The transition toward energy systems with a high share of renewable energy increases their vulnerability to rare extreme weather events, which can compromise the stability of power grids.

In the context of climate change, the resilience of these systems depends on our ability to anticipate, characterise, and quantify the uncertainty associated with these phenomena. This thesis develops advanced numerical methods to improve the modelling and prediction of extreme events with high energy impact: winter cold spells, compound events generating high peaks in residual demand, and, more specifically, heatwaves at mid-latitudes.

The study of these rare events faces a structural problem of data scarcity. Historical observations are limited, and traditional physical climate models require massive simulations to robustly sample very low-probability events. This leads to prohibitive computational costs and limits the accuracy of return period estimates.
To address these limitations, this work explores low-cost prediction and simulation approaches for rare events, building on recent advances in deep learning.

Three complementary directions are developed: (i) the design of direct prediction models able to learn from limited data, with particular attention to their physical interpretability; (ii) the design of climate model emulators able to efficiently reproduce their dynamics at reduced cost, examining in particular their ability to extrapolate beyond the training range for extreme events; (iii) the adaptation of rare event simulation algorithms, combining these emulators with physical models to guide the exploration toward trajectories leading to extreme weather events.

The methodological developments are first validated on climate models of intermediate complexity before being applied to state-of-the-art models, opening the way for future applications related to climate change adaptation.

Lieu
École normale supérieure – PSL
24 rue Lhomond – aile Erasme
salle Claude Froidevaux – E314

Visio
https://visio.numerique.gouv.fr/dbl-klbe-txl

Pierre Gentine (ENS-PSL) : rapporteur
Elizabeth Barnes (Boston University) : rapportrice
Laure Raynaud (Météo France) : examinatrice
Stéphane Mallat (ENS-PSL) : examinateur
Boris Bonev (Nvidia) : invité
Freddy Bouchet (ENS-PSL) : directeur de thèse
Laurent Dubus (RTE) : directeur de thèse