Enhancing short-term solar radiation forecasting with hybrid VMD and GraphCast-based machine learning models

Abstract

Accurate solar radiation forecasting is vital for photovoltaic power, agriculture, and weather prediction but faces complex nonlinear challenges. Recently, with the rapid development of artificial intelligence (AI), AI weather forecasting models based on machine learning (ML) have been proposed. The paper proposes a novel method by utilizing Variational Mode Decomposition (VMD) coupled with the global forecasting system Graph Neural Network model (GraphCast), combined with Support Vector Machine (SVM), Random Forest (RF), Convolutional Neural Network (CNN), and Long Short-Term Memory (LSTM). The method predicts daily solar radiation under limited meteorological data by leveraging five-fold cross-validation. Furthermore, the integration of GraphCast-based forecast meteorological data with ML technology effectively addresses the gap in GraphCast-based solar radiation forecasting. Using models with station-based meteorological variables as a benchmark, the prediction performance of hybrid ML models based on VMD was compared with that of individual models. Among all models, the VMD-based RF model provided the highest accuracy for solar radiation forecasting. Additionally, when more meteorological variables forecasted by GraphCast were used as model inputs, the VMD-based RF model showed improved prediction performance. The VMD-based RF model emerged as the best-performing predictive model in this study, demonstrating an average coefficient of determination (R²) of 0.928, mean absolute error (MAE) of 16.578 (W/m²), root mean square error (RMSE) of 24.32 (W/m²), and normalized root mean square error (NRMSE) of 5.6 (%) across different input combinations. The RMSE of the VMD-based RF model is 58.73% lower than that of the RF model based on station meteorological variables.