Multi‐Horizon Solar Irradiance Forecasting Using a Hybrid EMDELMTransformer Framework Across Diverse Climatic Regions

Solar irradiance is intermittent and nonstationary, creating a need for accurate multi‐horizon forecasts to support reliable grid operation. This study proposes a hybrid empirical mode decomposition–extreme learning machine–transformer (EMD–ELM–TF) framework. EMD separates the global horizontal irradiance (GHI) series into low‐frequency trend and high‐frequency residual components; ELM forecasts the trend component, while a transformer uses the residual history and meteorological covariates to model longer temporal dependencies. The framework was evaluated on 10‐min datasets from Karachi and Islamabad at 1, 3, and 5 h horizons and compared with seven baseline models. LightGBM achieved the lowest 1 h test RMSE, with 40.29 W/m ² for Karachi and 66.66 W/m ² for Islamabad. The proposed model achieved the lowest test RMSE at 3 h (65.95 and 93.61 W/m ² ) and 5 h (75.92 and 121.11 W/m ² ) for Karachi and Islamabad, respectively. Relative to the strongest baseline at the 5 h horizon, these values represent RMSE reductions of 23.9% for Karachi and 13.8% for Islamabad. The results indicate that separately modeling trend and residual components can improve longer‐horizon short‐term forecasting (STF) after site‐specific training. Computational cost and cross‐site transferability require further evaluation.