Abstract
Accurate prediction of solar irradiance is critical for optimizing solar energy systems, enhancing grid stability, and supporting sustainable energy transitions. While numerous studies have explored various methodologies for solar radiation prediction, challenges remain in achieving high accuracy across diverse geographic locations and temporal resolutions. This study presents a novel hybrid model combining temporal convolutional networks (TCN), Transformer encoders (TE), and artificial neural networks (ANN) to predict global horizontal irradiance (GHI) with high precision. Utilizing a comprehensive dataset from three significant U.S. solar energy sites-Desert Sunlight, Copper Mountain, and Solar Star-spanning 22 years at a 30-min temporal resolution, the proposed model demonstrated superior performance metrics, with R(2) ranging from 0.94768 to 0.97417, root mean square error (RMSE) between 0.04776 and 0.06543 W/m(2), and mean absolute error (MAE) between 0.02510 and 0.03526 W/m(2). By leveraging TCN's temporal feature extraction, TE's attention mechanisms, and ANN's dense layer refinements, the model demonstrates significant advancements over existing methods.