From my personal experience, you need to fit the value of U using experimental data (for example, use the value of U that most accurately reproduces the band dispersion).

While there are ab initio methods to compute it, using for example cRPA/DMFT (Phys. Rev. B 86, 165105) or the Coccocioni-de Gironcoli method (Phys. Rev. B 71, 035105), these methods fail in many particular cases so you should use experimental data if available.

Thanks,