In this paper, a new multi-model approach is proposed for identification of nonlinear systems. In similar identification methods, the operating space is partitioned and a local model is suggested for each partition. In such approaches, since the same linear structure is often used for all local models; huge number of local linear models is usually required to reasonably model an operating region with severely nonlinear dynamics. Therefore the size of the global model may exponentially increase; and as a result model robustness may decrease. In the proposed approach the best model structure is selected for the particular nonlinear study system in an iterative approach. At each iteration, a choice is made to increase number of local models and/or increase the local model complexity. Furthermore, it determines the complexity of local models based on increasing the model accuracy and ensuring the model robustness. In order to optimize the model approximation capability and model robustness, a model term selection approach based on a forward orthogonal least squares algorithm and a criterion that minimizes the sum of the variance of the parameter estimates is applied. Simulation results show that the proposed method results in an excellent validation performance with fewer parameters.