A multi-model integration method for monthly streamflow prediction: modified stacking ensemble strategy

• Published in: Journal of hydroinformatics Vol. 22; no. 2; pp. 310 - 326
• Main Authors: Li, Yujie, Liang, Zhongmin, Hu, Yiming, Li, Binquan, Xu, Bin, Wang, Dong
• Format: Journal Article
• Language: English
• Published: London IWA Publishing 01.03.2020
• Subjects: Access; Canyons; Climate prediction; Computer applications; Forecasting; Hydrology; Integration; Learning algorithms; Machine learning; Mathematical models; Methods; Monthly; Precipitation; River basins; Rivers; Simulation; Stacking; Stream discharge; Stream flow; Streamflow forecasting; Support vector machines; Training
• DOI: 10.2166/hydro.2019.066
• ISSN: 1464-7141

In this study, we evaluate elastic net regression (ENR), support vector regression (SVR), random forest (RF) and eXtreme Gradient Boosting (XGB) models and propose a modified multi-model integration method named a modified stacking ensemble strategy (MSES) for monthly streamflow forecasting. We apply the above methods to the Three Gorges Reservoir in the Yangtze River Basin, and the results show the following: (1) RF and XGB present better and more stable forecast performance than ENR and SVR. It can be concluded that the machine learning-based models have the potential for monthly streamflow forecasting. (2) The MSES can effectively reconstruct the original training data in the first layer and optimize the XGB model in the second layer, improving the forecast performance. We believe that the MSES is a computing framework worthy of development, with simple mathematical structure and low computational cost. (3) The forecast performance mainly depends on the size and distribution characteristics of the monthly streamflow sequence, which is still difficult to predict using only climate indices.