The invention discloses a dynamic state estimation method considering noise statistical characteristic unknown, and the method is used for realizing the accurate estimation of a dynamic state variableof a power system generator. In the framework of a traditional unscented Kalman filter (UKF), system noise and measurement noise statistical characteristic estimators based on innovative informationsequences are established, covariance matrixes which are satisfied by the system noise and the measurement noise can be dynamically adjusted and corrected according to environmental changes. A proposed method (adaptive unscented Kalman filter, AUKF) can effectively solve the problem of the state estimation performance degradation caused by the improper setting of a noise matrix in a traditional UKF method, and the state estimation accuracy of a dynamic state estimator is improved. The algorithm takes into account an actual engineering background and is simple and convenient to implement, and the algorithm has a high engineering application value.