Akademik Veri Yönetim Sistemi
Theoretical and Applied Climatology, cilt.157, sa.3, 2026 (SCI-Expanded, Scopus)
Clean and renewable solar energy holds great potential for supporting the United Nations' 7th Sustainable Development Goal. However, accurately forecasting global horizontal irradiance (GHI) remains a challenge due to strong nonstationarity, nonlinearity, and multivariate dependencies in climate data. This study develops and evaluates several novel hybrid forecasting models that combine multistage feature extraction with advanced decomposition and sequence learning. The principal proposed model integrating Correlation analysis—Grey Wolf optimization—Improved complete ensemble empirical mode decomposition with additive noise—Sample entropy -Spatio-temporal attention—Sequence2sequence (CA-GWO-ICEEMDAN-SE-STA-S2S) first filters candidate predictors via correlation and optimization, decomposes the selected inputs with ICEEMDAN, reconstructs informative modes using sample entropy, and finally forecasts with an attention-enhanced seq2seq network that captures spatial and temporal dependencies. We compare the proposed framework against 15 alternative methods (SVR, KRR, XGBoost, HGBR, ANN, GRU, LSTM, S2S variants, and other hybrids) using standard error and efficiency metrics. The CA-GWO-ICEEMDAN-SE-STA-S2S model achieves the best performance in our experiments (RMSE = 13.861 W/m2, MAE = 7.431 W/m2, dr = 0.976, CV = 11.165%), significantly improving over the strongest baseline (LSTM). These results demonstrate that systematic multistage feature extraction together with spatio-temporal attention yields robust and accurate short-term GHI forecasts. The proposed approach is readily adaptable to other environmental time series forecasting problems.