Product development and advanced control applications require models with good predictive capability. In some cases, it is not possible to obtain phenomenological models with good quality because of the lack of information about the phenomena which govern the process. The use of empirical models requires less investment in modeling but implies the need for a greater number of experimental data points to generate models with good predictive capability. In this work. a nonlinear phenomenological model and a set of empirical models were built based on industrial data and compared with respect to the capability of predicting the melt index and polymer yield rate of a low-density polyethylene production process constituted by two fluidized-bed reactors connected in series. Deterministic and stochastic optimization algorithms were used to adjust the phenomenological model, increasing the probability of finding the global minimum of the weighted least-squares problem. The optimization algorithm based on flexible polyhedrons of Nelder and Mead showed the best performance. Among the tested empirical models, the quadratic partial least-squares (QPLS) model was more appropriate to describe the polymer yield rate and melt-index behavior and also presented better results than the phenomenological model. The QPLS model for the melt index was successfully used as a virtual analyzer of all advanced control strategy, improving the controller action and the polymer quality by reducing significantly the process variability.