Comparative analysis of rainfall-runoff modelling using HEC-HMS and machine learning models in the gumara watershed, Upper Blue Nile Basin

Accurate rainfall–runoff modeling is essential for effective water resources management, particularly in data-scarce regions. This study evaluates HEC-HMS and five machine learning (ML) models, ANN, CNN, LSTM, XGBoost, and a hybrid XGBoost–LSTM–CNN, for rainfall–runoff simulation in the Gumara watershed, Upper Blue Nile Basin, Ethiopia. Daily rainfall from four meteorological stations and observed discharges at the watershed outlet were used. The dataset was divided into calibration (2001–2012) and validation (2013–2018) periods, with monthly simulations performed. Model performance was assessed NSE, R², and PBIAS. HEC-HMS showed very good performance (calibration: NSE = 0.81, R² = 0.83, PBIAS = −8.43%; validation: NSE = 0.82, R² = 0.83, PBIAS = 12.58%); however, it underestimated peak flows. The ML models improved simulation accuracy, with the hybrid XGBoost–LSTM–CNN model performing best (calibration: NSE = 0.98, R² = 0.98, PBIAS = 2.57%; validation: NSE = 0.89, R² = 0.88, PBIAS = −6.9%). For peak flows with exceedance probability below 10%, the hybrid model substantially outperformed HEC-HMS (NSE: 0.98 vs. 0.30; R²: 0.99 vs. 0.87; PBIAS: 0.10% vs. 18.12%). These findings demonstrate the potential of hybrid ML models to improve rainfall–runoff and peak-flow simulation in data-limited watersheds.