Deep Learning Models for Downscaling of Metereological Variables

Reanalysis datasets play an important role in meteorological and climate research, offering a consistent and long-term record of atmospheric conditions by assimilating past observations with modern forecast models. These datasets are of great utility in various applications, including weather forecasting, climate change research, renewable energy prediction, resource management, air quality risk assessment, and the forecasting of rare climatic events. Among the most prominent reanalysis datasets is the Copernicus Regional Reanalysis for Europe (CERRA), which stands out due to its high-resolution coverage of the European domain. CERRA has demonstrated significant utility across multiple climate-related tasks, providing detailed insights that are essential for precise and localized studies. Despite its advantages, the availability of CERRA lags two years behind the current date, primarily due to the intensive computational demands and the complexities involved in acquiring the necessary external data. To address this temporal gap, this thesis proposes a novel method employing several deep neural models to approximate CERRA downscaling in a data-driven manner without the need for additional external information other than ERA5. By leveraging the lower resolution ERA5 dataset, this research frames the problem as a super-resolution task. The study focuses on downscaling wind speed data over Italy, utilising a model trained on existing and freely available data. The results are encouraging, as the model produces outputs closely resembling the original CERRA data, with validation against in-situ observations confirming its accuracy in approximating ground measurements. This innovative approach not only demonstrates the potential of deep neural models in overcoming the computational and data acquisition constraints associated with high-resolution reanalysis datasets but also offers a viable solution to improve the timeliness and accessibility of such data.