Comparison of Model-Based and Non-Model-Based Strategies for Nonlinear Control of a Three-Tank System Martin A. Capcha, William Ipanaqué Laboratory of Automatic Control Systems Piura University Piura, Perú martin.capcha@posgrado.udep.edu.pe, william.ipanaque@udep.pe Robin De Keyser Department of Electrical Energy, Systems and Automation Ghent University Gent, Belgium Robain.DeKeyser@UGent.be Abstract— To achieve optimal performance in the regulation of nonlinear systems, more advanced control techniques are needed. In this paper, three of these techniques are applied and compared in a system composed of three tanks in cascade. The first one is the Nonlinear Extended Prediction Self-Adaptive Control (NEPSAC). NEPSAC is a nonlinear predictive controller and therefore, is based on solving an optimization problem using the nonlinear model of the process to find a feedback control law. The second one is a gain scheduling controller using an array of autotuned PID controllers. This will be the non-model-based strategy. Finally, the gain scheduling approach is used again but with an array of linear model predictive controllers. In this case the predictive controller is EPSAC, the linear version of NEPSAC controller. Comparing all these control strategies, this work deepens in the characteristics, performance, advantages and disadvantages of these techniques to deal with nonlinear processes. It is intended to be a reference guide to choose - from a practical point of view - the most appropriate controller when dealing with a soft nonlinear system. Keywords— Nonlinear predictive control, Adaptive control, Output regulation I. INTRODUCTION Most of the relevant processes in the industry present nonlinearities in their behavior and in many cases, these nonlinearities cannot be avoided. The use of the very well- known linear control techniques have demonstrated their limitations in dealing with these kind of processes, and the development and applications of nonlinear control techniques are necessary for a good performance in the processes. One alternative is Nonlinear Model Predictive Control (NMPC), which is a specialized set of control strategies that are part of the more general Model Predictive Control (MPC) family of controllers. NMPC is based in the on-line calculation of the control action in the attempt to optimize a cost function using the mathematical model of the system [1]. However, the special feature of NMPC controllers is that they can handle nonlinear models of the different processes and, in this way, they are not limited to processes with a single point of operation, or processes with low nonlinear behavior; their use is justified when the process is strongly nonlinear or when the system needs to work in different operating zones and regimes [2]. Although NMPC theory had reached a certain degree of maturity in the last years, it is still limited to a certain kind of applications. This restriction is in part because of the lack of a guaranteed real-time solution of the resulting nonlinear optimization problem due to the multiple local minima the problem could have, and also because of the difficulty to create a nonlinear model of the system, together with nonlinear state estimators of the same [3]. In this paper, the NMPC technique applied will be Nonlinear Extended Prediction Self-Adaptive Control (NEPSAC) [4]. NEPSAC is a nonlinear control technique, used in different works as in [5] or [6]. It is characterized by using a nonlinear model of the system instead of a linear approximation around an operating point, and it uses an iterative method for the optimization of the predicted behavior of the system, making its implementation computationally friendly compared to other nonlinear predictive control techniques. Another popular choice is the use of gain scheduling control, which has long acceptance to deal with nonlinear systems [7]. This method uses an array of linear controllers optimized around specific operating points of the system, and a gain scheduling signal, which will control and generate the control signal based on the individual responses of each linear controller. This way, the gain scheduling control can handle the variations of the dynamics of the system in its entire operation range. The second method used in this work is thus gain scheduling with an array of autotuned PID controllers (GS-PID). This technique does not need any model of the process because the parameters of each PID controller are tuned automatically based on the response of the system to the application of simple tests. The third technique, called here GS-EPSAC, will also be a gain scheduling approach but using an array of linear model predictive controllers, specifically the Extended Prediction Self- Adaptive Control (EPSAC), presented in [4] and [8]. With this technique, the facilities to design linear controllers is combined