TL;DR
Baguan-solar is a novel two-stage AI framework that fuses global weather foundation models with high-resolution satellite imagery to provide kilometer-scale solar irradiance forecasts. By decoupling cloud evolution from irradiance calculation, it achieves a 16.08% reduction in error (RMSE) over state-of-the-art baselines and is currently supporting real-world power grid operations in China.
What Problem Does This Paper Address?
As the world shifts toward renewable energy, power grids face a massive "stability" headache: solar power is notoriously fickle. To keep the lights on, grid operators need accurate Global Horizontal Irradiance (GHI) forecasts 24 hours in advance.
Currently, we have two main ways to predict this, but both have "blind spots":
- Numerical Weather Prediction (NWP) & Weather Foundation Models (WFMs): These are great at predicting large-scale weather patterns over days, but they are "blurry" (low resolution, typically 25km) and often miss the fine-grained cloud structures that cause sudden solar drops.
- Satellite Extrapolation: This uses high-res imagery to "slide" current clouds forward in time. It’s excellent for the next hour (nowcasting) but fails miserably at the 24-hour mark because it doesn't account for the complex physics of clouds forming or dissipating.
The authors bridge this gap with Baguan-solar, a model that combines the "big picture" physics of global models with the "sharp eyes" of geostationary satellites.
Methods
The researchers proposed a two-stage multimodal framework powered by Swin Transformers (a type of hierarchical vision AI):
- Stage 1: Cloud Evolution Modeling. Instead of predicting solar power directly, the model first predicts what the clouds will look like 24 hours from now. It takes historical satellite data and fuses it with forecasts from Baguan (Alibaba’s global weather foundation model). This stage is "day-night continuous"—it tracks clouds even at night when there is no solar power to measure.
- Stage 2: Irradiance Inference Modeling. This stage takes the predicted clouds from Stage 1 and combines them with "Clear-Sky GHI" (the theoretical maximum solar energy if there were zero clouds) and other meteorological variables (like humidity and temperature).
- Modality Fusion: The model uses a cross-attention mechanism to let the high-res satellite "texture" be guided by the large-scale "constraints" of the global weather model.
Datasets
The study utilizes a massive multimodal dataset covering East Asia:
- Himawari-8/9 Satellite: Multi-spectral imagery at 0.05° resolution (approx. 5km).
- Baguan Forecasts: Alibaba's operational WFM providing global weather variables at 0.25° resolution.
- CLDAS (China Land Data Assimilation System): Used as the high-resolution "ground truth" for training and evaluation.
- ERA5: Global reanalysis data used for historical context during training.
- Pyranometer Network: Real-world measurements from 246 physical ground stations were used for final operational verification.
Key Insights and Ideas
- The "Night Shift" Strategy: GHI is zero at night, which usually creates a "discontinuity" that confuses AI models. By forcing the model to predict clouds (Stage 1) before irradiance (Stage 2), the model maintains physical consistency through the night, allowing it to be "ready" the moment the sun rises.
- Temporal Hand-off: Through an importance analysis (Integrated Gradients), the authors found that satellite data is the primary driver for the first 6 hours of a forecast. However, after 6 hours, the global weather model (ERA5/Baguan) becomes the dominant factor. Baguan-solar learns this "hand-off" automatically.
- Vectorized Physics: To make the model fast enough for production, the team re-implemented the standard Ineichen–Perez clear-sky physics model using vectorized NumPy routines, achieving a 240x speedup in calculating theoretical solar maximums.
Key Results
- Superior Accuracy: Baguan-solar reduced the average RMSE by 16.08% compared to the strongest baseline (a two-stage Swin Transformer using only satellite data) and by 28.02% compared to the vanilla Baguan weather model.
- Better Transients: Qualitatively, the model is much better at capturing "vortex" cloud patterns and sharp gradients that other models tend to "smooth out" into blurry blobs.
- Operational Deployment: Since July 2025, the model has been live in an eastern Chinese province, supporting a regional grid with a solar capacity of over 900 GW.
- Robustness: Even when using real-time forecasts (which have their own errors) instead of perfect historical data, the model's performance remained stable.
Limitations and Future Work
- Geographic Coverage: The current evaluation is focused on East Asia (the footprint of the Himawari satellite). Expanding this globally would require integrating other geostationary satellites like GOES (Americas) or Meteosat (Europe/Africa).
- Point Forecasts: The model currently provides a single "best guess" value. The authors suggest that future versions should include uncertainty quantification (e.g., "there is a 90% chance the irradiance will be between X and Y") to help grid operators manage risk.
- End-to-End Co-training: Currently, the global weather model and the solar model are separate. Training them together from scratch could potentially unlock even higher accuracy.
Summary updated