Due to a variability and uncertainty of photovoltaic (PV) output power, PV operators may be subject to significant penalties on forthcoming energy markets. Thus, the accurate prediction of PV output power plays a very important role in energy market. This paper proposes a novel solar prediction scheme for one-hour ahead prediction of solar irradiance based on various meteorological factors including the cloud cover and support vector machine (SVM). A k-means clustering algorithm is applied to collect meteorological data and the entire data are classified into three clusters based on similar daily weather types. The same cluster data are used for the SVM regression in the training stage. We also investigate the prediction error analysis. It is shown that the solar irradiance prediction errors of each prediction scheme can be categorized to be leptokurtic and a t location-scale distribution is proposed as a distribution fitting for the prediction errors. In addition, the power and energy capacities of an energy storage system (ESS), which can absorb the prediction errors, are estimated from the probability density functions. Numerical results show that the proposed SVM regression scheme significantly improves the prediction accuracy and reduces the ESS installation capacity.