Forecasting short- and medium-term streamflow using stacked ensemble models and different meta-learners

Abstract

Streamflow forecasting holds a pivotal role in the effective management of water resources, flood control, hydropower generation, agricultural planning, and environmental conservation.

This study assessed the effectiveness of a stacked Multilayer Perceptron-Random Forest (MLP-RF) ensemble model for short- to medium-term (7 to 15 days ahead) daily streamflow forecasts in the UK. The stacked model combines MLP and RF, enhancing generalization by capturing complex nonlinear relationships and robustness to noisy data. Stacking reduces bias and variance by aggregating predictions and addressing differing sources of bias and variance in MLP and RF. Furthermore, this ensemble model is computationally inexpensive. The study also examined the impact of different meta-learner algorithms, Elastic Net (EN), Isotonic Regression (IR), Pace Regression (PR), and Radial Basis Function (RBF) Neural Networks, on model performance.

For 1-day ahead forecasts, all models performed well (Kling Gupta efficiency, KGE, from 0.921 to 0.985, mean absolute percentage error, MAPE, from 3.59 to 13.02%), with minimal impact from the choice of meta-learner. At 7-day ahead forecasts, satisfactory results were obtained (KGE from 0.876 to 0.963, MAPE from 11.53 to 24.55%), while at the 15-day horizon, accuracy remained reasonable (KGE from 0.82 to 0.961, MAPE from 18.31 to 34.38%). The RBF meta-learner generally led to more accurate predictions, particularly affecting low and peak flow rates. RBF consistently outperformed in predicting low flow rates, while EN excelled in predicting flood flow rates in many cases. For estimating total discharged water volume, all models exhibited low relative error (< 0.08).