GraphDiffWeather: A Graph-Augmented Diffusion Transformer for Global Weather Forecasting

Abstract

Timely and effective weather predictions are important for managing climate-related phenomena, disasters, and risks; for agriculture and aviation; and for making informed decisions in regard to the environment. Existing weather prediction systems have great difficulty addressing the needs for understanding the complex relationships among the variable elements of the atmosphere; for referring to the long-term behavior of the atmospheric system; and for making uncertainty forecasts. To address these needs, we present in this paper a new forecasting system called GraphDiffWeather. It is a new Graph-Augmented Diffusion Transformer Framework that builds upon the integration of deep learning, Graph Neural Networks (GNNs), and Transformers, Diffusion Models, and GNNs. First, we convert atmospheric data into a graph structure for representing the spatial interactions and teleconnections across the different regions of the atmosphere. Next, these graph structure embeddings are processed through a Transformer for the temporal progression, and the Diffusion Model for uncertainty forecasting. To test the effectiveness of the proposed framework, we used the ERA5 and WeatherBench datasets and compared the results with the state of the art of weather prediction, which includes several NWP techniques, CNN-LSTM, Transformer, GNNs, Pangu-Weather, and GenCast models. With respect to these techniques, the proposed framework was superior with a MAE of 1.52, RMSE of 2.13, and R2 of 0.967. In addition, the framework successfully performed and demonstrated the effectiveness of having graph-based spatial learning, transformer-based temporal learning, and diffusion-based probabilistic forecasting techniques. Finally, we believe GraphDiffWeather is the future of global weather forecasting. It is accurate and scalable, and it incorporates the element of forecasting uncertainty.