I'm trying to do MD on water vapor. As I know there exists some water models for liquid water, such as SPC,SPC/E,TIP3P, but will they also apply to vapor state of water? And what's the difference of simulations of liquid and vapor water?
It is certainly true that most of the standard water models are "tuned" according to properties of the liquid phase—they are intended, after all, to be used as models for bulk water, rather than modeling interfacial properties.
That said, you can use these liquid-phase models to simulate vapor-liquid equilibrium, so long as you're willing to live with the substantial inadequacies that are associated with them. A few years back, I did a study of surface tension of common water models (arXiv link), and found that none of them gave surface tensions higher than 62 mN/m (the experimental value is closer to 72 mN/m). Even the best water model currently out there, the TIP4P/2005 model, only gives values of approximately 65 mN/m (according to Gromacs, NAMD, and LAMMPS simulations).
The principal thing to remember about the vapor phases is that at low temperatures, there are very few water molecules in the vapor phase, since the density is so low. That means you'll need to have either a large system or a long sampling time to get enough statistics for accurate calculations.
Generally, you should choose the forcefield that most accurately reproduces the quantity of greatest interest to you (dielectric constant, dipole moment, surface tension, etc.). This is an excellent guide to all of the different water models and their respective strengths and weaknesses.
There are several dozen different water models out there, and they all suck in their own special way, so eventually you just have to close your eyes and pick one. My experience is that most non-experts just end up using the defacto standard, i.e. TIP3P/TIP4P, which is probably just fine for most applications.
As for the difference in how accurate forcefields are in gas vs liquid phases, one of the weaknesses of all MM forcefields is that they are very context sensitive. A forcefield which is accurate for gas phase will always be less accurate for liquid phase, and vice versa. Some forcefield features, such as polarization, allow for some degree of dynamic response to changes in molecular context, but even so there will never be an MM forcefield that has anywhere near chemical accuracy across two phases. I think that most simulators learn to live with a certain level of these kinds of inaccuracies.
If you do want a very accurate multi-phase simulation, you'll really have to start looking into the quantum methods.