Nox emission estimation in gas turbines via interpretable neural network observer with adjustable intermediate layer considering ambient and boundary conditions

Abstract

Nitrogen Oxides (NOx) emission is an inevitable concomitant process of gas turbines operating in industrial scenes. Most of the existing NOx estimation methods have characteristics of strong mechanism dependence, low estimation accuracy and long estimation time, moreover, the parameters used for estimation are not easy to be measured. This work designs a novel neural network model with an adjustable intermediate layer to estimate NOx by taking the ambient and boundary parameters as model inputs to promote engineering applications. The easily measurable outlet temperature of the compressor is selected as the adjustable intermediate parameters to improve estimation accuracy. The proposed model is trained and verified by historical data collected from the various operating conditions of gas turbines. The influences of weight coefficient combinations, relative humidity and intermediate parameter on the NOx estimation accuracy are further analyzed and discussed. The results show that the proposed model has higher estimation accuracy and convincing estimation results compared with the traditional estimation method. Adjusting the intermediate layer parameter can further improve the estimation performance. The estimation accuracy is improved by about 2.23%, and estimation error is reduced by half, which is more suitable for estimating NOx emission of gas turbines in industrial scenarios with smaller sample sizes.