Spatio-Temporal Graph Neural Network for Solar Irradiance Prediction: A Case Study in Nganjuk, Indonesia

Solar energy utilization in tropical regions is strongly influenced by the accuracy of solar irradiance estimation, which is affected by temporal variability and spatial atmospheric interactions. Conventional forecasting approaches commonly rely on single-station time-series models, limiting their ability to capture regional dependencies. This study proposes a spatio-temporal modeling framework based on Graph Neural Networks (GNN) to estimate solar irradiance by explicitly incorporating spatial relationships among observation sites. The study focuses on Sawahan Subdistrict, Nganjuk Regency, Indonesia, using solar irradiance data collected from five Automatic Weather Stations (AWS) operated by the Indonesian Agency for Meteorology, Climatology, and Geophysics (BMKG) during 2024. Each station is represented as a graph node, with spatial connections constructed based on geographical distance, while temporal dependencies are modeled using Long Short-Term Memory (LSTM). Experimental results show that the proposed model achieves a Mean Absolute Error (MAE) of 102.64 W/m², a Root Mean Squared Error (RMSE) of 166.76 W/m², and an R² value of 0.6446 for the target location. These findings demonstrate that GNN-based spatial aggregation improves estimation stability and accuracy, providing practical support for localized solar energy assessment in tropical regions.