Issues #10 and #13 suggest that it's a good idea to give the model the NWP cloud data from all the grid boxes in the line of sight between the PV system and the Sun. Maybe we can minimise the amount of NWP data per example by selecting just a rectangle of PV data, where the long axis of the rectangle is parallel to the angle between the PV system and the Sun (so the rectangle is pointing towards the Sun). (Why a rectangle? Because it's fairly easy to compute. See issue #17 for an implementation idea of how to select the rectangle).
UPDATE: Issue #19 suggests that, actually, NWPs sometimes get clouds really quite wrong. So let's use satellite imagery from the get-go. Maybe use a rectangle of "cheating" satellite imagery, adjusting for parallax (issue #13). Maybe each element of satellite imagery could be something like a 2x2 patch of HRV imagery, along with a pixel of IR imagery, so the model can roughly infer the height of the top of the clouds. For HRV, parallax makes a 10 km cloud appear 8 pixels (16 km) too far north. So the rectangle must always include a buffer of at least 8 km of data to the north.).
My current, broad, plan of work for the next month or so:
Very rough notes!
Issues #10 and #13 suggest that it's a good idea to give the model the NWP cloud data from all the grid boxes in the line of sight between the PV system and the Sun. Maybe we can minimise the amount of NWP data per example by selecting just a rectangle of PV data, where the long axis of the rectangle is parallel to the angle between the PV system and the Sun (so the rectangle is pointing towards the Sun). (Why a rectangle? Because it's fairly easy to compute. See issue #17 for an implementation idea of how to select the rectangle).
UPDATE: Issue #19 suggests that, actually, NWPs sometimes get clouds really quite wrong. So let's use satellite imagery from the get-go. Maybe use a rectangle of "cheating" satellite imagery, adjusting for parallax (issue #13). Maybe each element of satellite imagery could be something like a 2x2 patch of HRV imagery, along with a pixel of IR imagery, so the model can roughly infer the height of the top of the clouds. For HRV, parallax makes a 10 km cloud appear 8 pixels (16 km) too far north. So the rectangle must always include a buffer of at least 8 km of data to the north.).
My current, broad, plan of work for the next month or so:
Related issues:
17
16
13
10
7