Abstract

Flood early warning systems often rely on a single hydro-meteorological forecast, which can limit reliability. Recent advances in deep learning (DL) offer promising improvements due to their low computational cost, enabling the generation of ensemble forecasts. This study investigates how to process multiple weather-runoff forecasts to improve model performance in predicting extreme events. We applied DL-based weather-runoff forecasting in river stations located at the foot of the Andes Mountains in Chile. The models couple a near-future global weather forecast with short-range runoff forecasting systems based on Long Short-Term Memory (LSTM) cells. Meteorological and geomorphological input variables commonly used in hydrological models were selected. Training and validation used ERA5 data, while NCEP-GFS data were used for testing and real-time operation. Model performance was evaluated using the Kling-Gupta efficiency (0.6-0.8) and Nash-Sutcliffe efficiency (greater than 0.9). The threat score index, which assesses the model's ability to predict threat peak flow exceedance, ranged between 0.6 and 0.8. The best-performing models were analyzed probabilistically to quantify uncertainty. Finally, we introduced the concept of conditional probability to estimate the likelihood of exceeding a threat peak flow, providing a basis for raising alerts and improving decision-making under uncertain conditions.