⚡ electrical engineering

Reliability-Gated GRU-Based Secure Control of a Nonlinear Floating Offshore Wind Turbine Under False Data Injection Attacks

This paper proposes a reliability-gated GRU-based secure control framework that effectively detects false data injection attacks and compensates for corrupted sensor measurements in nonlinear floating offshore wind turbines, significantly improving rotor speed and power regulation accuracy while maintaining platform stability.

Original authors: Nadir Abbas, Li-Ying Hao, Majid Ayoubi

Published 2026-09-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nadir Abbas, Li-Ying Hao, Majid Ayoubi

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Offshore wind farms are among the most promising sources of clean energy, capable of harnessing powerful, consistent winds far out at sea. However, the machines that capture this energy are complex. A floating wind turbine does not sit on a solid foundation; it floats on a platform that sways and tilts with the waves. To keep the turbine spinning at the right speed and generating steady power, the control system relies entirely on a constant stream of information from sensors. These sensors measure how fast the blades are turning, how much electricity is being made, and how the floating platform is moving. If the control system receives false information, it can make dangerous mistakes, much like a driver trying to steer a car while looking at a broken speedometer. As these systems become more connected to the internet and digital networks, they face a new kind of threat: cyberattacks where an intruder secretly alters the sensor data to trick the machine.

Researchers have developed a new method to protect these floating turbines from such deception. In a study focused on a nonlinear model of a floating offshore wind turbine, the team created a security system that acts like a smart filter for the machine's senses. The system uses a type of artificial intelligence known as a gated recurrent unit, or GRU, which is a computer program trained to understand the normal patterns of how a wind turbine behaves. By learning from hours of safe, uncorrupted data, this program can predict what the sensor readings should be at any given moment based on the wind speed, wave conditions, and the commands the turbine is already following. When the actual sensor data arrives, the system compares it against this prediction. If the numbers match, the system trusts the sensor. If they diverge significantly, the system flags the data as potentially corrupted by an attack.

The innovation in this work lies not just in detecting the lie, but in how it fixes the problem. Simply replacing a bad sensor reading with a computer prediction can be risky, because the computer might not be equally good at predicting every single type of movement. For instance, the system might be very accurate at predicting the speed of the spinning blades but less accurate at predicting the subtle tilting of the floating platform caused by complex wave interactions. If the system blindly replaced all sensor data with its own predictions, it could accidentally introduce new errors in the areas where its prediction was weak. To solve this, the researchers introduced a "reliability gate." This is a mechanism that decides how much weight to give to the computer's prediction versus the original sensor reading for each specific measurement. If the computer is very confident in its prediction for a specific channel, like rotor speed, the system leans heavily on the prediction to correct the data. If the computer is less certain, such as with the platform's motion, the system remains conservative and relies more on the original sensor, blending the two sources carefully to avoid making things worse.

The researchers tested this framework in a simulated environment designed to mimic a 5-megawatt floating wind turbine. They subjected the system to a specific type of cyberattack called a false-data injection, where they artificially added biases to the readings for rotor speed, platform pitch, and generated power for a period of ninety seconds. The results showed that the system was highly effective at identifying the attack, correctly flagging the intrusion with an accuracy of 97.27 percent. More importantly, the method successfully recovered the corrupted data. At the sensor level, the system reduced the error in the rotor speed readings by 85.71 percent and the error in the generated power readings by 86.33 percent. The error in the platform pitch was also reduced, though to a lesser degree of 50.01 percent, reflecting the greater difficulty in predicting the complex motion of the floating platform.

When these corrected signals were fed back into the turbine's control system, the machine was able to maintain stable operation despite the attack. The study found that the protected turbine kept its rotor speed and power generation much closer to their intended targets than an unprotected turbine would have. Specifically, the error in rotor speed and generated power was reduced by 26.96 percent compared to a scenario where the turbine had no protection against the attack. The researchers also tested the system against different levels of attack intensity and added random noise to simulate real-world signal interference. The method remained robust, maintaining high detection accuracy and consistently improving control performance across these varied conditions.

A crucial part of the study involved comparing this new "reliability-gated" approach against a simpler method where the system would completely replace any attacked sensor data with the computer's prediction, regardless of how good that prediction was. The simpler method did a good job of fixing the rotor speed and power data, but it caused the platform's motion control to become less stable. This confirmed that the researchers' decision to be selective—trusting the computer more where it was strong and less where it was weak—was essential for the overall safety of the machine. The study concludes that while this method is not a perfect fix for every possible problem, such as reducing the physical stress on the floating platform itself, it provides a vital layer of security. It allows floating wind turbines to continue operating safely and efficiently even when their sensors are being actively deceived, ensuring that the transition to clean energy remains resilient against digital threats.

Technical Summary: Reliability-Gated GRU-Based Secure Control of a Nonlinear Floating Offshore Wind Turbine Under False Data Injection Attacks

Problem Statement
Floating Offshore Wind Turbines (FOWTs) operate in highly coupled cyber-physical environments where aerodynamic, hydrodynamic, and structural dynamics are intrinsically linked. These systems rely heavily on accurate sensor feedback (e.g., rotor speed, platform pitch, generated power) for closed-loop control, particularly in Region II (torque control) and Region III (pitch control). However, FOWTs are vulnerable to False Data Injection (FDI) attacks, where adversaries corrupt sensor readings without disrupting communication or causing obvious hardware failures. Such attacks can lead to misjudgments of the turbine state, resulting in inefficient operation, safety hazards, or misleading supervisory decisions. Existing control strategies for FOWTs primarily address physical perturbations (turbulence, waves) but often assume accurate sensor data, leaving a gap in secure control under cyber-physical sensor attacks. Furthermore, while data-driven methods like Recurrent Neural Networks (RNNs) can predict sensor values, directly replacing corrupted measurements with neural predictions can be unsafe if the prediction accuracy varies significantly across different sensor channels (e.g., platform motion is harder to predict than rotor speed).

Methodology
The paper proposes a Reliability-Gated GRU-Based Secure Control Framework designed for a nonlinear reduced-order FOWT model based on the NREL 5-MW reference turbine. The framework consists of four integrated components:

GRU-Based Safe Sensor Prediction: A Gated Recurrent Unit (GRU) network is trained offline on non-attacked operational data. It predicts "safe" sensor vectors ( $\hat{y}_k$ ) using operational inputs such as wind speed, wave excitation features, control commands, and operating region indicators. The network weights are fixed during online operation to prevent adaptation to corrupted data.
Residual-Based Attack Detection: A detector calculates the residual ( $\rho_k$ ) between the received (potentially attacked) measurements ( $y^a_k$ ) and the GRU predictions. A normalized detection index ( $D_k$ ) is computed; if $D_k$ exceeds a threshold derived from validation data, an attack flag ( $\delta_k$ ) is triggered.
Reliability-Gated Compensation: This is the core innovation. Instead of blindly replacing all attacked measurements with GRU predictions, the framework employs a channel-wise reliability gate.
- A reliability gain ( $g_i$ ) is calculated for each sensor channel based on its validation Root Mean Square Error (RMSE) and signal scale.
- The compensated measurement ( $y^c_k$ ) is a weighted blend of the attacked measurement and the GRU prediction: $y^c_k = (1-\delta_k)y^a_k + \delta_k [(I_m - G)y^a_k + G\hat{y}_k]$ , where $G$ is the diagonal reliability matrix.
- Channels with high prediction reliability (e.g., rotor speed, power) receive strong correction, while channels with lower reliability (e.g., platform pitch) receive conservative correction to avoid introducing new errors.
Secure Feedback Control: The compensated vector is fed into a baseline Region II/III controller. The controller logic remains unchanged, but it operates on the secured feedback, ensuring that rotor speed and power regulation are maintained even during attacks.

Key Contributions

Secure Sensor Validation Framework: The authors developed a control-oriented secure sensor validation framework using a reduced-order FOWT model. Unlike high-fidelity aero-hydro-servo-elastic models (e.g., OpenFAST), this model provides a clear nonlinear control environment for testing data-driven compensation strategies.
Physics-Informed GRU Predictor and Multi-Channel Detector: A GRU predictor was developed to estimate attack-free responses for generator speed, platform pitch, power, blade pitch, and torque. Coupled with a residual-based detector, this system identifies abnormal sensor deviations caused by FDI attacks.
Detector-Driven Reliability-Gated Compensation Law: The paper introduces a compensation strategy that avoids the unnecessary replacement of healthy signals. The gate activates only during detected attacks and applies channel-specific weighting. This balances the trade-off between correcting sensor errors and maintaining closed-loop safety, particularly for channels with high prediction uncertainty like platform motion.
- Note: The paper reports a detection accuracy of 97.27% and a total sensor-level RMSE reduction of 85.95% for the proposed framework in the main results, with specific channel reductions of 85.71% (rotor speed), 50.01% (platform pitch), and 86.33% (generated power).

Results
Simulation results were conducted over a 900-second period with FDI attacks activated between 120 and 210 seconds on rotor speed, platform pitch, and generated power channels.

Sensor-Level Performance: The reliability-gated GRU compensation achieved a total sensor-level RMSE reduction of 85.95%. Specifically, rotor speed error was reduced by 85.71%, generated power by 86.33%, and platform pitch by 50.01%. The lower reduction for platform pitch is attributed to the difficulty of predicting platform dynamics due to wave excitation and nonlinear coupling.
Closed-Loop Performance: The secure controller successfully maintained stable operation. Compared to the unprotected attacked case, the RMSE for rotor speed and generated power decreased by 26.96%. The detection accuracy was 97.27%. The platform pitch response remained bounded, with a negligible change in RMSE (approx. -0.86%), indicating that the conservative gating prevented the introduction of new instability.
Robustness: The framework was tested against nine scenarios with varying attack magnitudes and noise levels. The average detection accuracy remained high at 95.07%, with mean RMSE reductions for rotor speed and power of 24.10%, demonstrating resilience to perturbed conditions.
Ablation Study: A comparison with a "no-gate" strategy (where all attacked measurements are fully replaced by GRU predictions) revealed that while the no-gate approach offered aggressive correction for rotor speed and power, it led to larger deviations in platform pitch due to prediction errors in that channel. The reliability-gated approach provided a more balanced and safer closed-loop response.

Significance and Claims
The paper claims that the proposed reliability-aware learning compensation offers a practical compromise between attack rejection and prediction uncertainty. It demonstrates that secure reconstruction of feedback is possible without requiring a complete redesign of the baseline controller. The method is particularly significant for enhancing rotor-speed and power regulation while avoiding the degradation of control performance caused by over-reliance on weakly predicted channels (specifically platform motion).

The authors explicitly state that the framework is a sensor-recovery and rotor-speed/power control mechanism, not a platform-load-reduction controller. They acknowledge limitations, noting that the model is a reduced-order representation and has not yet been validated on high-fidelity tools like OpenFAST or experimental platforms. The work is positioned as a foundational step toward secure, learning-assisted FOWT operations, with future work planned to address high-fidelity models, actuator-side attacks, and coordinated multi-channel cyberattacks.

Technical Summary: Reliability-Gated GRU-Based Secure Control of a Nonlinear Floating Offshore Wind Turbine Under False Data Injection Attacks

More like this