Stick a few calibrated thermocouples on the PV cell surface and there you go.
I did it that way and it works perfectly. I have a dataset of one year (2012 - 2013) of minutely (real) power output, voltage and current as well as Cell temperature data from a photovoltaic panel setup.
Besides I have minutely incoming global radiation, UVA and UVB from a nearby pyranometer. That means that the dataset contains 1,500,000 data receords for a one year period. I fitted a Fourier series through all the datsets to obtain a factor of 100 in data compression. Also works fine