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

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

Problem Statement
The integration of Wind Energy Conversion Systems (WECS) into transmission networks is hindered by the random and non-linear nature of wind turbulence. These fluctuations cause vibration, mechanical stress on the drive-train, and instability in output power. While pitch-controlled variable-speed turbines are standard for megawatt-class applications, maintaining output power within an optimal and nominal range during gusty conditions remains a challenge. Traditional control methods, such as PID or generalized predictive models, often struggle with the adaptive requirements of highly turbulent environments. The paper posits that intelligent, online control strategies are necessary to effectively identify and reject wind turbulence disturbances.

Methodology
The study proposes a comparative analysis of three intelligent control strategies designed to regulate the blade pitch angle, thereby stabilizing output aerodynamic power and torque. The system is modeled using a 2 MW variable-speed turbine with a 37.5m rotor radius, where wind speed is simulated using an ARMA (Auto-Regressive Moving Average) model incorporating mean wind speed and Gaussian noise turbulence.

1. Artificial Neural Network (ANN) Controllers:

   • Multi-Layer Perceptron (MLP): A predictive controller with two inputs (wind speed and output power) and one output (pitch angle). It utilizes a single hidden layer with five neurons and a sigmoid activation function. The network is trained via back-propagation to map optimal pitch angles to wind conditions.
   • Radial Basis Function (RBF): A controller with three inputs and one output, utilizing a hidden layer with ten neurons and Gaussian radial basis functions. Like the MLP, it is trained on credible optimal input-output data to determine the appropriate pitch angle.

2. Genetic Fuzzy System (GFS):

   • The authors propose a new Genetic Fuzzy System to extract fuzzy if-then rules directly from the neural network data.
   • Approach: The system employs the "Michigan approach," where a population of fuzzy rules evolves. Rules are coded as numerical strings representing linguistic values (e.g., "very small" to "very large").
   • Fitness Function: The fitness function has two components: one comparing generated rules against optimal values, and another calculating the rule's effectiveness in regulating power based on the wind turbine power formula.
   • Genetic Operations: Selection, uniform crossover, and mutation are applied to evolve the rule set. The system replaces the worst-performing rules with new offspring until a stopping condition is met, ultimately selecting five optimal rules for the control system.

Key Results
Simulations were conducted over a 100-second period with varying wind speeds (including turbulence) to evaluate the controllers.

• MLP vs. RBF: Both neural network controllers successfully regulated pitch angles to maintain power near the rated value. However, the MLP controller demonstrated superior performance compared to the RBF controller. The authors attribute this to the MLP's ability to handle the "relatively large scatter" of the wind speed data, whereas the RBF results were less effective when the training data did not exhibit a cluster mode.
• GFS Performance: The proposed GFS controller successfully regulated output power and torque. When compared to other methods, including Linear Quadratic Gaussian (LQG) algorithms and BP neural network-based PID controllers, the GFS approach showed fewer sudden drops and rises in power. Specifically, during intense wind speed reductions (e.g., 45–70 seconds), the GFS maintained more stable power output compared to the LQG method, which exhibited dramatic fluctuations.
• Rule Extraction Efficiency: A comparison with the FRENGA (Fuzzy Rule Extraction using Neural Networks and Genetic Algorithms) method showed that while FRENGA extracts rules faster, the proposed GFS approach achieves almost equal accuracy in error minimization (lowest error of 0.1226 for GFS vs. 0.1445 for FRENGA without NN).

Significance and Claims
The paper claims that the proposed intelligent controllers effectively address the challenges of wind turbulence rejection. The primary significance lies in the ability of these "smart" controllers to:

1. Regulate Output: Maintain aerodynamic power and torque within a nominal range despite random wind fluctuations.
2. Reduce Mechanical Stress: By absorbing rapid power alterations through dynamic pitch angle adjustments, the controllers help escape mechanical pressures on the turbine.
3. Adaptability: The methods rely on initial learning data rather than specific environmental conditions, allowing them to be applied to wind farms in different locations without rewriting programs.

The authors conclude that while both MLP and GFS approaches yield acceptable results, the MLP-based controller provided the best response among the neural network types tested, and the GFS controller offered the most robust performance when compared to traditional and other intelligent algorithms. The study acknowledges a limitation: the proposed methods require a computer with high speed and memory to record and process previous data for training. Future work is identified as tuning membership functions to further improve accuracy and processing speed.