OCR Text |
Show Utilizing this tool, the boiler operating conditions were referenced to the stack emission data derived from portable continuous emissions monitors. The result was a high fidelity computer model capable of predicting boiler NOx emissions with the relative accuracy of a continuous emission monitor. Compliance testing of the computer model was completed on June 9, 1993 by Ramcon Environmental Corporation of Memphis, Tennessee utilizing Relative Accuracy Test Audit ( R A T A ) protocol. The testing was witnessed by A D P C & E personnel. A n operating permit was granted on July 13, 1993. Technology Employed The Software CEM emissions model is developed by multivariate nonlinear empirical regression of process operating data against stack test emissions data. In order to develop the empirical model, historical data demonstrating the relationship between monitored variables and emissions must be captured. The process data from all key operating parameters were already being captured through the plant distributed control system and data historian. To capture the emissions data, the emissions are measured and electronically recorded during an extended stack test (five days) utilizing conventional C E Ms installed in Eastman Chemical's own portable emissions monitoring trailer. Emission levels were measured over an extensive range of process operating rates. The process data and the emissions data files were extracted onto flat ASCII files for provision to the Process Insights tool. Process Insights was used to preprocess, analyze, and construct the empirical nonlinear model. This model was automatically designed by Process Insights using fuzzy logic and chaotic systems technologies to derive the optimal neural network structure for the model. The software removes the complexity of these advanced technologies from the user, and provides a friendly graphical user interface. The software provides tools that allow the developer to begin the project with A L L process variables currently being monitored included in the model, and glean from that large list of variables the few key non-linear parameters correlated to emissions. The list of parameters used by the model is quickly reduced to the minimal set. A final run-time model is then generated. The high accuracy of this model is shown in attached Figure III-15-1. This is a instantaneous (now averaging) strip chart showing minute by minute readings of the Software CEM model versus a hardware C E M . A similar figure (Figure III-15-2) is also attached for a gas fired turbine. In addition to development of the emissions estimation model, a sensor validation model was also developed for each application, and has become a standard feature of all Software CEM applications. The purpose of the sensor validation model is to detect the failure of critical process sensors and maintain the accuracy of the emissions prediction, thereby allowing the 9 0 % up time even though P E M S systems are dependent on a higher number of sensors. Process Insights is used to derive the interrelationships and dependencies of the monitored variables. The sensor validation model detects when measurements of process variables are no longer consistent with peer inputs. Upon recognition of a failed sensor, an alarm is generated in the plant database system. The model then automatically replaces the failed raw sensor values with a reconstructed estimate. The reconstructed value is accurate enough to maintain mandated relative accuracy of the Software CEM while the failed sensor is out for repair. In the case of Arkansas Eastman, five of the twenty-one final sensors could be simultaneously out for repair before the relative accuracy would fall below mandated limits. A similar sensor failure analysis is provided on each new project. Pg-3 III-15 |