NormaFlood: Closed-Form Flood Forecasting Without Backpropagation

Résumé fourni par la source

NormaFlood is a backprop-free daily streamflow forecasting and flood early warning system evaluated across 229 basins in CAMELS-IND (2016–2020) and 531 basins in CAMELS-US. By combining closed-form ridge regression, Box-Cox transformed flow-lag alignment, and online Recursive Least Squares (RLS) streaming, the model achieves a median test NSE of 0.562 (KGE 0.655) across all 229 Indian basins and 0.600 NSE (KGE 0.701) on the strict 163-basin benchmark, matching and outperforming published regional LSTMs without neural network training. The framework analytically derives flood exceedance probabilities P(Q > Q_95) directly from the online parameter covariance matrix without running ensembles, provides stable multi-day forecasts up to lead day 7, and correctly flagged the August 2018 Godavari flood on the day danger levels were crossed. All experiments run in seconds and are fully reproducible.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.