Hybrid data‐driven physics model‐based framework for enhanced cyber‐physical smart grid security

Abstract

This paper presents a hybrid data-driven physics model-based framework for real time monitoring in smart grids. As the power grid transitions to the use of smart grid technology, it's real time monitoring becomes more vulnerable to cyber attacks like false data injections (FDI). Although smart grids cyber-physical security has an extensive scope, this paper focuses on FDI attacks, which are modeled as bad data. State of the art strategies for FDI detection in real time monitoring rely on physics model-based weighted least squares state estimation solution and statistical tests. This strategy is inherently vulnerable by the linear approximation and the companion statistical modeling error, which means it can be exploited by a coordinated FDI attack. In order to enhance the robustness of FDI detection, this paper presents a framework which explores the use of data-driven anomaly detection methods in conjunction with physics model-based bad data detection via data fusion. Multiple anomaly detection methods working at both the system level and distributed local detection level are fused. The fusion takes into consideration the confidence of the various anomaly detection methods to provide the best overall detection results. Validation considers tests on the IEEE 118 bus system.

Ensemble CorrDet with adaptive statistics for bad data detection

“…The ECD algorithm extended the strategies from the CD algorithm and applied it to a dataset that has a drifting load profile in [26]. The drift was modelled by the Ornstein–Uhlenbeck process [44], which is a stochastic process similar to a random walk, but has a tendency to move back towards the original load.…”

“…In previous work [25, 26], the anomaly detection threshold is a fixed value estimated from the initialisation stage and not updated with new incoming data. Specifically, the standard deviation and mean of squared Mahalanobis distance values of initial k normal samples are calculated first.…”

“…These studies are mainly based on neural network, deep learning and fuzzy clustering approaches [21–24]. More recently, physics‐model based solutions have been integrated with data‐driven solutions [25, 26] to take advantage of the temporal characteristics of real‐time data. However, these solutions struggle with the fast‐changing state of power systems.…”

“…As both loads and generation change on a power system, the statistics used by data‐driven solutions can change rapidly. In [26], the authors introduce the Ensemble CorrDet (ECD) algorithm, a hybrid physics‐model based and data‐driven method of FDI detection. However, the ECD algorithm described and leveraged in [26] uses fixed estimated mean and covariance of normal samples after training rather than adapting to the changing state of power systems.…”

“…In [26], the authors introduce the Ensemble CorrDet (ECD) algorithm, a hybrid physics‐model based and data‐driven method of FDI detection. However, the ECD algorithm described and leveraged in [26] uses fixed estimated mean and covariance of normal samples after training rather than adapting to the changing state of power systems. This means that when applied to a realistic load profile for power systems, ECD statistics will quickly become outdated and the performance of the algorithm will decrease.…”