Nonlinear Model Predictive Control of a coagulation chemical dosing unit for water treatment plants ⋆ Oladipupo Bello ** Yskandar Hamam *** Karim Djouani **** Electrical Engineering Department/F’SATI, Tshwane University of Technology, Pretoria, South Africa ** (e-mail: engroobello@gmail.com) *** (e-mail: hamama@tut.ac.za) **** (e-mail: djouanik@tut.ac.za) Abstract: The need for processes to be operated under tighter performance specifications and satisfy constraints have motivated the increasing applications of nonlinear model predictive control (MPC) by the process industry. Nonlinear MPC conveniently meets the higher product quality, productivity and safety demands of complex processes by taking into account the nonlinearities and constraints in the processes. This paper examines the application of a nonlinear MPC to a multi-variable coagulation chemical dosing unit for water treatment plants. A nonlinear model of the dosing unit based on mechanistic modelling and identified by nonlinear autoregressive with external input (NLARX) estimator was developed. The simulation of the MPC based control system showed very good performance for set-point tracking and disturbance rejection. The closed loop performance of the nonlinear MPC (NMPC) compares favourably with the unconstrained and linearised nonlinear MPC (LTIMPC). The results of this study shows the suitability of nonlinear MPC for process control in the water treatment industry. 1. INTRODUCTION Coagulation in water treatment plants is a complex and nonlinear process requiring addition of optimum quantity of chemical reagents to raw water to meet the desired standards. One of the key issues in coagulation process is that water quality parameters vary unexpectedly and cannot be manipulated easily. These variations act as per- turbation or disturbance to the control loop of the system. The control objective is therefore targeted at manipulating the flow of the coagulants and pH adjustment chemicals to track the set-point signals in the presence of these possibly fast acting disturbances. The traditional control system for coagulation control has been found to have a number of limitations such as inaccurate process model to describe the behaviour of the system, slow responses to longer system delay time, variations in water quality parameters and loop interaction effects within the system. One of the commonly used control strategies is the feed- forward control. It involves adjusting the levels of chemical coagulants added to a process stream as a result of sensory information measured from the raw water variable(s). This is achieved by changing the feed rate of the coagulant metering pump according to the measured flow rate of the raw water [American Water Works Association & American Society of Civil Engineers, 2005]. This approach however becomes inappropriate, when the flow rates vary rapidly and there are large changes in other water quality variables. To address these problems in the feedforward control strategy, several models such as multi-linear re- ⋆ This work was supported by Tshwane University of Technology, Pretoria, South Africa. gression equations, artificial neural networks and fuzzy inference system algorithms have been proposed to predict the accurate amount of coagulants under varied conditions to replace the influent flowmeter response. For instance, Evans et al. [1998] proposed a feedforward controller based on adaptive neuro-fuzzy networks for Huntington water treatment works in North West Eng- land. In Baxter et al. [2002], the integration of neural network models with the supervisory control and data acquisition (SCADA) system through a number of process optimisation interfaces to optimise the chemical costs and doses online in real-time is presented according to varia- tions in influent water quality parameters. In Fletcher et al. [2002], a feedforward control was developed using mod- els based on nonlinear transformation of variables, multi- layer perceptron (MLP) and radial basis function (RBF) network to improve coagulation process. The findings of these research works are positive. However, the applica- tion of data-based models with feed-forward controllers to control coagulation process depends on the availability of a perfect model and accurate data from the plant opera- tional records. Another option in the literature for coagulation control is the application of feedback control strategy. This involves the use of sensors such as streaming current detector to measure the surface charge of the water after the coag- ulation process, compares the process value with the set point and adjusts the coagulant dosage pump accordingly to correct any deviation from the expected results. It is characterised by a system delay or dead time. During the seasons when the raw water quality changes frequently and widely, the control system may not function effectively re- Preprints of the 19th World Congress The International Federation of Automatic Control Cape Town, South Africa. August 24-29, 2014 Copyright © 2014 IFAC 370