Um framework AutoML para transferência de aprendizado aplicado à predição de irradiação solar em Minas Gerais

Abstract

This work presents the development of an Automated Machine Learning (AutoML) framework integrated with Transfer Learning techniques for the prediction of solar irradiation in the state of Minas Gerais, Brazil. The primary application lies in forecasting irradiation values in regions with a scarcity of meteorological data, enabling efficient predictions even in locations lacking physical measurement stations through the transfer of knowledge from regions with abundant data. The methodology utilized a database composed of 134 meteorological stations, with 67 in Minas Gerais and 67 in neighboring states, covering more than 20 years of measurements across 129 cities. The framework automates critical steps such as feature selection, hyperparameter optimization, and model selection, highlighting the efficiency of the Extreme Learning Machine (ELM) algorithm, which outperformed models such as ANN, SVR, EN, MARS, and XGBoost in preliminary tests. Different approaches were explored, including point-based models for each station, geolocated models with the addition of neighboring data, and transfer learning, where training is performed without data from the target site. The results demonstrated that the addition of geolocated data from nearby cities to train models at points with existing stations was, for the most part, detrimental or irrelevant. However, the application of transfer Learning for prediction in locations without stations proved to be viable and effective, with performance increasing as the number of cities in the training set expanded. The use of a genetic algorithm for the automatic selection of the cities that make up the training base yielded substantially superior results compared to selection based purely on geographic distance. The work concludes that the proposed framework represents a significant advancement for photovoltaic energy system planning and agricultural management, providing accurate predictions in locations with limited monitoring infrastructure.