Predicting Flood Water Levels in Data-scarce Meso-Scale River Basins for Support Early Warning Systems

This study proposes a methodology for predicting maximum river water levels during rainfall events in meso-scale river basins with limited data availability. The approach reformulates flood peak prediction as a multi-horizon conditional mapping problem, thereby reducing the dependence on long historical time series and recurrent architectures commonly employed in the literature. The proposed methodology estimates flood levels based on prior observations of river stage, precipitation data, and the elapsed time since the onset of the event, aiming to support early flood warning systems. For the proposed approach, water level and accumulated precipitation data were used for forecast horizons of 1, 2, 6, 12, and 14 hours prior to the occurrence of the predicted level. The methodology is based on an artificial neural network model, with the dataset divided into training (70.0%) and validation (30.0%) subsets. The model was applied to the Mascarada River basin, located in northeastern Rio Grande do Sul State, Brazil. The results demonstrated strong predictive performance, particularly for forecast horizons equal to or shorter than 6 hours, for both training and validation datasets. Nash–Sutcliffe efficiency values exceeded 0.75, while RSR values remained below 0.50. For forecast horizons of 12 hours or longer, performance ranged from good to satisfactory. Overall, the proposed methodology and forecasting model demonstrated high potential for water level prediction in data-scarce river basins, representing a promising tool for the anticipation of extreme hydrological events under such conditions.