Short-Term Solar Forecasting for San Cristóbal Island Using Ground Meteorological Instruments

Isolated island microgrids historically rely on unstable, carbon-intensive thermal energy generation via fossil fuel imports. While integrating renewable energy resources can mitigate severe ecosystem risks, high short-term meteorological variability complicates grid stabilization and planning. The Galapagos Islands, which lack a connection to the Ecuador national grid, face this fundamental tension as they continue to deploy more wind and solar infrastructure.

We propose a solar nowcasting framework trained exclusively on ground-based meteorological data from the El Junco weather station, located 700 m above sea level on San Cristóbal Island, Galapagos. Utilizing a 49-feature architecture engineered entirely from pre-existing instruments, we developed six independent Extreme Gradient Boosting (XGBoost) regression models. Each regressor is trained to forecast changes in solar radiation across 30- to 180-minute horizons.

The experimental framework significantly outperforms both standard persistence and climatology baselines across all tested horizons. The models yield a Mean Absolute Error ranging from 34.57 W/m² at the +30-minute horizon to 61.74 W/m² at +180 minutes. Furthermore, feature importance evaluations via SHAP analytics verify that the underlying machine learning models base their predictions on physically coherent dynamics, heavily prioritizing cyclical temporal features alongside current and lagged solar radiation values.

Accurate short-term solar forecasting for isolated environments is fully achievable using an existing ground weather station network, reducing the dependency on satellite imagery, numerical models, or costly additional sensor deployment. This strategy provides a replicable tool that empowers grid operators to anticipate solar generation changes, facilitating the broader renewable energy transition of isolated microgrids worldwide.