A dynamic neural network for nonlinear process modeling and control

Andre McArthur Shaw, Purdue University

Abstract

A novel approach, which uses intrinsically dynamic neurons inspired from biological control systems, is presented in this work for input/output modeling of nonlinear dynamical processes. The network structure containing these dynamic nodes, with nonlinear weights in a feedback architecture, is called the Recurrent Dynamic Neuron Network (RDNN). For SISO applications the RDNN is shown to have arbitrary dynamic order $n \le N$ (where N is the number of neurons) and relative degree $r = 1$. CSTR case studies show that the RDNN does an excellent job of predicting nonlinearities such as asymmetric dynamic response and significantly outperforms linear models and more traditional neural network models in open-loop simulations. Input-output linearization (IOL) techniques are used to globally linearize the RDNN for use within the internal model control (IMC) framework. Closed-loop simulations with these CSTR examples show that the RDNN performs robustly when used within this control framework. For MIMO applications the RDNN is shown to have arbitrary dynamic order $n \le N$, vector relative degree $\underbrace{(1\cdots1)}\limits\sb{1{\times}M}$ (where M is the number of outputs), and is able to represent systems with input multiplicities. A binary distillation column case study demonstrates that the RDNN performs well in both open- and closed-loop simulations. For this 2 x 2 MIMO application, open-loop simulations show that the RDNN predicts the asymmetric nonlinear output responses. The RDNN is shown to be easily implemented in MIMO model-based control applications including model predictive control (MPC) and IOL/IMC. Simulations show that a combination of closed-loop and open-loop identification for the RDNN model results in a model-based controller which achieves robust control performance. Linearized optimal control and nonlinear optimal control MPC applications are implemented, with both performing comparably. Finally, the RDNN is used to model and control the reactor/regenerator section of the Amoco model IV FCCU. For this 4 x 4 control problem, the RDNN model performed well in both open- and closed-loop simulations. For closed-loop simulations the linearized optimal MPC approach is implemented. The performance is excellent for both disturbance rejection and setpoint tracking simulations.

Degree

Ph.D.

Advisors

Doyle, Purdue University.

Subject Area

Chemical engineering|Mechanical engineering|Artificial intelligence

COinS