Advanced Machine Learning Ensembles for Improved Precipitation Forecasting: The Modified Stacking Ensemble Strategy in China

Abstract

Accurate and reliable precipitation forecasting is vital for effective water resource management and disaster mitigation, especially in geographically diverse and climatically complex regions like China. This study proposes an advanced methodology for medium- and long-term hydrological forecasting by integrating multiple machine learning models through a modified stacking ensemble strategy (MSES). We developed and compared five deterministic precipitation forecasting models, including elastic net regression (ENR), support vector regression, random forest, extreme gradient boosting, and light gradient boosting to provide forecasts with lead times ranging from 0 to 5 months at a spatial resolution of 0.5$^\circ$. The MSES was then evaluated against the traditional Bayesian model averaging (BMA) approach. Our comprehensive evaluation, based on deterministic forecasting metrics such as the anomaly correlation coefficient (ACC), mean squared skill score (MSSS), and Graded Precipitation Score (Pg), demonstrated the MSES outperformed individual models and the BMA method. The MSES achieved ACC scores between 0.6 and 0.9 for lead time (LT) $= 0$ month, with an average of around 0.8 for LT $= 2$ months. The MSSS for MSES was above 0.5 in more than half of the evaluations, and the Pg score was consistently above 80, indicating high accuracy in precipitation magnitude prediction. These findings highlight the promise of advanced machine learning strategies like MSES in improving the accuracy and robustness of precipitation forecasts, addressing critical needs in water resource management and disaster mitigation in China.

Keywords

• Ensemble methods
• machine learning
• multimodel machine learning integration
• precipitation forecasting