REAL-TIME OPTIMIZATION WITHOUT MODEL IDENTIFICATION

Throughout the petroleum and chemicals industry, the control and optimization of many large-scale systems is organized in a hierarchical structure. At the real-time optimization level, decisions are made on a time scale of hours to a few days by a so-called real-time optimizer (RTO) that determines the optimal operating point under changing conditions. The RTO is typically a nonlinear program (NLP) minimizing cost or maximizing economic productivity subject to constraints derived from steady-state mass and energy balances, reaction kinetic relationships, thermodynamic equilibrium equations, physical property relationships, and physical equipment constraints [1, 2]. Because mathematical models with high accuracy are unavailable for most industrial applications, RTO classically proceeds by a two-step iterative approach, namely an identication step followed by an optimization step; variants of this two-step approach, such as ISOPE [3], have also been proposed for improving the synergy between the identication and optimization steps. In order to avoid the troublesome task of identifying a model on-line, xed-model methods can be used. The idea therein is to utilize both the available measurements and a possibly inaccurate model to drive the process towards a desirable operating point: the process model is used for estimating the gradients of the cost and constraint functions, while the measurements are used to correct the constraint and/or cost function