Artificial neural network-based estimation of mercury speciation in combustion flue gases

• Published in: Fuel processing technology 2004-06, Vol.85 (6), p.451-462
• Main Authors: Jensen, Robert R, Karki, Shankar, Salehfar, Hossein
• Format: Article
• Language: English
• Subjects: Air pollution caused by fuel industries; Applied sciences; Energy; Energy. Thermal use of fuels; Exact sciences and technology; Metering. Control; Multilayer perceptron; Neural network correlation; Speciated mercury emissions
• ISSN: 0378-3820; 1873-7188
• DOI: 10.1016/j.fuproc.2003.11.020

Artificial neural networks (ANN) are powerful tools that can be used to model and investigate various highly complex and nonlinear phenomena. In this study, neural networks (NN) were used to investigate the complex relationships between mercury speciation emissions associated with coal-fired utility boilers and their air pollution control systems. The relationships between mercury speciation (elemental, oxidized and particulate) in the flue gas leaving the boiler and coal characteristics and boiler type have been examined using several different neural network architectures. A multilayer perceptron (MLP) form of the neural networks was identified to be the structure that can accurately capture and model the nonlinear relationship that exists between mercury speciation and coal characteristics and boiler type. Data from 76 power plants across the country have been used to train and develop the neural network models. After a number of iterative trainings of the model, the prediction and estimation results from the neural network turned out to be quite good. A comparison of the estimation of the mercury emissions from selected utility boilers with the observed values shows a high accuracy (i.e., a correlation coefficient of 0.9750). Neural network-based, three-dimensional graphical representations of the relationships between predictive and explanatory variables turn out to be quite nonlinear and complex. These three-dimensional graphs reveal valuable information that is difficult (if not impossible) to obtain from standard regression analyses.