A simple multi-model prediction method

Abstract

The present work introduces a new multi-model state-space formulation called simultaneous multi-linear prediction (SMLP), which is suitable for systems with significant gain variation due to nonlinearity. Standard multi-model formulations usually make use of a partitioned state-space, i.e., a state-space that is divided into regions to shift parameters of the state update equation according to the current location of the state, with a view to having a better approximation of a nonlinear plant on each region. This multi-model framework, also known as linear hybrid systems framework, makes use of different boundaries or partition rules concepts, which vary from systems of linear inequalities, propositional logic rules, or a combination of these. This standard approach inevitably introduces discontinuities in the output prediction as the state update equation parameters shift noticeably. Instead, the SMLP is built by defining and updating multiple states simultaneously, thus eliminating the need for partitioning the state-input space into regions and associating with each region a different state update equation. Each state's contribution to the overall output is obtained according to the relative distance between their identification (or linearisation) point and the current operating point, in addition to a set of parameters obtained through regression analysis. Unlike the methods belonging to the hybrid systems framework, no discontinuities are introduced in the output prediction