I calculated the reaction path of dehydrogenation of ethylene on
Pd surfaces with NEB method implemented in VASP.
But I always find a smaller activation energy barrier
comparing with the available calculation results(1.0 vs 1.5 eV).
I firstly run with general NEB, and converged to one
MEP. With the configurations, I also run with C-NEB.
But, the results don't improve. What should I
do to get a reasonable result?

First off, are you using our implementation of NEB (available at http://theory.cm.utexas.edu/vtsttools/) and what are the other calculations that you are comparing with? Are your calculations converged w.r.t. number of k-points, system size, spin, etc. What about pseudo-potentials and planewave energy cut-off? There are just so many variables that can affect that final results. And of course there is no guarantee that the two results should even give the same answer, depending on what kind of calculations you are comparing to. Another possibility is that you have converged to a different path.

Secondly, then I don't think it is necessarily a good strategy to run first a regular NEB and then switch to a CI-NEB. A more reasonable thing is to run CI-NEB from the start or regular NEB and then launch a min-mode (dimer or lanczos) search for the saddle point from a close by image.

Thanks, Andri.
Sure, I use your current code from that websit.
When calculated the H decomposition of ethylene, I
set up 3x3 cell for Pd(111) or Pd(100) surfaces.
And four layers are involved. 3x3x1 k-points grids are
generated by MP method. The USPP potential was chosen with
350 eV for energy cutoff. My referred results was done with
19-metal atom cluster model with Dgauss program. They also
confirmed it with slab model with DACAPO program.
I tried the dimer method, but it is too difficult to converge.

It is a bit hard to know why the dimer is converging slowly without all the input parameters. Maybe you could send us your run so we can take a look at it to see if there is something obviously wrong. Also make sure to use exactly the same setup as described for the DACAPO calculations, including using the same psp.

Well, there are couple of problems that see by quickly scanning over your INCAR file. The dimer uses a finite difference scheme where the local curvature is estimated by the force difference in two points, that very close to each other. It is paramount then to calculate the force as accurately as possible to minimize numerical error in the force difference.