MapAnalysisPaper

Forecast Performance Analysis

Can AI outperform traditional physics-based models in forecasting complex atmospheric variables? This web-tool allows for a performance comparison of three weather forecast models: Google Deepmind's GraphCast (AI-based), ECMWF Integrated Forecasting System (IFS) (physics-based), and ECMWF AIFS (AI-based). The comparison is based on verification results between each forecasting model against the ERA5 reanalysis dataset. Verification metrics include: RMSE (root mean square error), MAE (mean absolute error), MBE (mean bias error), and R (correlation coefficient). The evaluation covers the first 10 days of each month in 2024, with a focus on the performance of the forecast models across the lead times (10 days in 6 hour increments). The following weather variables have been assessed for their accuracy across the selected regions and lead times:
  • 2-meter temperature T2M
  • Mean sea level pressure MSL
  • 10-meter u-component of wind U10
  • 10-meter v-component of wind V10
  • Specific humidity at 1000 hPa pressure Q
The main objective of this comparative analysis is to evaluate whether the forecasting models perform differently in different geographic regions. Four geographic regions are examined:
  • Global
  • Tropics 23.5°S – 23.5°N
  • Subtropics 23.5° – 35° N/S
  • Northern Temperate 35°N – 60°N
  • Southern Temperate 35°S – 60°S
  • Polar >60°N and >60°S
  • Africa 37°N – 35°S, 17°W – 51°E
/
QUICK LOOK
Forecast Models Compared:
GraphCast (ML-model)
ECMWF IFS (NWP-model)
ECMWF AIFS (ML-model)

Verification Dataset:
ERA5 Reanalysis

Key Geographic Regions:
Global
Tropics + Subtropics
Temperate Zones
Polar + Subpolar
Africa

Metrics Evaluated:
RMSE
MAE
MBE
Correlation Coefficient (R)

Purpose:
• Compare weather forecasting accuracies of ML and NWP models in different climate zones

Region Comparison
Global
Tropics
Subtropics
N. Temperate
(Sub-)Polar
Africa

Lead Times Spatial Performance
(g/kg)

Lead Times Spatial Performance
(g/kg)
X,Y: 0000,0000