Near-real-time estimation of global horizontal irradiance from Himawari-8 satellite data

Accurate estimation of global horizontal irradiance (GHI) is not only an essential requirement of setting up a photoelectric power generation system but also critical information for terrestrial ecological models. Current methods to estimate GHI are mostly focused on hourly or daily scales and very often also require additional meteorological data. In comparison, few works have ever estimated GHI on the near-real-time scale, as the dynamically changing clouds pose great challenges for instantaneously estimating it. In this study, we adopt the Himawari-8 satellite data as the sole input without any supplementary meteorological parameters, to estimate the near-real-time GHI based on four machine learning algorithms and their ensemble. All models achieved similarly good performance, with R2 being about 0.81, and nRMSE being within the range of 25.22%–26.34%. Ground validations revealed that our result outperform the official Himawari-8 shortwave radiance product. Further analyses revealed that different machine learning models behave differently under different weather conditions, while all performed badly under overcast conditions, suggesting an in-depth investigation is required to improve the model performance. Even so, we foresee that this efficient way, which relies solely on the Himawari-8 geostationary satellite data, can be applied widely to estimate near-real-time GHI in the future.