文摘
In this paper, a modified rule generation approach with self-constructing property for neuro-fuzzy system modelling is proposed. In structure identification stage, input–output patterns are divided into the clusters and interval type-2 membership functions are generated roughly. Interval type-2 Takagi–Sugeno–Kang (TSK) neuro-fuzzy structure is fine tuned by quantum inspired bacterial foraging algorithm (QBFA) in parameter identification stage to achieve higher precision, a recursive least squares (RLS) estimator is used to update consequent parameters. Comparisons with two type-1 neuro-fuzzy systems on three nonlinear functions and chaotic Mackey-Glass time series show that the proposed systems can approximate the target with little error. Experiments are also executed involving the proposed systems for modelling flue gas denitrification efficiency of a thermal power plant. It is verified by the results that interval type-2 neuro-fuzzy structure can learn knowledge from input–output data set with the aid of QBFA and hybrid training progresses are able to improve its performance.