Gas Concentration Reconstruction for Coal-Fired Boilers Using Gaussian Process

Abstract

The goal of combustion optimization of a coal-fired boiler is to improve its operating efficiency while reducing emissions at the same time. Being able to take measurements for key combustion ingredients, such as O2, CO, H2O is crucial for the feedback loop needed by this task. One state-of-the-art laser technique, namely, Tunable Diode Laser Absorption Spectroscopy (TDLAS) is able to measure the average value of gas concentration along a laser beam path. A active research direction in TDLAS is how to reconstruct gas concentration images based on these path averages. However, in reality the number of such paths is usually very limited, leading to an extremely under-constrained estimation problem. Another overlooked aspect of the problem is that how can we arrange paths such that the reconstructed image is more accurate? We propose a Bayesian approach based on Gaussian process (GP) to address both image reconstruction and path arrangement problems, simultaneously. Specifically, we use the GP posterior mean as the reconstructed image, and average posterior pixel variance as our objective function to optimize the path arrangement. Our algorithms have been integrated in Siemens SPPA-P3000 control system that provides real-time combustion optimization of boilers around the world.