TL;DR
This paper introduces a novel two-stage framework to detect "Fault-Masking Attacks" (FMAs) on Line Current Differential Relays (LCDRs) in smart grids. By combining a physics-based mismatch index with a neural network classifier, the system accurately identifies these stealthy cyberattacks, which aim to hide physical faults, without causing false alarms, thus enhancing grid security.
What Problem Does This Paper Address?
Smart grids rely on advanced communication and control systems, making them vulnerable to cyberattacks. Line Current Differential Relays (LCDRs) are crucial for protecting critical transmission lines due to their speed and sensitivity. However, these relays can be targeted by "Fault-Masking Attacks" (FMAs). In an FMA, an attacker manipulates the remote measurements sent to an LCDR, making it believe the line is healthy even when a physical fault has occurred. This is particularly dangerous because it prevents the fault from being cleared, potentially leading to cascading failures, equipment damage, and instability. Existing detection methods often focus on false tripping attacks (where a healthy line is incorrectly identified as faulty) and are ineffective against FMAs, which intentionally keep the relay from tripping. This paper addresses the critical need for a robust and real-time solution to detect these covert FMAs on LCDRs.
Methods
The proposed solution is a two-stage framework:
- Mismatch Index (MI): This physics-based module leverages the healthy equivalent model of the protected transmission line. It compares the LCDR's local measurements (voltage and current) with the expected values derived from the model, considering the remote current measurement. A significant discrepancy between the measured and calculated values, especially when the LCDR itself is not triggered, signals a potential FMA. This module uses a "Trailing Moving Mean" (M) and a dynamic threshold (Lu) to smooth out minor fluctuations and detect significant deviations.
- Zone-Confirmation Classifier (ZCC): If the MI is triggered, this module, implemented as an Artificial Neural Network (ANN), takes over. The ZCC uses only the local measurements (I₁ and V₁) to confirm whether the detected anomaly is an internal fault on the protected line or an external disturbance. This reliance on local data makes it resilient to attacks targeting remote measurements. The ANN is trained offline on datasets of internal masked faults and external disturbances.
The combined framework aims to achieve high detection accuracy while maintaining a low false positive rate, ensuring the selectivity and security of the protection system.
Datasets
The paper uses two primary datasets for evaluation:
- Comprehensive FMAs Dataset: This dataset was generated by simulating a wide range of faults (single-phase to ground, phase-to-phase, three-phase, and combinations) on the IEEE 39-bus system's critical line (11-6). These faults were simulated at various locations (10% to 90% of the line) and resistances (0.001 Ω to 300 Ω). FMA cases were then generated from these simulated faults using a specific attack model. A total of 5,346 FMAs were generated.
- External Disturbances Dataset: This dataset consists of 5,000 simulated faults on the five lines adjacent to line 11-6, covering different types, locations, and resistances.
Measurements were sampled at 1K sample/sec. Additive white Gaussian noise with 35 dB SNR was also used to model measurement errors.
Key Insights and Ideas
- Stealth of FMAs: The core challenge is that FMAs intentionally prevent the LCDR from tripping, making traditional fault detection methods useless. The proposed MI specifically looks for inconsistencies when the relay is not triggered.
- Physics-Based Detection: The MI's reliance on the physical model of the transmission line provides a strong foundation for detecting deviations caused by manipulated measurements.
- Local Measurement Reliance: The ZCC's use of only local measurements for confirmation is a crucial design choice for resilience against attacks that compromise remote data.
- Two-Stage Approach: Combining a physics-based filter (MI) with a data-driven classifier (ZCC) effectively reduces false alarms, a common issue with single-stage detection systems.
- Real-time Capability: The framework is designed to operate within the tight time constraints of LCDRs, ensuring it can detect faults as quickly as a non-attacked system.
Key Results
- High Detection Accuracy: The complete proposed framework (MI + ZCC) achieved an accuracy of 99.845%, with a precision of 100% and a recall of 99.69%.
- No False Alarms: The combined framework successfully eliminated false alarms, a significant improvement over the MI-only approach (which had a 14.55% false positive rate on outside-zone faults).
- Fast Detection: Detected FMAs occurred within 1.5 power cycles, meeting the real-time requirements.
- Robustness: The framework demonstrated robustness against various system conditions, including different loading levels, CVT and CT transients, measurement noise, and even simultaneous internal faults and joint cyberattacks.
- Real-time Verification: The framework's performance was validated on an OPAL-RT real-time simulator, confirming its practical applicability.
Limitations and Future Work
- Limitations:
- The paper notes that the MI might struggle with unsymmetrical very high impedance faults located close to the far end of the protected line, which are less common.
- While the ZCC is resilient to remote measurement attacks, it relies on the integrity of local measurements.
- Future Work:
- Exploring strategies for detecting and mitigating evolving, adaptive attacker strategies through frequent re-training, continuous model monitoring, and online learning.
- Investigating explainable AI techniques to understand the ANN's decisions better.
- Developing physics-only-based protective relays that can detect both physical faults and cyber FMAs without relying on communication.
Summary updated