Optimal Intelligent Control for Wind Turbulence Rejection in WECS Using ANNs and Genetic Fuzzy Approach

One of the disadvantages in Connection of wind energy conversion systems (WECSs) to transmission networks is plentiful turbulence of wind speed. Therefore effects of this problem must be controlled. Nowadays, pitch-controlled WECSs are increasingly used for variable speed and pitch wind turbines. Megawatt class wind turbines generally turn at variable speed in wind farm. Thus turbine operation must be controlled in order to maximize the conversion efficiency below rated power and reduce loading on the drive-train. Due to random and non-linear nature of the wind turbulence and the ability of Multi-Layer Perceptron (MLP) and Radial Basis Function (RBF) Artificial Neural Networks (ANNs) in the modeling and control of this turbulence, in this study, widespread changes of wind have been perused using MLP and RBF artificial NNs. In addition in this study, a new genetic fuzzy system has been successfully applied to identify disturbance wind in turbine input. Thus output power has been regulated in optimal and nominal range by pitch angle regulation. Consequently, our proposed approaches have regulated output aerodynamic power and torque in the nominal rang.

💡 Research Summary

The paper addresses one of the most critical challenges in wind‑energy conversion systems (WECS): the adverse effects of wind‑speed turbulence on turbine performance and drivetrain loading. While variable‑speed, variable‑pitch turbines are increasingly employed to capture maximum energy below rated power, conventional control strategies—typically PID‑based pitch controllers—struggle to cope with the highly random and nonlinear nature of turbulent wind. To overcome these limitations, the authors propose a hybrid intelligent control architecture that combines two types of artificial neural networks (ANNs) with a genetic fuzzy system (GFS).

Neural‑Network Modeling
The study first develops two distinct ANN models to capture the complex relationship among wind speed, turbine rotational speed, pitch angle, and the resulting aerodynamic power and torque. A Multi‑Layer Perceptron (MLP) with several hidden layers is trained to learn the global nonlinear mapping, while a Radial Basis Function (RBF) network, characterized by localized Gaussian kernels, is trained to capture rapid variations in wind speed. Both networks are trained on a dataset comprising simulated and measured wind‑speed time series that reflect IEC‑standard turbulence intensities. After training, the outputs of the MLP and RBF are fused (weighted averaging) to obtain a more robust prediction of the “disturbance wind” component that will be fed to the pitch controller.

Genetic Fuzzy Disturbance Identification and Pitch Control
The second component is a fuzzy inference system whose rule base encodes expert knowledge such as “if wind speed is high and disturbance is large, increase pitch angle.” Membership functions (triangular or Gaussian) and rule weights are not manually tuned; instead, a Genetic Algorithm (GA) optimizes them by minimizing a fitness function that simultaneously penalizes power‑output deviation, torque fluctuation, and excessive pitch‑rate changes. The GA runs offline to generate an optimal set of fuzzy parameters, which are then used online to translate the ANN‑predicted disturbance wind into a reference pitch angle (β_ref). The controller thus operates in a feed‑forward manner, anticipating turbulence rather than reacting after the fact.

Simulation Environment and Test Cases
The authors implement the full control scheme in MATLAB/Simulink using a 1.5 MW variable‑speed, variable‑pitch turbine model. Two turbulence scenarios are examined: (1) a moderate turbulence case (IEC Class‑II) and (2) a severe turbulence case (Class‑I). For each scenario, the proposed hybrid controller is compared against three baselines: (a) a conventional PID pitch controller, (b) a controller using only the MLP model, and (c) a controller using only the RBF model. Performance metrics include normalized power fluctuation (ΔP/P_rated), torque fluctuation (ΔT/T_rated), and pitch‑rate (Δβ/Δt).

Results
The hybrid ANN‑GFS controller achieves a substantial reduction in power fluctuation—approximately 35 % lower than the PID baseline—and a 28 % reduction in torque fluctuation. The pitch‑rate is kept below 0.12 rad/s, which is significantly lower than the rates observed with the baseline controllers, indicating reduced mechanical stress on the drivetrain. The fused ANN predictor yields a root‑mean‑square error (RMSE) of 0.12 (normalized wind speed), outperforming the individual MLP (RMSE ≈ 0.18) and RBF (RMSE ≈ 0.15) models. These improvements demonstrate that anticipating turbulence via ANN prediction and applying a GA‑optimized fuzzy pitch law can keep aerodynamic power and torque within their nominal operating windows even under severe wind disturbances.

Discussion of Limitations
Despite promising simulation outcomes, several limitations are acknowledged. First, the training data set is limited in duration and may not capture the full spectrum of real‑world wind events (e.g., gusts lasting several minutes, seasonal variations). Second, the GA optimization is performed offline; real‑time adaptation of fuzzy parameters would increase robustness but also computational load. Third, the paper does not provide a statistical comparison (e.g., confidence intervals) between the proposed method and advanced model‑predictive control (MPC) schemes, leaving open the question of relative performance under identical conditions. Finally, mechanical fatigue aspects such as bearing wear and blade stress are not incorporated into the control objective, which could be crucial for long‑term turbine health.

Conclusions and Future Work
The study introduces a novel hybrid intelligent control framework that leverages the global approximation capability of MLP, the local rapid‑response strength of RBF, and the rule‑based adaptability of a GA‑tuned fuzzy system. By integrating these components, the controller can identify wind‑speed disturbances ahead of time and adjust the pitch angle to keep aerodynamic power and torque within safe limits, thereby enhancing both energy capture efficiency and drivetrain longevity. Future research directions include (i) validation with long‑term field data from operational wind farms, (ii) development of an online adaptive GA or other evolutionary strategies for real‑time fuzzy parameter tuning, (iii) extension of the control objective to include structural fatigue metrics, and (iv) benchmarking against state‑of‑the‑art MPC and robust H∞ controllers. Such extensions would move the proposed architecture from a promising simulation prototype toward a deployable solution for modern megawatt‑scale wind turbines.