Comparison of supervised learning techniques for atmospheric pollutant monitoring in a Kraft pulp mill

In this paper, supervised learning techniques are compared to predict nitrogen oxide (NOx) pollutant emission from the recovery boiler of a Kraft pulp mill. Starting from a large database of raw process data related to a Kraft recovery boiler, we consider a regression problem in which we are trying to predict the value of a continuous variable. Generalization is done on the worst case configuration possible to make sure the model is adequate: the training period concerns stationary operations while test periods mainly focus on NOx emissions during transient operations. This comparison involves neural network techniques (i.e., multilayer perceptron and NARX network), tree-based methods and multiple linear regression. We illustrate the potential of a dynamic neural approach compared to the others in this task.