Predicting Gas Turbine NOx Emissions With Machine Learning

Abstract Gas Turbine operations result in production of Nitrogen Oxides (NOx) and Carbon Monoxide (CO), both of which are pollutants that are to be strictly monitored and controlled. To control NOx within permissible limits, Combustion Emissions Monitoring System (CEMS) readings are to be continuously monitored and flagged for any anomaly in NOx emissions. This paper covers the method for developing a proof of concept of a Machine Learning (ML) model that predicts NOx from a selected pair of General Electric 7FA (7FA) and Siemens Westinghouse 501F (SW501F) gas turbines that are actively monitored in real time. The Keras library and the open-source machine learning framework TensorFlow are used for constructing, developing, testing, and tuning all respective models in Python. The authors show that working with quantitative, time-series data, a deep neural network (DNN) using a rectified linear unit activation function is deemed the most appropriate for the given issue at hand. Using this framework in conjunction with the domain knowledge from a monitoring and diagnostic team and gas turbine experts, a DNN for each frame type is successfully constructed. The 7FA test site and SW501F test site produced NOx predictions with a mean absolute error of 0.55 and 0.7 ppm respectively, hence demonstrating the capability of a machine learning model to accurately predict NOx emission for any typically trained scenario.