Aiming at the low accuracy of harmonic disturbance detection in power system, the residual analysis using Kalman Filter based on the Maximum Likelihood (KF-ML) is proposed in this paper. In this proposed method, the optimal error covariance matrices and the initial condition parameters are adaptively optimized by using the maximum likelihood method. Thus, it can realize the prediction and correction of the harmonic amplitude, which overcomes the low precision and easy divergence of the traditional Kalman filter. The estimated residuals, exhibiting mutation at the starting and the ending of disturbances, are proposed to detect the disturbances. Simulation results show that the method can accurately detect the amplitude of the dynamic and static harmonic signals. Also, the residuals can exactly determine the starting and ending instants of the disturbances. This work is supported by National Natural Science Foundation of China (No. 51507015 and No. 61673388), Natural Science Foundation of Hunan Province (No. 2018JJ2439), and Excellent Youth Fund of Hunan Education Department (No. 18B130).