A Sparse Robust Model for a Linz-Donawitz Steel Converter

Steelmaking with a Linz-Donawitz converter is a typical example of a complex industrial process where due to the lack of exact mathematical (physical-chemical) models a construction of a black-box behavioral model is required, based on noisy and imprecise data. To construct a good model, a large number of such input-output samples should be used, which calls for a method that is sparse, in a sense that the resulting model complexity is independent of the sample number; and robust to reduce the effects of noise. Lately kernel based methods, like SVMs have been successfully applied to a number of such problems. The main problem with the traditional SVM is its high algorithmic complexity which makes it infeasible for really large databases. LS-SVM solves this problem, but the resulting model is not sparse. Our solution uses a sparse and robust extension of LS-SVM, which leads to good results compared to other methods (such as MLPs) applied to the same problem.