An extreme forecast index-driven runoff prediction approach using stacking ensemble learning

Runoff prediction plays a crucial role in hydropower generation and flood prevention, enhancing prediction accuracy in hydrology. This study proposes an extreme forecast index (EFI)-driven runoff prediction approach using stacking ensemble learning to improve prediction performance. EFI is introduced as an input into four machine learning models (Support Vector Regression, Multi-layer Perceptron, Gradient Boosting Decision Tree, and Ridge Regression) for runoff prediction with lead times of 24 h, 48 h, and 72 h. The stacking ensemble learning framework comprises four base-models and a meta-model, and model hyperparameters are re-optimized using the particle swarm optimization algorithm. The approach focuses on predicting the inflow processes of the Geheyan Reservoir in the Qing River using EFI and runoff time series. Results demonstrate that the EFI-runoff simulation can improve runoff prediction capability due to EFI's higher sensitivity to observed runoff, and the proposed stacking ensemble learning model outperforms the individual model in predicting runoff with all lead times. The relative flood peak error, mean relative error, root mean square error, and Nash-Sutcliffe efficiency coefficient of the model's one-day-ahead prediction are 7.987%, 22.421%, 632.871 m3/s, and 0.771, respectively. Therefore, this approach can be effectively utilized to improve accuracy in short-term runoff prediction applications.