Diffusion generative models for weather forecasting

In recent years, traditional numerical methods used for accurate weather prediction have faced increasing challenges from deep learning techniques. The historical datasets commonly employed for short and medium-range weather forecasts are typically structured in a regular spatial grid format. This arrangement closely resembles images, with each weather variable akin to a map or, when considering the temporal axis, as a video. Several classes of generative models, including Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), and the recent Denoising Diffusion Models (DDMs), have demonstrated their effectiveness in tackling the next-frame prediction problem. Consequently, it is only natural to assess their performance in the context of weather prediction benchmarks. DDMs, in particular, hold strong appeal in this domain due to the inherently probabilistic nature of weather forecasting. This methodology aims to model the probability distribution of weather indicators, with the expected value representing the most likely prediction. This thesis is dedicated to investigating the application of diffusion models in the realm of weather forecasting, with a specific focus on precipitation nowcasting. To achieve this, a specific subset of the ERA5 dataset has been leveraged, encompassing hourly data for Western Europe spanning the years 2016 to 2021. Within this context, the effectiveness of diffusion models has been rigorously assessed in the challenging domain of precipitation nowcasting. The research is conducted in direct comparison to the well-established U-Net models, as extensively documented in existing literature. The proposed approach, referred to as Generative Ensemble Diffusion (GED), harnesses a diffusion model to generate a diverse set of potential weather scenarios. These scenarios are subsequently amalgamated into a probable prediction through the application of a sophisticated post-processing network. In direct contrast to recent deep learning models, the GED approach consistently demonstrated superior performance across multiple performance metrics, underscoring its significant advancement in the field of weather forecasting.

Recommended citation: Paparella, Alberto (2023) Diffusion generative models for weather forecasting. [Laurea magistrale], Università di Bologna, Corso di Studio in Informatica [LM-DM270].
Download Paper | Download Slides