By dividing the redshift region under consideration into two bins, we estimate the limitations in determining the equation of state $w_{de}$ of dark energy at high redshift from current and future observational data including supernovae, baryon acoustic oscillation and observational Hubble data. It is found that the constraints of $w_{de}$ from current data are weak (2$\sigma(w_{de})\sim1$) even beyond redshift $z=0.3$. For simulated future observational data, it is shown that $\sim2300$ supernovae data from a SANP-like JDEM survey give $2\sigma(w_{de})\sim1$ beyond $z=0.6$. We consider the effects of the divided point, the number of supernovae data and the error in the distance modulus on constraining $w_{de}$ at high redshift. It is shown that the increase of number of supernovae data seems not efficient on improving the constraints of $w_{de}$ at high redshift, while the improvement of observational error in distance modulus seems to be a better way.