Satellite-Based Clear-Sky Index Nowcasting with U-Nets to Improve Solar Irradiance Forecasting in the Indian Ocean

Tropical regions exhibit high solar potential, making them well suited for photovoltaic deployment. Moreover, in non-interconnected zones, such as islands, solar power generation can contribute substantially to electricity supply. In this context, accurate intra-hour forecasts of surface solar irradiance facilitate the operation of photovoltaic plants and their smooth integration into the power grid. However, short-term (0-1h) irradiance forecasting remains particularly challenging on tropical islands due to the high variability of cloud cover.

Cloud cover forecasting techniques based on satellite observations are of great value for ground-level solar irradiance forecasting. Among them, optical flow algorithms for deriving Cloud Motion Vectors (CMVs) are now considered a standard in the field. Recently, deep learning approaches, especially U-Net architectures, have shown strong promise in improving the short-term anticipation of cloud positions. However, the extent to which these models can enhance subsequent irradiance forecasting in tropical environments remains to be investigated.

This study introduces a lightweight U-Net-based model designed to simultaneously forecast four future clear-sky index (K_c) maps, with lead times of up to one hour. The model takes four past K_c maps as input, which are processed from Meteosat-9 visible channels (Indian Ocean Data Coverage, 45.5°E). First, distinct models are trained using 2.5 years of satellite data. While sharing the same architecture and training procedure, each model is tailored to a specific territory in the South-West Indian Ocean through dedicated training data. Then, independent evaluations are performed on a separate year. Beyond the assessment of forecast K_c map quality, surface global horizontal irradiance (GHI) is derived from model outputs at specific station locations. These values are then compared with ground-based GHI measurements from the Indian Ocean Solar Network (IOS-net).

The results demonstrate that our U-Net-based models improve spatial K_c map forecasts by approximately 10-15% over standard CMV methods, in terms of mean absolute error (MAE) and root mean square error (RMSE). Consequently, local GHI forecasts are typically enhanced by 5-10% across all territories tested. In addition, the models exhibit promising cross-island generalisation, suggesting that they capture robust location-invariant patterns. These findings point to the potential for developing a unified model for tropical islands, as well as for leveraging transfer learning approaches, which could simplify the practical deployment of such models from an operational perspective.  Overall, this study establishes U-Net-based models as valuable candidates for improving intra-hour GHI forecasts over tropical islands in the South-West Indian Ocean, with the potential to further increase photovoltaic penetration in the local energy mixes.