This study investigates the application of machine learning, deep learning, and hybrid approaches for predicting solar radiation and meteorological variables. Using a dataset of 6,421 hourly observations across eight features, the study compared traditional models, including Extreme Gradient Boosting, Support Vector Machine, and Least Squares regression, with advanced models such as Recurrent Neural Network, Long Short-Term Memory network, and hybrid frameworks combining these two types of models. The results demonstrate that hybrid models, particularly the Extreme Gradient Boosting–Recurrent Neural Network and Extreme Gradient Boosting–Long Short-Term Memory models, consistently outperform other approaches, achieving coefficients of determination values above 0.999 with the lowest Root Mean Square Error and Mean Absolute Error. Deterministic parameters such as solar zenith angle, clear-sky surface downward shortwave radiation, and all-sky clearness index were predicted with high accuracy, while stochastic variables such as wind speed at 10 meters and surface albedo exhibited lower predictive accuracy due to their higher variability. Feature importance and local interpretable model-agnostic explanations analysis confirmed the dominance of physical constraints in predictive accuracy. The findings highlight the strong potential of hybrid machine learning–deep learning models for renewable energy forecasting, atmospheric analysis, and climate-related applications. This study not only advances methodological understanding but also offers practical insights for operational deployment of photovoltaic systems in Hue City.
Le et al. (Mon,) studied this question.