Solar irradiance variability, left unmitigated, will threat the stability of grid system, and might incur significant economical impacts. This paper focuses on a pipeline to predict solar irradiance from 30 minutes to 5 hours using geostationary satellite. It consists of two parts: 1) cloud motion estimation and 2) solar irradiance prediction using the estimated satellite images. The main challenge is image noise at all levels of processing from motion estimation to irradiance prediction. To overcome this problem, we propose to use optical flow motion estimation, and subsequently combine multiple evidences together using robust support vector regression (SVR). Our systematic evaluation shows significant improvements over the baseline in both motion estimation and irradiance prediction.