Missing data imputation using climate data reanalysis in solar energy generation time series

• Dealing with solar energy generation missing data using reanalysis data. • Framework combining clustering, climate reanalysis, and imputation techniques. • Selecting relevant climate variables to improve missing data imputation. • Machine learning models outperform statistical methods in data imputation. • Convex combination of top models improves imputation. The global effort to reduce carbon emissions has led to more investment in renewable energy, particularly solar energy. However, missing values in historical series still pose a significant obstacle to forecasting models, affecting the reliability of estimates for solar energy generation. This study, therefore, proposes a framework for analyzing the characteristics of solar power generation time series, as well as for incorporating climate variables such as temperature, surface humidity, and global horizontal irradiance into imputation models. Using data from Brazilian solar power plants and the MERRA-2 reanalysis dataset, the framework evaluates statistical and machine learning techniques, including column median imputation, Random Forest, and LightGBM. The analysis suggests that simpler statistical techniques have relevant potential in certain contexts, while machine learning approaches, especially tree-based algorithms, tend to outperform traditional methods. A convex combination of the top-performing models is also tested, which improves imputation performance across different time series patterns. By applying time series clustering and variable selection strategies, the framework provides a more comprehensive understanding of generation patterns, enabling model adjustments that lead to a refined imputation performance and data reliability. By addressing the challenges posed by missing data, this study makes valuable contributions to enhancing the quality of solar energy generation time series and supporting the development of more accurate forecasting tools.