The paper describes a heuristic method for the ultra-short-term computation of prediction intervals (PIs) for photovoltaic (PV) power generation. The method allows for directly forecasting the AC active power output of a PV system by simply extracting information from past time series. Two main approaches are investigated. The former relies on experimentally observed correlations between the time derivative of the PV AC active power output and the errors caused by a generic point forecast technique. The latter approach represents an improvement of the first one, where the mentioned correlations are clustered as a function of the value of the AC active power. The work is framed in the context of microgrids and inertialess power systems control, where accounting for the fastest dynamics of the solar irradiance can become extremely valuable. We validate the proposed model using one month of AC active power measurements and for sub-second time horizons: 100, 250 and 500 ms.