<p>Solar energy is a sustainable and environmentally friendly energy source; however, its intermittent and highly dynamic nature makes accurate Global Horizontal Irradiance (GHI) forecasting a challenging task. Reliable GHI prediction is crucial for efficient photovoltaic power generation, grid stability, and renewable energy integration. To address this challenge, this study proposes a hybrid EEMD-CSA-VTX-GRU forecasting framework that combines Ensemble Empirical Mode Decomposition (EEMD), Cuckoo Search Algorithm (CSA)-based hyperparameter optimization, and a Transformer-enhanced Gated Recurrent Unit (VTX-GRU) model for solar irradiance prediction. The EEMD technique is employed to decompose nonlinear and non-stationary irradiance signals into intrinsic mode functions, improving feature representation and reducing noise effects. The Transformer module captures long-range temporal dependencies using an attention mechanism, while the GRU network effectively models short-term sequential patterns. Furthermore, CSA is utilized to optimize model hyperparameters, enhancing convergence stability and forecasting accuracy. The proposed framework was evaluated using the Jodhpur solar irradiance dataset under different forecasting scenarios. Its performance was compared with conventional GRU and other benchmark models using RMSE, MAE, nRMSE, MAPE, and R² metrics. Experimental results demonstrated superior forecasting performance, achieving an MAE of 31.48, RMSE of 46.64, nRMSE of 0.0442, and MAPE of 10.56%. Sensitivity and statistical analyses further confirmed the robustness and stability of the proposed model across varying training configurations and forecasting horizons. Although the framework was validated on a single geographical region, the results indicate its effectiveness for short-term solar irradiance forecasting. Future work will focus on multi-regional validation, lightweight optimization strategies, and improved forecasting robustness under highly dynamic atmospheric conditions.</p>