Model Training
After the process and emissions data were collected and preprocessed, Callidus developed NOx emissions prediction models using a computer program for modeling. The initial model training used all of the available input sensors. Then a sensitivity analysis was performed to determine the dominant input sensors for emissions prediction. By using the top ranked inputs along with combustion engineering expertise, the model training process was repeated using a smaller set of inputs. Sixteen emissions prediction models were trained and tested. Project engineers then selected final NOx models that gave the best predictions and utilized a set of input sensors that represents the 54-F-1 operations from the combustion viewpoint. The final NOx emissions prediction model included six inputs, and the appropriate input sensor lower and upper bounds were set. The six sensor inputs from the emissions prediction model were then used for training the related sensor validation model. During real-time execution of the PEMS, the input variables must stay within the range defined by their lower and upper bounds in order for the model prediction results to be reliable.