Decentralized Modular Nonlinear Physics-Informed Neural Network (mnPINN) for Synchrophasor Data Anomaly Detection

Abstract

Digital automation and advanced sensors providing high resolution measurements are enabling reliable and efficient operation of the electric grid. Inaccurate measurements caused by anomalies can deteriorate the performance of grid operation. It is critical to detect these anomalies in the sensor measurement and flag or replace them to maintain the data integrity. The source of anomalies may include sensor failures, communication failures, firmware problems, database corruption, software bugs, and cyber intrusions. Given large amount of sensor data, decentralized approaches reduce the burden of data transfer for long distances and may run faster on edge devices. Relying solely on data-driven approaches with no system context may lead to inaccuracies in the anomaly detection results. This can be significantly improved by exploiting the knowledge of the underlying physic of the system. In this paper, we have proposed a decentralized approach involving overlapping Physics Informed Neural Networks (PINNs) covering different key components of the power system. Detailed generator dynamics, network power flow, load models, solar cells, and wind turbines are implemented in the PINN along with a deep learning layer to complement known dynamics with supplemental data driven computations. Both linear and nonlinear models of generator dynamics are implemented in modular nonlinear PINNs (mnPINNs) for approximating different generators as Single Machine Infinite Bus (SMIB) models with varying details. The performance of the mnPINN is evaluated using specific metrics for changing levels of anomalies in the presence of physical events like load change, and faults. Results demonstrate the superior performance of the proposed mnPINNs.