Previsão de radiação solar com métodos de machine learning em séries temporais

Abstract

This work aims to develop neural network models, using machine learning architectures, that estimate solar radiation in order to remedy the lack of data caused, for example, by breakdowns in meteorological stations, reading errors (gaps) and/or lack of collection equipment in some locations. The neural network models used to estimate radiation were, specifically, neural networks of the type Gated Recurrent Unit (GRU) and Multi-Layer Perceptron (MLP). The Sinop and Alta Floresta datasets used in this work were collected from a micrometeorological tower and a meteorological station, respectively. Temperature, relative humidity, precipitation and soil moisture were used as predictor variables for Sinop, while for Alta Floresta temperature, relative humidity, precipitation and wind speed were used. The algorithm was developed in Python, mainly using libraries such as pandas for data manipulation, scikit-learn for data normalization, Keras and TensorFlow for construction neural networks, scipy for statistical analysis of data and evaluation of models, among others. To validate the algorithm, metrics such as Mean Absolute Error (MAE), Mean Square Error (RMSE), Coefficient of Determination (R2), Will mott Index and Spearman and Pearson Correlations were used. These metrics were applied to different time intervals, such as hourly average, daily average and monthly average. The study demonstrated that both neural networks presented satisfactory results. For the Sinop group, the coefficient of determination (R2) had a minimum of 88% for monthly averages, reaching 95% model adjustment for the GRU and MLP network for hourly averages. The Pearson correlation presented a minimum of 94% adjustment for monthly averages and reached close to 98% for hourly averages. For the Alta Floresta data set, both networks obtained good results for hourly intervals, reaching a coefficient of determination of 92% for MLP and 93% for GRU, however, due to the reading errors presented, the network did not obtain good results for daily averages. and monthly, thus identifying that the quality of the data substantially impacts the quality and final adjustment of the model.