ABSTRACT Conceptual workflow linking sensor-network observations and meteorological data with flood typology, machine-learning and deep-learning modelling, scenario-based flood simulations, model performance evaluation, and SHAP feature-contribution analysis. This study presents a scenario-based framework for rainfall–runoff modelling that evaluates classical machine learning, tree-based ensembles, and deep learning architectures across distinct flood types in a montane basin. Using 15 years of hourly hydrometeorological and sensor-network observations from the Upper Vydra Basin (2008–2023), we assessed eight models and an equal-weight ensemble to link predictive skill to flood-generating processes. Model performance varied across six flood typologies. The extended LSTM achieved the highest individual accuracy for long-duration and multi-peak events, whereas Random Forest and XGBoost were most effective for short-duration floods under contrasting antecedent wetness. Transformer models showed systematic overprediction, and support vector regression performed weakest. An equal-weight ensemble combining eight ML/DL architectures provided the most robust overall performance, with the highest prediction accuracy (NSE = 0.955) and reduced error variance. SHAP analysis highlighted the dominant influence of precipitation and snowmelt and the value of distributed water-level observations for representing catchment storage dynamics. These findings show that predictive skill depends strongly on the match between model architecture and hydrometeorological context. Scenario-based evaluation combined with ensemble integration offers a practical pathway toward reliable and operationally efficient flood forecasting in montane environments.
Ba et al. (2026) studied this question.