Abstract: The NARMAX model identification for nonlinear dynamical systems much depends on the correction of the system structure and the precision of the parameter estimate. The algorithms existed at present for determining the model's structure are all based on the \ selecting rule, by which the spare terms can easily be firstly introduced with incorrectness, resulting the true system structure can never be obtained. Furthermore, the numerical stability of these algorithms is also unsatisfied, often resulting an unacceptable precision for the parameter estimation. In this paper, the following modification work has been done: a new selecting rule (\ rule) has been proposed to judge the importance of each term compared to all the other terms, which has been selected by the forward process, deleting spare terms and obtaining the true optimal structure for the nonlinear system. Meanwhile, the modified identification algorithm overcomes the storage difficulty resulted by the number explosion of the terms of the assumed NARMAX whole model while applying the traditional MGS algorithm, avoids the backward substitution calculating process, improving to a great extent the numerical stability of the algorithm, and thus obtaining a high precision for the parameter estimation. Theoretical analysis and the simulation results indicate its superiority and effectiveness.