Sebastian Engell

Model predictive control using neural networks

In this article, we present the application of a neural-network-based model predictive control scheme to control pH in a laboratory-scale neutralization reactor. We use a feedforward neural network as the nonlinear prediction model in an extended DMC-algorithm to control the pH-value. The training data set for the neural network was obtained from measurements of the inputs and outputs of the real plant operating with a PI-controller. Thus, no a priori information about the dynamics of the plant and no special operating conditions of the plant were needed to design the controller. The training algorithm used is a combination of an adaptive backpropagation algorithm that tunes the connection weights with a genetic algorithm to modify the slopes of the activation function of each neuron. This combination turned out to be very robust against getting caught in local minima and it is very insensitive to the initial settings of the weights of the network. Experimental results show that the resulting control algorithm performs much better than the conventional PI-controller which was used for the generation of the training data set. >

Gain-scheduling trajectory control of a continuous stirred tank reactor

Abstract The control of continuous stirred tank reactors is often a challenging problem because of the strong pronounced nonlinearity of the process dynamics. Exact feedback linearization and gain-scheduling are two well-known approaches to the design of nonlinear process control systems. The basic idea in this paper is to combine these techniques to obtain a control structure which preserves the advantages and overcomes some of the problems of the two concepts. In a first step, a nonlinear state feedback controller is computed by exact linearization of the process model to shape the nominal closed-loop system. The required unmeasurable state variables are obtained by simulation of the process model. This part of the controller thus is a pure nonlinear feedforward compensator for the nominal plant. To act against disturbances and model uncertainty, a nonlinear gain-scheduled controller is designed by approximately linearizing the process model not for a number of operating points as in the standard gainscheduling approach but around the nominal trajectory generated by the nonlinear feedforward controller. The design approach is applied to a non-trivial concentration control problem in a continuous stirred tank reactor with nonminimum phase behaviour, unmeasurable states, and model uncertainties as well as unknown disturbances. The nonlinear control structure is compared to a linear controller and to a pure gain-scheduling controller and shows excellent performance even for worst case disturbances and model uncertainties.

Sequencing of batch operations for a highly coupled production process: Genetic algorithms versus mathematical programming

In this contribution the application of a genetic algorithm to a real-world scheduling problem with highly coupled production is presented. Since the schedules are highly constrained, special attention is paid to the handling of constraints at the different levels of the implementation of the algorithm. The quality of the result and its numerical performance is discussed in comparison with a mathematical programming algorithm.

The Robust Performance Number: A new tool for control structure design

The Robust Performance Number (RPN) and the Robust Performance Number of a Plant Set (RPPN) are introduced as measures to determine the controllability of a system. These numbers indicate how potentially difficult it is for a given system to achieve the desired performance robustly. The RPN and the RPPN reflect both the attainable performance of a system and its degree of directionality. The RPN is used for the analysis of the nominal system, whereas the RPPN can be used to include the effects of the nonlinearities and uncertainties of the model. It is shown how the attainable performance can be determined. As appropriate scaling is crucial for the analysis and for the design of multivariable systems, a strategy for system scaling based on the RPN is also presented.

A case study for control structure selection: air separation plant

Abstract The control structure selection for an air separation plant is discussed using the robust performance number (RPN) as a controllability index. The RPN indicates how potentially difficult it is for a given system to achieve the desired performance robustly. It reflects both the attainable performance of a system and its degree of directionality. The RPN criterion is compared with other well-known and new methodologies. These other criteria do not take the dynamic characteristics of the plant into account, hence they can be misleading. The predictions made by RPN are verified by closed loop simulations. The analysis was performed using a linear nominal model, but can be extended to include nonlinearities and uncertainties.

Decentralized vs. model predictive control of an industrial glass tube manufacturing process

For the glass tube manufacturing process considered in this paper, one of the controlled variables can only be measured after a considerable time delay. For this process a decentralized (diagonal) PI-controller is compared to a dynamic matrix control (DMC) scheme. The DMC controller shows superior performance in set-point tracking. However, its disturbance rejection properties are comparable. By analyzing the DMC scheme in state-space form, it is shown that the time delay limits the achievable performance of the DMC controller regardless of its tuning parameters. To overcome this limitation, at least partly, it is suggested to use a suitable disturbance model.

An open software architecture for batch process simulation

Simulation of recipe driven batch processes is a complex demanding task. It can be performed either with general purpose simulators, causing a high modelling effort, or with specific tools which often lack flexibility. Neither of these approaches is appropriate in all cases. In this contribution, we outline a software architecture for batch process simulation which is flexible enough to be adapted to meet the particular requirements of the simulation problem at hand. It consists mainly of the elements model creation, output clients, simulator, and user interface. All these components are exchangeable to offer the software solution which is best suited for the specific problem. The software concepts are presented in terms of design patterns.

FUZZY LOGIC CONTROLLER DESIGN FOR PH-CONTROL IN A CSTR

Abstract In this paper two different approaches for the design of fuzzy logic controllers for the control of pH in a laboratory-scale CSTR are shown and compared. First a fuzzy controller was designed based on pure heuristics; the design process was fairly demanding due to the complexity of the system. Second, a fuzzy controller was identified based on a time-optimal controller derived from a coarse mathematical model. After addition of some heuristic rules to the rulebase which increase the robustness, this fuzzy controller shows an excellent performance.

Dynamics and control of a process with recycle streams

Abstract The dynamic behavior and the control of a heating system with recycle streams are analyzed using a nonlinear grey-box model. The analysis led to the conclusion that the system can become unstable at low frequencies. The instability was verified in experiments and possible solutions for this problem are discussed. A control strategy that works for a wide range of distilling mixtures and operating conditions is derived. Finally, startup procedure is presented.

Using Model-Checking for Timed Automata to Parameterize Logic Control Programs

In this contribution we describe how the modeling and analysis framework of Timed Automata can be used to determine valid parameter ranges for timers in logic control programs. The procedure is illustrated by means of a simple process engineering example for which the complete Timed Automata model is presented. To analyse the model, the tool HyTech is used which provides routines to determine values of model parameters depending on reachability conditions.