An AI-Based Supervisory Measurement Integrity Validation Layer for Cyber-Resilient AC/DC Protection in Inverter-Based Microgrids

Technical Summary: An AI-Based Supervisory Measurement Integrity Validation Layer for Cyber-Resilient AC/DC Protection in Inverter-Based Microgrids

1. Problem Statement

Modern power systems, particularly inverter-based microgrids, are increasingly reliant on digitally communicated, time-synchronized measurements. While this enables high-speed protection via Line Current Differential Relays (LCDRs), it introduces a critical vulnerability to False-Data Injection Attacks (FDIAs).

In these attacks, adversaries manipulate remote current measurement streams (magnitude or phase) to satisfy the relay's tripping criteria without a physical fault occurring. This leads to unnecessary false trips, which can disrupt power flow and potentially trigger system-wide instability or collapse. Traditional detection methods struggle in inverter-based systems because fault currents are actively limited by converter controls, meaning the large magnitude changes typically used to distinguish faults from attacks are absent.

2. Methodology

The authors propose a Measurement Integrity Validation Scheme (MIVS) that acts as a "supervisory instrumentation layer" sitting above the existing LCDR logic.

• Core Approach: Instead of modifying the protection logic or adding hardware, the MIVS intercepts the short windows of synchronized instantaneous current measurements (both local and remote) recorded during a relay triggering event.
• AI Architecture: The scheme utilizes a Recurrent Neural Network (RNN). Unlike Convolutional Neural Networks (CNNs) or standard Multi-Layer Perceptrons (MLPs), RNNs are specifically chosen for their ability to learn temporal correlations and the hierarchical structure of time-series data. This allows the model to distinguish the subtle physical consistency of a genuine fault from the mathematically crafted but physically inconsistent trajectories of an FDIA.
• Data Modality: The model operates directly on raw, sampled-value time-domain waveforms (multi-phase for AC; multi-pole for DC). This eliminates the need for hand-crafted feature extraction (like phasor estimation), which reduces computational latency.
• Operational Flow: When an LCDR is triggered (by either a fault or an attack), the MIVS analyzes the preceding and succeeding milliseconds of current data. It only permits the trip command if it classifies the event as a "genuine fault."

3. Key Contributions

• Universal Applicability: A single framework applicable to both AC and DC LCDRs without requiring structural modifications to the relay or the network.
• Hardware-Agnostic & Model-Free: The method requires no additional sensors, auxiliary protection elements (like reactors or capacitors), or prior knowledge of the network topology.
• Temporal-Structure Exploitation: It shifts the detection paradigm from magnitude-based detection (which fails in low-inertia inverter systems) to temporal-consistency detection.
• Scalability: The authors demonstrate that a single MIVS can be trained for an entire class of relays (e.g., one model for all AC LCDRs in a microgrid) rather than requiring individual training for every single relay.

4. Results and Validation

The proposed scheme was validated through extensive PSCAD/EMTDC simulations and real-time Hardware-in-the-Loop (HIL) testing using an OPAL-RT simulator.

• Detection Accuracy: The MIVS achieved high accuracy across both AC and DC systems. For the AC LCDR, it blocked 98.3% of FDIAs while maintaining a 99.5% dependability rate (correctly validating genuine faults). For the DC LCDR, it achieved 99.17% FDIA detection and 100% fault validation.
• Robustness: The system remained highly effective under measurement noise (40 dB SNR), with accuracy only marginally degrading.
• Real-Time Feasibility: HIL testing confirmed a detection latency of approximately 1.2 ms. This is well within the operational time limits of modern digital relays, ensuring that the security layer does not compromise the speed of protection.
• Comparative Performance: The MIVS outperformed existing machine-learning approaches (such as those based on current magnitudes or phase angles) which are less effective in the low-fault-current environments of inverter-based microgrids.

5. Significance

This research provides a critical layer of cyber-physical security for the next generation of resilient energy infrastructures. By treating measurement integrity as a temporal pattern-recognition problem rather than a simple threshold problem, the authors offer a practical, low-cost, and high-speed solution to protect microgrids from sophisticated cyber-adversaries. It bridges the gap between cybersecurity and power system protection, ensuring that the transition to inverter-dominated grids does not come at the cost of increased vulnerability to digital attacks.