It'd be great to automatically test if the "full data pipeline" correctly spatially aligns GSPs with NWPs.
(Just checking the the coordinates have the same values is necessary but not sufficient. Coordinates might be incorrectly assigned to GSPs. As happened in issue #187!)
Possible Implementation
Several possibilities:
Compute the correlation between the GSP PV and the NWP DSWRF (take the mean DSWRF across the GSP's area?). If the correlation is below some threshold for any GSP then the test fails. Or maybe hard-code different thresholds for each GSP?
Similar to above, but train a very simple ML model to predict GSP PV from NWP DSWRF. If the MAE is above some threshold for any GSP then the test fails.
Any other possibilities?
One complication with the above is that the CI/CD infrastructure will need relatively fast access to the NWPs and GSP data. Maybe we could include just, say, 1 month of just DSWRF (no other NWP variables) with the tests, and 1 month of GSP? Both as uint8?
It'd be great to automatically test if the "full data pipeline" correctly spatially aligns GSPs with NWPs.
(Just checking the the coordinates have the same values is necessary but not sufficient. Coordinates might be incorrectly assigned to GSPs. As happened in issue #187!)
Possible Implementation
Several possibilities:
Any other possibilities?
One complication with the above is that the CI/CD infrastructure will need relatively fast access to the NWPs and GSP data. Maybe we could include just, say, 1 month of just DSWRF (no other NWP variables) with the tests, and 1 month of GSP? Both as uint8?