Artificial neural network sensitivity analysis and process modeling with application to Kingston Fossil Plant net unit heat rate optimization

The objectives of this research are the development of a general purpose sensitivity analysis method utilizing an artificial neural network (ANN) as a model, and the application of this general method to investigate the net unit heat rate (NUHR) of the Kingston Fossil Plant.

Analog signals corresponding to 140 system variables measured in the extensively instrumented coal furnace and associated steam supply system were sampled and supplied by TVA. This data set, collected at one minute intervals over the period of June 2127, 1993, was provided along with many calculated quantities including NUHR evaluated at each measurement interval. Sensitivity coefficients were obtained from the ANN model, in both the scaled regimes of the ANN's transfer functions and in real world units. These coefficients were utilized to rank the importance of the sampled analog signals with respect to the NUHR.

To improve model prediction of the ANN, data smoothing of the erratic NUHR was performed with a Savitzky-Golay digital filter. The filter is found to satisfactorily smooth the signal, producing what is maintained to be a more meaningful signal. However, prediction results are not significantly changed.

Additional modeling was performed over a smaller data segment, representing a more local state of the system. Model prediction was significantly improved for this reduced data set.