A novel physical process-ensemble learning model framework with residual error decomposition to upskill daily runoff prediction,Journal of Hydrology

当前位置： X-MOL 学术 › J. Hydrol. › 论文详情

Our official English website, www.x-mol.net, welcomes your feedback! (Note: you will need to create a separate account there.)

A novel physical process-ensemble learning model framework with residual error decomposition to upskill daily runoff prediction
Journal of Hydrology ( IF 5.9 ) Pub Date : 2024-12-21 , DOI: 10.1016/j.jhydrol.2024.132565
Yan Kang, Yue Xiang, Zishang Zhang, Ruyi Wan, Wanxue Li, Shuo Zhang, Lingjie Li, Songbai Song

Accurate and reliable daily runoff prediction plays a crucial role in water resources management. Physical-based hydrological models generally perform unsatisfactory in semi-arid areas. Current studies have attempted to utilize data-driven learning method to correct the residual errors of physical models. However, the highly non-stationarity of residual series is ignored in the conventional approach. In this paper, we propose a novel model framework that integrates Physical Process models (PP), Residual Decomposition (RD) and Ensemble Learning algorithms (EL), namely PP-RD-EL framework, to upskill daily runoff prediction in semi-arid areas. In this framework, HBV and SIMHYD as basic PP models, incorporating storage-excess and infiltration-excess runoff mechanism, were employed to simulate precipitation-runoff process, respectively. Meanwhile, EL models were presented to predict the residual errors from PP models. In particular, Stacking EL models are developed to verify the framework effectiveness by comparing to Random Forests (RF), Adaptive Boosting (AdaBoost), Gradient Boosted Decision Tree (GBDT), Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM) and Categorical Boosting (CatBoost). Furthermore, we introduced Residual Decomposition into PP-EL models to decompose the nonstationary residual errors into multiple relatively stationary sub-series for enhancing the stability of input data, including wavelet transform (WT), Variational mode decomposition (VMD), Robust local mean decomposition (RLMD) and Time varying filtering based empirical mode decomposition (TVF-EMD). Here, we verify PP-RD-EL framework by predicting the daily runoff of Wei River, China. The results showed that PP-RD-EL framework achieves the best performance in daily runoff prediction, and HBV-VMD-Stacking3 and SIMHYD-VMD-Stacking3 in the framework showed distinct improvements when compared to HBV and SIMHYD model in terms of both NSE (31%- 47%) and KGE (117%-265%), and improved NSE by 3.8–17.9% and KGE by 0.3–20.2% over the PP-EL models. Especially, the proposed framework significantly improved the simulation accuracy for peak and low flow. The study demonstrates that the proposed PP-RD-EL framework with a “Runoff Simulation-Residual Decomposition-Residual Simulation-Runoff Restructure” modeling process, has proven to be a very promising approach, which enables efficient and accurate nonstationary daily runoff prediction.

更新日期：2024-12-21

点击分享查看原文

点击收藏

阅读更多本刊新发论文本刊介绍/投稿指南