NormaFlood: Closed-Form Flood Forecasting Without Backpropagation

Résumé fourni par la source

Code and manuscript for **NormaFlood**, a backprop-free daily streamflow forecasting and flood early warning model tested on CAMELS-IND (229 Indian basins, 2016–2020) and CAMELS-US (531 basins). Instead of training a neural net, NormaFlood fits a closed-form ridge regression per basin and updates weights online using Recursive Least Squares (RLS). The core trick is expressing autoregressive flow lags in the exact same Box-Cox transformed space as the target. That alignment lifts median nowcast NSE on CAMELS-IND from 0.519 to 0.658 (and 0.709 on the 163 long-record basins), closing most of the performance gap to published regional LSTMs. A few key features in the repo:- **Probabilistic flood alarms:** Exact exceedance probabilities ($P(Q > Q_{95})$) are computed analytically from the RLS covariance matrix without running ensembles (Brier 0.014–0.016, AUC ~0.97, max economic value 0.84).- **Multi-day forecasting (1–7 days):** Iterating the 1-day head recursively with self-generated Box-Cox flow estimates holds day-7 median NSE above 0.34 on CAMELS-IND and 0.36 on CAMELS-US, while direct per-lead heads collapse to ~0.11–0.16.- **Godavari 2018 case study:** Evaluates the 1-day online alarm through the August 2018 Godavari flood, tracking the rise to $39{,}500\,\text{m}^3/\text{s}$ and firing $P(Q > Q_{95})$ from 0.01 to 1.00 on the exact day danger level was crossed.- **Fast & deterministic:** All experiments (E1–E10) solve in under 30 seconds for all 229 basins on a GPU (JAX), or under 5 minutes on CPU.

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