Reliable engine and emission models allow for an online monitoring of commercial gas turbine operation and help the plant operator and the original equipment manufacturer (OEM) to ensure emission compliance of the aging engine. However, model development and validation require fine-tuning on the particular engines, which may differ in a fleet of a single design type by production, assembly and aging status.
For this purpose, Artificial Neural Networks (ANN) offer a good and fast alternative to traditional physically-based engine modeling, because the model creation and adaption is merely an automatized process in commercially available software environments. However, ANN performance depends strongly on the availability of suitable data and a-priori data processing. The present work investigates the impact of specific engine information from the OEM’s design tools on ANN performance. As an alternative to a strictly data-based benchmark approach, engine characteristics were incorporated into ANNs by a pre-processing of the raw measurements with a simplified engine model. The resulting ‘virtual’ measurements, i.e. hot gas temperatures, then served as inputs to ANN training and application during long-term gas turbine operation. When processed input parameters were used for ANNs, overall long-term NOx prediction improved by 55%, and CO prediction by 16% in terms of RMSE, yielding comparable overall RMSE values to the physically-based model.