The invention discloses an improved MRAS control method based on a limited memory least square method. A limited memory least square method identification module is configured in front of an adjustable model of a traditional MRAS; online identification is carried out on PMSM nonlinear parameters, the characteristics of old data elimination and updating of a limited memory least square method identification algorithm are utilized, the defects of data saturation and identification result direction deviation in the identification process are effectively avoided, and the identification accuracy and rapidity are improved. And the stator resistance, the stator inductance and the rotor flux linkage of the PMSM obtained by identification are input into the MRAS adjustable model in real time, and the nonlinear quantity of the adjustable model is updated in real time, so that the stator current vector error is more accurate, and the accuracy of self-adaptive rotating speed estimation is furtherimproved. The method has the advantages of being high in control accuracy, good in dynamic performance, high in robustness, low in cost, simple in control algorithm, high in estimation speed and precision of the rotating speed and the position and the like.