Variabilidade climática e estimativa da geração fotovoltaica: análises observacionais e modelagem em diferentes escalas no Brasil

Abstract

Climate change may impact the performance of renewable energy systems, as the efficiency of photovoltaic power generation is highly sensitive to variations in meteorological variables. In this context, understanding local climatic conditions becomes essential for the proper sizing, planning, and operation of these systems. Thus, this thesis aims to analyze the influence of climatic variables on photovoltaic power generation in locations across Brazil, through the analysis of historical meteorological data and the evaluation of different generation estimation methods. As a contribution, it incorporates an approach based on cluster analysis to investigate climatic patterns, enabling a more detailed assessment of temporal variability. The thesis is structured into chapters in the format of scientific articles. In Chapter 2, trend analysis was performed using the Mann–Kendall test, associated with Sen’s slope estimator, in conjunction with a method based on cluster analysis. The results indicated a significant increase in mean air temperature at 36 meteorological stations in the Brazilian semi-arid region. The cluster analysis made it possible to identify that more recent years are associated with higher temperature values, as well as to highlight the periods of the year in which these increases were more pronounced. This approach also enabled the detection of variations not identified by traditional trend analysis. In Chapter 3, a comparison was carried out between observed and estimated photovoltaic generation using five empirical models. The results indicated that all models exhibited a tendency to overestimate generation, with relative errors ranging from approximately 19% to 21%, although they showed similar performance in representing data variability. Among the evaluated models, Model 3 presented the best performance, with lower bias and dispersion. In Chapter 4, multiple linear regression, Random Forest, and an empirical model were applied to estimate monofacial photovoltaic generation, using temporal validation with TimeSeriesSplit. The Random Forest model showed superior performance, with an average coefficient of determination of approximately 0.66 and a mean squared error of 7.64 (kWh/day)², while multiple linear regression presented an R² of around 0.61. Hyperparameter optimization led to further improvements, achieving an R² of 0.67 and an MSE of 6.42 (kWh/day)². In Chapter 5, bifacial photovoltaic generation was estimated using empirical equations and the Random Forest method. The results indicated a gain of approximately 12.6% in effective irradiance, resulting in an increase of about 83.42 MWh in energy generation. The Random Forest model showed better performance (RMSE ≈ 175 kWh and R² = 0.76), while the empirical model exhibited a tendency to underestimate generation. Overall, the results highlight the importance of a detailed analysis of local climatic conditions and demonstrate that the combination of statistical methods, machine learning techniques, and cluster analysis enables a more robust evaluation of photovoltaic power generation.