If you have a rating curve, from a gage for instance, then it would generally be preferable to Normal Depth.
Normal Depth is what you use when you don’t have any other information.
You can check the sensitivity of the downstream boundary condition.
Use your best estimate of Normal Depth and run the model.
Increase the slope (maybe double it), run the model.
Decrease the slope (maybe by half), run the model.
Then look to see how much the results changed. Are the results different in the particular areas you are interested in? If they are, you may need to extend your model further downstream.