A method for short term forecast of all-sky radiance distribution based on diurnal pattern sky classification

The spatial distribution and temporal fluctuation characteristics of solar irradiance are of crucial importance for power generation in photovoltaic power plants and energy assessment in building-integrated photovoltaics. The all-sky radiance distribution can reflect sky conditions and is a significant factor influencing solar irradiance. However, existing studies lack temporal continuity in sky type classification and mostly focus on point data, making it difficult to support irradiance calculations for different orientations and tilted surfaces. Based on the observational data of the MS_321LR sky scanner in Xi'an from 2021 to 2022, this paper proposes a daily pattern matrix representation method for sky radiance, taking into account both temporal fluctuation and spatial distribution information. It also puts forward a sky classification method based on the K_means approach, namely the daily pattern sky classification method. Taking the VGAE_LSTM model as an example, a model framework of “sky classification + short-term radiance prediction” is constructed. On the basis of classifying the sky into four types: clear sky, circumsolar sky, parhelic sky, and meridional sky, the observational data set is reorganized and divided into four subgroups of data sets to train the prediction models respectively. Experiments show that the prediction accuracy of the prediction model with sky classification is improved by about 1 %–2 % compared with that of the unclassified model. Among them, the VGAE_LSTM model has the highest improvement amplitude, reaching 1.9 %. This model provides a new reference for solar irradiance, photovoltaic power generation, and meteorological research.