Cloud cover can cause dramatic fluctuations in photovoltaic (PV) power, which poses a great challenge to the sTable operation of the power system. In order to address the problems of large fluctuation of PV power with cloud cover and low accuracy under highly complex weather conditions, we propose an ultra-short-term PV power prediction model based on two-stream feature enhancement network (TSFE-LSTM). Firstly, the groundbased cloud images are processed by fish-eye image correction algorithm and optical flow estimation algorithm, and the spatiotemporal feature of the ground-based cloud image is extracted by the two-stream convolutional network. Secondly, an ultra-short-term PV power prediction model based on two-stream feature enhancement network is built to predict the PV power in the next 15 min. The results of the experiment show that the MAE and RMSE of the TSFE-LSTM model for PV power prediction in the next 15 min are 6.49% and 10.76% with the input of ground-based cloud image and optical flow, respectively, which are improved by 5.81% and 4.61%, respectively, compared with the single input data. In sunny conditions, the MAE and RMSE of TSFE-LSTM model improved 7.23% and 14.99% than the CNN model, respectively. In cloudy conditions, the accuracy of the model is slightly de-creased, but the MAE and RMSE are improved by 3.53% and 1.26%, respectively, compared with the CNN-MLP model. This provides new ideas for PV power prediction with cloud cover.