Thesis:

Estudio sobre la inclusión de información probabilística en la optimización en tiempo real de procesos

This work aims to present a theoretical framework to include information of variables with stochastic behavior in real-time optimization (RTO) with modifier adaptation (MA), ensuring the feasibility of the operation with a given level of certainty. In every process, there are uncertainties intrinsically associated with the random behavior of variables involved and the partial understanding of certain phenomena. Model, process and market uncertainties, such as the use of imperfect models, disturbance variables and uncertain conditions of the process economy, are always present and make difficult the task of the control and optimization system. The management of these uncertainties will be fundamental to implement an optimization system capable of reaching a condition that meets the specifications and aims to maximize the economic benefit of the process. The proposed methodology was implemented in two examples: In simulation in a convex benchmark problem and experimentally, and at a laboratory scale rougher flotation circuit, aiming to increase economic efficiency and constraint feasibility. The inclusion of exogenous stochastic information allowed the system to reach an optimal point of operation, with a substantial improvement in the satisfaction of the problem constraints. However, a decrease in economic profit was observed compared to a deterministic optimization solution. A reduction in profit was observed in about 2%, while the restrictions satisfaction increased from 50% to 98% by including the probabilistic information for those restrictions with exogenous effect of stochastic variables. Additionally, the performance of the algorithm was evaluated by including disturbances in feed variables. The optimization layer presented poor performance by including perturbations in the process. The fast variation of the feed variable compared to the long stabilization times of the system implied a poor estimation of gradients, product of which the operation did not converge to an optimum point.Including probabilistic information of uncertain variables allowed to increase considerably the satisfaction of the restrictions, with the cost of decreasing in the obtained profit. This type of solution will be of importance when non-compliance with the restrictions leads to finer or higher costs of reprocessing, for example, non-compliance with quality or quantity of product. It is recommended to investigate other methodologies that explicitly incorporate the perturbation variables, to improve the convergence of the process under perturbations.