Numerical substructures adopt predetermined hysteretic models for conventional substructure testing (CST). It is difficult to achieve accurate results in CST due to errors from modelling of numerial substructures. To diminish the negative effect caused by the assumed restoring model, online identification with least square method using experimental data was adopted to update the model of numerical substructure. The effectiveness of least square method are verified with pseudo-dynamic substructure testing for single story frame with a buckling-restrained brace. The results indicate that the on-line identification algorithm based least square method has good accuracy, stability, and efficiency.