Enhancing Global Solar Irradiance Forecasts in Brazil Using Numerical Models and Machine Learning
Abstract
This study aims to improve forecasting of global solar irradiance (GHI) in Brazil by integrating ECMWF numerical weather prediction (NWP) data with machine learning techniques for bias correction. Observational data from 17 meteorological stations in the SONDA network were used for training and validation, covering 2021 and 2022. Two models were implemented: Multiple Linear Regression (MLR) and a Multilayer Perceptron (MLP) neural network. The MLP model was optimized through hyperparameter tuning and feature selection, incorporating 28 predictor variables related to radiative, thermodynamic, and cloud parameters. The results indicate that the MLP model generally provided slightly improved performance compared with both the raw ECMWF forecasts and the MLR model in terms of error metrics and correlation. In May and October 2022, the MLP achieved the lowest NRMSEs (26% and 28%, respectively) and the highest correlation coefficients (R ≈ 0.93). These results suggest that neural networks can capture nonlinear atmospheric relationships and contribute to improving the representation of solar irradiance variability. The findings also align with benchmark studies in the literature, reinforcing the robustness of the proposed approach. The methodologies developed here can contribute to more accurate solar forecasting, supporting operational energy management and the integration of renewable sources in Brazil’s electricity matrix.
Keywords
Solar forecasting, machine learning, ECMWF, bias correction, irradiance modeling, Brazil