Depending on your exact usage profile, adding a couple of smaller SSDs (in a mirrored configuration) for the OS and anything which changes a lot might enable you to set your NAS to spin the HDDs down for a lot of the time - they draw a lot less in that state (c1W each) until you need them. ZFS generally seems to handle the longish delays in spinning up before returning data in my experience (though annoyingly it will spin disks up one after the other instead of in parallel, extending the wait time by quite a bit - I keep meaning to write a simple shell script to detect if one drive is spun up and spin the rest up to fix this - openSeaChest is worth knowing about for stuff like this).

You’ll also need to do some sleuthing work to find out what background stuff is likely to keep using them though (don’t forget stuff like ZFS auto snapshot tools) - it took me a while to find everything when I did this (I had to keep a few instances of fatrace running and watch the timing of file accesses for a couple of weeks to find everything).

If the data isn't changing often, could you use an SSD read cache to keep the 20TB array mostly spun down?