With the current tendency to larger photovoltaic power plants the demand for accurate forecasts of solar radiation increases steadily. In contrast to the traditional, controllable power generation that can handle the power grid variability, renewables must cope with natural power resource fluctuations. To maintain the grid stability with the increasing penetration of renewables, the variability of natural power sources has to be accounted for in the operation of the renewables by accurate minutes ahead to days ahead resource forecasts. We present a semi-empirical model to predict solar radiation up to 15 minutes ahead using ground-based measurements of solar irradiances and sky cameras. Our method forecasts projected cloud fields and solar radiative fluxes at the surface on the dimension of PV strings (a few tenths of square meters) every minute up to 15 minutes ahead. Three forecasting models are applied, the classic persistence and advection models as well as a pyramidal matcher model. Our integrated cloud classification algorithm differentiates between advective stratified clouds and convective cumulus-type clouds so that we are able to choose the most appropriate forecasting method. We demonstrate that this approach makes it possible to achieve better results than with persistence models.