Forecasting Solar Irradiance in Southern Bangladesh using Transformer-Based Approaches

Paper in proceeding, 2025

The rapid increase in power consumption in Bangladesh, driven by population growth and urbanization, has intensified the need for alternative energy sources. Conventional options like coal and gas are depleting, making renewable sources such as solar vital for long-term sustainability. The traditional method focuses more on the non-renewable energy sources. However, forecasting energy output remains challenging due to the inherent variability in solar irradiance, influenced by complex atmospheric dynamics. Unlike prior studies that treat solar forecasting in isolation, this paper simultaneously proposes machine learning (ML), deep learning (DL), transformer learning, and hybrid learning simultaneously to forecast solar irradiance for the underexplored southern coastal region of Bangladesh. Using 15 years of NASA-derived meteorological data, we evaluated five ML models (Linear regression, Random Forest, Gradient Boosting, Decision Tree, and XGBoost), three DL models (MLP, ANN, and LSTM), three Transformer-based models (Autoformer, Transformer, and Informer), and hybrid models (LSTM + GRU, GRU + Autoformer, and LSTM + Autoformer) to identify the robust model. Informer obtained an R² value of 99.3%, which outperformed all other models in terms of R² value. In addition, the error was also the minimum for the Informer model, with MSE, RMSE, MAE, and MAPE values of 0.004, 0.060, 0.036, and 0.025, respectively. Early prediction of solar irradiance can mitigate the pressures on non-renewable energy sources.