VMD-LSTM-transformer: A hybrid approach for robust solar irradiance forecasting in complex terrains with application to photovoltaic systems

Accurate global horizontal irradiance (GHI) forecasting in complex terrains remains a major challenge owing to strong non-stationarity, rapid cloud transitions, and highly variable atmospheric conditions, all of which directly affect photovoltaic (PV) operation and grid integration. This study proposes and evaluates VMD–LSTM–Trans , a decomposition-enhanced hybrid deep learning (DL) framework that integrates variational mode decomposition (VMD), long short-term memory (LSTM) networks, and Transformer-based attention mechanisms for solar irradiance forecasting under highly variable mountainous conditions. VMD is employed to decompose the irradiance signal into intrinsic multiscale components, thereby enhancing signal representation under highly variable conditions, while the hybrid recurrent–attention architecture captures both local temporal dynamics and long-range dependencies. The proposed framework was validated using real high-resolution measurements collected from a PV system installed at the National University of Loja, Ecuador, a mountainous tropical region characterised by pronounced climatic variability. A comprehensive preprocessing pipeline, including quality control, temporal alignment, signal conditioning, and feature engineering, was implemented to ensure robust model development. The hybrid model family was evaluated across temporal resolutions of 10 min, 30 min, and 1 h, and forecasting horizons of 1 h, 6 h, and 12 h. The results show that the proposed hybrid configurations provide competitive performance, particularly for short-term and selected medium-term forecasting scenarios. The best 1 h forecasting case achieved an R 2 value close to 0.93, while longer horizons, especially 12 h, remained more challenging due to the increasing uncertainty of solar irradiance dynamics. These findings suggest that decomposition-enhanced recurrent–attention models can be useful for irradiance forecasting in complex mountainous environments, while further validation across additional sites and climates is required.