The balanced truncation method for reducing the size of a model was originally developed for linear systems. When extended to nonlinear systems, some considerations must be faced. First of all, the calculation of the balancing transformation matrix is not unique. This may results in nonphysical values for the reconstructed states, which may lead to failure, for example, in thermodynamic routines. To reduce this problem, it is recommended to include all the states in the balancing outputs. To further reduce the effect of nonlinearties in the original model, it is recommended to use a linearizing static transformation of the states, if available. In this paper, distillation column models are used as a case study, and, in this case, a logarithmic transformation of the compositions is beneficial.