Since cloud motions are deeply connected with solar irradiance and photovoltaic power (PV) output of any solar grid, motion tracking and finding out the pattern are crucial for energy management. In this paper, we introduced a novel method that uses fisheye sky images taken from the solar, and extracts inter-correlated motion features and forecasts the short-term motion features. Firstly, 54,698 images are processed using lens correction, grayscale conversion, sky masking, and Contrast Limited Adaptive Histogram Equalization (CLAHE). After an extensive pattern analysis, key motion features that fluctuate with the solar irradiance are: optical flow magnitude, centroid displacement, variations in cloud density, and velocity magnitude. After we applied Variational Mode Decomposition (VMD) to normalize the pattern of features. For multiple feature forecasting, a hybrid transformer model that integrates an autoencoder is proposed and performs better than other traditional models. This model achieved high R2 scores in predicting Cloud Density Change (R2=0.913), Optical Flow Magnitude (R2=0.902), Centroid Shift (R2=0.896), and Velocity Magnitude (R2=0.884), Glow Contrast (R2=0.942. To sum up, the proposed framework provides an emerging solution for energy prediction that is significantly implementable for renewable energy production and grid stability.