Hourly UV Index Forecasting Across Climatic Zones in Sri Lanka: A Residual-Correction Hybrid Machine Learning and Deep Learning Approach

Ultraviolet (UV) radiation reaching the Earth's surface has shown a long-term increasing trend, raising concerns due to its adverse effects on human health, including an elevated risk of skin cancer. The Ultraviolet Index (UVI) is a standardized measure of UV radiation intensity and plays an important role in public health protection and climate adaptation planning. However, UVI forecasting studies in Sri Lanka remain limited, particularly those comparing traditional statistical, machine learning (ML), deep learning (DL), and hybrid forecasting approaches across different climatic zones. This study develops and evaluates forecasting frameworks for hourly UVI prediction in three representative climatic regions of Sri Lanka: Colombo (wet zone), Badulla (intermediate zone), and Jaffna (dry zone). Hourly UVI observations from January 2018 to December 2023 were obtained from the NASA POWER database. Since UVI values between 18:00 and 05:00 were consistently zero, the analysis was restricted to daylight hours (06:00–17:00). Traditional time-series, ML, and DL models were evaluated for one-hour-ahead forecasting using a univariate framework. In addition, a residual-correction hybrid approach was proposed, in which residuals from the best-performing standalone model were modeled using a secondary learning algorithm to improve prediction accuracy. The results show that ML and DL models consistently outperform traditional statistical approaches, while hybrid models achieve the highest predictive performance across all climatic zones. The CNN–RF hybrid achieved the best performance in Colombo (RMSE = 0.3967, R² = 0.9881), whereas the FNN–RF hybrid performed best in Badulla (RMSE = 0.4265, R² = 0.9847) and Jaffna (RMSE = 0.4291, R² = 0.9844). Furthermore, lag analysis revealed that both short-term dependencies and recurring diurnal patterns contribute substantially to forecasting performance. These findings demonstrate the effectiveness of residual-correction hybrid forecasting for regional UVI prediction and its potential application in public health early warning systems.