From real-time optimization techniques to an autopilot for steady-state processes
Most industrial systems have objectives to meet (e.g. economic performance, production quality), are subject to disturbances (e.g. market fluctuations, aging, variations in the processed raw material), and almost never follow exactly the models that are supposed to predict their behavior. In order for these systems to be usable, it is often necessary to use controllers to stabilize (reject the effects of disturbances) some of the plant variables as close as possible to setpoints assigned to them. The selection and tuning of these controllers is a challenge for control engineers that is not discussed in this thesis. Once these controllers are installed in a plant, their setpoints become the interface through which the engineers supervising the plant can drive it, just as the steering wheel, pedals, and gearshift are in a car the interface between (the engineer) the driver and (the pant) the engine and wheel orientations. Plant piloting combines the expertise of the engineers that can be represented by a (rough) model of the plant, real time observations, and the production objectives. If the example of a car is reused, then the piloting of the plant can be assimilated to the combination of the driver's skills (his expertise) with what he sees of his environment (data received in real time) and the destination to reach (objectives). The comparison with the car stops here since driving involves social phenomena (other cars) that are not encountered in the industrial systems considered in this thesis. The automation of the piloting of a plant corresponds to a research field called real time optimization (RTO). This thesis builds on a RTO method published in 2009, called modification adaptation (MA), to which it adds methodological improvements: • (Chapter 2) The number of experiments needed to identify the optimum of the plant is reduced thanks to a novel way of adjusting the only free parameter of MA (its filter). The efficiency of this method is supported by an empirical analysis and simulation results. • (Chapter 3) A great methodological weakness of MA implying the possibility of violating plant constrains even when the model predicts such violations is identified and illustrated. A simple variant of MA is proposed to remove this methodological weakness. • (Chapter 5) It is shown that MA can be applied in a distributed way and that such an application generally enables a better decision making thanks to a more relevant combination of the model with the data. Also, the limits of distributed approaches is discussed. Then, an autopilot that responds to each type of event that can affect a plant while offering an intelligent management of the collected data is proposed. The description of this autopilot is divided into two chapters: • First, chapter 4 presents all its functionalities and the way they interact with each other. To this end, a simplified autopilot is introduced to present as clearly as possible the functioning of an autopilot. • Second, in chapter 6, for each of an autopilot functions a practical way of fulfilling its objective is proposed. The outcome of this chapter is an autopilot that provides an answer to most of RTO-related problems. Note that we do not claim at any point that each of these answers is "optimal''. Finally, each contribution is illustrated by means of a case study from chemical engineering, such as the Williams-Otto reactor, the Tennessee Eastman challenge process, and the Williams-Otto process. Moreover, all codes are made available (in open source) and are commented so that they can be used to, e.g. reproduce and/or improve the results presented in this manuscript, or also to apply them to industrial processes.