A novel ultra-short-term power forecast (UST-PF) approach for solar photovoltaic (PV) power plant is presented. Micro-scale weather condition of the closest time horizon to the next forecast is identified through support vector machine (SVM) classification model whose input vector is composed of four characteristic parameters defined using irradiance data after the previous forecast of the last time. According to the classification results, the appropriate model for the next forecast is determined among multiple forecast models constructed corresponding to different micro-scale weather conditions. Simulation results verify the effectiveness of the proposed approach