Model-based Detection of Data-Injection Cyber-Attacks on Wind Turbine Controllers

Álvaro Martín Gómez, León Navarro-Hilfiker, Rafal WisniewskiCoDIT 2024 · DOI

Abstract

As wind energy becomes a cornerstone of the global power grid, wind turbines have emerged as high-value targets for cyber-attacks. This paper presents a robust model-based framework for detecting "stealthy" False Data Injection Attacks (FDIA) on wind turbine grid-side controllers. By integrating high-fidelity electrical modeling with an Unscented Kalman Filter (UKF), the proposed system acts as a digital twin to identify subtle manipulations that traditional network-level security often misses. Through statistical evaluations using CUSUM and GLR algorithms, the study demonstrates rapid detection of sensor corruption, setpoint attacks, and gain manipulations before they lead to physical damage or grid instability.

Introduction

Traditional cybersecurity measures, such as firewalls and Intrusion Detection Systems (IDS), primarily focus on network-level protection. However, these methods are often ill-equipped to detect False Data Injection Attacks (FDIA), where an attacker subtly manipulates sensor data, setpoints, or control signals. These manipulations can be engineered to resemble normal system noise or minor transients, gradually pushing the turbine toward mechanical failure or instability without triggering network alarms.

This research addresses the critical need for a physics-aware detection benchmark. By leveraging the turbine's own physical and electrical dynamics, the proposed architecture can "sniff out" anomalies that deviate from expected behavior, providing a second line of defense at the control layer.

Proposed Methodology

The authors propose a detection architecture that treats the wind turbine as a dynamic system rather than a mere network node. The framework consists of three primary components:

  1. High-Fidelity Modeling: A detailed non-linear state-space model of the wind turbine's electrical subsystem was developed, incorporating the Phase Locked Loop (PLL), DC-link voltage control, and the LC filter.
  2. Unscented Kalman Filter (UKF): The UKF serves as a "digital twin," predicting real-time system behavior. By comparing these predictions against actual measurements, the system generates residuals—the delta between expected and observed states. The UKF was specifically selected over standard Kalman Filters for its superior ability to handle the non-linearities inherent in power electronics, thereby minimizing false alarms during normal transients.
  3. Statistical Evaluation: To convert residuals into actionable alerts, the framework utilizes two statistical tests: CUSUM (Cumulative Sum), which is optimized for detecting small, persistent drifts, and GLR (Generalised Likelihood Ratio), which is highly sensitive to abrupt changes.
Detection architecture combining physical modeling with UKF-based residual generation.

Experimental Setup and Datasets

The study utilizes a high-fidelity industrial simulation model of a wind turbine converter system. This model is grounded in real-world industrial settings used for multi-frequency stability analysis, ensuring that the synthetic data used for evaluation accurately reflects the complexities and noise profiles of actual hardware.

Analysis and Results

The Importance of Non-Linearity

A significant finding of the study is the necessity of non-linear observers. As shown in the comparison between a standard Kalman Filter (KF) and the UKF, linear models often produce discrepancies during normal system transients. These discrepancies can be easily mistaken for cyber-attacks, leading to high false-positive rates. The UKF’s ability to accurately track non-linear dynamics is essential for reliable operation in a real-world power plant environment.

Comparison of UKF and KF performance during system transients.

Detection Performance

  • Setpoint Attacks: Unauthorized changes in AC and DC voltage setpoints (0.1 pu) were detected within 0.38 to 1.10 seconds.
  • Sensor Corruption: Detection times for sensor measurement attacks ranged from nearly instantaneous (0.01s) for abrupt changes to several seconds for highly stealthy, filtered injections.
  • Gain Manipulation: The detector successfully identified attacks on controller gains (e.g., $K_{pq}$, $K_{pc}$) before the system reached instability. For instance, an attack on the PLL gain was flagged in 2.07 seconds.

Stealthiness and Visibility

The research highlights a critical "stealthiness" trade-off: as attackers make their injections more gradual (using low-pass filters), the time required for detection increases. Furthermore, while the detector is most effective when the attacked variable is a direct input to the observer, it can still detect attacks on internal control signals if they are large enough to cause secondary ripple effects in measured states.

Conclusion

Model-based detection provides a vital layer of security for modern wind energy systems. By utilizing the physical laws governing the turbine, operators can detect sophisticated FDIAs that bypass traditional IT security. Future work will focus on improving the localization of attacks—identifying exactly which sensor or signal is compromised—and exploring "residual profiling" to assist operators in rapid incident response.

Summary updated

BibTeX
@inproceedings{gomez2024model,
  author = {Álvaro Martín Gómez and León Navarro-Hilfiker and Rafal Wisniewski},
  title = {{Model-based Detection of Data-Injection Cyber-Attacks on Wind Turbine Controllers}},
  booktitle = {CoDIT},
  year = {2024},
  doi = {10.1109/codit62066.2024.10708323},
}
Figure