Estimação de dados faltantes em séries temporais de irradiação solar: uma abordagem baseada em aprendizado de máquina

Abstract

This Master’s thesis investigates missing data imputation in solar irradiance time series using machine learning-based approaches. Missing values pose a major challenge for analyses and applications in meteorology and solar energy, where data continuity and quality are essential for climate forecasting and photovoltaic system planning and operation. This study evaluates and compares statistical methods and machine learning models, including Multivariate Imputation by Chained Equations, Autoencoders, Long Short-Term Memory, and Transformers, applied to meteorological data provided by the National Institute of Meteorology. The methodology includes a characterization of missingness patterns and an analysis based on bioclimatic zones, which supports the selection of stations of interest and the incorporation of spatial and climatic context into the imputation process. Experiments were conducted under a controlled masking protocol with evaluation restricted to the missing positions, and results were compared against classical baselines such as linear interpolation. Overall, the findings indicate that models capable of capturing temporal dependencies and leveraging neighborhood information achieve higher robustness, particularly under longer contiguous gaps.