Improved dynamic state estimation of power system using unscented Kalman filter with more accurate prediction model

Abstract

Power system dynamic state estimation plays an important role. However, rapid changes in states cause state estimation to become very hard. To reduce the residual between pseudo and real measurement, prediction models are adopted, which are strongly associated with the convergence rates and accuracies of estimation methods. In this paper, to improve the estimation accuracy, a prediction model that consists of the convolutional neural network and long short-term memory (CNN-LSTM) is employed and then integrated into the unscented Kalman filter (UKF). In the proposed UKF with CNN-LSTM, state vectors are considered as time-series data, so CNN performs feature extraction for data pre-processing first, and then the features go through LSTM to improve its forecast accuracy in real-time. Next, online training and error normalization are introduced to UKF, which increases the estimation accuracy effectively. Finally, simulations are carried out on the IEEE 33-bus system. Simulation results show that the accuracies of the CNN-LSTM prediction model are much higher than those of conventional methods. Furthermore, compared to widely used state estimation methods, our method decreases RMSE and MAPE by about 2 multiples.