Transfer learning for solar irradiation prediction in Minas Gerais, Brazil.

Minas Gerais is, in economic and population terms, the third-largest state in Brazil, where significant investments have been made in producing photovoltaic energy, primarily in distributed micro and mini-generation. Despite the growing demand, no computational model in the literature efficiently predicts future values of solar irradiation, the source of photovoltaic energy, for the entire state. The vast majority of works use empirical models for specific regions of the state. A few papers that use computational models, even those dealing with broader regions of the state, create local models for each point of data observation. This work presents the methodology and development of an optimized computational model for predicting solar irradiation applied throughout Minas Gerais. The presented methodology can be applied to various machine-learning models and used in different locations. Here, we utilize data from 67 meteorological stations distributed throughout all state regions, spanning 20 years of measurements. Different approaches are studied, manipulating the data used in the training and test databases, so that we can answer two main questions: Does the addition of geolocation data improve the prediction of solar irradiation? Is it possible to efficiently predict solar irradiation values in a place without meteorological stations? Experiments have shown that using data from neighboring cities is either detrimental or irrelevant to the results. However, using data from neighboring cities for transfer learning yields good results, generally comparable to those obtained in models that utilize the city's database.