Utilizing Neural Networks to minimize NOx-emissions in a Steam Cracking Furnace

Abstract

Emissions of nitrogen oxides (NOx) is detrimental to human health and the environment. The industrial sector is the second largest emitter of NOx emissions in Sweden. NOx is formed through complex processes during combustion, but formation can be reduced through optimizing the combustion process, which often is a cost-effective method to reduce emissions. The combustion process is however highly integrated with the main process in the steam cracking furnace, making the product quality the primary target for the combustion optimization. NOx formation is complex and difficult to predict. Reliable methods for predicting NOx emissions depending on operational modes for the combustion chamber is therefore vital to minimize NOx emissions. One method is to use artificial neural networks to create models for NOx formation depending on the process control variables. Artificial neural networks is a type of machine learning that are created to predict the outputs of a specific process using data collected for the process in question. This thesis presents a method for find operational modes of a cracking furnace that would reduce NOx while not affecting the main purpose of the furnace i.e. maintaining the product quality. The furnace in question is a steam cracking furnace operated by Borealis AG in Stenungssund. The process of finding optimal operational modes where started with collecting real-time data for all parameters that was available and identified as important, followed by data cleaning. Multiple networks were trained to account for three parameters of interest, i.e. emissions of NOx and carbon monoxide (CO), as well as temperature of cracking. The trained networks were then incorporated into a genetic algorithm to find operational modes that would minimize NOx emissions while keeping CO low and the temperature of cracking within an acceptable range. Two operational modes where proposed, which would reduce emissions by 10% and 50% compared to current levels, respectively. The operational mode with 50 % reduction carries the risk of leaving the furnace too cold and should be further tested before implemented. Using real-time data for training neural networks to predict NOx proved comparable to other methods, with a correlation coefficient of 0.927. Using neural networks to predict CO emissions and temperature of cracking was not as successful. Especially the networks for CO predictions showed poor performance.