Global control of nonlinear dynamical systems using neural networks. Case study: continuous stirred tank reactor

A neural network model is proposed to simulate dynamic behaviour of a continuous stirred tank reactor. The network with two hidden layers, seven inputs, and two outputs was trained by a backpropagation learning algorithm. The transfer functions of neurons in hidden layers are sigmoidal, and the output neurons are linear. The network is capable to learn the global behaviour of the reactor and can be used as a one-step-ahead predictor in the whole region in which it was trained. This predicting capability of the network is later used by a "topological" control algorithm that can drive the reactor even into an unstable steady state and stabilize it there. The "topological" control algorithm uses the network predictor to create a set of reachable states; then finds a state that is closest to the target, and eventually calculates corresponding controls. This control strategy has a global character.