Nonlinear dynamic predictive models are essential for optimal operation and control of many processes. This work presents the predictive modeling of a simulated high purity distillation column using a continuous-time block-oriented modeling methodology developed by Rollins et al. (1998). The proposed methodology is based on a closed-form exact solution to Hammerstein structure and, hence, referred to as the Hammerstein Block-oriented Exact Solution Technique or H-BEST. The Hammerstein structure consists of a nonlinear static gain function that feeds a linear dynamic function. Thus, input vector u passes through the static gain function to obtain f(u) and then passes through the linear dynamic map G(s) and produces the output vector y.The use of Hammerstein structure for modeling this particular column had previously been investigated by Eskinat et al. (1991). We will refer to the model proposed by Eskinat et al. as the EJL model in this work. This work compares the EJL model to the H-BEST model and demonstrates the advantages of H-BEST in model development and its higher prediction accuracy for these types of processes. The EJL model is a discrete-time Hammerstein model, where the nonlinear static gain is given by a polynomial in the input u and the dynamics are modeled using a discrete transfer function. Chen and Rollins (2000) have shown that discrete-time models (DTM) can exhibit critical drawbacks when sampling is infrequent, nonconstant, or not online, but the continuous-time H-BEST method does not suffer from these limitations.The steady-state or ultimate response of the distillate composition shows dramatically different behavior for positive and negative changes in reflux rate, the input. A critical advantage of H-BEST is that it has the ability to account for this variation by using separate models for the steady-state distillate response over the input space. More specifically, since the static gain function f appears explicitly in the model formulation of H-BEST, it can be changed over the input space, as shown in this study.The procedure that H-BEST relies on for model identification is also very different from the conventional approach used by Eskinat et al. The input signal in the traditional approach is usually a pseudo-random sequence design (PRSD) with varying switching times. In contrast, H-BEST uses an input change sequence in the form of sequential step tests for model development derived from the statistical design of experiment (SDOE). Thus, this methodology is able to obtain very accurate ultimate response models over a well designed input space. Rollins et al. (2003a) show the superiority of SDOE over PRSD through the use of a quantitative measure of information content based on the D-optimality criterion (Bates and Watts, 1988) and Bhandari and Rollins (2003) demonstrates the superiority of SDOE over PRSD in a study involving a continuous stirred-tank reactor.The main objective of this article is to demonstrate greater accuracy of H-BEST in modeling this high purity distillation column as...