Desaturated Probability Integral Transform for Normalizing Power System Measurements in Data-Driven Manipulation Detection
Jingyu Wang, Sheng Su, Yinhong Li, Jinfu Chen, Dongyuan Shi · 2019
Data-driven solutions to detecting power system measurement manipulation rely on thorough understandings of the patterns behind the data. From a probabilistic viewpoint, power system measurements inherently follow multimodal distributions usually with heavy tails. Existing normalization methods, both linear and nonlinear, suffer from some drawbacks when being applied to this kind of distribution. Using inappropriately normalized data to train the detectors may undermine their pattern recognition capability. In this paper, a tradeoff between the linear and nonlinear normalization is investigated to propose a variant of the widely used probability integral transform (PIT) by introducing an approximate linear compensation. When being employed to normalize power system measurements, the socalled desaturated PIT can have a better fidelity towards the measurements collected under infrequent operation conditions. Experiments show that the proposed normalization method is promising in improving the detection performance.