Previsão de geração de energia fotovoltaica baseada em aprendizado de máquina com monitoramento de degradação e aprendizado incremental

Abstract

The increasing participation of photovoltaic energy in the electrical grid highlights the need for forecasting methods capable of providing accurate estimates of power generation, supporting the planning and operation of power systems. However, predictive models may experience concept drift over time due to changes in data characteristics, requiring continuous adaptation. In this context, this study proposes an adaptive architecture for photovoltaic power forecasting based on machine learning, integrating concept drift monitoring and incremental learning. Initially, different preprocessing strategies, predictive models, and hyperparameter optimization methods were evaluated using historical data from a photovoltaic power plant located in Minas Gerais, Brazil. The evaluated models included Persistence, ARIMA, Ridge Regression, MLP, and LSTM networks. Subsequently, a second dataset was used to evaluate the effects of cyclical features, the removal of weakly correlated variables, and the exclusion of nighttime data. Finally, four adaptive configurations combining the CUSUM and ADWIN concept drift detectors with the finetuning and EWC incremental learning strategies were evaluated. The results showed that mean aggregation, Min-Max normalization, and LSTM models achieved the best predictive performance, particularly for forecasting horizons longer than 15 minutes. Excluding nighttime data improved model learning, whereas the inclusion of cyclical features did not provide consistent performance gains. Among the evaluated adaptive architectures, the ADWIN + EWC combination achieved the best performance, reducing the frequency of model updates while preserving predictive accuracy over time. Furthermore, incremental learning enabled models trained with limited data to progressively adapt, demonstrating its potential for real-world applications.