On zero-shot learning in neural state estimation of power distribution systems