Pre-processing of Partition Data for Enhancement of LOLIMOT

Abstract

The Local Linear Model Tree (LOLIMOT) algorithm is a versatile tool for black-box identification of nonlinear complex systems with a set of local linear models. In this work two methods for pre-processing of the partition data for this algorithm are presented. These methods aim at reducing the number of LLMs while improving the global model fit. The proposed methods are a (linear or nonlinear) principal component analysis and a rotational transformation of the input space. Both methods aim at mitigating the limitations of the axis-orthogonal splits in the partition space that LOLIMOT performs. The application to real data from industrial processes and the efectiveness is demonstrated on a grate-fired biomass plant and the thermal model of a large office building.