Abstract: Power fluctuations of photovoltaic (PV) plants present challenges to the safe and stable operation of power grids. Particularly under cloudy conditions, the position and movement of cloud clusters can cause drastic fluctuations in PV power output. To improve the accuracy of ultra-short-term power prediction, a hybrid prediction framework integrating dual-view point cloud cluster position and motion information was proposed. First, cloud cluster features— including height, velocity, direction and cloud cover— were extracted using ground-based cloud image segmentation, cloud motion estimation, and dual-view point geometric inversion technologies. Next, meteorological and historical power data were incorporated, with principal component analysis applied for dimensionality reduction and multi-source feature fusion. Finally, four sub-models—long and short term memory (LSTM), Informer, gated recurrent unit (GRU), and convolutional neural network (CNN)—were employed for parallel prediction, and a voting ensemble strategy was designed to optimize the output. Experimental results demonstrate that incorporating cloud cluster information reduces the prediction normalized root mean square error (R_NRMSE) to 0.117 9, an improvement of 4.3% over traditional methods. Heterogeneous model fusion further enhances performance, with accuracy surpassing that of any single model by 0.8%-5.6%.