The control community usually sees the primary purpose of automatic feedback control in keeping the controlled variables as close as possible to their set-points or in nicely tracking dynamic set-point changes. From the perspective of process engineering, however, the goal is to operate the plant in such a way that the net return is maximized in the presence of disturbances and uncertainties while respecting constraints on product specifications, conversion, environmental constraints and the operating limits of the equipment. This cannot be achieved by simply implementing the nominal operating point determined in plant design, as models used for plant design do not represent the real process exactly and multiple internal and external disturbances act on the plant, from varying feed characteristics to varying temperatures of cooling media, catalyst ageing and heat exchanger fouling. The issue of varying feeds applies particularly to the “green” plants of the future processing bio-based feed streams.

An operating regime that was optimized for the plant model will not lead to an optimal operation of the real plant; it may even be infeasible. Feedback control, automatic or manual, is indispensable to handle the inaccuracies and uncertainties present in the design process, and to make full use of the capacity of the equipment.

In highly automated plants, the goal of an economically optimal operation is usually addressed by a two-layer structure . On the upper layer, the operating point of the plant is optimized based upon a rigorous nonlinear but stationary plant model (real-time optimization, RTO). The optimal operating point is characterized by set-points for controlled and manipulated variables that are passed to lower-level controllers that keep the chosen variables as close to these set-points as possible by manipulating the available degrees of freedom of the process within certain bounds.

Recent progress in algorithms for numerical simulation and optimization creates the potential to move from the two-layer architecture to direct online optimizing control, e.g., (Engell, 2007) , (Engell, 2009)  (also called one-layer approach (Zanin, et al., 2000)  or full optimizing control (Rolandi and Romagnoli, 2005) , or dynamic RTO). In this approach, the available degrees of freedom of the process are directly used to optimize an economic cost functional over a certain prediction horizon based upon a rigorous nonlinear dynamic process model. The regulation of quality parameters, which is usually formulated as a tracking or disturbance rejection problem, can be integrated into the optimization by means of additional constraints that have to be satisfied over the prediction horizon.

• • Fast reaction to disturbances, no need to wait for convergence to a steady state as in the case of steady-state optimization of the operating point, and the problem to detect stationarity is avoided;
  • No regulation of constrained variables to set-points, which necessitates a safety margin between these set-points; the exact constraints can be implemented for variables which are measured and only the model error has to be taken into account for unmeasured constrained variables;
  • Over-regulation is avoided, no variables are forced to fixed set-points and all degrees of freedom can be used to optimize process performance;
  • No inconsistency arises from the use of different models on different layers;
  • Economic and ecologic goals and process constraints are formulated clearly and naturally and do not have to be represented indirectly by parametrizing a control cost function where optimality would be lost and the trade-offs are not transparent;
  • The overall scheme is structurally simple and the tuning requires less trial-and-error.

 

The project focuses on those aspects of the application of dynamic online optimization in process control that urgently need to be addressed in order to make this extremely versatile and promising approach an industrial reality. It has been shown in simulations and in lab-scale and pilot-scale experimental applications in university labs that it is feasible to replace the conventional tracking-type control schemes by the online optimization of the degrees of freedom in order to optimize an economic cost function. But for long-term real-life applications, huge obstacles still have to be overcome. The major issues are:

• • Reliability and robustness of online nonlinear optimization based control schemes;
  • Improved interaction with and acceptance by the operators;
  • Making use of the data collected during real operation to reduce the cost of modeling and to maintain the accuracy of the models when the plant and the feeds change;
  • Extension to the handling of discrete control variables (in particular during start-up and shut-down).

 

We will combine research in numerical algorithms, problem formulations, model adaptation and human-machine interaction in order to pave the ground for the practical application of optimizing control of industrial processes. The results will be validated and demonstrated by the application to a challenging complex and innovative process in pilot scale at TU Dortmund which is very similar to an industrial application but does not impose as many constraints to experimental work.

Model-based optimizing control is a versatile innovative technology with an enormous potential to improve the economic performance and the sustainability of processing plants. The need for online adaptation will drastically increase if more bio-based feed streams are used to produce chemicals or fuels, as their properties cannot be kept as constant as in the classical petrochemical routes.

The results of this work are relevant not only for the chemical industry but also for all industries where complex transformations are performed with relatively slow dynamics that are controlled interactively by skilled operators, e.g. iron and steel, glass, food and beverages, treatment of water and waste water.

In this project, we focus on issues that are decisive for the transfer of the general concept of optimizing control to real applications and require a large effort. In addressing these problems, we will explore new ideas to improve the reliability and the performance of the optimization algorithms, to exploit the data collected during operation to update the process models in a safe way, to improve the interaction between the operators and the optimizing control scheme and to extend the scope to handling discrete manipulated variables. These issues have significant connections and therefore require an integrated interdisciplinary approach.

The concept of optimizing control can be combined with the idea of future modularized chemical production systems that is currently developed in several large EU-funded projects, e.g. F3 FACTORY and COPIRIDE. In the processing plant of the future, “intelligent” autonomous modules will be equipped with optimizing control schemes that can be parameterized externally such that the overall plant is operated in a competitive and environmentally benign fashion.

In the long term, the optimization-based approach will also lead to an increasing integration of plant design and control design. In the design stage, not only stationary models but also dynamic models are increasingly developed, mostly for operator training simulators that are frequently built in parallel to the design and commissioning of the plant itself. Employing the optimization-based approach, the behavior of the controlled plant under the influence of disturbances can already be studied in the design phase without a time-consuming design of tailored control schemes. The optimizing controllers can then be implemented without large modifications on top of a simple standard regulatory control layer, first in the operator training systems and then at the real plant.