AI-Machines in weather forecasting

Abstract

AI integrated into meteorology opened new frontiers in weather forecasting with prospects of enhanced accuracy and efficiency compared to traditional numerical weather prediction models. This work is concerned with the new/backend improvements done by AI-driven models, particularly from the European Centre for Medium-Range Weather Forecasts: FourCastNet, GraphCast and panguweather. These models make use of the latest deep learning techniques at their core, like the convolutional neural network and graph neural network, to improve resolutions and reliability for weather forecasts. Therefore, in this paper, the performance of the models is evaluated using the WeatherBench benchmark. Setup in a standard framework of input-output, it puts light on the climate models, which seek to predict weather conditions for significant atmospheric variables: temperature, wind, rainfall, mean sea level pressure—contrast of outputs from AI models with traditional NWP methods. This thesis is organized into seven chapters, including an introduction of the motivation of AI in meteorology, a state-of-the-art overview of the current weather forecasting technologies, a detail presentation about the models from ECMWF that used AI, the assessment methodology using WeatherBench, experimental results from that assessment, a comparative analysis of model performance, and conclusion discussing the implications and future directions. In this work, it is demonstrated that AI has huge transformational potential in weather forecasting by rigorous evaluation and proper visualization of experimental results. This work contributes to improving general understanding of the role of AI in meteorology.