Nonlinear model predictive control strategy based on soft computing approaches and real time implementation on a coupled-tank system

In order to effectively implement a good model based control strategy, the combination of different linear models working at various operating regions are mostly utilised since a single model that can operate in that fashion is always a difficult task to develop. This work presents the use of soft computing approaches such as evolutional algorithm called simulated annealing (SA), a genetic algorithm (GA) and an artificial neural network (ANN) to design both a robust single nonlinear dynamic ANN model derived from an experimental data driven system identification approach and a nonlinear model predictive control (NMPC) strategy. SA is employed to give an initial weight for the training of the ANN model structure while a gradient descent based Levenberg–Marquardt Algorithm (LMA) approach is used to optimise the ANN weights. The designed NMPC strategy is optimised using a stochastic GA optimisation method and is tested first in simulation and then implemented in real time practical experiment on a highly nonlinear single input single output (SISO) coupled tank system (CTS). An excellent control performance is reported over the conventional proportional-integral-derivative (PID) controller and results show the effectiveness of the approach under disturbances. The nonlinear neural network model proved very reliable in different operating regions. The SISO system can be upgraded to multi-input multi-output (MIMO) system while the whole NMPC approach can easily be adapted to other industrial processes