Weather Phenomena Monitoring: Optimizing Solar Irradiance Forecasting With Temporal Fusion Transformer

Global climate change has spurred the search for renewable energy sources, with solar power being a cost-effective option for electricity generation. Accurate energy generation forecasting is crucial for efficient usage planning. While various techniques have been introduced, transformer-based models are effective for capturing long-range dependencies in data. This study proposes an hour-ahead solar irradiance (SI) forecasting framework based on variational mode decomposition (VMD) for handling the meteorological data and modified temporal fusion transformers (TFT) for forecasting solar irradiance. The proposed model decomposes the raw solar irradiance sequence into intrinsic mode functions using VMD and optimizes the TFT using the variable screening network and GRU-based encoder-decoder. The resulting deep learning model provides interpretable outputs such as importance ordering of solar irradiance sub-sequences, and multi-head attention analysis of different forecasting window sizes. Empirical study shows that transformers have high performance on the long-term dependencies as compared to other time series models such as ANN and LSTM. On the e National Solar Radiation USA SI dataset, the proposed TFT achieved an MAE of 19.29 and an R2 of 0.992 for a one-hour ahead forecast, while on the Pakistan SI dataset, it achieved an 11.46% reduction in MAE and a 7.12% improvement in MSE compared to the original TFT.