Process control of a laboratory combustor using artificial neural networks

Active process control of nitric oxide (NO) emissions from a two-stage combustor burning ethylene (doped with ammonia) in air is demonstrated using two clusters of feed-forward multi-layer-perceptron neural networks. Steady-state experimental data are used for static back-propagation network training. The first cluster consists of two neural networks. The first network identifies the amount of ammonia in the feed. Based on that value and the NO set point, the second network adjusts the first-stage fuel equivalence ratio φ₁. The second cluster also consists of two neural networks. It is the process emulator and serves as a Smith time-delay compensator. A VISUAL BASIC interface control program accepts incoming concentration and flow rate data signals, accesses the neural networks, and outputs feedback control signals to selected electronic valves. Closed-loop results are compared to the open-loop results. The neural network-based controller successfully brought NO emissions into control after a step disturbance in the feed composition stream (ammonia dopant). The neural network-based controller shows a superior performance over the conventional proportional-integral-derivative controller.