Solar Irradiance Forecasting: Methods, Challenges, and Emerging Trends
Kusum Dalal *
Department of Electronics and Communication Engineering, Deenabandhu Chhotu Ram University of Science and Technology, Murthal, India.
*Author to whom correspondence should be addressed.
Abstract
Solar irradiance forecasting connects atmospheric observations with decisions about photovoltaic generation, storage and electricity-system operation. Yet comparisons between forecasting methods are frequently weakened by differences in radiation components, forecast horizons, spatial support, information availability and verification procedures. This critical narrative review examines physical, statistical and learning-based approaches, with particular attention to the relationship between predictive information and defensible evidence of improvement. Literature published from 2010 to 1 August 2026, supplemented by foundational contributions, was selected through targeted scholarly searches and appraised for measurement quality, chronological integrity, benchmark strength, uncertainty treatment and operational relevance. The synthesis indicates that short-horizon performance depends strongly on observing cloud structures before they affect the target, whereas longer horizons require credible representation of atmospheric evolution. Machine learning can improve feature extraction, model combination and forecast calibration, but architectural complexity alone does not establish transferable skill. Comparative studies provide stronger support for combining complementary observations and forecasts than for a universally superior model family. Probabilistic methods broaden the information available to decision-makers, although calibrated marginal distributions do not guarantee realistic spatial or temporal scenarios. Emerging diffusion models, multimodal networks and physics-informed approaches offer promising capabilities, but their evaluation must distinguish satellite retrieval from forecasting and genuine advance information from future observations used as experimental upper bounds. Persistent challenges include geographically uneven observations, component-specific measurement errors, changing cloud and aerosol regimes, data leakage and inadequate testing of deployment constraints. A decision-oriented research agenda should prioritise independent multi-site benchmarks, availability-aware forecast archives, conditional uncertainty verification and prospective operational trials. Progress is therefore best judged by reproducible gains under realistic information constraints, rather than by isolated improvements in average error on a single dataset.
Keywords: Cloud nowcasting, numerical weather prediction, satellite imagery, probabilistic prediction, forecast verification, multimodal learning, operational decision support