Towards the development of time series foundation model for weather forecasting: a UK case study

Abstract

This dissertation delves into the burgeoning realm of artificial intelligence (AI) with a specific focus on the potential of foundation models in weather forecasting. As AI continues to transform diverse domains, the research seeks to harness the capabilities of foundation models, notable for their expansive neural networks and adaptability, to enhance meteorological predictions in the United Kingdom. By meticulously analysing historical weather data from Guildford and Woking, the study aims to model intricate weather patterns, such as temperature and wind speed. The process of model development is explored in depth, emphasizing data pre-processing, the innovative adoption of the Transformer architecture, rigorous training, fine-tuning strategies, and a comprehensive evaluation approach. The findings underscore the potential of foundation models in weather forecasting, suggesting they may offer enhanced accuracy and efficiency compared to traditional methods with more development. The results, while promising, highlight the need for more extensive data integration in terms of geographical scope and variables. They also underscore the adaptability of these models to various regions and their proficiency in predicting a broader range of weather variables. This research's implications extend beyond meteorology, impacting sectors heavily reliant on weather forecasts, such as agriculture and emergency response. While the study marks significant strides, it also acknowledges its limitations, presenting a roadmap for future research encompassing data expansion, model refinement, and techniques for real-time adaptation. Overall, this research represents a significant advancement in meteorological prediction, bridging traditional methods with the cutting-edge innovations of artificial intelligence.