Spatiotemporal Neural Forecasting for Solar Power Generation Across Distributed Energy Networks

Abstract

Accurate renewable energy forecasting is essential for improving power system stability and supporting the integration of distributed solar generation into modern electricity networks. However, solar power output is strongly influenced by rapidly changing weather conditions and spatial correlations among geographically distributed facilities. This study develops a spatiotemporal neural forecasting framework that combines graph-based spatial representations with attention-driven temporal modeling. The proposed model integrates historical power generation, solar irradiance, temperature, cloud coverage, and geographical information to predict short- and medium-term solar energy production. The illustrative experimental dataset consists of four years of 15-minute operational measurements from 92 solar farms distributed across multiple climatic regions. For 24-hour forecasting, the proposed method achieves a mean absolute percentage error of 7.6%, compared with 10.1% for the strongest recurrent neural network baseline. The root mean squared error decreases from 0.57 to 0.43 MW, while the empirical coverage of nominal 90% prediction intervals reaches 89.6%. These findings indicate that spatial dependency modeling can improve forecasting reliability across distributed renewable energy systems.