Application of Joint Time-Frequency Methods and Artificial Neural Networks for Monitoring Co-Firing Burners

Abstract

Conventional coal-red burners are designed to operate within specific limits that, in part, result from the need to efficiently burn the fuel, ensure stable combustion and result in the lowest emissions. However, recent requirements to reduce CO2 emissions from coal-red boiler plant as part of the drive to
reduce the carbon footprint of energy suppliers has focussed on the co-ring of biomass, primarily wood, either by delivering the pulverised biomass with the coal or through separate burners. Typically this approach has taken place at substitution levels of around 10% or less by mass and at these levels the operation of the burner and boiler are not adversely affected. However, as the proportion of biomass is increased in a boiler designed for coal, the fuel characteristics of the blend moves further away from the burner design parameters. This can lead to combustion instabilities and in extreme cases extinction of the flame. In order to co-fire higher concentrations of biomass a system or technique is required that can detect the onset of these instabilities and warn before the combustion conditions become dangerous.

This research presents an investigation of a system that monitored the combustion flame using photodiodes with responses in the Ultra Violet (UV), Infrared (IR) and Visible (VIS) bands. The collected data was then processed using the Wigner-Ville Distribution (WVD) joint time-frequency method and subsequently classified using a Self-organising Map (SOM). It was found that it was possible to relate the classification of the sensor data to operational parameters such as the burner airflow rate, CO and NOx emissions. Then using a simple rule based approach the developed system was successfully tested at pilot scale (500kWth) where the ability of the system to optimise the combustion for a variety of unseen coal/biomass blends was demonstrated.

With wide range of operating conditions and the inherent complex nature of coal combustion, a set of stable conditions were recorded to be processed and train the developed system. Such a trained system was subsequently used for monitoring and was able to classify the conditions accurately. The system was
also able to relate the sensor data to varying combustion conditions and hence suggest changes to bring the combustion back to a desired state. This system is capable of making such predictions correctly even with some variation in fuel
properties and has been demonstrated to do so in both pilot and full scale power plants.

Date of Award	Nov 2016
Original language	English
Supervisor	Steven Wilcox (Supervisor), Zyh Chong (Supervisor), CK Tan (Supervisor) & Shee Meng Thai (Supervisor)