Radiation-constrained AI weather model beats ECMWF HRES on sunlight and cloud forecasts

Why it matters

Most AI weather models do not predict clouds, radiation or albedo, yet these matter for solar power and the energy budget. This work shows one way to add them while keeping the speed of machine learning. The gains also carried into day-ahead solar power forecasts at four of five farms.

What they did

The authors built MLRTM, a neural network that imitates the RRTMG shortwave radiation scheme and can pass gradients back. They coupled it to the FuXi weather model during training, creating FuXi-RTM. They compared it with FuXi-base, FuXi-base+ (which adds radiation targets but no gradients) and ECMWF HRES. Tests used ERA5, BSRN ground stations, CERES satellite products and five solar farms, for forecasts from January 2024 to February 2025.

Key findings

• FuXi-RTM beat ECMWF HRES on radiation and cloud cover in both RMSE and ACC across the 10-day forecasts, and beat the unconstrained variants.
• FuXi-RTM beat FuXi-base in 97.6% of 4360 variable-lead time combinations, and FuXi-base+ in 96.8%.
• Adding radiation targets without gradients (FuXi-base+) gave almost no benefit over FuXi-base, so the gradient constraint drives the gain.
• FuXi-RTM had lower irradiance RMSE than HRES at most of 26 BSRN stations.
• For day-ahead solar power, FuXi-RTM beat HRES at four of five farms; HRES was better at Guangdong_01 (91.9% against 91.4%). HRES also stayed better for surface albedo.

Limitations

• ERA5 verification mainly tests consistency with the reanalysis and RRTMG-based targets, and it favors the AI models at short lead times for accumulated variables.
• FuXi-RTM is worse than HRES for albedo over complex terrain such as the Tibetan Plateau, and the surrogate covers only shortwave radiation, leaving out aerosols, ozone and extra hydrometeors.
• The solar power test covers only five farms and two months, so broader validation is needed.

Glossary

• Radiative transfer model: A model that calculates how sunlight and heat radiation move through, and are absorbed or scattered by, the atmosphere.
• Differentiable surrogate: A neural network that imitates a physical model and lets error signals flow back through it during training.
• Surface albedo: The fraction of incoming sunlight that the ground reflects.
• SEDI: The symmetric extremal dependency index, a score for how well rare, extreme events are forecast.