MULTIVARIABLE MODELLING AND CONTROL OF CONTINUOUS PROCESSES: Wastewater treatment plants and complex fluids production as case studies

In this thesis, the multivariable modelling and control of continuous processes is discussed. Two main lines of research were followed: the multivariable system identification for processes subjected to disturbances and the multivariable modelling and control of a continuous production of complex fluids. For the first topic, wastewater treatment plants were used as case study. The goal of the work was to develop a method to implement multivariable variations of the manipulated inputs chosen for the identification phase, in order to obtain as much information as possible on the system in the shortest time. Signals for manipulated inputs were randomly generated according to the Generalized Binary Noise approach and inputs combinations were selected on the basis of the D-Optimal Design criterion. The Benchmark Simulation Model No. 1 was used as process simulator. Non-linear autoregressive neural networks were implemented to evaluate transfer functions of linear models. The procedure allowed to obtain good results as regards the estimation of gain constants of such models. For the second topic, the production of non-Newtonian water-free detergents was considered as case study, with the goal to develop control strategies for such process. Rheological characterization of the product was addressed by means of rheometers and a viscometer. The Carreau model was chosen for the description of the rheological behaviour. The process was first modelled relating the parameters of the Carreau model with the mass flow rate of one ingredient. A single-input single-output feedback Proportional-Integral controller was designed with the purpose to control a point on the viscosity curve of the product. The main outcome was that a viscosity curve was controllable with such control configuration, but the selection of the right controlled variable needs particular care. A second modelling attempt was made exploiting a multi-input multi-output control configuration. A process simulator based on a non-linear neural network was built. A double feedback controller was implemented with the objective to control two separate points of the viscosity curve using two manipulated variables. A Model Predictive Control was designed with the purpose to control more than two points on the viscosity curve using the same manipulated variables. The second controller returned faster responses in terms of dynamics with respect to the double feedback controller. Finally, the possibility to control the detergent production process by using an on-line ultrasound rheological sensor was explored. A data-driven approach was applied by means of Partial Least Squares technique and neural networks, in order to obtain a model capable to relate ultrasound variables with off-line rheological measurements of viscosities of the product. Fittings of experimental data by the neural network were better than those obtained with the Partial Least Squares model. A "smart operator" action was implemented as a control system, by means of a second neural network model. Thus, the control system was based on two data-driven models based on neural networks. Simulated tests of this control algorithm returned satisfactory results, proving the possibility of a real-time control of the viscosity curve of a complex fluid during its continuous production.