20 12 International Conference on Power and Energy System s Lecture N otes in Inform ation Technology , Vol.13 Smart Pitch Control Strategy of Wind Generation System Using Differential Evolution and Neural Network S.A.Raza, A.H.M.A Rahim Department of Electrical Engineering, King Fahd University of Petroleum and Minerals, Dhahran, Saudi Arabia sraza@kfupm.edu.sa, ahrahim@kfupm.edu.sa, Keywords: Doubly Fed Induction Generator, Pitch Control, Neural Network, Differential Evolution. Abstract. A smart pitch control strategy for a variable speed doubly fed wind generation system is presented in this article. Non-linear as well as linearized dynamic models of the wind system pitch controller and the doubly fed induction generator including the drive train are developed. A PI controller is employed to generate the appropriate pitch angle for varying wind speed conditions. An artificial neural network (ANN) is trained to produce PI gain settings for various wind speed conditions. The training data, on the other hand, was generated through differential evolution intelligent technique (DEIT). Simulation studies show that the DEIT based ANN can generate the appropriate control to deliver the wind power to the generator efficiently with minimum transients. The data used was collected from the wind generator located at the King Fahd University beach front. 1. Introduction Because of diminishing oil reserves and environmental concerns, renewable energy research is getting momentum. Wind energy in itself encompasses various engineering fields and is rapidly developing into a multi-disciplinary area of research and experimentation [1]. In wind generation systems, the variable speed doubly fed induction generator (DFIG) is preferred over the permanent magnet synchronous generator (PMSG) type for its versatility [2]. The control of blade pitch angle is a necessary part of variable speed wind turbines since by controlling the pitch angle, the aerodynamic power that flows through to the generator can be adjusted. The system dynamics using the PI controller have described in [3-5]. Fuzzy logic was used in [6] to find the pitch controller parameters. A self tuning fuzzy-PID controller was proposed in [7]. Robust controllers for output power leveling of variable speed variable pitch wind turbine generator systems are also available in literature. The use of generalized predictive control has been reported in [8, 9]. A neural network capable of self tuning for during different operating conditions has been reported in [10]. The proposed controller consists of neural networks inverse and forward identifiers for modeling the dynamics of the system. Generation of test data efficiently encompassing all possible scenarios of wind system operation is a challenging job. This article proposes a method of generating a neural network test data for the wind system pitch controller using a fast evolutionary method. 2. System modeling A schematic diagram of the DFIG system connected to the power grid equipped with pitch control is shown in Fig.1. The induction generator is driven by a horizontal axis wind turbine through its 978-1-61275-0 11-8/ 10 / $ 25.0 0 © 20 12 IERI ICPES20 12 202 gear boxes. The converters are located between the rotor terminals and the grid. The dynamic model of the system includes the wind turbine, pitch controller and the generator with its converters.  Fig.1 DFIG system configuration 2.1 Wind turbine aerodynamics The amount of power extracted from wind is a function of air density and is given by, Pm  1  R 2V W 3 C p   ,   2 (1) Here, V W is the wind velocity, R is the radius of the rotor blades and C p   ,   is the power coefficient that is dependent upon the tip speed ratio  and the pitch angle  . The power coefficient C p is a non-linear function of i and  given as, 116    0.4   5 e i  0.0068 C p   ,    0.5176   i  1 1 0.035   i   0.08  3  1 The tip speed ratio  is related to wind speed through,  R  T VW 21 (2) (3) 2.2 The induction generator model In dynamic modeling the DFIG is normally represented by a 4th order model of stator and rotor currents along the d-q axes and are given as i ds i qs i dr i qr  . These equations are dependent on generator slip. The converter dynamic model is normally represented by a second order model containing d-q components of converter currents i ad i aq  in addition to the converter DC capacitor voltageVc . The drive train model consists of the high inertia turbine coupled to relatively lower inertia generator and are expressed in terms of a third order model with states [ωt, ωg, s] which are turbine and generator speeds, and torsion angle, respectively. 2.3 Pitch angle control As can be observed from (1) and (2), control of pitch angle β provides an effective means for controlling the power input to the generator under varying wind speeds. To put the blades into the necessary position, pitch servos are employed, which may be hydraulic or electrical systems. Conventional pitch angle control uses generator rotor speed error to drive a PI controller to generate appropriate  , while a more popular method is to compare the available power from the wind with actual generator power. Fig. 2 shows the pitch angle controller configuration 203 i Fig. 2. Pitch control strategy using generator power A PI controller is used as pitch controller. The composite model of the system including the PI controller is written as, x  f [x , u ] (4) Here, ‫ ݔ‬is the vector of the states [ i ds i qs i dr i qr t  g  s V c i da i qa  i ] and u is the pitch control. In this work the gains of the PI controller are obtained from a trained artificial neural network (ANN). The wind data for the training was collected from the 5-kW wind system installed at the King Fahd University beach front. The training data was generated by a differential evolution technique. A brief outline of the ANN and DEIT procedures are given in the following. 2.4 The Back-propagation Neural Network Fig.3 shows the layout of a three-layer perceptron – the input, the hidden and the output layers with activation functions in the hidden and output layers. The number of neurons in these layers is assumed to be p, r and m, respectively. The training starts by arbitrarily assuming weighting function wji and the signals at the hidden and output layers are computed as, v jn (n )   w ji (n )x i (n ) p i 0 y j (n )   j (v j (n )) (5) Here,  j is a logistic activation function of the sigmoid type. For neuron k at the output layer, the net internal activity level is, v k (n )   w kj (n ) y j(n ) r (6) j 0 v1 w j1 w k1 y1 x1 xp ym w jp vr wkr Fig.3 The back propagation net layout In the training process, the network is presented with a pair of patterns – an input pattern and a corresponding desired output pattern. In the back-propagation algorithm, there are two distinct passes of computation. In the forward pass, the outputs are computed on the basis of selected weights and the error e k (n )  d k (n )  k v k (n )  is computed. In the backward pass the weights are updated so as to minimize the sum of the squares of errors, E av  204 1 m 2  ek . The synaptic weights wji 2 k 1 at any layer l is updated through the steepest descent technique. The solution is accelerated through a proper choice of momentum constant α and learning rate parameter and is finally expressed as, l 1 (7) w ji (n  1)  w ji (n )   [w ji (n ) w ji (n  1)]    j (n ) y i (n ) where, ' (8)  j (n )   (v j (n ))  ik 1(n )w lkj1(n ) k In the pitch control algorithm, the input to the network is the set of wind speeds collected for a sample time and the output trained variables are the controller gains, which are determined from the following algorithm. 2.5 Differential evolution intelligent technique The DEIT is a evolutionary search algorithm which finds the optimum value of an objective function subject to satisfying the system constraints. The steps involved in the algorithm are briefly given in the following [11]. Step a: Define the dimension of the problem. Since in this case the control variables are, KP and KI, dimension is 2. Also, set Maximum and minimum range of the variables Step b: Within the upper and lower bounds for KP and KI, create the population members through the relation, x i , j  x j ,min  rand  0,1  x j ,max  x j ,min  i  1, NP , j  1, D Here, NP=100, D=2 and rand  0,1 is a random number selected between 0 and 1. (9) Step c: To change each member of the target generation X i(G ) , a donor vector V i (G 1) is produced given by mutation as. V i (G 1)  X r 1(G )  F  X r 2 (G )  X r 3(G )  (10) X r 1(G ) , X r 2 (G ) and X r 3(G ) are randomly selected solution vectors from the target generation, F is the mutation factor. Step d: To enhance the diversity, a crossover operation a binomial type crossover is applied on each variable defined by, u i(G, j )   i(G, j ) if rand  0,1  CR ; x i(G, j ) otherwise (11) CR is the crossover factor, u i(G, j ) ,  i(G, j ) and x i(G, j ) is the jth component of the trail vector, donor vector and target vector respectively in the ith population members. Step e: To keep the generation size constant select which is going to survive in the next generation. by using Survival of the Fittest concept using, X i(G 1) U i(G ) if J U i(G )  X i(G )  ; X i(G 1)  X i(G ) if J  X i(G )  U i(G )  (12) J is the objective function to be minimized which is difference in damping ratio obtained from the dominant eigenvalues of linearized system of the original nonlinear equation (3). U i(G ) is the current trial vector and X i(G ) is the current target vector. Step f: The best solution corresponds to the minimum value of the objective function. Repeat steps 'a-e', to get the global best values of KP and KI, or stop when the maximum number of iterations has been reached and restart the procedure. 205 3. Simulation results Fig. 4. (a) Output power, (b) Generator speed, (c) Stator current and (d) Terminal voltage variation for 12m/s to 11m/s step change in wind speed The pitch controller designed through the ANN and DEIT was tested for various wind speed conditions. Figs. 4 and 5 show the DFIG power output, speed, stator current and terminal voltage variations for two test cases. These are for wind speed variations from 12 m/s to 11m/s, and 12 m/s to 14 m/s, respectively. The responses show a comparison with no control cases. It can be observed that the proposed DEIT based ANN pitch controller transfers the wind power to the generator with minimum transients. The steady state response is also very good. Fig. 5. (a) Output power, (b) Generator speed, (c) Stator current and (d) Terminal voltage variation for 12m/s to 14m/s step change in wind speed 206 4. Summary A smart pitch controller which adjusts its parameters depending on the variation of wind speed is proposed. The controller uses a back propagation neural network backed by a highly efficient adaptive evolutionary technique. The neural network uses the nonlinear dynamic model of the DFIG wind generation system, while the DEIT uses the quasi-linearized model for computational efficiency. Simulation results clearly show the effectiveness of the proposed smart control in transferring the wind power to the grid through the doubly fed generator smoothly and with very little transients. While the background work requires some computational effort, the controllers can be tuned adaptively with varying wind speed. 5. Acknowledgement The work was done as part of KFUPM research group projects RG 1202-1 & 1202-2. The authors wish to acknowledge the support of the King Fahd University of Petroleum and Minerals. References [1] Lucy Y. Pao and Kathryn E. Johnson, "A Tutorial on the Dynamics and Control of Wind Turbines and Wind Farms", American Control Conference, St. Louis, June 10-13, 2009. [2] A.H.M.A. Rahim and I.O. Habiballah, "DFIG rotor voltage control for system dynamic performance enhancement", Electric Power Systems Research, vol. 81, 2010, pp. 503-509. [3] V. Akhmatov, "Variable-speed wind turbines with doubly-fed induction generators Part I: Modelling in dynamic simulation tools", Wind Engineering, Vol. 26, No.2, 2002, pp. 85-108. [4] D. Hansen, P. Sorensen, F. Iov, F. Blaabjerg, "Control of variable speed wind turbines with doubly-fed induction generators", Wind Engineering, Vol. 28, No. 4, 2004, pp. 411-434. [5] T. Senjyu, R. Sakamoto, N. Urasaki, H. Higa, K. Uezato, T. Funabashi, "Output power control of wind turbine generator by pitch angle control using minimum variance control", Electrical Engineering in Japan, Vol. 154, No. 2, 2006, pp. 10-18. [6] Z. L. Dou, M. Z. Cheng, Z. B. Ling, and X. Cai, "An Adjustable Pitch Control System in a Large Wind Turbine Based on a Fuzzy-PID Controller", International Symposium on Power Electronics, Electrical Drives, Automation and Motion, 2010, pp. 391-395. [7] H. Geng and G. Yang, "Robust pitch controller for output power leveling of variable-speed variable-pitch wind turbine generator systems", IET Renewable Power Generation, vol.3, issue 2, 2009, pp. 168-179. [8] T. Senjyu, R. Sakamoto, N. Urasaki, T. Funabashi, H. Fujita, and H. Sekine, “Output Power Leveling of Wind Turbine Generator for All Operating Regions by Pitch Angle Control,” IEEE Trans. on Energy Conversion, vol. 21, no. 2, 2006, pp. 467-475. [9] D. W. Clarke, C. Mohtadi, and O. S. Tuffs, “Generalized predictive control Part I. The basic algorithm,” Automatica, vol. 23, no. 2, 1987, pp. 137– 148. [10] F. Bati, and S. K. Leabi, "NN Self-Tuning Pitch Angle Controller of Wind Power Generation Unit", PCSE, 2006, pp.2019-2026. [11] Vesterstrom, J and Thomsen R, "A comparative study of differential evolution, particle swarm optimization, and evolutionary algorithms on numerical benchmark problems". Congress on Evolutionary Computation CEC2004, vol. 2, 19-23 June 2004, pp. 1980 – 1987. 207