Staged combustor control using artificial neural network-based process models

Process controllers using trained, feed-forward, multi-layer-perceptron (FMLP) neural networks as complex process models have been successfully demonstrated for the active, on-line control of selected species emitted from a two-stage combustion reactor. In the first case, as compared to a proportional-integral-derivative controller, faster control of exhaust oxygen content with nearly no offset was achieved using a proportional controller with a variable bias value as determined by an FMLP. In the second case, effective and rapid control of exhaust nitrogen oxide, after a separate feed stream disturbance and a set point change, was achieved using a controller comprised of two clusters of FMLP neural networks. The first cluster identified the process disturbance and adjusted the manipulated variable. The second cluster served as a Smith time-delay compensator. All the FMLP networks used were trained off-line using steady-state data obtained from both experiments and from direct combustor simulations based on detailed chemical reactions. Copyight