Efeito do viés geográfico em uma rede neural profunda através da transferência de aprendizado entre as estações de Hanford (EUA) e Petrolina (Brasil)

Abstract

Due to the worsening of global warming, the development of new strategies aimed to damper the rise of Earth’s temperature to a maximum of 1,5◦C by the end of 21st century. To comply with this goal, the development of technologies focused on renewable and less polutant energies is crucial. However, the stochastic and sporadic behaviour makes this type of energy hard to merge together with the already implemented energy mix, making a direct impact on its implementation costs. Therefore, it is imperative to previously know the solar irradiance at a place of interest, so better management strategies for this type of energy could be employed. In this way, the present study seeks to forecast the solar energy analysing the geographic bias effect in the learning transfer by the assessment of the capacity of a convolutional neural network to estimate solar global horizontal and solar direct normal irradiances using geostationary satellite imagery for the location of Hanford, California. The hybrid CNN-LSTM model, when trained and tested with data from Hanford, achieved values for RMSE of 88,17 W/m2 e 178,40 W/m2 respectively. These results are in accordance with those obtained by the reference models Solcast and PSM. However, when the CNN-LSTM model is trained and test with data from different locations, the geographic bias showed to be a hindrance, preventing the convolutional network to achieve satisfactory results, when compared with those from the reference models.