Local Model Networks for nonlinear system identification

Local Model Networks represent a nonlinear dynamical system by a set of locally valid submodels across the operating range. Training such feedforward structures involves the combined estimation of the submodel parameters and those of the interpolation functions. The paper describes a new hybrid learning approach for local model networks that uses a combination of singular value decomposition and second order gradient optimization. A new nonlinear Internal Model Control scheme is proposed which has the important property that the controller can be derived analytically. Simulation studies of a pH neutralization process confirm the excellent modelling and control performance using the local model approach.