The exciting Bouc-Wen model with fixed-parameters can not characterize the frequency-dependent and time-varying properties from the hysteresis of piezoelectric actuators and easy to generate simulation errors. In order to accurately describe these characteristics, the Bouc-Wen model was established and a recursive least square online identification method was proposed to identify the parameters of the Bouc-Wen model in real-time. Meanwhile, the limited memory method was used to limit the data sets to avoid the data saturation phenomenon. To verify the availability of the identification method, an experimental system was set up and the performance of the identification method was experimentally verified. Experimental results show that the limited memory recursive least square identification method makes the Bouc-Wen model show the frequency shift and time-varying characteristics. When the drive voltage is set to be 100 Hz, the largest absolute simulation error decreases from 1.38 μm to 0.51 μm, and reduced by 63.7%. Compared with the traditional off line parameter identification, the online identification effectively improves the modeling accuracy of the Bouc-Wen model.