A neurofuzzy scheme is proposed to perform an on-line identification of non-linear systems that can be represented by a transfer function with varying parameters. The parameter variation case due to one external (measurable at each sample interval) variable has been studied. The proposed scheme is composed of two blocks. The first one involves a fuzzy partition of the external variable universe of discourse. This partition is used to smoothly commute between several linear models. In the second block, a recurrent linear neuron with interpretable weights performs the identification of the models by means of supervised learning. The resulting identifier has two main advantages: interpretability, because the weights of the neuron can be assimilated to coefficients of transfer functions, and learning speed, due to the local behavior imposed by the fuzzy partition. The proposed scheme has been tested on a real laboratory plant as an on-line identifier on an adaptive predictive control structure, showing a good performance.