Resumen de A contribution to chemical process operation support: new machine learning and surrogate models based approaches for process optimization, supervision and control
Ahmed Shokry Abdelaleem Taha Zied
In the chemical process industry, the decision-making hierarchy is inherently model-based. The scale and complexity of the considered models (e.g., enterprise, plant or unit model) depend on the decision-making level (e.g., supply-chain management, planning, scheduling, operation) and the allowable time slot (weeks, hours, seconds) within which model simulation runs must be performed and their output is analyzed to support the decision making. The use of high-fidelity models, which include detailed physics-based description of the process, is attracting wide interests of the process engineers. Since, these First Principle Model (FPMs) are able to accurately predict the real behavior of the process, leading to realistic optimal decisions. However, their use is hindered by practical challenges as the high computational time required for their simulation and the unguaranteed reliability of their consistent convergence. The challenges become prohibitive at lower levels of the decision-making hierarchy (i.e., operation), where decisions are required online within time slots of minutes or seconds entailing lots of simulation runs using such complex and highly nonlinear FPMs. Surrogate modelling techniques are potential solution for these challenges, which relies on developing simplified, but accurate, data-driven or machine learning models using data generated by a FPM simulations, or collected from a real process. Although, there are progressive developments of surrogate-based methods in the chemical engineering area, they are concentrated in process design and steady-state optimization areas.
This Thesis presents a framework for the proper and effective use of surrogate models and machine learning techniques in different phases of the process operation. The objective is to provide efficient methodologies, each supports the decision making in a specific phase of the process operation, namely; steady-state operation optimization, Model Predictive Control (MPC), multivariate system identification and multistep-ahead predictions, dynamic optimization, Fault Detection and Diagnosis (FDD) and soft-sensing. Each developed methodology is designated according to careful State-Of-Art (SOA) review that identifies the gaps and missing requirements to be covered. The SOA, identified gaps and the contributions of each methodology are summarized in Chapter 1 and detailed in the introduction of each of the following chapters.
In this context, Chapter 3 presents a surrogate-based methodology for steady-state operation optimization of complex nonlinear chemical processes modelled by black-box functions. Chapter 4 proposes machine learning-based methodologies for multiparametric solution of complex operation optimization problems subjected to uncertainty. Chapter 5 presents a data-based multiparametric MPC methodology that enables simple implementations of explicit MPC for nonlinear chemical processes. Chapter 6 proposes a data-driven methodology for multivariate dynamic modelling of nonlinear chemical processes and for multistep-ahead prediction. Chapter 7 suggests a dynamic optimization methodology for solving optimal control problems of complex nonlinear processes based on data-driven dynamic models. Chapter 8 shows a hybrid methodology to improve FDD of chemical processes run under time-varying inputs based on multivariate data-driven dynamic models and classification techniques. Chapter 9 presents data-driven soft-sensing methodologies for batch processes operated under changeable initial conditions. The effectiveness of the developed methodologies is proved by comparing their performances to those of classical solution procedures existing in the SOA, via their applications to different benchmark examples and case studies. The promising results and their sound analysis allowed to publish many papers in top-ranked journals and proceedings, and to present them at several top-ranked international conferences including two Keynote presentations.
