Predictive Emissions Monitoring Using a Continuously Updating Neural Network Available to Purchase

In the European Union, power plants of more than 50 MW (thermal energy) need to comply with the Large Combustion Plant Directive (LCPD, 2001) implying that flue gas emissions need to be measured continuously. Traditionally, emissions from power plants are measured using Automated Measuring Systems (AMS). The LCPD states that no more than 10 days of emission data may be lost within one year including days needed for maintenance. This is the reason why more and more power plants are currently installing a second, back-up AMS since they have problems with the availability of their AMS. Since early 1990’s, Predictive Emissions Monitoring Systems (PEMS) are being developed and accepted by some local authorities within Europe and the United States. PEMS are in contrast to AMS based on the prediction of gaseous emissions (most commonly NOx and CO) using plant operational data (eg. fuel properties, pressure, temperature, excess air, …) rather than the actual measurement of these emissions. The goal of this study is to develop a robust PEMS that can accurately predict the NOx and CO emissions across the entire normal working range of a gas turbine. Furthermore, the PEMS should require as little maintenance as possible. The study does not intend to replace the AMS by a PEMS but rather to use the PEMS as a backup for the AMS. Operational data of a gas turbine, acquired over a long period, was used to identify inputs with a high influence on the NOx and CO formation. Consequently, simulations were done testing different model structures and calibration methodologies. The study shows that a static model failed to predict the emissions accurately over long time periods. In contrast, a dynamic or self-adapting algorithm proved to be most efficient in predicting the emissions over a long time period with a minimum of required intervention and maintenance. The self-adapting algorithm uses measured AMS data to continuously update the neural network. Since the PEMS is developed as a backup for the AMS, these data are readily available. The study shows that in case of a failing AMS, the developed model could accurately predict the NOx emissions for a duration of several weeks. Although not discussed in detail in this study, a quality assurance system of the PEMS is also developed since the PEMS needs to comply to the EN14181 (as does any AMS). The PEMS as a backup of the AMS instead of a second AMS is cost and time saving. Not only is the purchase of a second AMS avoided (between 40 and 100 k€) but equally important and of the same order of magnitude are the cost and time savings with respect to the Quality Assurance of the second AMS.