Abstract: The main reason that the ultra-short-term output power of photovoltaic cells changes comes from the random movement of clouds, which will significantly affect the output power within a time scale of 1-2 minutes. Therefore, a sky cloud chart prediction is proposed to improve the accuracy of the ultra-short-term power prediction of photovoltaics. Firstly, the cloud shape gray discrimination is used to extract the cloud shape, the cloud transmittance and other information from the cloud images. Then, the cloud motion is restored through the cloud point tracking to obtain information such as the cloud layer displacement and its velocity. Next, the CFA-LSTM model is proposed. By adding the cloud feature association to the LSTM model, the ultra-short-term photovoltaic output power is associated with the sky cloud images. Finally, based on the cloud enhancement, a CEM-LSTM switching model with clear sky irradiance as the standard is proposed. Experiments show that the CEM-LSTM switching model can not only meet the accuracy requirements of the photovoltaic ultra-short-term power prediction with high accuracy under all weather conditions. It can also meet the reliability requirements of high precision and high stability of PV ultra-short-term power prediction, which provides the possibility of efficient and economical operation of photovoltaic power plants.