Pitch control of wind turbine based on deep neural network

This paper analyzed the input and output data of wind farm based on deep neural network, developed intelligent model, and realized the predictive modeling of important parameter variables and control of wind turbine. By establishing the Deep Extreme Learning Machine(DELM), the higher-order nonlinear model is simplified. In this structure, unsupervised hierarchical ELM is conducted for feature extraction, and the features of the lower layer are transferred to the higher layer through layer by layer coding to form a relatively complete feature representation. Finally, the Extreme Learning Machine (ELM) is used to complete the mapping of feature representation to target output to minimize the loss of information in the transmission process. The target output is used as reference data for Pitch control of wind turbine, which is proposed by using a radial basis function (REF) neutral network. Simulation results from GH-Bladed show that proposed control algorithm can mitigate the loads effectively. The algorithm provides a practical reference for the design of wind turbine controller.

controller of wind turbine. In order to obtain the optimal data set of training RBF neural network, PSO evolutionary algorithm is adopted.
In this design, wind speed value is predicted by Deep Extreme Learning Machine(DELM), which is used as the input of the RBF neural network, and the proportional and integral gain of PI controller is used as the output of RBF neural network. In order to obtain the optimal training data set, this design uses PSO algorithm to optimize the proportional gain and integral gain at a specific wind speed greater than the rated wind speed. By adjusting the blade Angle, the trained controller can minimize the difference between the measured and rated value of generator speed.

Extreme Learning Machine
Different from traditional gradient-based algorithms, ELM's input weights and single-hidden layer biases are arbitrarily chosen without iterative adjust, and the only parameters to be learned in training are the output weights which can be calculated by solving a single linear system.
, where x j and t j are the j-th input and target values respectively. To seek a regressor function from the input to the target [7][8][9], the standard Single Hidden Layer Feed-forward network can be mathematically modeled as: where o j is the output vector of the j-th training sample, w i is the input weight vector connecting the input nodes to the i-th hidden node, b i is the bias of the i-th hidden node, g(.) denotes hidden nodes nonlinear piecewise continuous activation functions. The above N equations can be written compactly as: The matrix H is the hidden layer output matrix, which can be randomly generated independent of the training data. β = [β1, β2, ... , βnh]T is the output weight matrix between the hidden nodes and the output nodes. Thus, training SLFNs simply amounts to getting the solution of a linear system (2) of output weights β [10][11][12].
A simple representation of the solution of the equation (2) is given explicitly by Huang et al. [12] as † H T β ∧ = (5) where † H is the Moore-Penrose generalized inverse of the hidden layer output matrix H.
To improve generalization performance and make the solution more robust, we can add a regularization term [13], as shown in the equation (6) and equation (7), Thus, the ELM tends to reach the solutions straightforward without the issue of over-fitting. These two features make ELM more flexible and attractive than traditional gradient-based algorithms.

Deep Extreme Learning Machine(DELM)
The Deep Extreme Learning Machine integrates the idea of Auto-Encoder in the algorithm, and codes the output by minimizing the reconstruction error, so that the output can infinitely approximate the original input [14][15][16].This structure can capture relevant higher-level abstractions of the input.  W , W ,..., W h L + = denotes the parameters of the network that need to be learned.
In our paper, the Deep ELM is applied to learning the parameters L, which is designed by using the encoded outputs to approximate the original inputs by minimizing the reconstruction errors.The output weights β can be analytically determined by the equation (6) or (7) depending on the number of nodes in the hidden layer.
Each layer in the network can be decoupled as an independent ELM, and the target output T of each ELM can be equal to the input of the ELM [17][18].In this way, you can get a low-dimensional representation of the input data, that is, the hidden output of ELM, which is the input of the next ELM. By means of the trained DELM model, the wind speed at the moment t+s can be predicted.

Pitch control based on RBF
Artificial neural network and genetic algorithm are widely used in the design of the control system, while the pitch control of wind turbine uses the Radial Basis Function (RBF) neural network and Particle Swarm Optimization (PSO) evolutionary algorithm.

Radial basis function (REF) neutral network Artificial Neural Networks (ANN) is a mathematical model which is inspired by Biological Neural
Network and proposed to simulate the process of human brain information processing. It can realize features learning, classification and regression and other functions. The mapping relationship between the input and output of neurons is: where xj and wj are the input signal and weight of the j-th neuron respectively. g( ) ⋅ is the activation function, the common activation functions are traingd( ) ⋅ 、 RBF( ) ⋅ 、 sigmoid( ) ⋅ etc. And the neural network using RBF( ) ⋅ as the activation function is RBF neural network. A typical RBF neural network is a three-layer structure: the input layer, the hidden layer with nonlinear RBF activation function and the linear output layer. Its structure is shown in fig.4

PSO evolutionary algorithm
Particle swarm optimization (PSO) is a kind of evolutionary computing technology, which derives from the research on the predation behavior of birds.The basic idea of particle swarm optimization algorithm is to find the optimal solution through the cooperation and information sharing among individuals in the group.
PSO is initialized to a group of random particles (random solutions), and then iterated to find the optimal solution.In each iteration, the particle updates itself by tracking two "extreme values" ( k ipbest y , k gbest y ).After finding these two optimal values, the particle updates its velocity 1 Where ω is inertia weight, 1 2 , c c are acceleration constants, and 1 2 , (0,1) r r ∈ is random number.

Pitch control process
A new pitch control method for wind turbines is proposed in this paper, as shown in the fig.1. In this method, based on the Extreme Learning Machine, the effective predictive value of wind speed is first obtained, which is detailed in section 3. Then, taking the predictive wind speed as an important input of the turbine control system, the optimal PI controller is designed to minimize the difference between the measured value and the rated value of the generator speed [19][20]. In order to achieve this goal, the PI controller should provide an appropriate pitch Angle reference value ref In order to design the optimal performance of the PI pitch controller, the RBF neural network must be trained with the optimal training data set, so that the RBF neural network can provide the optimal PI gain.In order to obtain the optimal training data set, PSO evolutionary algorithm is used in this design. PSO algorithm is used to optimize the proportional gain and integral gain at a specific wind speed greater than the rated wind speed, as shown in the fig.5. The whole optimization process is completed by minimizing the cost function, namely the integral absolute error: After initializing the population size and particle position and velocity, PSO calculates the PI controller gain of each particle and stores its IAE value. According to IAE value, the optimal position of each particle can be calculated, and the optimal particle can also be selected from all the population particles.The whole process iterates until the number of cycles reaches the maximum number of iterations. After the last cycle, the optimal particle is selected. In fact, the position information of the optimal particle contains the proportion and integral gain corresponding to the minimum IAE value.
PSO can only provide a pair of gains for constant wind speed, but we can obtain an optimal data set in this way.In this design, RBF neural network is selected to calculate the optimal PI gain for each wind speed based on the above optimal data set.When the network is trained, the constant wind speed in the optimal data set is taken as input and the corresponding PI gain is taken as output.After training, the RBF neural network can be used to calculate the optimal PI gain corresponding to any wind speed in the full load area with its general approximation characteristics.

Simulation results
In order to evaluate the performance of the proposed wind speed prediction model based on DELM , it is applied to turbulence wind prediction of 2MW wind turbine, as shown in fig. 6 in which 10 m/s average wind speed is used. This wind speed profile is obtained based on kaimal wind model. It can be seen from fig. 6 that the use of DELM for wind speed prediction can effectively track the real-time change of wind speed, which provides the possibility for the advance pitch control.   Therefore, according to the simulation results, compared with conventional PI controller, the proposed controller with RBF neural network has more effective performance in pitch angle control. In this way, the ultimate load of the blade can be effectively reduced and the service life of the wind turbine can be prolonged.

Conclusion
This paper analyzes the input and output data of wind farm through the design of intelligent control algorithm, develops intelligent model, and realizes the predictive modeling and control of important parameter variables of wind turbine. The new intelligent method of deep neural network can reduce the computational complexity of the whole model, reduce the consumption of computing resources, and realize the modeling and prediction of wind turbines or important parameter variables.

References
[1] Xing G, Guo W. Method for collective pitch control of wind turbine generator system[J]. [Nm]