ABSTRACT
As part of an ongoing effort at the Pittsburgh Energy Technology Center to identify new techniques for the improved operation and control of coal-fired utility boilers, advanced approaches for data analysis, such as neural networks, are being explored. The availability of an existing set of highly interrelated data on ash deposition facilitated an initial comparison of standard statistical and neural network techniques for data analysis. A previously published database of fouling deposit weights formed while firing various low-rank coals in the Energy and Environmental Research Center's pilot-scale combustion facility at the University of North Dakota was used for this study. A feed-forward back-propagation neural network algorithm was used to predict fouling deposit weight based on standard coal analyses. The completed study indicates neural networks were substantially better than standard statistical techniques at generating useful relationships using both independent and dependent variables based on coal ash chemistry. The completed study also demonstrates the utility of neural networks at analyzing complicated processes, such as the formation of ash deposits, and provides the incentive for their broader application for power plant studies.
