| Abstract: |
Solar energy forecasting has emerged as a critical technology for enabling grid integration of renewable power sources and optimizing demand-supply management in smart grids. This review synthesizes contemporary advances in machine learning methodologies applied to solar irradiance prediction and photovoltaic power generation forecasting across temporal horizons ranging from very-short-term (minutes) to seasonal scales. We examine supervised learning paradigms including support vector machines, neural networks, ensemble methods, and deep learning architectures, alongside emerging graph neural networks and transformer-based approaches. The review evaluates data integration strategies incorporating satellite imagery, weather variables, and temporal features, while critically analyzing forecasting accuracy metrics and model evaluation frameworks. Recent advances demonstrate that hybrid ensemble methods combining multiple algorithms achieve mean absolute percentage errors below 15% for hour-ahead forecasting in favorable atmospheric conditions. However, significant challenges persist in handling cloud transience effects, rare extreme weather events, and transferability across geographic locations with dissimilar solar climatologies. This paper identifies research gaps in explainable artificial intelligence for forecasting models, uncertainty quantification methodologies, and cost-benefit analyses for large-scale deployment. Future directions emphasize physics-informed neural networks, federated learning for distributed data, and integration with grid-scale energy storage optimization frameworks to maximize renewable energy penetration in global electrical networks. |
