Previsão intervalar de geração fotovoltaica por regressão quantílica em regimes de irradiância

Abstract

The growing share of solar photovoltaic energy in the Brazilian electricity mix intensifies the need for reliable forecasts that quantify the uncertainty inherent to this intermittent source. This work investigates hourly photovoltaic power prediction intervals in a humid tropical climate, focusing on the Tucuruí region in the state of Pará. We propose a methodological framework that com- bines a deterministic hybrid model, based on Multilayer Perceptron (MLP) and XGBoost with automatic selection by irradiance threshold, and a probabilistic extension grounded in quantile regression stratified by irradiance regimes, implemented via LightGBM. Hourly meteorological data, obtained from the Open-Meteo API for the period 2018–2023, underwent preprocessing that included multicollinearity removal via the Variance Inflation Factor (VIF), resulting in six predictor variables. Prediction intervals are evaluated using coverage (PICP) and normalized width (PINAW) metrics, and subsequently calibrated by a scaling factor λ to approximate the empirical coverage to the nominal 80% level. Results demonstrate that the regime-based ap- proach improves interval calibration, especially under intermediate irradiance conditions, where variability is more pronounced. This study contributes to advancing probabilistic photovoltaic forecasting in tropical regions, where the literature remains scarce.