FuXi-2.0: Advancing machine learning weather forecasting model for practical applications

TL;DR

Comparisons between ML-based 1-hourly forecasts and those from the high-resolution forecast (HRES) of the European Centre for Medium-Range Weather Forecasts (ECMWF) for various practical scenarios demonstrate that FuXi-2.0 consistently outperforms ECMWF HRES in forecasting key meteorological variables relevant to these sectors and proves its efficacy as a reliable tool for scenarios demanding precise weather forecasts.

Abstract

Machine learning (ML) models have become increasingly valuable in weather forecasting, providing forecasts that not only lower computational costs but often match or exceed the accuracy of traditional numerical weather prediction (NWP) models. Despite their potential, ML models typically suffer from limitations such as coarse temporal resolution, typically 6 hours, and a limited set of meteorological variables, limiting their practical applicability. To overcome these challenges, we introduce FuXi-2.0, an advanced ML model that delivers 1-hourly global weather forecasts and includes a comprehensive set of essential meteorological variables, thereby expanding its utility across various sectors like wind and solar energy, aviation, and marine shipping. Our study conducts comparative analyses between ML-based 1-hourly forecasts and those from the high-resolution forecast (HRES) of the European Centre for Medium-Range Weather Forecasts (ECMWF) for various practical scenarios. The results demonstrate that FuXi-2.0 consistently outperforms ECMWF HRES in forecasting key meteorological variables relevant to these sectors. In particular, FuXi-2.0 shows superior performance in wind power forecasting compared to ECMWF HRES, further validating its efficacy as a reliable tool for scenarios demanding precise weather forecasts. Additionally, FuXi-2.0 also integrates both atmospheric and oceanic components, representing a significant step forward in the development of coupled atmospheric-ocean models. Further comparative analyses reveal that FuXi-2.0 provides more accurate forecasts of tropical cyclone intensity than its predecessor, FuXi-1.0, suggesting that there are benefits of an atmosphere-ocean coupled model over atmosphere-only models.

Figures (6)

• Figure 1: Comparison of globally-averaged and latitude-weighted root mean squared error ($\textrm{RMSE}$) (first row), anomaly correlation coefficient ($\textrm{ACC}$) (second row), and forecast/observation activity (third row), as well as normalized differences in $\textrm{RMSE}$ (fourth row), $\textrm{ACC}$ (fifth row), and activity (sixth row) of ECMWF HRES (blue lines), FuXi-2.0 (red lines), Pangu-Weather (purple lines), and ERA5 (black lines represent observation activity) for 3 surface variables: 2-meter temperature ($\textrm{T2M}$) (first column), mean sea level pressure ($\textrm{MSL}$) (second column), and 10-meter wind speed ($\textrm{WS10M}$) (third column), in 90-hour forecasts at a temporal resolution of 1 hour using testing data from 2018. The $\textrm{RMSE}$ and $\textrm{ACC}$ are calculated against ERA5, and normalized differences in $\textrm{RMSE}$, $\textrm{ACC}$, and activity are calculated using ECMWF HRES as the baseline model.
• Figure 2: Comparison of globally-averaged and latitude-weighted root mean squared error ($\textrm{RMSE}$) (first row), anomaly correlation coefficient ($\textrm{ACC}$) (second row), and forecast/observation activity (third row), as well as normalized differences in $\textrm{RMSE}$ (fourth row), $\textrm{ACC}$ (fifth row), and activity (sixth row) of HRES (blue lines), FuXi-2.0 (red lines), and Pangu-Weather (purple lines), and ERA5 (black lines represent observation activity) for 3 upper-air variables, including 850 hPa temperature ${\textrm{T850}}$ (first column), 500 hPa geopotential ${\textrm{Z500}}$ (second column), 700 hPa specific humidity $\textrm{Q700}$ (third column), in 90-hour forecasts at a temporal resolution of 1 hour using testing data from 2018. The $\textrm{RMSE}$ and $\textrm{ACC}$ are calculated against ERA5, and normalized differences in $\textrm{RMSE}$, $\textrm{ACC}$, and activity are calculated using ECMWF HRES as the baseline model.
• Figure 3: Comparison of normalized differences in globally-averaged and latitude-weighted root mean squared error $\textrm{RMSE}$ (first column), anomaly correlation coefficient $\textrm{ACC}$ (second column), and forecast activity (third column) of FuXi-2.0 compared to ECMWF HRES, in 90-hour forecasts at a temporal resolution of 1 hour using testing data from 2018. The evaluation encompasses a broad range of variables relevant to different sectors. For wind and solar energy forecasting (first row), variables include 100-meter u wind component (${\textrm{U100M}}$), 100-meter v wind component (${\textrm{V100M}}$), 100-meter wind speed (${\textrm{WS100M}}$), surface net solar radiation (${\textrm{SSR}}$), surface solar radiation downwards (${\textrm{SSRD}}$), and total sky direct solar radiation at surface (${\textrm{FDIR}}$). In the aviation sector (second row), variables include low cloud cover (${\textrm{LCC}}$), medium cloud cover (${\textrm{MCC}}$), high cloud cover (${\textrm{HCC}}$), and total cloud cover (${\textrm{TCC}}$). For marine shipping (third row), variables evaluated include sea surface temperature (${\textrm{SST}}$), mean direction of total swell (${\textrm{MDTS}}$), mean direction of wind waves (${\textrm{MDWW}}$), mean period of total swell (${\textrm{MPTS}}$), mean period of wind waves (${\textrm{MPWW}}$), significant height of total swell (${\textrm{SHTS}}$), and significant height of wind waves (${\textrm{SHWW}}$). All variables are evaluated at a temporal resolution of 1 hour using testing data from 2018. The $\textrm{RMSE}$ and $\textrm{ACC}$ are calculated against ERA5, and normalized differences in $\textrm{RMSE}$, $\textrm{ACC}$, and activity are calculated using ECMWF HRES as the baseline model.
• Figure 4: Comparison of wind speed and wind power at Kelmarsh and Yeongheung wind farms. The top row illustrates one-year averaged 10-meter wind speed ($\textrm{WS10M}$) and 100-meter wind speed ($\textrm{WS100M}$) as a function of forecast lead times for Kelmarsh wind farm. The second rows shows one-year averaged wind power forecasts as a function of forecast lead times, for both Kelmarsh and Yeongheung wind farms. The third and fourth row present example two 15-day time series of observed and predicted wind power from November 16th to December 1st, 2018, for Kelmarsh, and from November 5th to November 20th, 2018, for Yeongheung, respectively. The forecast horizon for these day-ahead wind power forecasts ranges from 28 to 51 hours.
• Figure 5: Comparison of the average mean absolute error (MAE) for tropical cyclone (TC) track forecasts (first column) and root mean squared error (RMSE) for mean sea-level pressure (${\textrm{MSL}}$) (second column) and 10-meter wind speed ($\textrm{WS10M}$) (third column) in TC intensity forecasts for HRES (blue lines), FuXi-1.0 (green lines), FuXi-2.0 (red lines), Pangu-Weather (purple lines), International Best Track Archive for Climate Stewardship (IBTrACS) (black lines with solid circles), and ERA5 (black lines with squares), as a function of forecast lead times. The evaluation covers forecasts for 90 TCs, and is performed against the IBTrACS dataset. The $\textrm{MSL}$ and $\textrm{WS10M}$ forecast comparisons are dual Y axis figures, here the secondary Y axis shows the IBTrACS data (black lines with solid circles) in the first row and ERA5 (black lines with squares) in the second row. Numbers in brackets on the x-axis indicate the sample size for each forecast lead time. For example, ‘(734)’ means 734 samples from FuXi-1.0, FuXi-2.0, and ECMWF-HRES were averaged from 734 initial points where the TC persisted for at least 24 hours.