Enhancing Climate Projections: A Multivariate Probabilistic Weather Generator Using a Denoising Diffusion Model

Many Global Climate Models (GCMs) provide climate data at limited spatial and temporal resolutions, posing challenges for evaluating climate change on scales relevant to local stakeholders. This has led to the development of dynamic and statistical downscaling techniques, which typically provide deterministic solutions. However, quantifying uncertainty in both the underlying data and the downscaling process is crucial for robust assessments of downscaled climate projections. To address this, we introduce a multivariate probabilistic weather generator based on a denoising diffusion probabilistic model (DDPM). This model represents a data-driven generative approach to simulating the probability distribution of high-resolution climate data with uncertainty. The DDPM based weather generator can produce realistic time-series of hourly surface weather data, addressing the temporal limitations of GCMs. These generated time-series are used to evaluate climate extremes and their potential changes in future climate scenarios, such as the number of days with precipitation exceeding a certain amount or temperatures surpassing a specific threshold. Preliminary results demonstrate the model's effectiveness in reproducing historical climate metrics and its potential for downscaling future climate projections.