KNOU-Flood: an LSTM model trained using GraphCast improves global river flow forecasting for up to ten days

Out of 2,099 catchment areas tested between 2021 and 2023, this open probabilistic model provides more accurate predictions of floods with return periods of 1.5 and 2 years than the AIFL method on which it is based.

Flow forecasting models trained on reanalyses lose quality when fed with real weather forecasts, and all the more so as the time frame gets longer. KNOU-Flood, proposed by Minjong Cheon, takes the open recipe of AIFL (a two-step transfer learning around an LSTM) without touching its recurring core, but changing everything around it: a 424-day entry window, forecast data that takes into account the deadline, GraphCast as a second source next to the ECMWF HRES model, and a probabilistic set that provides uncertainty.

On 2,099 test basins and three initializations, the median efficiency (KGE') reached 0.65 at one day and 0.48 at ten days, compared to 0.62 and 0.34 for the reimplemented reference, with significant gains at all deadlines. The author also shows that a large part of the degradation attributed to AIFL is due to the way in which the forecasts are injected: with the same weights, the ten-day score increases from 0.18 to 0.34 depending on the protocol. The sets are almost well calibrated and better detect floods with a return period of 1.5 and 2 years; for rarer floods, the benefit still depends on the initialization.