Neural networks for control of NOx emissions in fossil plants

We discuss the use of two classes of artificial neural networks, multilayer feedforward networks and fully-recurrent networks, in the development of a closed-loop controller for discrete-time dynamical systems. We apply the neural system to the control of oxides of nitrogen (NOx) emissions for a simplified representation of a furnace of a coal-fired fossil plant. Plant data from one of Commonwealth Edison's (ComEd) fossil power plants were used to build a recurrent neural model of NOx formation which is then used in the training of the feedforward neural controller. Preliminary simulation results demonstrate the feasibility of the approach and additional tests with increasingly realistic models should be pursued.