Previsão da irradiação solar na cidade de Curitiba a partir de redes neurais artificiais

Abstract

The main goal in this work is make the forecast about solar irradiation in the city of Curitiba. Along this work, will be introduced forecast solar irradiation models using artificial neural networks trained by the dataset of Instituto Nacional de Meteorologia (INMET). This idea is an alternative for the existent models, which, the more accurate the results are desired, bigger the number os data necessary for the model, resulting an increase in processing demand. The training of the neural networks in this work happen using different groups of climatic characteristics subdivided by seasons climatic (summer, fall, winter and spring) to verify which one present the best forecast. The architecture used in this neural networks is named feedforward multilayer with two hidden layers. As internal parameters to the models, Sigmoidal activation functions are applied to the hidden layers and Rectified Linear functions in the output layer. The weight optimization is done by the Stochastic gradient descent algotithm. As a result of these trainings using different features, the dataset 4 perform better in the forecast of irradiation in the summer and the dataset 5 had lower erros in fall, winter and spring. Therefore, the split made between the seasons show more efficiency resulting in minor errors for each different setting. The best result, was a error of 124,93 𝑊ℎ/𝑚2 in the winter for the dataset 5, by MAE and the biggest global error was 246,71 𝑊ℎ/𝑚2, in the spring for the setting dataset 2, by RMSE, showing very different results depending of the dataset used for training.