Fuzzy clustering, genetic algorithms and neuro-fuzzy methods compared for hybrid fuzzy-first principles modeling

Hybrid fuzzy-first principles models can be a good alternative if a complete physical model is difficult to derive. These hybrid models consist of a framework of dynamic mass and energy balances, supplemented by fuzzy submodels describing additional equations, such as mass transformation and transfer rates. Identification of these fuzzy submodels is one of the main issues in constructing hybrid models. In this paper, a new approach to constructing hybrid fuzzy-first principles models is presented, which uses a Kalman filter for parameter estimation. In addition, a comparison between three classes of identification techniques for fuzzy submodels is presented: fuzzy clustering, genetic algorithms and neuro-fuzzy methods. The comparison is illustrated for a penicillin fed batch-reactor test case. Fuzzy clustering proved to be the most suitable technique, with genetic algorithms being a good alternative.