Learning Voltage Stability with Missing Data Using Phasor Measurements

Abstract

Recently, much effort has been taken to apply learning algorithms for voltage stability in power systems. Even though these algorithms obtained remarkable performance, an evident disadvantage is that they generally depend on a fixed-length inputting data, where even slight data loss will cause them completely invalid. To overcome this shortcoming, this paper proposes an adaptive neural network based approach which is tolerant for missing data. In this approach, the temporal measurement sequence from each PMU is treated equally by the identical multi-layer neural network to generate abstract representations. Then multiple abstract representations of different PMUs are aggregated by a symmetrical function, followed by an output block to conduct either regression or classification tasks of voltage stability. Massive experiments using different testing systems show that our method is almost unaffected by missing data as well as maintains a comparable accuracy with previous learning algorithms.