Wind turbines are generally designed to deliver maximum output at wind speeds around 15 m/s [3] . According to equation (1), for a given wind speed, the only parameter that allows to control extracted power is the Cp parameter (Figure 2). This parameter depends on the type of turbine and the type of design from the manufacturer. It is also a function of the pitch angle control, were mainly used to limit the mechanical power during strong wind speed or to adapt power production (Figure 3). The aerodynamic behavior of the wind turbine can be controlled actively by varying the turbine blade angle around their longitudinal axis [4] .

In low and medium wind speed, the pitch angle control is rarely actuated and only set to the optimum value. In high wind speeds, the pitch angle control effects to regulate the extracted wind power so that the design limits of the turbine will not be exceeded. Rotor speed, torque or power control can be taken to regulate the pitch angle [5] . In this paper, the power error criteria are used. The error between the reference and the measured power at the output of the PMSG is sent to a PI controller to generate the reference value β_ref of the pitch angle [6] .

The control system structure used to generate the pitch angle reference is depicted in Figure 4.

The feed forward term assumes that all the components are ideal and does not account for the losses in the system. The feedback path compensates for the losses by decreasing the pitch angle, if the output power is less than the desired power, to increase the power captured.

Fuzzy logic control is one of the most powerful control methods where based on fuzzy set theory and associated techniques. The fuzzy logic algorithm is suitable for wind turbine control with complex nonlinear models and parameters variation. The fuzzy controller (Figure 5) uses the perturbation and observation to track the optimal pitch angle without knowledge of wind turbine characteristic.

Basic structure of the fuzzy logic controller consists of three important stages: Fuzzification, Decision Making Unit and Defuzzification Unit. The inputs of the Fuzzification are selected as error and rate of change of error and the output is the pitch angle reference. The two input variables are calculated at every sampling instant say k [6] .

where, P_ref(k) and P_g(k) are respectively the reference and generated power at instant n.

In Fuzzification stage the crisp variables e(k) and De(k) are converted into fuzzy variables which can be identified by membership function (Figure 6). The proposed controller uses following linguistic labels cited in Table 1.

NB (Negative Big), NS (Negative Small), Z (Zero), PM (Positive Small), PB (Positive Big).

The input and output variables map into the following fuzzy set (Figure 6).

The triangular membership functions are used to define both inputs (error, rate of change of error) and output (pitch angle).

For evaluating the robustness of the fuzzy logic controller compared with the conventional PI controller at variable wind speed, the simulations (Figure 7) have been carried out using MATLAB/Simulink software.

Figure 7 shows the results of the conventional pitch angle control strategy where power is used as the controlling variable. The comparisons of different strategies show that proposed fuzzy logic control strategies is more robust in pitch angle control.

The fuzzy logic controller for the DC bus has three parameters (Figure 8). The first input is the error of DC voltage, the second is the rate of change of error and the output is the power controlling current component. The inputs variables are calculated at every sampling instant say k [9] .

where, V_DCref(k) and V_DC(k) are respectively; the reference and measured voltage at instant n.

In Fuzzification stage the crisp variables e(k) and De(k) are converted into fuzzy variables which can be identified by membership function. The fuzzification maps the error and change in error to linguistic labels of fuzzy sets.

The proposed controller uses following linguistic labels in Table 2: NB (Negative Big), NM (Negative Medium), NS (Negative Small), NVS (Negative Very Small), Z (Zero), PVS (Positive Very Small), PM (Positive Medium), PB (Positive Big).

In this section a procedure for constructing a neural network model predictive controller for the voltage DC is presented. In practice, an operational load shedding scheme for bulk power systems should be able to incorporate an infinite number of possible system states that would be mapped to a finite number of actions. The training set should contain both the “standard” normal operational conditions and the state of parameters necessary to implement the appropriate action for system stability. Once the training process is completed, the benefits of the systems are immediately evident in the speed of response to faults and the seamless integration with existing power system controls. Load shedding neural networks are composed of an input layer, two hidden layers and one output layer. As mentioned earlier, using the model reference control tool in neural networks, it is possible to include a network control (Figure 9) with the transformer model to monitor the performance of the system [10] .

The activation function of both the output layer are linear and unbiased while that of the hidden layer is sigmoid and biased.

Training the neural network is often done in epochs or cycles that are based largely on the inputs into the system. So for various inputs, the neural network basically compares each input to other known inputs and plots a graph of the error gradient between these two values. The process of comparison is carried out for as many random inputs as possible while each result and the corresponding response of the Plant are stored for future reference.

In order to evaluate the performance of the control system for the variable DC bus reference concept with PMSG, we set reference to 10 V. In the first next we shoes 20 V and finally we shoes 40 V.

Figures 10-12 show the responses of the system when using respectively IP, fuzzy logic and neural controllers.

From Figure 13, showing the comparison of the three responses; we note that the use of fuzzy provides a response with a smaller overshoot and faster compared to the neural controller and conventional IP controller.