Enhancing multi-step short-term solar radiation forecasting based on optimized generalized regularized extreme learning machine and multi-scale Gaussian data augmentation technique

Accurate solar radiation forecasting is crucial for renewable energy, agriculture and building design. For this reason, a hybrid short-term solar radiation forecasting model based on optimized Generalized Regularized Extreme Learning Machine (GRELM) and multi-scale Gaussian Data Augmentation (GDA) technique is proposed in this paper. First, original data is decomposed into multi-scale high, medium, and low-frequency data by Variational Mode Decomposition (VMD) in data preprocessing. GDA is then applied to the training set of the multi-scale data, generating a new sample training set which helps to maintain the data distribution characteristics and increase the data diversity. Next, Partial Autocorrelation Function (PACF) is employed for feature extraction, reducing redundant information and computational complexity. Finally, Improved Transit Search (ITS) algorithm is used to optimize the hyperparameters of GRELM, enhancing the model's performance and generalization ability in both training and prediction stages. Four solar radiation datasets in the U.S. laboratory are selected for validation. Experimental results of adding GDA to the data at different scales show that adding GDA to low-frequency data resulted in excellent performance of the proposed model in both single- and multi- step predictions, with an increase in R2 ranging from 9.8 % to 20.8 %. The proposed prediction model for solar radiation forecasting demonstrates higher accuracy and generalization than traditional models, offering more possibilities and flexibility for future research.