| Nonlinear processes, such as a pH system, can not be controlled using conventional methods. Nonlinear processes either have gains, dynamics, or both that change depending on the operating conditions. Control of nonlinear processes can be achieved by using neural networks and nonlinear model predictive control. Neural networks provide a model of the system and nonlinear model predictive control controls the process.; The process examined is a single input single output simulated pH nonlinear system provided by a Dual Process Simulator KI 100. It consisted of three continuous line segments, each with a different slope. The manipulated variable and process variable range from 0–100%. Two types of neural networks are used to model the process. The first network is a feedforward network that uses error back-propagation to train the network. The second is a recurrent network using the Random Optimization Method to adjust the weights. Each model requires a different nonlinear model predictive control program to control the process.; The effectiveness of the control strategies is evaluated by implementing set point changes and adding various magnitudes of noise to the process input. Both methods are nearly equivalent in their ability to control the simulated process. When controlled variable bounds are set between 9–10, the recurrent network is able to control the process better than the feedforward NMPC. The response appears to be less oscillatory with the recurrent NMPC at this set point. |
