Optimizing Intra-Hour Solar Forecasts with Satellite-Driven Deep Learning Models

With the increasing integration of large-scale solar energy into electricity grids, anticipating fluctuations in solar power supply has become critical. Very short-term solar forecasting, particularly under rapidly changing cloud conditions, remains a significant challenge. Conventional forecasting systems that rely on Numerical Weather Prediction (NWP) models like the High-Resolution Rapid Refresh (HRRR) often fail to capture these rapid changes unless supplemented with current cloud observations from satellite or ground-based imagers. Projecting cloud evolution using methods like kinematic extrapolation with cloud motion vectors (CMVs) has significantly enhanced forecast accuracy in many previous studies. For instance, the NCAR Sun4Cast system, which blends several observation-based nowcasting technologies, achieved irradiance forecast improvements of 36-56% for lead times under 6 hours.

Over the past five years, deep learning techniques have been increasingly applied to satellite imagery for enhanced solar energy forecasting. Initial efforts employed common architectures like convolutional neural networks (CNNs) and encoder-decoders (U-Nets). Subsequent models, integrating encoders with recurrent units, demonstrated superior performance by capturing both spatial covariances and temporal evolution. Inspired by these successes, our approach aimed to address their limitations by leveraging both visible and infrared radiances to improve handling of the night-to-day transition and to mitigate the smoothness of predictions where possible.

After testing two candidate architectures, we developed the CoaT-GRU model, which combines an image transformer (CoaT) with a convolutional gated-recurrent unit (ConvGRU). The model was trained on GOES-16 satellite images from 2022, predicting image sequences out to three hours with four modalities: visible, near-infrared, infrared, and binary cloud mask. We employed both L1 and SSIM loss functions, incorporating lead time, elevation, and solar angles as auxiliary input variables. We experimented with tile size to optimize the perceptive field of the model, ensuring it effectively captures relevant spatial features for improved forecast accuracy. Performance of our deep learning model was benchmarked against strong optical flow baselines.

The CoaT-GRU model exhibited significant improvements over optical flow baselines, for lead times beyond 10 minutes extending to 3 hours. At 30-minute lead times, the model achieved a 24% improvement in infrared channel predictions and a 50% improvement in visible channel predictions compared to the best optical flow method. Regionally, infrared performance improvements ranged from 20-41% across various U.S. climate zones.

Encouraged by these results, we deployed the CoaT-GRU model in Vaisala Xweather’s operational environment in late 2023, introducing it as a new data source for site-specific solar power forecasts. At selected U.S. utility solar projects, including this model improved 5-minute solar power forecast accuracy by up to 20% in Mean Absolute Error (MAE) out to 1-hour lead times. In 2024, we expanded the model’s application to the GOES-18 field of view and fine-tuned it for use with Meteosat-11 Rapid Scan Service (RSS) images in Europe. The model performed notably well during the night-to-day transition, though challenges remain, particularly the tendency to smooth predictions at longer lead times.

Future efforts will explore generative AI and improved training methods to retain finer-scale features and enable fast ensemble predictions, while also addressing the increased computational demands.