It depends on how you're actually interacting with your data in practice.
If you have no overnight holding positions and forcefully liquidate at the end of bar 10 minutes before the market close, then you truncate the last three returns since you can't actually act on them.
If you have overnight holding positions, and the market has continuous trading in non-US business hours, then this is merely a challenge in data acquisition because those samples actually exist. You may want to address the seasonal regime shift in the data in this case.
If you have overnight holding positions, and the market halts trading for some nontrivial amount of time $\gg$ 5 min between two sessions, but you aren't actually trying to predict the jumps, you can just leave out those data points.
If you have overnight holding positions, and the market halts trading as above, but you are actually trying to predict this gap, it is reasonable to model this separately.