Avaliação de técnicas de Deep Learning para prognósticos de energia solar.

Abstract

In recent years, solar energy has gained significant relevance due to its renewable nature, considered one of the sources with almost no impact on the environment and economically viable. Due to the geographic conditions of the NEB, the use of photovoltaic energy becomes a great option for the diversification of energy sources. In this context, this research is motivated by the analysis of the application of Deep Learning techniques, known as "Deep Learning" DL, for the prognosis of solar energy generation in the NEB, aiming to expand knowledge and facilitate study applications in the area. For this, ERA5 reanalysis data was used from 2009 to 2019, including information such as SSRD, GHI, T2M, TP and TCC, in addition to data from INMET of Rg. Power generation data, including GEV, CGE and CI, were obtained from ONS. When analyzing meteorological data, correlations were observed between T2M, Rg and other meteorological variables. With the efficiency data in solar energy generation, it was identified that the plants with the highest IC potential are concentrated in Piauí and Bahia, regions where the GHI has high hourly average values, reaching approximately 200 Wh/m². These regions also present higher T2M and SSRD values, and lower TCC and TP values. The efficiency of the forecast models was examined in more detail for the Bom Jesus, Horizonte, Nova Olinda, Rio Alto and Sol do Futuro Complexes, located in Bahia, Piauí, Paraíba and Ceará, highlighting the GRU method and combinations that involve it, such as superior in terms of RMSE and MAE metrics, maintaining consistency for different step sizes. Validation, through comparison of predicted with observed data, reinforced that the GRU, or combinations that involve it, present more effective results in the prediction task for the data sets studied.