Solar Irradiance Forecasting across Climatic Zones Using Machine Learning for Smart Energy Systems

As global demand for renewable energy grows, accurate forecasting of solar irradiance is increasingly critical for efficient energy management in smart systems. Solar irradiance (W/m2) reflects the instantaneous solar radiation received and serves as a key parameter for solar power generation. This study evaluates machine learning (ML) models for short-term solar irradiance forecasting in terms of global horizontal irradiance (GHI). Multi-year datasets from diverse climatic zones such as, Penang, Malaysia (tropical rainforest); Manaus, Brazil (tropical monsoon); and London, United Kingdom (temperate oceanic), were obtained from the Solcast API, each containing 12 variables at hourly intervals. The five-year datasets were partitioned into four years for training/validation and one year for testing. Two machine learning models, long short-term memory (LSTM) networks and extreme gradient boosting (XGBoost), were implemented for both single-step and multi-step forecasting using a sequence-to-sequence approach. Results show that XGBoost outperformed LSTM in single-step forecasting for the Penang dataset, while LSTM achieved better accuracy in Manaus and London. For multi-step forecasting, XGBoost consistently outperformed LSTM across all locations. XGBoost performed best for Penang; with MAE = 45.60 and RMSE = 88.66, and LSTM for London with MAE equal to 33.94 and RMSE equal to 70.21. A domain adaptation analysis further demonstrated the models’ ability to generalize across regions, where models trained on the PG dataset and tested on the MAN dataset performed well with minimal retraining, highlighting their robustness under varying climatic conditions. Overall, both models effectively captured solar irradiance trends, supporting reliable solar power integration into smart energy systems.