Synchrophasor Data Reliability Monitoring in Power System Operational Mode Analysis

(WAMS) and provide an unprecedented level of observability of operating parameters in large power systems. As the need to minimize the time required to obtain analysis results grows, the scale of the tasks increases, as does the volume of synchrophasor data that must be processed at high speed. Under these conditions, the issue of maintaining high “data quality” in the context of data validation and functional correctness becomes especially important, since the increasing complexity of the measurement infrastructure amplifies the influence of factors that may occasionally produce faulty or anomalous data in different locations. Even short, unfiltered anomalies can distort analysis outcomes. This paper illustrates this effect through examples demonstrating how individual outliers in raw signals can affect the accuracy of identifying low-frequency oscillatory-mode parameters. Thus, “data quality” monitoring emerges as a separate and necessary stage of the analysis.

Recent years have seen significant progress in applying machine-learning methods across many fields. One such application is time-series forecasting. Various neural-network models are capable of learning complex patterns in data by accounting for the dependence of each new measurement sample on sequences of previous samples obtained both from the same device and from related devices. In this work, forecasting using recurrent neural networks is investigated as a means of detecting anomalous fragments in PMU output signals. To this end, a series of computational experiments was performed in which hidden artifacts were injected into the signals. These artifacts preserve the general statistical and harmonic properties of the data within an analysis window but do not reflect the physical nature of the signals, the power-system model, or related characteristics. Such artifacts cannot be detected as outliers in the time or frequency domains or through statistical-distribution analysis. A pretrained neural network was used to forecast both the original and artifact-containing signals. By comparing the forecasting results, we conclude that it is possible to build a tool for detecting hidden artifacts in the data using relatively simple and computationally efficient neural-network models. The experiments used real WAMS recordings from the Unified Power System of Russia.