OCR Text |
Show the furnace 0 2 and the air preheater inlet temperature. The best M L R model for C O included just the natural gas flow and the furnace 02. The neural network analysis was performed on this data with what is termed a holographic neural network (HNet). The parameters most significantly effecting the N o x model were the natural gas flow, the landfill gas flow, 0 2 at the furnace outlet, windbox to furnace differential pressure, air preheater inlet temperature and the amount of flue gas recirculation. The same set of parameters were chosen for the C O model, however, the non-linearity of these parameters were significantly higher for the C O model compared with the N O x model. Table 6. Gas-Fired Boiler Results Parameter NOx CO Measured Min 16 1.2 Max 225 3.8 R2 Stats (MLR) 0.84 0.67 N.N. (BPN) 0.91 0.81 RMS-Error Stats (MLR) 21.1 0.26 N.N. (BPN) 15.9 0.2 IL? o M K J SoiLt-/C Boiler applications appear to have more variables and be more non-linear than the turbine and reciprocating engine applications discussed previously. Many newer boilers m a y be equipped with combustion systems that have many different air registers and each different permutation may have significant affects on the N O x and the C O . It is important that these configurations be understood and tested over the extent of the complete operating ranges. It was apparent when testing that some CO spikes could not be predicted by the system, however, these occurrences generally happened for short durations during transient conditions when for example burners were taken in and out of service. The steady state C O was quite stable, however, these transient spikes are difficult to detect using purely predictive systems without relying on additional instrumentation. A comparison of predicted verses actual NOx is given in the following two figures for both the gas turbine and the gas-fired boiler: -14- |