Adaptive GRNN for the modelling of dynamic plants

An integrated General Regression Neural Network (GRNN) adaptation scheme for dynamic plant modelling is proposed in this paper. It possesses several improved features compared to the original GRNN proposed by Specht (1991), such as flexible pattern nodes add-in and delete-off mechanism, dynamic initial sigma assignment using a nonstatistical method, automatic target adjustment and sigma tuning. These adaptation strategies are formulated based on the inherent advantageous features found in GRNN, such as highly localised pattern nodes, good interpolation capability, instantaneous learning. Good modelling performance is obtained when the GRNN is tested on a linear plant in a noisy environment. It performs better than the well-known extended recursive least squares identification algorithm. Analysis on the effects of some of the adaptation parameters involving a nonlinear plant is also investigated. The results show that the proposed methodology is computationally efficient and exhibits several attractive features such as fast learning, flexible network sizing and good robustness, which are suitable for the construction of estimators or predictors for many model-based adaptive control strategies.