Transferência de aprendizado profundo para previsão da irradiação solar: uma análise para diferentes bioclimas e escassez de dados

Abstract

The intermittent nature of solar irradiation has been a major challenge in planning and operating electrical power systems. In Brazil, due to the distinct bioclimatic zones, each with its own peculiarities and climatic behaviors, the sce- nario becomes complex. For this reason, a methodology based on transfer learning combined with state-of-the-art solar irradiation forecasting can bring advantages, especially when there is a low amount of data from certain climatological stations. In this sense, transfer learning can be used both between data from stations be- longing to the same bioclimatic zone (aiming to solve the problem of missing data) and between data from stations located in different bioclimatic zones (seeking to solve the problem of climatic disparity). This work proposes, based on a baseline constructed with the seasonal autoregressive integrated moving average model, to evaluate the performance of LSTM (long-short term memory) neural networks and their hybridization with convolutional neural networks (CNNs) in order to perform hourly solar irradiance forecasting. The results demonstrate that, in relation to the test of models constructed solely by LSTMs, compared to the CNN-LSTM hybrid models, the former showed slightly better performance, by approximately 1% on av- erage. Regarding the learning transfer tests, the feature extraction and fine-tuning approaches performed reasonably well when compared to zero-shot adaptation, be- ing comparable to the results of the LSTM models for each climatological station, especially in the configuration of stations by distance. Regarding the configuration of stations with a low amount of data, the approach performed well, mainly within the MAE metric.