Hybrid Multivariate Machine Learning Models for Streamflow Forecasting: A Two-Stage Decomposition–Reconstruction Framework

Abstract

Robust and accurate streamflow forecasting holds significant importance for flood mitigation, drought warning and water resource management. On account of the intricate nonlinear and nonstationary nature of streamflow time series, numerous decomposition-based approaches have been proposed and integrated with other architectures. However, directly decomposing the entire streamflow data set introduces future information into the decomposition and reconstruction processes, while decomposing calibration and validation sets independently can result in undesired boundary effects. Besides, the signal decomposition techniques tend to generate a large number of decomposed modes. Using all these modes directly as input variables results in intricate forecasting models and is prone to overfitting. To address these challenges, we developed a novel two-stage decomposition reconstruction forecasting (TSDRF) framework by coupling sequentially decomposition technique, sample entropy and multivariate machine learning methods in this study. This newly proposed TSDRF framework is assessed at three hydrologic stations from Yellow River, China. Furthermore, the TSDRF framework is also compared with the two-stage decomposition reconstruction hindcasting (TSDRH) framework under different lead times. The findings suggest that TSDRF framework based on variation mode decomposition (VMD) algorithm outperform other models in terms of mitigating boundary effects, minimizing computational costs, and enhancing generalization capabilities across various lead times.