基于PSO-SAE-XGB的多模型水文预报

祝宾皓; 李向东; 李涛; 窦身堂; 李珂; 孙周亮; 谷少闯

doi:10.13476/j.cnki.nsbdqk.2026.0079

南水北调与水利科技（中英文） ›› 2026, Vol. 24 ›› Issue (4) : 842 -852. DOI: 10.13476/j.cnki.nsbdqk.2026.0079

水文水资源

基于PSO-SAE-XGB的多模型水文预报

祝宾皓 ¹^,³^,⁴ ,
李向东 ² ,
李涛 ¹^,³^,⁴ ,
窦身堂 ¹^,³^,⁴ ,
李珂 ¹^,³^,⁴ ,
孙周亮 ¹^,³^,⁴ ,
谷少闯 ¹^,³^,⁴

作者信息 +

Multi-model hydrological forecasting based on PSO-SAE-XGB

Binhao ZHU ¹^,³^,⁴ ,
Xiangdong LI ² ,
Tao LI ¹^,³^,⁴ ,
Shentang DOU ¹^,³^,⁴ ,
Ke LI ¹^,³^,⁴ ,
Zhouliang SUN ¹^,³^,⁴ ,
Shaochuang GU ¹^,³^,⁴

Author information +

文章历史 +

PDF (2146K)

摘要

为解决水文预报中多源径流驱动机制复杂、模型结构不确定性突出等问题，采用5种机器学习模型对径流结果进行预测，利用堆叠自编码器（stacked autoencoder, SAE）对多模型预测结果进行深层特征提取，基于极端梯度提升树（eXtreme gradient boosting, XGB）进行非线性集成预报，引入粒子群算法 (particle swarm optimization, PSO)对SAE和XGB的超参数进行联合优化，构建PSO-SAE-XGB多模型耦合预报方法。以湟水河流域宝库河牛场断面为研究对象，使用逐日降雨径流数据进行训练与验证。结果表明：PSO-SAE-XGB耦合模型在验证期的纳什效率系数（Nash-Sutcliffe efficiency coefficient，E_NS）达到0.976，均方根误差（root mean square error，E_RMS）为0.840，优于单一模型及未引入SAE的PSO-XGB模型，尤其在中高流量段表现出更强的拟合能力与泛化性能。该预报方法有效融合了多模型互补信息与深度学习特征提取优势，研究成果可为复杂环境下的径流预报提供可靠技术方法。

Abstract

To address problems in hydrological forecasting, such as complex multi-source runoff generation mechanisms and significant model structural uncertainties, a novel multi-model coupled forecasting framework was developed. The goal of this study was to improve prediction accuracy and robustness by incorporating multiple machine learning models, advanced feature extraction, and intelligent optimization algorithms. The Niuchang section of the Baoku River in the Huangshui River basin was chosen as the case study. Daily rainfall and runoff data were collected over a 23-year period (2002-2024), with the first 17 years used for model training and the remaining 6 years for validation. The PSO-SAE-XGB forecasting methodology, which combines multiple models, was proposed and implemented. Using their complementary strengths in bias-variance trade-off and nonlinear fitting, five well-known machine learning models - AdaBoost, Bagging, Histogram Gradient Boosting (HistGradient), random forest (RF), and back propagation neural network (BP) - were used as base predictors to produce a variety of runoff forecasts. A stacked autoencoder (SAE) was then used to extract deep features from the concatenated predictions of these five base models. By reconstructing the input, the SAE sought to learn robust, high-level sparse representations that would improve the feature space for the final integration and filter out model-specific noise. The extracted latent features from the SAE were then fed into an extreme gradient boosting (XGB) model for nonlinear ensemble forecasting. To optimize the overall framework, a particle swarm optimization (PSO) algorithm was introduced to jointly calibrate the key hyperparameters of both the SAE (e.g., number of units in each hidden layer) and the XGB model (e.g., max depth, learning rate). This joint optimization ensured synergistic performance between the feature extraction and regression stages. The model's performance was evaluated using the Nash-Sutcliffe efficiency coefficient (E_NS), root mean square error (E_RMS), and mean absolute error (E_MA). The results demonstrated the effectiveness of the proposed approach. All of the base models were evaluated and found to produce reasonable forecasts, with HistGradient outperforming the others (validation E_NS = 0.921). The SAE successfully reconstructed the outputs of all base models with high fidelity (validation E_NS > 0.859 for all), effectively extracting essential features and proving particularly effective for ensemble-based models like RF and AdaBoost. The integrated PSO-SAE-XGB model significantly outperformed both the individual base models and a counterpart model without SAE (PSO-XGB). During the validation period, the PSO-SAE-XGB model achieved an E_NS of 0.976 and an E_RMS of 0.840, surpassing the PSO-XGB model's E_NS of 0.952 and E_RMS of 1.484. Importantly, the PSO-SAE-XGB model exhibited superior generalization capability, as indicated by a much smaller gap between its training and validationE_NS (0.003) compared to the PSO-XGB model (0.019). The PSO-SAE-XGB model produced more accurate and stable predictions, particularly during medium to high flow periods when other models displayed greater scatter and deviation, according to a visual analysis of the hydrographs. The accuracy and dependability of runoff forecasting in the examined basin were found to be significantly improved by the PSO-SAE-XGB multi-model coupling framework. The integration of SAE for deep feature extraction was identified as a critical factor in improving model generalization and stability by processing the multi-model predictions into a more informative and denoised feature set. The utilization of PSO for the joint hyperparameter optimization of SAE and XGB ensured an optimal configuration of the entire system. This study presents a robust and advanced technical methodology for hydrological forecasting in complex environments, which is useful for flood control and water resource management.

关键词

水文预报 / 多模型耦合 / 机器学习 / 自编码器 / 深度学习

Key words

hydrological forecasting / multi-model coupling / machine learning / autoencoder / deep learning

引用本文

引用格式 ▾

祝宾皓,李向东,李涛,窦身堂,李珂,孙周亮,谷少闯. 基于PSO-SAE-XGB的多模型水文预报[J]. 南水北调与水利科技（中英文）, 2026, 24(4): 842-852 DOI:10.13476/j.cnki.nsbdqk.2026.0079

登录浏览全文

4963

注册一个新账户忘记密码

参考文献

原文顺序 | 出版日期 | 本文引用

[1]	张建云. 中国水文预报技术发展的回顾与思考[J]. 水科学进展, 2010, 21(4): 435-443. DOI: 10.14042/j.cnki.32.1309.2010.04.019.

[2]	赵丽平, 刘晓阳, 任明磊, 等. 垂向混合产流模型在三峡水库洪水预报中的应用[J]. 南水北调与水利科技（中英文）, 2024, 22(增刊 1): 1-9.

[3]	郭园, 赵忠峰, 杨党锋, 等. 基于梯度下降法的大房郢水库洪水预报模型参数率定及适用性研究[J]. 南水北调与水利科技（中英文）, 2024, 22(增刊 1): 10-16.

[4]	李强, 赵铜铁钢. 基于大语言模型的洪涝灾害统计调查[J]. 水利水电技术（中英文）, 2025, 56(9): 60-75. DOI: 10.13928/j.cnki.wrahe.2025.09.005.

[5]	赵人俊. 流域水文模拟: 新安江模型与陕北模型[M]. 北京：水利电力出版社, 1984.

[6]	陆旻皎. 新安江模型研究的回顾和展望[J]. 水利学报, 2021, 52(4): 432-441. DOI: 10.13243/j.cnki.slxb.20200695.

[7]	Sugawara M. Tank model with snow component[J]. Study Report of National Research Center for Disaster Prevention, 1984: 293.

[8]	冯旭松, 施伟, 杨雪, 等. 基于堆栈自动编码器的泵站机组故障分析[J]. 人民长江, 2018, 49(8): 99-102. DOI: 10.16232/j.cnki.1001-4179.2018.08.019.

[9]	Kao I, Liou J, Lee M, et al. Fusing stacked autoencoder and long short-term memory for regional multi-step-ahead flood inundation forecasts[J]. Journal of Hydrology, 2021, 598: 126371. DOI: https://doi.org/10.1016/j.jhydrol.2021.126371.

[10]	龙岩, 谢新文, 雷晓辉, 等. 基于PLO-Transformer-LSTM模型的径流式电站入库流量预测[J]. 南水北调与水利科技（中英文）, 2025, 23(6): 1390-1400. DOI: 10.13476/j.cnki.nsbdqk.2025.0138.

[11]	何立新, 伍玉龙, 龙岩, 等. 基于CEEMDAN分解与混合神经网络的径流预报模型[J]. 南水北调与水利科技（中英文）, 2025, 23(6): 1379-1389. DOI: 10.13476/j.cnki.nsbdqk.2025.0137.

[12]	张挺, 马睿佳, 程泳铭, 等. 基于VMD-LSTM模型的流域径流预报技术[J]. 南水北调与水利科技（中英文）, 2025, 23(5): 1196-1203. DOI: 10.13476/j.cnki.nsbdqk.2025.0119.

[13]	武少振, 任智慧, 赵雪花, 等. 基于多模型比选耦合的降水预测[J]. 南水北调与水利科技（中英文）, 2024, 22(1): 99-109. DOI: 10.13476/j.cnki.nsbdqk.2024.0012.

[14]	翟晓燕, 李淼, 刘昌军, 等. 基于机器学习和水文模拟的不同山洪类型行为特征指标预测[J]. 南水北调与水利科技（中英文）, 2025, 23(6): 1468-1479. DOI: 10.13476/j.cnki.nsbdqk.2025.0145.

[15]	王磊, 牛庚, 桑学锋, 等. 基于径流组分约束的机器学习融水径流模型[J]. 南水北调与水利科技（中英文）, 2025, 23(5): 1173-1184. DOI: 10.13476/j.cnki.nsbdqk.2025.0117.

[16]	陈浩, 王贝, 何锡君, 等. 深度学习在河湖生态流量预报预警中的应用研究[J]. 水力发电学报, 2023, 42(8): 10-20. DOI: 10.11660/slfdxb.20230802.

[17]	Chen Tianqi, Guestrin C. XGBoost: A scalable tree boosting system[C]// Proceedings of the 22nd Acm Sigkdd International Conference on Knowledge Discovery and Data Mining, 2016: 785-794. DOI: 10.1145/2939672.2939785.

[18]	闫宝伟, 常建波, 孙明博, 等. 考虑遥相关因子的月降水数据偏差校正方法[J]. 水科学进展, 2025, 36(1): 50-61. DOI: 10.14042/j.cnki.32.1309.2025.01.005.

[19]	余红玲, 王晓玲, 任炳昱, 等. 土石坝渗流性态分析的IAO-XGBoost集成学习模型与预测结果解释[J]. 水利学报, 2023, 54(10): 1195-1209. DOI: 10.13243/j.cnki.slxb.20230240.

[20]	杨兴豪, 刘晓阳, 赵丽平, 等. 集合预报降雨信息在寸滩−三峡区间的精度评价和可利用性评估[J]. 南水北调与水利科技（中英文）, 2024, 22(增刊 1): 17-22.

[21]	田济扬, 刘荣华, 刘含影, 等. 基于不同物理参数化方案的中小流域降雨径流集合预报[J]. 南水北调与水利科技（中英文）, 2022, 20(2): 281-296. DOI: 10.13476/j.cnki.nsbdqk.2022.0030.

[22]	李大洋, 姚轶, 梁忠民, 等. 基于变分贝叶斯深度学习的水文概率预报方法[J]. 水科学进展, 2023, 34(1): 33-41. DOI: 10.14042/j.cnki.32.1309.2023.01.004.

[23]	丁诚, 王兆才, 丁伟杰, 等. 基于可解释多源数据特征融合的深度学习集合径流预测[J]. 水科学进展, 2025, 36(4): 581-595. DOI: 10.14042/j.cnki.32.1309.2025.04.004.

[24]	许月萍, 陆豪楠, 于欣廷, 等. 基于深度学习集合预报的水库闸门防洪优化调度[J]. 水科学进展, 2024, 35(6): 900-913. DOI: 10.14042/j.cnki.32.1309.2024.06.004.

[25]	黄靖涵, 王兆才, 吴俊豪, 等. 基于深度学习集合优化模型的径流区间预测研究[J]. 水利学报, 2025, 56(2): 240-252. DOI: 10.13243/j.cnki.slxb.20240168.

[26]	廖胜利, 王在能, 程春田, 等. 耦合气象信息和可解释深度学习方法的日径流预报[J]. 水利学报, 2025, 56(7): 831-843. DOI: 10.13243/j.cnki.slxb.20240737.

[27]	雍斌, 张建云, 王国庆. 黄河源区水文预报的关键科学问题[J]. 水科学进展, 2023, 34(2): 159-171. DOI: 10.14042/j.cnki.32.1309.2023.02.001.

[28]	赵泽谦, 黄强, 明波, 等. 基于多模型随机组合的水文集合预报方法研究[J]. 水力发电学报, 2021, 40(1): 76-87. DOI: 10.11660/slfdxb.20210108.

[29]	Dion P, Martel J, Arsenault R. Hydrological ensemble forecasting using a multi-model framework[J]. Journal of Hydrology, 2021, 600: 126537. DOI: https://doi.org/10.1016/j.jhydrol.2021.126537.

[30]	Wang Jingyang, Li Xiang, Wu Ruiyan, et al. A runoff prediction approach based on machine learning, ensemble forecasting and error correction: A case study of source area of Yellow River[J]. Journal of Hydrology, 2025, 658: 133190. DOI: https://doi.org/10.1016/j.jhydrol.2025.133190.

[31]	Liu Jiadong, Xu Teng, Lu Chunhui. VMDI-LSTM-ED: A novel enhanced decomposition ensemble model incorporating data integration for accurate non-stationary daily streamflow forecasting[J]. Journal of Hydrology, 2025, 653: 132769. DOI: https://doi.org/10.1016/j.jhydrol.2025.132769.

[32]	EL Bilali A, Taleb A, Bahlaoui M A, et al. An integrated approach based on Gaussian noises-based data augmentation method and AdaBoost model to predict faecal coliforms in rivers with small dataset[J]. Journal of Hydrology, 2021, 599: 126510. DOI: https://doi.org/10.1016/j.jhydrol.2021.126510.

[33]	Han Xin, Wei Zheng, Zhang Baozhong, et al. Crop evapotranspiration prediction by considering dynamic change of crop coefficient and the precipitation effect in back-propagation neural network model[J]. Journal of Hydrology, 2021, 596: 126104. DOI: https://doi.org/10.1016/j.jhydrol.2021.126104.

[34]	Chen Wei, Hong Haoyuan, Li Shaojun, et al. Flood susceptibility modelling using novel hybrid approach of reduced-error pruning trees with bagging and random subspace ensembles[J]. Journal of Hydrology, 2019, 575: 864-873. DOI: https://doi.org/10.1016/j.jhydrol.2019.05.089.

[35]	Islam K I, Gilbert J M. Identifying the influence of hydroclimatic factors on streamflow: A multi-model data-driven approach[J]. Journal of Hydrology, 2025, 652: 132684. DOI: https://doi.org/10.1016/j.jhydrol.2025.132684.

[36]	Li Chuan, Bai Yun, Zeng Bo. Deep feature learning architectures for daily reservoir inflow forecasting[J]. Water Resources Management, 2016, 30(14): 5145-5161. DOI: 10.1007/s11269-016-1474-8.

[37]	Bai Yun, Bezak N, Sapač K, et al. Short-term streamflow forecasting using the feature-enhanced regression model[J]. Water Resources Management, 2019, 33(14): 4783-4797. DOI: 10.1007/s11269-019-02399-1.

[38]	祝宾皓, 周建中, 方威, 等. 基于机器学习的多模型耦合径流预报研究[J]. 中国农村水利水电, 2023(5): 119-123. DOI: 10.12396/znsd.221329.

[39]	Kennedy J, Eberhart R. Particle swarm optimization[C]// Proceedings of ICNN'95-International Conference on Neural Networks. IEEE, 1995, 4: 1942-1948. DOI: 10.1109/ICNN.1995.488968.

[40]	秦梦春, 刘军政, 白玉川, 等. 黄河下游典型河段断面调整规律及河槽断面预测[J]. 水科学进展, 2025, 36(4): 693-706. DOI: 10.14042/j.cnki.32.1309.2025.04.013.