The Kalman filter prediction step of traditional dynamic harmonic state estimation usually constructs the state space model with a unit matrix and assumes the system noise covariance matrix as a constant matrix. This reduces the accuracy of the estimation and affects the results of the dynamic state estimation model. In order to establish the spatial model of harmonic state accurately, this paper proposes a time series prediction method based on a long short-term memory network. The complex state transfer process is simulated by off-line training of a large number of historical data, and the harmonic state at the present moment is predicted based on the filtering estimation of historical moments. This effectively improves the accuracy of the prediction model in an unscented Kalman filter. The method in this paper is tested and analyzed on the improved IEEE 34-node three-phase unbalanced system. Compared with the traditional method, the results show that the proposed method performs better in both precision and robustness of harmonic state estimation. This work is supported by the National Natural Science Foundation of China (No. U1966205).