Jua Vilhelm: A Novel End-to-End AI System for Global Weather Prediction

In this study, we present Jua Vilhelm, an innovative end-to-end artificial intelligence (AI)-powered weather forecasting system that exhibits significant improvements over existing state-of-the-art numerical models in terms of accuracy and efficiency for up to 48-hour forecasts. The Jua Vilhelm system demonstrates superior performance in predicting near-surface parameters, such as air temperature and wind, when compared to the well-established Integrated Forecast System (IFS) model. Leveraging a combination of cutting-edge deep learning techniques and in-depth physics knowledge, Jua Vilhelm can generate high-resolution global predictions on a 1x1km grid. The system is especially adept at predicting critical surface parameters like precipitation, wind, and air temperature. This study focuses on its performance in wind and temperature prediction. Notably, Vilhelm achieves rapid forecast run times in a matter of seconds, enabling the execution of hundreds of ensemble runs within minutes, an accomplishment previously unattainable. To assess the performance of Jua Vilhelm against medium-range Numerical Weather Prediction (NWP) models, we selected the IFS model of the European Centre for Medium-Range Weather Forecasts (ECMWF) as a suitable comparison. The ERA5 reanalysis dataset was employed as a benchmark for evaluating the Jua Vilhelm model on a global scale, using its full 0.25-degree resolution. We benchmarked both models at 6-hour intervals for multiple initial conditions per day for all seasons. For the evaluation of air temperature and wind speed (zonal and meridional) forecasts, we utilized several metrics, including Mean Bias Error (MBE), Root Mean Square Error (RMSE), and Kling Gupta Efficiency (KGE) (Figure 1). The results of this study indicate that Jua Vilhelm and IFS models exhibit comparable RMSE levels for wind speed; however, Jua Vilhelm outperforms IFS significantly in terms of MBE. Consequently, Jua Vilhelm presents better KGE due to its low bias and standard deviation closely aligned with the ground truth. In terms of air temperature, Jua Vilhelm not only outperforms IFS in RMSE but also demonstrates reduced bias, resulting in superior KGE. Furthermore, the IFS model occasionally generates inconsistent outputs, while Jua Vilhelm maintains complete consistency. This research builds upon prior work in utilizing noisy Internet of Things (IoT) data for weather forecasting. The Jua Vilhelm model was trained on an extensive global dataset, incorporating observations from satellites, weather stations, radiosondes, and other sensors for various weather variables. The development of Jua Vilhelm marks a significant advancement in the field of weather forecasting, offering unprecedented accuracy and speed.