Testing machine learning algorithms as post-processing tools for hydro-meteorological modelling over a small river basin

Abstract

Accurate runoff forecasting in downstream urban areas of small river basins is crucial for flood warning and risk management. This study proposes a hybrid runoff forecasting framework that integrates a process-driven hydro-meteorological model with data-driven post-processing using ensemble machine learning (ML) algorithms. First, the MOLOCH meteorological model is used to predict weather conditions, and it is coupled with the FEST hydrological model, which translates meteorological output into hydrological responses. Finally, multiple ML algorithms are employed for error correction, and the results are integrated using the Stacking ensemble strategy. The case study indicates that: The Stacking model consistently outperforms traditional autoregression (AR) and long short term memory (LSTM) models in both overall accuracy and flood-event-specific performance metrics. Particularly, the post-processing framework exhibits comparable effectiveness when applied to both the coupled hydro-meteorological forecasting chain (CHMFC) and the FEST model, confirming its flexibility and potential to improve forecast lead time. While forecast performance naturally degrades over longer lead times, the Stacking model maintains better robustness and slower performance decay. The proposed hybrid framework combines the interpretability of process-driven models with the nonlinear capture capabilities of data-driven models, offering a promising solution for enhancing real-time runoff forecasting in flood-prone small river basins.