Due to the intermittency and fluctuation of solar energy, its exponential growth presents serious challenges to the power system. Therefore, photovoltaic (PV) power forecasting, including solar irradiance forecasting, has become a necessary prerequisite for the grid connection of photovoltaic power stations. However, traditional 2D convolution networks are less effective in extracting spatial features, especially limited in handling long-term dependencies. To address these problems, in this paper, a novel ultra-short-term solar irradiance forecasting method based on a 3D Convolutional Long Short-Term Memory and 3D Convolutional Neural Networks (3D ConvLSTM-CNN) hybrid model is proposed by processing multiple consecutive all-sky images with various color channels, the spatial information of different color channel images can better extract different types of cloud information, and the 3D ConvLSTM-CNN can take into account the temporal information. The temporal and spatial features of the sky image are extracted from multiple images at different times, simultaneously, and the textual meteorological features of the corresponding images fused via the LSTM hybrid network input model to finally establish the model for forecasting the next moment. All-sky image data and irradiance data collected by Yunnan Normal University are used to test and verify the model. The experimental results indicate that the proposed method has a promising performance and achieves 28.2%, 34.8%, 19.9%, 42.7%, and 68.3% improvement on nRMSE, MAPE, SMAPE, MedAPE, and R2 over the persistence model for 5-min ahead global horizontal irradiance (GHI) prediction.
