Influência do El Niño Oscilação Sul (ENOS) na previsão da irradiação global horizontal de Fortaleza

Abstract

The introduction of solar energy in the Brazilian energy matrix is increasing annually, including in Ceará. In this context, issues related to the complementarity of supply are raised, since the solar source is intermittent. Irradiation forecasts can help decision making by the electric network driver, avoid power outages and reduce the variation of this matrix. Proper planning can be facilitated by predicting irradiation using machine learning methods. The present work analyzed the performance of 4 models of global horizontal irradiation prediction - neural networks, Boosting, Bagging and persistence model - for a city of Fortaleza, Ceará, in eight different time horizons, analyzing the influence of El Niño and La Niña, in the form of the Oceanic Niño Index, ONI, predictor, in these circumstances. In addition to ONI, meteorological information (ambient temperature, relative temperature, air speed, wind direction and precipitation level), irradiation and time data and information acquisition data ere used. The performance of the models is evaluated considering three situations: the complete database, the database subdivided between the years with the occurrence of La Niña and the years with the occurrence of El Niño (database La Niña and database El Niño), and the database subdivided between the seasons (database Winter, database Summer, database Spring and database Autumn). Cross-validation 5-fold is applied, as well as selection of parameters for neural networks, Boosting and Bagging. The calculation of the global horizontal irradiation variability allows the classification of this predictor as having low variability. The results point to a reduction in RMSE between 0.11% to 2.2% when the ONI predictor is added to the complete database. The database El Niño obtains nRMSE between 0.03% to 1.3% higher than the database La Niña. There is a reduction of up to 5.7% in the nRMSE due to the addition of ONI in the database subdivided by stations. Boosting has the smallest errors among the models considered, and Bagging is the model least sensitive to the presence of ONI, presenting zero variation of nRMSE in six time horizons due to the addition of this predictor. The use of the Winter and Spring databases for forecasts at the respectives times of the year presents fewer errors than with a complete database.