High-temporal-resolution solar irradiance data are essential for calculating and assessing high-performance buildings. However, limited access to measurement equipment often restricts the availability of such data. To address this challenge, this study proposed a highly accurate, interpretable, and convenient method for ultra-short-term global horizontal irradiance (GHI) prediction. Firstly, a dataset containing six types of conventional meteorological parameters and corresponding irradiance values was prepared, and its feasibility for model development was investigated through correlation analysis. Then, the eXtreme Gradient Boosting (XGBoost) model, combined with Bayesian optimization (BO) algorithm, was developed to predict 1-minute GHI based on the selected meteorological parameters. Finally, the prediction mechanism was revealed by analyzing the feature importance and the effect of key features using interpretation techniques. The results show that the BO-XGBoost model outperforms the other state-of-the-art models, with coefficient of determination (R²) of 0.907 and root mean square error (RMSE) of 76.199, especially in clear sky conditions, the R² and RMSE can be 0.990 and 24.077. The model interpretation results further indicate that GHI prediction heavily relies on the solar elevation angle and relative humidity, along with their interactions. This study provides a cost-effective solution for obtaining irradiance data critical for designing and optimizing solar-based low-carbon buildings.