IRENE: a deep learning model predicts rainfall at 1 km resolution and 5-minute intervals across the whole of Italy, outperforming STEPS and DGMR

Trained on the Italian Civil Protection’s radar composite, this probabilistic system for short-term precipitation forecasting outperforms the two industry benchmarks on the CRPS score across all time horizons.

Researchers presented IRENE (Italian Radar Ensemble Nowcasting Experiment), a probabilistic precipitation nowcasting model covering the entire Italian territory, with a resolution of one kilometer and a five-minute step. The encoder-forecaster architecture is based on multi-scale convolutional recurrent units (ConvGRU), trained on the national radar composite of the Department of Civil Protection. Sampling by importance focuses learning on the rainfall episodes that are really useful.

More reliable, but a sharpness defect to correct

Two variants complete the basic model: one adversarial (IRENE-GAN) to obtain sharper fields, the other adding a constraint on the spatial spectrum. The three versions obtain a better probabilistic score than the STEPS extrapolation method and the DGMR model at all time frames, with better calibrated sets. On the average error, however, the advantage is limited to the first 90 minutes. Adversarial training corrects the progressive loss of small structures, but adds too much to long time frames, an excess that the spectral penalty only partially compensates.