Temporally Validated Machine Learning for Solar Radiation as a Foundation for UV-Risk Modeling in High-Altitude Puno

High-altitude UV-risk services require predictive models whose target, validation design, and atmospheric predictors are physically aligned. This study establishes a reproducible machine-learning foundation for that objective by modeling the quantity actually measured at the Puno automatic station: hourly global solar radiation. We analyzed 43,739 records from 2020–2024 and formulated prediction as supervised regression under a prospective temporal holdout. Calendar and wind direction were mapped to cyclic coordinates; predictor gaps were imputed using training-only medians; and 2020–2023 observations were used to estimate a month–hour climatology, ordinary least squares, ridge regression, and gradient-boosted decision stumps. Performance was evaluated exclusively on 2024 using MAE, RMSE, and R². The climatology yielded the best full-day generalization (MAE 57.6 W/m², RMSE 116.9 W/m², R² 0.896), outperforming boosting in RMSE by 3.0%. Boosting reduced daytime RMSE by only 1.0% but increased daytime MAE by 10.8%. These results identify deterministic solar geometry as the dominant predictable component and missing optical variables as the principal barrier to short-term accuracy. The framework is therefore a validated benchmark for future UV-risk modeling, not a surrogate UV-index predictor; operational extension requires colocated erythemal UV, cloud, ozone, aerosol, and albedo measurements.