Abstract
Local model networks represent a complex nonlinear dynamical system by a weighted sum of locally valid, simpler sub-models denned over small regimes of the operating space. Training such networks requires the determination of the appropriate regimes and the local model parameters. This paper compares a hybrid training algorithm, which combines nonlinear structural optimisation and linear parameter estimation, with a tree construction approach which recursively determines the best structure. Rather than optimising for one-step-ahead prediction, the parallel model prediction error is minimised in each modelling approach, producing good generalisation from the identified local model networks. The modelling performances are evaluated using practical, noisy data from a pilot plant of a pH neutralization process. Results show comparable prediction performance but the construction algorithm requires considerably less computational effort and initial knowledge.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have