LSTM-Transformer with decomposition–reconstruction for enhanced solar irradiance forecasting incorporating meteorological variables

Solar power is growing rapidly worldwide, but its intermittency and volatility threaten grid stability, while the accurate irradiance forecasts are vital to reduce dispatching costs and maintain system security. This paper proposes a VMDFE-LSTM-Transformer forecasting framework optimized by Bayesian optimization (BO) for high-precision solar irradiance prediction. The framework employs variational mode decomposition (VMD) and fuzzy entropy (FE) to decompose and reconstruct solar irradiance data, effectively extracting multi-frequency components from the data. The resulting three Cooperative Intrinsic Mode Functions (Co-IMFs) are combined with meteorological variables and fed into an LSTM-Transformer predictor optimized through BO. To validate the model under different climatic conditions, systematic experiments are conducted on seasonal datasets from China’s four major climate zones. Performance comparisons with multiple forecasting models, including advanced architectures such as Crossformer, confirm the superior single-step and multi-step forecasting capability of the proposed model. Additionally, the differentiated contribution of meteorological features to forecasting accuracy under various climatic conditions is investigated. To enable unified performance assessment across multiple datasets, the normalized MAE (nMAE) metric is introduced. Within the decomposition framework, adding meteorological variables reduces nMAE by 7.91–25.09%, with greater gains observed when sunny conditions are scarce and data volatility is high. These results can provide practical guidance for designing region-specific irradiance forecasting systems that balance accuracy and data acquisition costs.