ML models are increasingly used in weather forecasting, providing accurate forecasts and reduced computational costs compared to traditional numerical weather prediction (NWP) models. However, current ML models often have limitations such as coarse temporal resolution (typically 6 hours) and a narrow range of weather variables, which may limit their practical use. Accurate forecasts are crucial for the renewable energy, aviation and shipping sectors. Despite progress, ML models still struggle to ensure prediction continuity and temporal resolution. Although some models have made progress in terms of accuracy and efficiency, it remains difficult to improve their temporal granularity and include a broader set of meteorological variables.
Researchers from Fudan University and Shanghai Academy of Artificial Intelligence have introduced FuXi-2.0, an advanced ML model for global weather forecasting that provides hourly forecasts and covers a wide range of weather variables. FuXi-2.0 outperforms the European Center for Medium-Range Weather Forecasts (ECMWF) High Resolution Forecasts (HRES) in key areas such as wind energy forecasting and tropical cyclone intensity. The model integrates atmospheric and oceanic components, providing improved accuracy compared to its predecessor, FuXi-1.0, and other models like Pangu-Weather. The improved temporal resolution and comprehensive set of variables of FuXi-2.0 significantly advance practical weather forecasting applications.
The study uses the ECMWF ERA5 reanalysis dataset, which provides hourly meteorological data with a spatial resolution of approximately 31 km from January 1950. For this research, two subsets of ERA5 datasets were used : one covering the period 2012-2017 for the training of a 6- hourly forecast model and another from 2015 to 2017 for an hourly forecast model. FuXi-2.0 predicts 88 weather variables, including altitude and surface variables, with additional static and temporal encodings of geographic information. Model training involved resetting accumulated variables to match operational conditions and setting ocean variables to NaN where appropriate. Data from wind farms in the UK and South Korea have also been used for wind energy forecasting, incorporating quality control measures to ensure accuracy.
FuXi-2.0 introduces a dual-model system to provide continuous 1-hour forecasts, integrating a primary model for 6-hour forecasts and a secondary model for hourly interpolation. This architecture improves reliability and efficiency compared to previous models. The hourly model processes data through convolution layers and Swin Transformer blocks, while the hourly model generates hourly forecasts within a 6-hour window. The training used the robust Charbonnier loss function and involved many GPU cluster iterations. Wind energy forecasting was carried out using an MLP model focused on daily forecasts. Evaluation metrics included RMSE, ACC, and forecast/observation activity, with normalized differences used to compare model performance.
The study evaluates the hourly forecasts of FuXi-2.0 using 2018 data, comparing its performance with that of ECMWF HRES and Pangu-Weather. FuXi-2.0 shows superior accuracy in variables important for weather forecasting, such as temperature and wind speed, outperforming the ECMWF HRES in terms of root mean square error (RMSE) and anomaly correlation coefficient (ACC ) within most forecast times. Its forecasts are more detailed than those of Pangu-Weather and it has better activity measurements. Additionally, FuXi-2.0’s wind power forecasts for wind farms and tropical cyclone intensity forecasts are more accurate than those of ECMWF HRES, demonstrating its improved forecasting capabilities.
In conclusion, recent advances in ML for weather forecasting have led to models surpassing the ECMWF HRES in terms of global forecast accuracy. These ML models typically offer a temporal resolution of 6 hours and a spatial resolution of 0.25°, but are limited by their focus on basic weather variables. The FuXi-2.0 model addresses these limitations by providing hourly forecasts and including a wider range of crucial variables for sectors such as wind and solar power, aviation and shipping. FuXi-2.0 outperforms ECMWF HRES and integrates atmospheric and oceanic data to improve tropical cyclone forecasts. Future improvements include higher spatial resolutions, additional variables, and improved precipitation accuracy.
Discover the Paper. All credit for this research goes to the researchers of this project. Also don’t forget to follow us on Twitter and join our Telegram channel And LinkedIn Groops. If you like our work, you will love our bulletin..
Don’t forget to join our 50,000+ ML subreddit
Sana Hassan, Consulting Intern at Marktechpost and a dual degree student at IIT Madras, is passionate about applying technology and AI to address real-world challenges. With a keen interest in solving practical problems, he brings a fresh perspective to the intersection of AI and real-world solutions.
