A Two-Stage Feature Optimization Algorithm Based on Correlation and Redundancy for Solar Radiation Prediction

Global horizontal irradiance (GHI) is a key parameter for estimating solar background noise in Geiger-mode single-photon-counting LiDAR operating on a moving airborne platform, where direct irradiance measurement increases system size, weight, and cost and may be affected by platform motion. To enable efficient GHI prediction from meteorological variables, we propose a two-stage feature optimization method based on relevance and redundancy. First, first-order linear and second-order nonlinear fitting are combined to evaluate feature relevance; then, a recursive variance inflation factor (VIF)-based procedure is used to remove redundant variables. The method is evaluated on the NSRDB and NBSC public datasets using six prediction models, including LSTM, CNN, RNN, RFR, SVR, and MLP. Averaged over the two datasets and six prediction models, the proposed method achieves MAE, RMSE, R2, RRMSE, RSR, and WI values of 0.0220, 0.0351, 0.9683, 0.2042, 0.1533, and 0.7215, corresponding to improvements of 9.06%, 9.22%, 0.21%, 6.97%, 11.85%, and 0.23%, respectively, over the corresponding second-best competing results. A field experiment is further conducted as an auxiliary validation, yielding average improvements of 5.60%, 5.42%, 5.14%, 5.80%, 1.05%, and 0.90% for the same six metrics relative to the corresponding second-best results among VIF, Pearson, ARD, NGA, WRFE, PI, and Boruta.