Generative machine learning for extreme weather generation


As our climate warms, the frequency, duration, and intensity of extreme weather events has been increasing. For example, climate change leads to more evaporation that may exacerbate droughts and increase the frequency of heavy rainfall and snowfall events. That directly impacts various sectors such as agriculture, water management, energy, and logistics, which traditionally rely on seasonal forecasts of climate conditions for planning their operations. In this context, stochastic weather generators are often used to provide a set of plausible climatic scenarios, which are then fed into impact models for resilience planning and risk mitigation. Several weather generation techniques have been developed over the last few decades. However, they are often unable to generate realistic weather scenarios, especially for increasingly common extreme weather events.

In recent years, generative machine learning models have achieved notable success in modeling high-dimensional complex distributions, which has led to their application in stochastic weather generation. In this project, we investigate the use of generative machine learning techniques to improve weather generators' temporal and spatial downscaling capability, control synthetic weather scenarios given ‘what-if’ scenarios, such as climate projections or return periods derived from extreme value theory and enhance uncertainty quantification.