The obvious approach is to just train your numberplate recognition algorithm with blurred plates. Since the blur is almost equal across the whole plate, and nearly all cars are moving in the same direction, you aren't really losing much information. Sure, it might be hard for a human to read, but for a deep learning algorithm I don't think it's actually any harder.

But there are other approaches too - like putting a 99 cent novelty zoom lens on the front of your camera to capture more light for your region of interest, allowing you to use shorter exposure times. Or an infrared strobe light that flashes once per frame (most numberplates are retroreflective, so IR strobes work really well).