Optimization on pollutants emission control with developing machine-learning algorithms for advanced statistical performance monitoring

Abstract

Nitrogen Oxides (NOx) emission from power generation is one of the major air pollution sources in Hong Kong. In support of the Government’s environmental policy and the transition from coal-fired to gas-fired generation, CLP Power Hong Kong Limited (CLP Power) has increased the proportion of natural gas, which is a relatively clean fossil fuel, to around 50 percent in its fuel mix from 2020 and met the stringent emission cap set by the Government and achieved substantial decreases in air emissions compared with the 2019 levels. Current CLP Power NOx emission forecast methodology for the power generation machines adopted calculating the planned annual fuel consumption and the NOx emission factor. The NOx annual emission can always be achieved within the targeted cap. However, the emission gap existed between the actual and the year-end projection amount which the difference affected by the operating factors such as actual load demand and fuel consumption. This study paper is to develop a NOx emission forecast model based on the power generation operating parameters and are specific for the gas-fired Black Point Power Station (BPPS) of CLP Power. The machine learning model eXtreme Gradient Boosting (XGBoost) is adopted to predict NOx emission performance of gas-fired generation unit C1 at Black Point Power Station, which the model is trained with historical operating parameters of high correlation. The model examined the relationship between NOx emission and gas turbine contribution factors possibilities as the pathway forward. The model showed that inclusion of gas turbine operating parameters such as Gas Control Valve, Fuel Gas Temperature, Natural Gas Flow, Inlet Guide Vane Angle and Relative Humidity helped improving accuracy in NOx emission performance, but the study recognized that incorporation of those operating parameters has not been incorporated into the current forecast methodology used in CLP because of the actual load demand and fuel consumption challenges in incorporating those operating parameters into the existing practice. The translation of the model into part of a value chain in fuel cost and in emission reduction remains and outstanding issue in future work.

Date of Award	2024
Original language	English
Awarding Institution	The Hong Kong University of Science and Technology
Supervisor	David Chuen Chun LAM (Supervisor) & Ki On Ng (Supervisor)