Comparing the power shape - measured vs modeled

[riccardo adinolfi]

Hi everyone!! 

I'm trying to compare the output power simulated using PVlib with the output power measured from the field. Specifically, I'm interested in understanding which scenario, among many simulated scenarios, has the most similar output power to the measured one. I do this comparison considering a daily window but the data frequency that I have is 1-minute. I'm trying to use R-squared or RMSE but they both do not seem very robust. 

I would like to combine both metrics, but I'm questioning how to choose the weight for each. Have you ever had a similar problem? Are there works that I could relate to?? 

Thanks in advance to all

best regards

Riccardo AB 

[cwh...@sandia.gov]

Can you elaborate on what you mean by not "very robust"?

It may help to think of a PV model as transforming the geometric curve of the irradiance data to the curve of the measured power. A metric quantifying agreement between simulated and measured power is likely to perform similarly when comparing measured irradiance and measured power. On a clear-sky day, I would expect close agreement as measured by either R2 or RMSE. But on a day with variable irradiance conditions, at 1-minyute intervals the time offsets between shadows on the irradiance sensor and the extent of shadows on the arrays can result in large R2 or RMSE. That's not a problem with the metrics, but rather, of comparing two signals with unaligned shadow features at a 1-minute time scale.

If that's the case with your data, you can try smoothing it by time-averaging to e.g. 15 minutes.

Cliff