Previsão de potência gerada por sistema fotovoltaico na cidade de Foz do Iguaçu de acordo com dados meteorológicos

Abstract

Photovoltaic power is getting popular lately as a reneweable alternative to fossil fuel consumption. However, its power generation is linked to meteorological factors. The use of the regression algorithms Support Vector Machines, Decision Trees, Nearest Neighbors, and an Artificial Neural Network are proposed to predict the amount of power generated by a solar power plant located in Foz do Iguaçu city, based on meteorological data. The data was treated, and a feature selection was performed after an exploratory data analysis, evaluating the correlation coefficient and scatter plots. In this context, the selected variables were: solar radiation, temperature, humidity, dew point, rain, hour and month. After that, the regression models were trained and validated using cross validation. 10% of the dataset was designated to be used as a test set, using RMSE, MAPE and R2 as evaluation metrics. According to these metrics, the best algorithms are SVR and ANN. A Dunnett’s Test with a Confidence Interval of 95% was performed to compare the means obtained by each one of the algorithms and the real data, pointing out that there is not a significant difference between the algorithms and the real data, suggesting that all of them are fitted for the prediction.