PDF(9789 KB)
Prediction of Runoff in Yangtze River Source Region Based on VMD-KAN-LSTM Model
LIU Chen-yu, LI Wei, LOU Yu, ZHANG Zhen-song, CHEN Qi, HE Yi-fei
Journal of Changjiang River Scientific Research Institute ›› 2026, Vol. 43 ›› Issue (8) : 213-221.
PDF(9789 KB)
PDF(9789 KB)
Prediction of Runoff in Yangtze River Source Region Based on VMD-KAN-LSTM Model
[Objective] Due to glacial melt and permafrost degradation, runoff series in the Yangtze River source region exhibits high non-stationarity and nonlinearity, which constrains prediction accuracy. To address this challenge, we develop a hybrid long short-term memory (LSTM) model incorporating variational mode decomposition (VMD) and Kolmogorov-Arnold Network (KAN) to improve forecasting precision. [Methods] Daily runoff data from the Zhimenda Hydrological Station spanning 2000-2019 and concurrent meteorological observations from six meteorological stations (including daily precipitation, mean temperature, maximum temperature, and minimum temperature) were used. LSTM, gated recurrent unit (GRU), KAN-LSTM, and KAN-GRU, were constructed to comparatively evaluate the effect of introducing KAN. In KAN-LSTM and KAN-GRU, traditional fully connected layers were replaced with KAN structures featuring spline-parameterized activation functions, enabling adaptive nonlinear mapping of extracted temporal features. The VMD-KAN-LSTM and VMD-KAN-GRU models followed a “decomposition-prediction-reconstruction” framework: the raw runoff series was first decomposed into multiple component sequences via VMD; each component sequence, combined with meteorological inputs, was then fed independently into KAN-LSTM or KAN-GRU for prediction; finally, the predictions of all components were aggregated to obtain the final runoff forecast. Model performance was assessed using the Nash-Sutcliffe efficiency (NSE), coefficient of determination (R2), and Kling-Gupta efficiency (KGE). [Results] (1) The introduction of KAN significantly improved model performance relative to the baseline architectures. The KAN-LSTM model attained NSE, R2, and KGE values of 0.850, 0.856, and 0.924, respectively, corresponding to gains of 0.019, 0.024, and 0.048 over the standard LSTM (NSE=0.831, R2=0.832, KGE=0.876). Similarly, KAN-GRU exhibited consistent improvements over GRU. With KAN integration, KAN-LSTM effectively mitigated the systematic overestimation observed in LSTM during low-flow periods while preserving its accuracy in peak-flow prediction. KAN-GRU, in turn, showed reduced volatility and enhanced stability compared to GRU. (2) After VMD preprocessing, model accuracy was further elevated. The VMD-KAN-LSTM model achieved the optimal performance, with NSE=0.883, R2=0.885, and KGE=0.937, exceeding KAN-LSTM by 0.033, 0.029, and 0.013, respectively. Parallel improvements were also observed for VMD-KAN-GRU. (3) VMD-KAN-LSTM demonstrated the best overall predictive performance among all compared models. Nevertheless, during extremely low-flow periods, the model occasionally yielded negative predictions, a consequence of algebraic error superposition during the reconstruction of IMFs. Despite this physical inconsistency under extreme low-flow conditions, the model maintained high fidelity to the observed hydrograph and significantly improved dry-season prediction accuracy relative to the non-decomposed approach. (4) A parallel-architecture VMD-KAN-LSTM-GRU model was tested, which delivered slightly superior metrics (NSE=0.908, R2=0.909, KGE=0.939) but demanded exponentially greater computational resources and training time. Hence, VMD-KAN-LSTM offered a more favorable balance between efficiency and performance. [Conclusion] This study proposes a VMD-KAN-LSTM coupled model for daily runoff prediction in the Yangtze River source region. The results confirm that KAN enhances nonlinear fitting capability compared to traditional fully connected layers, while VMD effectively separates multi-scale oscillatory components, thereby further improving runoff forecasting accuracy. VMD-KAN-LSTM exhibits the best overall predictive performance among all models compared, demonstrating its enhanced capacity to handle the inherent complexity and nonlinearity of hydrological time series, which in turn delivers more reliable and robust forecasts. The proposed VMD-KAN-LSTM model provides a reliable new methodology for daily runoff forecasting in the Yangtze River source region.
runoff forecasting / long short-term memory / gated recurrent unit / Kolmogorov-Arnold Network / variational mode decomposition
| [1] |
|
| [2] |
|
| [3] |
徐宗学, 姜瑶. 变化环境下的径流演变与影响研究:回顾与展望[J]. 水利水运工程学报, 2022(1):9-18.
(
|
| [4] |
Rainfall–runoff modelling is complicated due to numerous complex interactions and feedback in the water cycle among precipitation and evapotranspiration processes, and also geophysical characteristics. Consequently, the lack of geophysical characteristics such as soil properties leads to difficulties in developing physical and analytical models when traditional statistical methods cannot simulate rainfall–runoff accurately. Machine learning techniques with data-driven methods, which can capture the nonlinear relationship between prediction and predictors, have been rapidly developed in the last decades and have many applications in the field of water resources. This study attempts to develop a novel 1D convolutional neural network (CNN), a deep learning technique, with a ReLU activation function for rainfall–runoff modelling. The modelling paradigm includes applying two convolutional filters in parallel to separate time series, which allows for the fast processing of data and the exploitation of the correlation structure between the multivariate time series. The developed modelling framework is evaluated with measured data at Chau Doc and Can Tho hydro-meteorological stations in the Vietnamese Mekong Delta. The proposed model results are compared with simulations of long short-term memory (LSTM) and traditional models. Both CNN and LSTM have better performance than the traditional models, and the statistical performance of the CNN model is slightly better than the LSTM results. We demonstrate that the convolutional network is suitable for regression-type problems and can effectively learn dependencies in and between the series without the need for a long historical time series, is a time-efficient and easy to implement alternative to recurrent-type networks and tends to outperform linear and recurrent models.
|
| [5] |
The diversity in hydrologic models has historically led to great controversy on the "correct" approach to process-based hydrologic modeling, with debates centered on the adequacy of process parameterizations, data limitations and uncertainty, and computational constraints on model analysis. In this paper, we revisit key modeling challenges on requirements to (1) define suitable model equations, (2) define adequate model parameters, and (3) cope with limitations in computing power. We outline the historical modeling challenges, provide examples of modeling advances that address these challenges, and define outstanding research needs. We illustrate how modeling advances have been made by groups using models of different type and complexity, and we argue for the need to more effectively use our diversity of modeling approaches in order to advance our collective quest for physically realistic hydrologic models.
|
| [6] |
李步, 田富强, 李钰坤, 等. 融合气象要素时空特征的深度学习水文模型[J]. 水科学进展, 2022, 33(6): 904-913.
(
|
| [7] |
Runoff simulations are of great significance to the planning management of water resources. Here, we discussed the influence of the model component, model parameters and model input on runoff modeling, taking Hanjiang River Basin as the research area. Convolution kernel and attention mechanism were introduced into an LSTM network, and a new data-driven model Conv-TALSTM was developed. The model parameters were analyzed based on the Conv-TALSTM, and the results suggested that the optimal parameters were greatly affected by the correlation between the input data and output data. We compared the performance of Conv-TALSTM and variant models (TALSTM, Conv-LSTM, LSTM), and found that Conv-TALSTM can reproduce high flow more accurately. Moreover, the results were comparable when the model was trained with meteorological or hydrological variables, whereas the peak values with hydrological data were closer to the observations. When the two datasets were combined, the performance of the model was better. Additionally, Conv-TALSTM was also compared with an ANN (artificial neural network) and Wetspa (a distributed model for Water and Energy Transfer between Soil, Plants and Atmosphere), which verified the advantages of Conv-TALSTM in peak simulations. This study provides a direction for improving the accuracy, simplifying model structure and shortening calculation time in runoff simulations.
|
| [8] |
邹红梅, 朱成涛. 基于LSTM和BP神经网络的水库入库径流中长期预测比较研究[J]. 水文, 2024, 44(4):27-31,37.
(
|
| [9] |
|
| [10] |
XuYuanhao,
|
| [11] |
|
| [12] |
胡庆芳, 曹士圯, 杨辉斌, 等. 汉江流域安康站日径流预测的LSTM模型初步研究[J]. 地理科学进展, 2020, 39(4): 636-642.
论文基于2003—2014年水文资料,采用长短期记忆神经网络(Long-Short Term Memory,LSTM),构建了汉江上游安康站日径流预测模型,评价了不同输入条件下日径流预测的精度。结果表明：当预见期为1 d时,在仅以安康站前期日径流量作为输入的条件下,LSTM模型在训练期和检验期的效率系数分别达到0.68和0.74;如再将流域前期面雨量和上游石泉站前期日径流量加入LSTM网络作为输入变量,安康站日径流量预测效果将更好,训练期和检验期的效率系数最高可达到0.83和0.84,均方根误差也有显著削减,且对主要洪峰流量的预测能力也有一定提高。此外,LSTM可以有效避免过拟合等问题,具有较好的泛化性能。但当预见期从1 d延长至2、3 d时,LSTM的预测精度显著降低。
(
Based on the hydrological data from 2003 to 2014, Long-Short Term Memory (LSTM) was used to construct a daily runoff prediction model for the Ankang discharge station in the upper reaches of the Hanjiang River. The accuracy of daily runoff prediction was evaluated under different input conditions. The result shows that when the foreseeing period is one day, the efficiency coefficient of the LSTM in the calibration period and the validation period can reach 0.68 and 0.74 respectively under the condition that only the previous daily runoff of the Ankang Station is used as input. When the previous areal rainfall of the catchment and the previous daily runoff of the upstream Shiquan Station were added to the LSTM model as input variables, the daily runoff prediction precision was improved. The efficiency coefficient of the training period and the validation period could reach 0.83 and 0.84, respectively. The root mean square error was also significantly reduced. The accuracy of the main flood peak flow forecasting also increased. The LSTM can effectively avoid the problem of over-fitting, and has better generalization performance. However, when the foreseeable period is extended from one day to two or three days, the performance of LSTM is significantly reduced. |
| [13] |
效文静, 周建中, 杨建华, 等. 基于机器学习的长江干流中长期径流预报[J]. 水电能源科学, 2022, 40(9): 31-34, 26.
(
|
| [14] |
张琴琴, 刘文强, 陈之鸿, 等. 基于LSTM-SVM模型的河流流量预测[J]. 天津师范大学学报(自然科学版), 2023, 43(6): 45-52.
(
|
| [15] |
|
| [16] |
|
| [17] |
|
| [18] |
|
| [19] |
|
| [20] |
|
| [21] |
丁占涛, 安杰, 吴国洋, 等. 基于EEMD-SE-LSTM组合模型的开都河日径流模拟研究[J]. 石河子大学学报(自然科学版), 2024, 42(3):335-341.
为提高开都河日径流模拟的精度和更科学地进行开都河水资源的管理与规划,在集成经验模态分解(EEMD)的基础上进行样本熵(SE)重构来完成长短期记忆网络(LSTM)组合模型的构建。采用集成经验模态分解提取开都河日径流序列中具有物理含义的信息,得到一系列本征模态分量(IMF)及一个趋势项(Res),计算每个分量的样本熵,复杂程度接近的子序列叠加为新序列,建立长短期记忆神经网络模型进行预测,叠加得到最终模拟值。结果表明:EEMD-SE-LSTM组合模型日径流模拟的精度得到提高,其确定系数R<sup>2</sup>=0.81、纳什效率系数NSE=0.73,均高于LSTM模型的R<sup>2</sup>=0.73、NSE=0.52和EEMD-LSTM模型的R<sup>2</sup>=0.64、NSE=0.63;EEMD-SE-LSTM组合模型的日径流模拟准确性更高,其评价指标(R<sup>2</sup>=0.81、NSE=0.73)高于其他单一模型SVM(R<sup>2</sup>=0.70、NSE=0.58)。EEMD-SE-LSTM组合模型提高了日径流模拟精度,可以更好地为开都河水资源管理与规划提供科学依据。
(
|
| [22] |
Hydrological time series forecasting is one of the most important applications in modern hydrology, especially for the effective reservoir management. In this research, an artificial neural network (ANN) model coupled with the ensemble empirical mode decomposition (EEMD) is presented for forecasting medium and long-term runoff time series. First, the original runoff time series is decomposed into a finite and often small number of intrinsic mode functions (IMFs) and a residual series using EEMD technique for attaining deeper insight into the data characteristics. Then all IMF components and residue are predicted, respectively, through appropriate ANN models. Finally, the forecasted results of the modeled IMFs and residual series are summed to formulate an ensemble forecast for the original annual runoff series. Two annual reservoir runoff time series from Biuliuhe and Mopanshan in China, are investigated using the developed model based on four performance evaluation measures (RMSE, MAPE, R and NSEC). The results obtained in this work indicate that EEMD can effectively enhance forecasting accuracy and the proposed EEMD-ANN model can attain significant improvement over ANN approach in medium and long-term runoff time series forecasting. Copyright © 2015 Elsevier Inc. All rights reserved.
|
| [23] |
包苑村, 解建仓, 罗军刚. 基于VMD-CNN-LSTM模型的渭河流域月径流预测[J]. 西安理工大学学报, 2021, 37(1): 1-8.
(
|
| [24] |
刘声洪,
(
|
| [25] |
袁旦, 谭尧耕, 朱艳霞, 等. 改进LSTM在金沙江流域中长期径流预报研究[J]. 水利水电技术(中英文), 2024, 55(增刊1): 28-38.
(
|
| [26] |
杜嘉妮, 蔡宜晴, 王岗. 长江源区径流变化归因分析[J]. 水文, 2021, 41(6): 73-78.
(
|
| [27] |
姜晓萱, 王文卓, 袁喆, 等. 基于嵌套主成分回归模型的长江源区径流重建[J]. 长江科学院院报, 2025, 42(3): 202-210.
基于嵌套主成分回归模型,重建了长江源区公元1433—2002年的天然年径流序列,探讨了其历史变化特征与干湿事件演变及周期性波动。利用树轮数据及径流观测数据,结合模型评估指标(CRSQ、VRSQ、RE、CE),证明了该模型的重建精确性。研究结果表明,在过去570 a中,长江源区年径流量整体呈显著波动,识别出6个湿润期与9个干旱期,其中1451—1510年和1596—1645年是最长的湿润期,1848—1903年是持续时间最长的干旱期。重建时期的干旱与长江流域青藏高原等地区的干旱时期较为吻合,表明长江源区重建径流变化能够反映大尺度的气候波动。此外,长江源区重建径流序列存在4~8、16~32、50~100、100~200 a的显著周期性波动,可能受ENSO(厄尔尼诺-南方涛动)、东亚夏季风(EASM)变化、太平洋年代际振荡(PDO)及大西洋多年代际振荡(AMO)共同驱动,也反映了全球气候长期变化及青藏高原冰川与积雪消融趋势的影响。
(
|
| [28] |
|
| [29] |
|
| [30] |
|
/
| 〈 |
|
〉 |
