Data-Driven Regional Weather Forecasting: A Comparative Study of FourCastNet and GraphCast

Regional weather forecasting is crucial for preparing for severe weather events, ensuring public safety, and efficiently managing resources by providing timely and localized weather information. Traditional numerical weather forecasting methods, while extremely useful in practice, face challenges such as high computational costs and the need for parameterization.
Recently, machine learning (ML) techniques have emerged as promising, data-driven alternatives to enhance weather forecast accuracy and reliability.
In this study, we investigate the effectiveness of two advanced ML models, FourCastNet and GraphCast, in regional weather forecasting. FourCastNet leverages the token-mixing scheme of the Adaptive Fourier Neural Operator with a Vision Transformer (ViT) backbone, while GraphCast utilizes a Graph neural network in an encoder-processor-decoder configuration. We apply these models to perform regional weather forecasts using High-Resolution Rapid Refresh (HRRR) data from 2015 to 2022, focusing on a 192 km × 192 km region in Tennessee. The models' forecast accuracy is evaluated against data from 2023.
In our study, we evaluated the forecast accuracy, computational efficiency, robustness and reliability, and potential for ensemble predictions of FourCastNet and GraphCast for regional weather forecasting. We assessed forecast accuracy by predicting various meteorological variables and comparing them with numerical weather prediction (NWP) models, such as Model for Prediction Across Scales Atmosphere (MPAS-A). Computational efficiency was analyzed by comparing training and inference times. Robustness and reliability were tested under different weather conditions. Additionally, we explored ensemble predictions by combining outputs from both models to enhance overall forecast reliability. These efforts aim to explore the potential of data-driven ML models in regional weather forecasting.