Automatically detect "bad" PV data

[JackKelly]

Some automatic checks for each PV system's PV power timeseries:

• Does it produce PV power at night?
• Does PV power roughly correlate with NWP irradiance (or, maybe even better: does PV power correlate with PVLib's estimate of PV power, produced from NWP irradiance & temperature)?
  • Maybe compute correlation on a week-by-week basis. Because sometimes a PV system might produce dud data for a short period, and we should throw away that dud period, not the entire PV system's data.
• Are there any crazy high peaks? If so, does the data look sensible again if we throw away, say, the top 0.1% of values?

[bndxn]

Another potential idea: what about comparing yield with other systems nearby? I guess they might have different angles/positions etc. This is what I did when looking at PV stations in Devon.

[JackKelly]

Another potential idea: what about comparing yield with other systems nearby

That is an interesting idea. I guess one challenge is: what happens in the worst case when all the "near neighbours" are bad? Then the PV system currently under consideration might be thrown away because it doesn't look like its bad neighbours :slightly_smiling_face:

To start with, we probably just want to throw away the "really obviously bad" data :slightly_smiling_face: Like PV systems which report zero power for an entire week. Or PV systems which always report 1 kW, even at night :slightly_smiling_face: Stuff like that :slightly_smiling_face:

[bndxn]

Thanks! Yeah that makes sense. I would guess that if you compare system k with n nearby systems, then as n increases, the chance that the majority of the n nearby systems are bad decreases. On the other hand if more systems are required for comparison they'd be more geographically dispersed so maybe the correlation between their yields would decrease.