大坝变形, 预测模型

Abstract:

To address the reliance on a single factor selection approach in traditional dam deformation prediction and the difficulty of comprehensively capturing complex inter-factor relationships among high-dimensional influencing factors, a factor screening method based on the fusion of multiple algorithms was proposed and applied to construct an LSTM model using optimally factors. Specifically, influencing factors were selected separately using the correlation coefficient method, neighborhood component analysis (NCA), and least absolute shrinkage and selection operator (LASSO) technique. The results from these individual methods were subsequently integrated. Since correlations exist among the factors, highly correlated ones were further eliminated using symmetrical uncertainty (SU), thereby an optimized factors set was obtained. The selected factor set was then used to develop a dam deformation prediction model via an LSTM network. A concrete-faced rockfill dam in Xinjiang was used as the case study. The model performance was evaluated using root mean square error (RMSE), mean absolute error (MAE), mean square error (MSE), and the coefficient of determination (R²). The results demonstrate that, compared with traditional factor selection methods, the proposed multi-algorithms are integrated and factors with significant influence on dam deformation are comprehensively and accurately identified. MSE is reduced by 20.11%-59.09%, RMSE by 10.61%-36.05%, and MAE by 9.95%-37.86%, and a superior predictive model was obtained, compared with models using conventional factor selection methods. For specific monitoring points, the maximum reductions in MSE, RMSE, and MAE reach 53.5%, 31.9%, and 34.7%, respectively, while the highest R² value attains 0.986 0.

Key words: multi-algorithm fusion, feature selection, long short-term memory(LSTM), dam deformation, prediction model

中图分类号:

X928.03事故预防与预测

尹孝君, 丁勇, 李登华. 多算法融合因子筛选的LSTM模型在大坝变形预测中的应用[J]. 中国安全科学学报, 2025, 35(12): 129-138.

YIN Xiaojun, DING Yong, LI Denghua. Application of LSTM model with multi-algorithm fusion factor screening in dam deformation prediction[J]. China Safety Science Journal, 2025, 35(12): 129-138.

图/表 11

图1

图2

图3

表1

表2

表3

图4

表4

图5

表5

表6

参考文献 22

[1]	吴中如. 水工建筑物安全监控理论及其应用[M]. 北京: 高等教育出版社, 2003:1.
[2]	周仁练, 苏怀智, 韩彰, 等. 混凝土坝变形的长期预测模型与应用[J]. 水力发电学报, 2021, 40(9):122-131.
	ZHOU Renlian, SU Huaizhi, HAN Zhang, et al. Long-term deformation prediction model of concrete dams and its application[J]. Journal of Hydroelectric Engineering, 2021, 40(9): 122-131.
[3]	WILLM G, BEAUJOINT N. Les mÉthodes de surveillance des barrages au service de la production hydraulique d'ElectricitÉ de France, problèmes anciens et solutions nouvelles[C]. IXth International Congress on Large Dams, 1967: 529-550.
[4]	GAMSE S, OBERGUGGENBERGER M. Assessment of long-term coordinate time series using hydrostatic-season-time model for rock-fill embankment dam[J]. Structural Control and Health Monitoring, 2017, 24(1): DOI:10.1002/stc.1859.
[5]	KARLUPIA N, ABROL P. Wrapper-based optimized feature selection using nature-inspired algorithms[J]. Neural Computing and Applications, 2023, 35(17): 12 675-12 689. doi: 10.1007/s00521-023-08383-6
[6]	赵佳蕊, 王玲芝, 李晨阳. 基于LEA-LSTM的光伏发电功率短期预测方法[J]. 电力信息与通信技术, 2024, 22(11):34-42.
	ZHAO Jiarui, WANG Lingzhi, LI Chenyang. A short-term prediction method of photovoltaic power generation based on LEA-LSTM[J]. Electric Power Information and Communication Technology, 2024, 22(11):34-42.
[7]	罗璐, 李志, 张启灵. 大坝变形预测的最优因子长短期记忆网络模型[J]. 水力发电学报, 2023, 42(2):24-35.
	LUO Lu, LI Zhi, ZHANG Qiling. Optimal factor set based long short-term memory network model for prediction of dam deformation[J]. Journal of Hydroelectric Engineering, 2023, 42(2):24-35.
[8]	郑好, 李登华, 丁勇. 基于集成因子优选算法的面板堆石坝变形预测模型[J]. 排灌机械工程学报, 2025, 43(2):178-186. doi: 10.3969/j.issn.1674-8530.23.0262
	ZHENG Hao, LI Denghua, DING Yong. Deformation prediction model of panel rockfill dam based on integrated factor optimization algorithm[J]. Journal of Drainage and Irrigation Machinery Engineering, 2025, 43(2): 178-186. doi: 10.3969/j.issn.1674-8530.23.0262
[9]	SU Yan, WENG Kailiang, LIN Chuan, et al. Dam deformation interpretation and prediction based on a long short-term memory model coupled with an attention mechanism[J]. Applied Sciences, 2021, 11(14): DOI:10.3390/app11146625.
[10]	蒙伟, 何川, 陈子全, 等. 岭回归在岩体初始地应力场反演中的应用[J]. 岩土力学, 2021, 42(4):1156-1169.
	MENG Wei, HE Chuan, CHEN Ziquan, et al. Application of ridge regression in the inversion analysis of the initial geo-stress field of rock masses[J]. Rock and Soil Mechanics, 2021, 42(4):1156-1169.
[11]	袁冬阳, 顾冲时, 顾昊. 严寒地区混凝土重力坝变形行为分析与预测模型[J]. 水利学报, 2022, 53(6):733-746.
	YUAN Dongyang, GU Chongshi, GU Hao. Displacement behavior analysis and prediction model of concrete gravity dams in cold region[J]. Journal of Hydraulic Engineering, 2022, 53(6):733-746.
[12]	徐韧, 苏怀智, 杨立夫. 基于GP-XGBoost的大坝变形预测模型[J]. 水利水电科技进展, 2021, 41 (5): 41-46,70.
	XU Ren, SU Huaizhi, YANG Lifu. Dam deformation prediction model based on GP-XGBoost[J]. Advances in Science and Technology of Water Resources, 2021, 41 (5): 41-46,70.
[13]	LIU Mingkai, WEN Zhiping, SU Huaizhi. Deformation prediction based on denoising techniques and ensemble learning algorithms for concrete dams[J]. Expert Systems with Applications, 2024, 238:DOI:10.1016/j.eswa.2023.122022.
[14]	何裕坤, 晁阳, 李同春, 等. 基于SBO-LSTM的大坝变形预测模型[J]. 水利水电技术:中英文, 2024, 55 (增1): 78-86.
	HE Yukun, CHAO Yang, LI Tongchun, et al. Dam deformation prediction model based on SBO-LSTM[J]. Water Resources and Hydropower Engineering, 2024, 55(S1): 78-86.
[15]	吉丽萍, 黄景涛, 李一凡, 等. 一种基于多信息融合的风电功率预测特征选择方法[J]. 兰州理工大学学报, 2024, 50 (5): 94-100.
	JI Liping, HUANG Jingtao, LI Yifan, et al. A wind power prediction feature selection method based on multi-information fusion[J]. Journal of Lanzhou University of Technology, 2024, 50 (5): 94-100.
[16]	XIONG Rui, SUN Yue, WANG Chenxu, et al. A data-driven method for extracting aging features to accurately predict the battery health[J]. Energy Storage Materials, 2023, 57: 460-470. doi: 10.1016/j.ensm.2023.02.034
[17]	高浩田, 李东喜, 陈泽华, 等. 基于交互式多特征融合算法的药物靶标预测[J]. 太原理工大学学报, 2024, 55(4): 751-758.
	GAO Haotian, LI Dongxi, CHEN Zehua, et al. Prediction of drug-target interactions based on interactive multi-feature fusion algorithm[J]. Journal of Taiyuan University of Technology, 2024, 55(4): 751-758.
[18]	王晓楠, 苏文浩, 董灵波. 应用特征选择和机器学习方法建立兴安落叶松单木树龄预测模型[J]. 东北林业大学学报, 2024, 52 (11): 64-71,82.
	WANG Xiaonan, SU Wenhao, DONG Lingbo. Establishing a prediction model for the age of Larix gmelinii individual tree using feature selection and machine learning methods[J]. Journal of Northeast Forestry University, 2024, 52(11): 64-71, 82.
[19]	毛杰, 陈宓舟, 许磊, 等. 基于双重选择LASSO模型的我国股市定价因子边际有效性研究[J]. 系统工程理论与实践, 2024, 44 (9): 2993-3008. doi: 10.12011/SETP2023-0831
	MAO Jie, CHEN Mizhou, XU Lei, et al. Research on the marginal effectiveness of Chinese stock markets'pricing factors: application of double-selection LASSO model[J]. Systems Engineering-Theory & Practice, 2024, 44 (9): 2993-3008.
[20]	刘天翼, 艾星星, 张九丹. 基于MLR-DE-LSTM的大坝变形串联组合预测模型[J]. 中国农村水利水电, 2025(2):207-212.
	LIU Tianyi, AI Xingxing, ZHANG Jiudan. Dam deformation serial combined prediction model based on MLR-DE-LSTM[J]. China Rural Water and Hydropower, 2025(2):207-212.
[21]	史云峰, 韩晓明, 王利红, 等. 基于特征选择和机器学习的锂电池寿命早期预测[J]. 现代电子技术, 2024, 47 (22): 90-98.
	SHI Yunfeng, HAN Xiaoming, WANG Lihong, et al. Lithium battery life early prediction based on feature selection and machine learning[J]. Modern Electronics Technique, 2024, 47 (22): 90-98.
[22]	YU Lei, LIU Huan. Feature selection for high-dimensional data: a fast correlation-based filter solution[C]. Proceedings of the 20^th International Conference on Machine Learning (ICML-03),2003: 856-863.

筛选方法	相关系数法	LASSO 回归	NCA法	合并的全部因子	再筛选后的最优因子集
J-06	H₁~H₃ T₁~T₃	H₃ T₁~T₄ I₅	H₁、H₃ T₃、T₄ I₂、I₇	H₁~H₃ T₁~T₄ I₂、I₅、I₇	H₁ T₁~T₄ I₂、I₅
J2-02-1	H₁、H₃ T₁~T₃ I₁	H₂、H₃ T₁~T₄ I₅	H₁、H₂ T₃、T₄ I₁、I₄	H₁~H₃ T₁~T₄ I₁、I₄、I₅	H₁~H₃ T₁~T₃ I₁、I₄

测点	SU阈值	保留因子数	RMSE	MAE	R²
J-06	0.7	9	0.721 4	0.592 1	0.964 0
	0.75	8	0.688 2	0.557 8	0.970 3
	0.8	7	0.665 0	0.532 4	0.974 8
	0.85	5	0.673 1	0.542 0	0.971 5
	0.9	4	0.702 8	0.562 3	0.965 9
J2-02-1	0.7	10	0.823 5	0.698 7	0.962 6
	0.75	9	0.793 2	0.670 1	0.967 1
	0.8	8	0.785 7	0.662 5	0.974 2
	0.85	6	0.791 4	0.668 9	0.970 4
	0.9	4	0.812 9	0.681 5	0.963 8

测点	筛选方法	MSE	RMSE	MAE	R²
J-06	未筛选	0.600 8	0.775 1	0.641 3	0.965 8
	相关系数法	0.507 2	0.712 2	0.505 9	0.971 1
	随机森林	0.442 3	0.665 0	0.532 4	0.974 8
	LASSO回归	0.416 3	0.64 52	0.524 5	0.976 3
	未再筛选	0.312 0	0.558 5	0.442 5	0.982 2
	本文方法	0.245 7	0.495 7	0.398 5	0.986 0
J2-02-1	未筛选	0.866 1	0.930 6	0.721 8	0.963 8
	相关系数法	0.753 3	0.867 9	0.758 3	0.968 5
	随机森林	0.617 3	0.785 7	0.662 5	0.974 2
	LASSO回归	0.612 1	0.782 4	0.635 5	0.974 4
	未再筛选	0.566 9	0.752 9	0.591 6	0.976 3
	文中方法	0.452 9	0.673 0	0.509 0	0.981 1

测点	预测模型	MSE	RMSE	MAE	R²
J-06	XGBoost	0.529 4	0.727 6	0.610 3	0.969 8
	随机森林	0.467 1	0.683 4	0.496 9	0.973 4
	LightGBM	0.430 9	0.656 4	0.503 4	0.975 5
	GRU	0.329 3	0.573 8	0.445 0	0.981 2
	LSTM	0.245 7	0.495 7	0.398 5	0.986 0
J2-02-1	XGBoost	0.764 3	0.874 2	0.635 8	0.968 1
	随机森林	0.600 7	0.775 0	0.627 1	0.974 9
	LightGBM	0.565 7	0.752 1	0.594 2	0.976 4
	GRU	0.547 1	0.739 6	0.590 4	0.977 2
	LSTM	0.452 9	0.673 0	0.509 0	0.981 1

数据来源	MSE	RMSE	MAE	R²
原文测试集	0.245 7	0.495 7	0.398 5	0.986 0
第1折	0.248 3	0.498 3	0.401 2	0.985 6
第2折	0.251 7	0.501 7	0.403 8	0.985 1
第3折	0.243 1	0.493 1	0.395 2	0.986 5
第4折	0.253 4	0.503 4	0.406 8	0.984 7
第5折	0.249 6	0.499 6	0.401 9	0.985 5
交叉验证平均	0.249 2± 0.003 9	0.499 2± 0.003 9	0.401 8± 0.004 4	0.985 5± 0.000 7