Criação de base de dados inédita de usinas solares e aplicação em previsão de geração de energia com redes neurais otimizadas por aprendizado por transferência

Abstract

The accelerated growth of solar photovoltaic generation in Brazil highlights the need for structured and high-quality databases capable of supporting the development of new technologies and improving the reliability of studies in the sector. Among the main challenges is the creation of datasets that enable the comparison and evaluation of generation forecasting models with greater scientific rigor. In this context, the primary objective of this work is the creation and public release of the BR-PVGen database, composed of historical records of electrical generation and meteorological variables from 51 photovoltaic power plants operating in real conditions across Brazil. The database was processed according to criteria of quality, completeness, and standardization, constituting an unprecedented repository designed to support research on forecasting, performance analysis, and optimization of photovoltaic systems. In addition, the study presents an application of machine learning models, particularly Artificial Neural Networks with architectures optimized through hyperparameter search and Transfer Learning techniques, applied to the same BR-PVGen database. The experiments demonstrated that Transfer Learning between plants with different data histories reduced the average forecasting error, outperforming reference models such as Gradient Boosting, Bayesian Ridge, and Kernel Ridge, with emphasis on the SMAPE and NRMSE metrics. The results show that BR-PVGen represents a significant advancement in the consolidation of photovoltaic data in Brazil, promoting reproducibility and standardization of studies in the field. It is concluded that the combination of a multivariate, standardized, and nationwide database with modern machine learning techniques constitutes a promising path for improving forecasting accuracy and for the intelligent integration of solar energy into the Brazilian power system.