Missing Data Recovery in Large Power Systems Using Network Embedding
Tong Wu, Ying–Jun Angela Zhang, Yang Liu, Wing Cheong Lau, Huanle Xu
Abstract
This paper proposes a novel network-embedding based method to recover the missing measurements in power systems. In particular, we first construct the spatial and temporal graphs to describe both the spatial correlation among the buses in a power flow network and the temporal correlation of the bus states over different time. Secondly, we propose a Softwork algorithm to map the spatial and temporal graphs to low-dimensional spatiotemporal features. Then, we train a regression neural network using the pairs of spatiotemporal features and observed matrix entries. The trained network can then predict the missing measurements. Furthermore, the proposed missing data recovery algorithm can be extended to an online version to recover the missing measurements from streaming data collected in power systems in real time. Numerical experiments on real-world power systems verify the effectiveness of the proposed method. In particular, the proposed method achieves (on average) -55.36 dB, -42.06 dB, -53.26 dB and -45.32 dB relative recovery errors (RREs) for random, row, column and block missing patterns of the voltage magnitude matrix, respectively, which are much smaller than those achieved by the existing methods.