Input Variable Selection in Modelling of Desulphurization Efficiency

Several methods are applied to find the input variables with predictive power to the degree of desulphurization modelling. The methods are applied on the data from a desulphurization plant processing flue gases coming from a coal-fired power plant. In non-linear and complex industrial processes, the nature of the relationships between the variables may be vague and a functional model based on a physical interpretation of the process may be difficult to define. Data-driven statistical modelling approaches are, therefore, reasonable alternatives. However, such models may become corrupted due to the inclusion of uninformative, weakly informative or redundant variables. Linear correlation coefficients, principal component analysis and regression, partial least squares regression, mutual information based algorithms and the general regression neural network are tested in the selection of the informative variables. The results obtained are relevant to desulphurization plant monitoring development.