Planning the power system for a changing climate needs fine-scale projections of several weather variables at once. Global climate models (GCMs) are too coarse for this. Supervised downscaling would learn to turn coarse fields into fine ones, using as its target a high-resolution reanalysis such as ERA5, a reconstruction of past weather. That requires pairs of coarse and fine fields describing the same day. A free-running GCM, however, simulates its own sequence of weather, so none of its days matches a day of the reanalysis, and no such pairs exist.
SerpentFlow gets around this for wind fields. It applies a low-pass filter to the reanalysis, keeping only the large scales it shares with the GCM, and pairs each filtered field with its original. A generative model learns from these pseudo-pairs to restore the fine scales and is then applied to the GCM. The energy system needs temperature and precipitation as well, so SerpentFlow-CLIM extends the approach to all three variables together. It shows that the low-pass filter fails to align the climatology, especially for temperature, and that the fix has to be tailored to each variable. Instead of hand-crafting such fixes, ChimeraFlow learns the degradation itself by searching over combinations of interpretable climate operators.
To compare these methods with statistical and generative baselines, we built an open benchmark that spans three GCMs and three regions: France, Australia and the central US. What matters in the end is the power system, so the benchmark also feeds every scenario into a net-load model. This shows that weather-space scores can rank methods differently from the residual-demand tails that matter to grid operators.

Lieu
École normale supérieure – PSL
24 rue Lhomond – aile Erasme
salle Serre – E509
accès par la cage d’escalier D