Inteligência artificial aplicada à previsão do potencial de geração de energia solar

Abstract

The growing concern with environmental impacts from fossil fuel-based energy generation, coupled with the worsening of global warming, has driven e!orts toward a more sustainable energy matrix transition. In this scenario, solar energy plays a strategic role, establishing itself as one of the main renewable sources. In this context, the present work aims to model and predict the generation potential of photovoltaic systems in the state of Goiás, Brazil, using meteorological data combined with real generation records from existing photovoltaic systems, involving the development of predictive models that will be compared to consolidated approaches. Based on this approach, we seek to contribute to the advancement of planning, optimization, and energy management strategies for photovoltaic systems in a regional context, strengthening the technical-scientific understanding of the integration between meteorological variables and solar generation e"ciency. The results showed that the Random Forest (RF) model performed best in most plants, o!ering more stable predictions with 24-hour windows. RF achieved R2 values above 0.90 and the lowest mean errors (MAE and RMSE) in several plants. On the other hand, Long Short-Term Memory (LSTM) obtained good results when it had access to more historical data, especially with 48-hour windows. Even with meteorological stations located at considerable distances from the plants, ranging from 151.9 km to 4.0 km, the use of climatic variables still contributed positively to the improvement of predictions.