Comparing solar forecasts across different contexts is essential to improve forecast performance and integrate solar energy into power systems. The skill score (SS), a standardized metric that accounts for baseline models, enables such comparisons. Yet, no meta-analysis has systematically examined forecasts across studies using this metric. This article presents the first comprehensive meta-analysis of deterministic solar forecasts based on SS, using multivariate adaptive regression splines, partial dependence plots, and linear regression. Results reveal horizon-specific patterns. Sky or satellite images and historical power data improve intra-hour and intra-day forecasts by 5–8 percentage points (pp) in SS. Spatiotemporal information is particularly useful for intra-hour forecasts (+5.55 pp), while locally measured meteorological data are more relevant for intra-day (+4.36 pp) and day-ahead (+6 pp). Day-ahead forecasts perform best with numerical weather prediction data (+11.5 pp). Ensemble–hybrid models have the most robust performance across horizons, increasing SS by 7–27 pp compared to time-series models, while many advanced machine learning methods show inconsistent gains. Climate conditions strongly affect forecast performance, highlighting challenges in transfer learning. These findings provide actionable guidance for researchers and practitioners and support robust benchmarking of solar forecasting models.
Nguyen et al. (Sun,) studied this question.