Novel Approaches to Online Process Optimization Under Uncertainty: Addressing the limitations of current industrial practice

In the face of growing competition, and increased necessity to focus on sustainability and energy efficiency, there is a clear need to optimize the day-to-day operation of many industrial processes. One strategy for online process optimization is to use model-based real-time optimization (RTO). Despite the motivation and the potential, real-time optimization is not as commonly used in practice as one would expect. This thesis takes a detailed look at the different challenges that impede practical implementation of real-time optimization and aims to address some of these challenges. In brief, this thesis presents novel algorithms and methods ranging from simple control structures, to model-free and model-based optimization and more complex scenario-based economic model predictive control. One of the fundamental limiting factors of traditional steady-state RTO is the steady-state wait time. This essentially discards transient measurements, which otherwise contains useful information. In part I of this thesis, we propose different approaches to use transient measurements for steady-state optimization, with the goal of minimizing the steady-state wait time. Moreover, different algorithms to real-time optimization that do not require the need to solve numerical optimization problems are proposed, thus alleviating many of the computational challenges which impede practical application of traditional RTO approaches. First, we propose a “hybrid” approach, where the model adaptation is done with transient measurements and dynamic models, and the optimization is performed using steady-state models. To further simplify the steady-state optimization, we then convert the hybrid RTO approach into a feedback RTO approach. Here, the transient measurements are used to estimate the steady-state gradient, which is controlled to a constant setpoint of zero using feedback controllers. The steadystate gradient is estimated using a novel method based on linearizing the nonlinear dynamic model around the current operating point. To address the cost of developing models, we demonstrate the use of classical controllers where the economic objectives are translated into control objectives. We also provide a systematic approach to switch between different active constraint regions using selectors. For the unconstrained degrees of freedom, we then propose a novel extremum seeking scheme using transient measurements for a class of Hammerstein systems, where the linear dynamics are fixed. We show that the proposed approach converges significantly faster than the classical extremum seeking scheme, and provide robust stability margins. Part I concludes by showing that the different methods work in different time scales, and by hierarchically combining the different approaches, one can handle a wider class of uncertainties.