Dimensionality reduction in solar radiation forecasting: a combined PCA-SVM framework for renewable energy applications

Introduction Solar radiation forecasting (SRF) faces significant challenges due to high-dimensional meteorological data that can affect model generalization and computational efficiency. This study presents a comprehensive framework integrating Principal Component Analysis (PCA) with Support Vector Regression (SVR) to address these challenges and improve prediction accuracy. Methods Meteorological data were collected throughout 2023 at VIT University’s campus in Vellore, Tamil Nadu, incorporating multiple parameters including ambient temperature, dew point, wind characteristics, and atmospheric conditions. PCA was employed to transform correlated variables into uncorrelated principal components through data normalization, covariance analysis, component extraction, and selective feature retention based on cumulative explained variance thresholds. The dimensionally reduced dataset was then fed into various SVR models with different kernel functions (linear, polynomial, tanh, and Gaussian). Model validation was rigorously performed using k-fold cross-validation to identify the optimal configuration for solar radiation prediction. Results Comparative analysis revealed that the SVR_Gaussian implementation demonstrated superior performance with RMSE of 9.111125, MAE of 3.765607, MAPE of 0.013853, and R 2 of 98.37%, outperforming all alternative models tested. Discussion The hybrid PCA-SVR approach effectively handles the inherent complexity of solar radiation patterns while maintaining computational efficiency, demonstrating that dimensionality reduction combined with advanced machine learning techniques can significantly enhance solar radiation forecasting accuracy.