Soutenance de Bastien Cozian

Climate change mitigation measures are expected to lead to a large increase in the share of variable renewable generation, mainly wind and solar photovoltaic (PV). As a result, future electricity systems are expected to become increasingly dependent on the weather. To manage this system, difficult challenges must be overcome, such as the management of high-impact weather-related events. Considering these events is essential to correctly design the future energy system, to understand its requirements in terms of flexibility sources such as backup and storage, and therefore, its cost. The most problematic events could occur in winter when low solar radiation is accompanied by low temperatures and wind speeds, leading to high demand for heating and low renewable production. Such events can lead to a major imbalance between supply and demand. If this imbalance were to occur on a European scale, imports of electricity from other countries to balance the electricity system would be limited.

In this thesis, we consider extreme events combining low renewable production and high demand over Europe, with a focus on persistent events. We use two approaches to study these extreme events. In the first approach, based on a very long, 1000-year climate model simulation, we study the statistics of these extreme events and propose statistical methods that relate moderately extreme events to more extreme events. These methods, based on statistical hypotheses, can be used to make extrapolations when little data is available, such as with reanalysis data. With a simple Gaussian stochastic process, we show that the probability can be correctly estimated at the scale of France for two-week events, and at the scale of Europe for events lasting from one day to the whole season. We also propose a hypothesis of joint Gaussian distribution between weather variables and the amplitude of extreme events of large supply-demand imbalance to explain why we observe similar average patterns for moderately extreme and more extreme events.

In the second approach, we use a rare event algorithm to improve the sampling of extreme events of low production and high demand. This algorithm favors the occurrence of rare events by modifying the sampling statistics of an ensemble of trajectories, each trajectory being a climate simulation. It allows us to study both the statistics and the dynamics of rare events, since they are actually simulated by the climate model. With the rare event algorithm, we sample a large number of extreme 1-in-100 and 1-in-1000 events, that would be impossible to study with a direct sampling approach. This allows us to study their meteorological conditions, which are found to be very similar to less rare events and are associated with hemispheric-scale teleconnection patterns. Finally, we briefly discuss a collaborative work on the prediction of extreme heatwave. Based on a very long, 8000-year simulation with a climate model of intermediate complexity, a neural network model is developed and skillfully forecasts extreme heatwaves over France. The main contribution of this work is to show that, in forecast applications with relatively little data, such as reanalyses, neural networks operate in a regime of lack of data.