An extension of the well-known process-reaction curve method to empirically determine reduced-complexity models aimed to the design and tuning of feedback controllers for nonstationary batch processes is reported. The basic idea is to isolate the dynamics associated with the manipulated variable from the main time-variable behavior that characterizes the operation, by taking the time evolution of a previous run as reference. One or more input-perturbed evolutions can then be compared to the previous dynamic pattern yielding referential reaction curves. This modeling approach cancels out most of the nonstationary behavior, allows capturing the dominant manipulated-variable dynamics, and the use of available tuning rules for integrating systems. The effectiveness of this procedure is illustrated by using a nonlinear model of a bioreactor that simulates the production of xanthan gum. (C) 2004 American Institute of Chemical Engineers.