Fast flood forecasting for Forlì, Italy
A geometry-informed neural operator surrogate of LISFLOOD-FP provides rapid flood forecasts.
A geometry-informed neural operator surrogate of LISFLOOD-FP was developed for the municipality of Forlì, Italy. This area was severely affected by the May 2023 Emilia-Romagna floods.
The surrogate is an autoregressive geometry-informed neural operator (GINO). It uses river stage gauges and rain gauges. A Voronoi mesh of 659 nodes is used to predict volume-conserving polygon-mean depths.
The surrogate was trained on 216 simulations of 24 observed events with scaled forcing. It was tested on 38 independent observed events. On these events, it reached a critical success index (CSI) of 0.934 at 0.05m and 0.844 at 0.3 m.
The depth RMSE is 0.008m and the is 0.984 for maximum-depth maps. On one NVIDIA A10G GPU, a 72-h forecast takes 1.7 s instead of 55 s with LISFLOOD-FP.
A 50-member ensemble takes 6.2 s instead of about 46 min. This makes probabilistic flood forecasting practical. These results show that neural-operator surrogates can deliver accurate pluvial-fluvial flood forecasts driven by river stage for operational early warning.
Key facts
- A 72-h forecast takes 1.7 s on one NVIDIA A10G GPU.
- A 50-member ensemble takes 6.2 s instead of about 46 min.
- The surrogate reached a critical success index (CSI) of 0.934 at 0.05m.