Energy Conversion and Management 157 (2018) 587–599 Contents lists available at ScienceDirect Energy Conversion and Management journal homepage: www.elsevier.com/locate/enconman Maximum power extraction for wind turbines through a novel yaw control solution using predicted wind directions T ⁎ Dongran Songa, Jian Yanga, Xinyu Fanb, , Yao Liuc, Anfeng Liua, Guo Chend, Young Hoon Jooe a School of Information Science and Engineering, Central South University, Changsha, China School of Automation, Beijing Institute of Technology, Beijing, China c Guangdong Power Grid Corp, Zhuhai Power Supply Bur, Zhuhai, China d School of Electrical Engineering and Computing, University of Newcastle, Callaghan, Australia e Department of Control and Robotics Engineering, Kunsan National University, Kunsan, Republic of Korea b A R T I C L E I N F O A B S T R A C T Keywords: Maximum power extraction Yaw control Wind direction prediction ARIMA-KF Model predictive control For modern horizontal axis wind turbines (WTs), a yaw drive mechanism is utilized to adjust the nacelle position to face the wind direction. Depending on historical signals from wind direction sensors, conventional yaw control methods could not provide sufficient performance in tracking winds, and thus result in a reduction of wind power extraction. This issue needs to be tackled using advanced control solutions. Taking advantage of predicted wind directions, a novel control solution is proposed in this study. Specifically, the proposed solution refers to a novel control structure that consists of a wind direction predictive model and a novel yaw control method. Under the proposed control structure, a hybrid autoregressive integrated moving average method-based Kalman filter (ARIMA-KF) model is used to predict the wind direction, and two novel yaw control methods are proposed: one created by using the predicted wind direction as the tracking reference, and the other based on a model predictive control (MPC) using a finite control set. To demonstrate the feasibility and the superiority of the proposed solution, two novel yaw controllers are developed and tested through some simulation tests using industrial data. Their performance is compared to the one of two industrial yaw controllers. Comparison results show that the two novel yaw controllers are capable of reducing yaw error, and thus increase wind power extraction for the WTs. Meanwhile, it is noticeable that the MPC-based controller has an advantage in the aspect of reducing yaw actuator usage. 1. Introduction As the increasing demands of wind energy, the focus of research today in wind turbines (WTs) lies in maximizing the power production per unit investment. To make wind energy more competitive with other sources of renewable energy, optimal solutions have been developed constantly for WTs [1], where the control technology plays an indispensable role that directly affects performance of the WTs in the both aspects of power production [2] and component loads [3]. Modern WTs with horizontal axis have three control actuators: pitch actuator, torque actuator, and yaw actuator. The former two actuators are considered as the two dominating ones, since they can provide a fast response that answers to the rapid variation of wind force. Accordingly, there are large quantities of literature that focus on control methods for the pitch and torque actuators. By comparison, the literature about the yaw system control is limited. Nevertheless, the function of the yaw system should not be neglected. ⁎ The operation of the yaw system may affect performance of the WT. On the one hand, a yaw misalignment leads to a decreased wind power capture. Theoretically, the wind power captured by a horizontal axis WT is decreased by the cube of the yaw error [4]. Although empirical data have shown that the relationship could be cosine-squared instead of cosine-cubed [5], it is obvious that the yaw error results in the power reduction of the WT. On the other hand, a yaw misalignment may bring about an increment of component loads. The impact of yaw misalignment on loads of the WTs has been investigated and validated by researchers using calculation and measurement methods. For instance, Schepers conducted a comparison investigation between calculations and measurements on a small WT with 10 m rotor diameter in yaw, which revealed that the yaw misalignment had effects on blade root and shaft loads on a sectional level [6]. Boorsma presented a report of power and loads for a 2.5 MW WT in yawed flow conditions, in which the edgewise fatigue equivalent loads were found to be increased along with the increasing yaw error [7]. Kragh et al. [8] showed the potential Corresponding author. E-mail address: fanzyzl@163.com (X. Fan). https://doi.org/10.1016/j.enconman.2017.12.019 Received 18 September 2017; Received in revised form 15 November 2017; Accepted 6 December 2017 0196-8904/ © 2017 Elsevier Ltd. All rights reserved. Energy Conversion and Management 157 (2018) 587–599 D. Song et al. Nomenclature N number of distribution zones of yaw error θyej averaged yaw error at the j th zone f j ∈ [0,1] distribution probability of θyej in the j th zone (cos(θye ))eq equivalent cosine of yaw error Pa wind power extracted by a horizontal axis WT Pred,Pideal reduced power extraction and ideal power extraction the kth sampling period k sampling period Ts Tc control period Ah1,Ah2,Ah3 amplitude thresholds predefined in yaw control algorithm Th1,Th2,Th3 time thresholds predefined in yaw control algorithm nacelle position measured at kth sampling period θnp (k ) ̇ (j ) θnp permissible yaw speed ̇ (k ) yaw speed during the kth control period θnp w1,w2,w3 weighting factors in the quality function θwd (k ) measured wind direction sampled at the kth sampling period θwd (k + 1|k ) predicted wind direction at the kth sampling period θye (k + 1|k ) θye (k + 1|k ) predicted yaw error at the kth sampling period 10s 30s 60s θye ,θye ,θye mean wind directions averaged at the sampling periods of 10 s,30 s,60 s θwd (k + 1|k )Ts = 10s,θwd (k + 1|k )Ts = 30s,θwd (k + 1|k )Ts = 60s predicted mean values of wind direction at the kth sampling period, Ts = 10 s,30 s,60 s θye (k + 1|k )Ts = 10s,θye (k + 1|k )Ts = 30s,θye (k + 1|k )Ts = 60s predicted mean values of yaw error at the kth sampling period, Ts = 10 s,30 s,60 s Abbreviations WT MPP HCM ARIMA KF MPC MY NREL MAE RMSE MAPE QF SCADA wind turbine maximum power point hill climbing method autoregressive integrated moving average Kalman filter model predictive control Ming Yang National Renewable Energy Laboratory mean absolute error root mean squared error mean absolute percentage error quality function Supervisory Control and Data Acquisition System Symbols ρ Ar Cp V0 θye θwd θnp tyaw Cyaw ξ air density rotor area aerodynamic power coefficient free stream wind speed yaw misalignment error wind direction nacelle position yaw action time yaw action count reduction factor of wind power extraction caused by yaw error Table 1. The employed techniques are broadly categorized into four types: free of measurement, normal measurement, advanced measurement and indirect measurement. Accordingly, relevant control methods can be also categorized into four types and they have the following features: of alleviating blade load variations induced by the wind shear through yaw misalignment for wind speeds above rated wind speed. From their studies, it is observed that the operation of the yaw system significantly affects the performance of the WTs. A study of operating WTs revealed a fact that there was a static yaw error of 10 degree for wind speeds below 20 m/s and 5 degree for wind speeds above 20 m/s, which unavoidably reduced wind power extraction of the WTs [9]. Besides, an early survey of failures in wind power systems showed that the portion of downtime caused by yaw failure comprised 13.3% of the total downtime, and the yaw system failure rate comprised 6.7% [10]; and a recent analysis for wind turbine reliability concluded that the failure rate of wind turbines was increased up to 12.5% [11]. Thus, the controls for the yaw system deserve more attention than they received. In the literature, the control methods for yaw systems are directly relevant to the measurement techniques which can be seen from • Controls without wind direction measurement, which originates from early WTs limited by the wind measurement technology. Because the main objective of yaw control system is to maximize wind power extraction, the mechanism for controls without wind direction measurement is to directly search the maximum power point (MPP). Hill climbing method (HCM) was proposed to find the desired yaw angle corresponding to the MPP [12], and bisectingplane algorithm was presented to enhance the efficiency and accurateness of conventional HCM [13]. Besides, a combined maximum Table 1 Summary of the yaw control methods recorded in the literature. Measurement Control objective Method WT capacity Refs. Free Searching optimal power Searching optimal power Tracking optimal rotor speed HCM Modified HCM PI < 50 kW 1.5 MW 1.1 kW/2.5 MW [12] [13] [14] Normal Tracking Tracking Tracking Tracking Fuzzy-PID Logic control Logic control Logic control Unclear 2 kW 600 kW 1.5 MW [15] [16] [17] [18] Advanced Tracking wind direction Tracking wind direction Logic control Conventional MPC 600 kW 5 MW [19–21] [22] Indirect Maximizing power production/minimizing structural loads Searching optimal power Conventional MPC Logic control 1 MW 1.1 kW [23] [24] wind wind wind wind direction direction direction direction 588 Energy Conversion and Management 157 (2018) 587–599 D. Song et al. • • • direction, which aims at extracting maximum wind power for the WTs with a moderate yaw actuator usage. With regards to the literature, the major contributions of this paper are threefold. power point tracking and yaw control technique aiming at tracking the optimal rotor speed was presented in [14]. In theory, this type of controls may provide better performance than normal measurement-based one which may suffer from the inaccurate measurements disturbed by operation of the WTs. However, the MPP of WTs is changing following the variation of wind speed, besides the wind direction. As a result, the real difficulty for controls without measurements consists in locating the MPP, which remains an open issue in wind energy research community. Controls with normal measurements, which are currently widely employed by modern WTs. This type of controls employs active yaw control strategies to face the turbine into the wind by acquiring signals from wind vanes installed at the rear of the nacelle. Although a fuzzy-PID strategy is introduced to track the wind direction, the motivation is unclear [15]. By comparison, most of control strategies employ some predefined logic controls [16–18], where the yaw actuators are activated when the yaw error measured by wind vanes exceeds some thresholds. Although the strategies are simple, the difficulty consists in obtaining a proper reference to adjust the nacelle position. The measurements from the wind vanes are always mixed with disturbing noises and outliers. Meanwhile, the wind direction constantly changing is different from the future wind direction. Consequently, the controls with normal measurements could not provide sufficient performance [9–11]. Controls with advanced measurements, which have been recently proposed in some advanced wind energy projects. To obtain accurate wind direction measurement, remote sensing instruments based on Lidar and hypersonic (Sodar) technologies have been employed. With the powerful remote measurement, performance of the yaw control system can be potentially enhanced with simple logic controls [19–21]. Under the assumption that the wind direction preview provided full information about wind direction over the future 60 s, a conventional model predictive control (MPC) can provide an increment of 8% wind power extraction and some fatigue load reductions during an extreme direction change [22]. Nevertheless, the solutions are expensive and thus, affordable only for high power WTs [14]. Controls with indirect measurements, which have gradually gained attentions by researchers. The short-term prediction of wind direction is incorporated into a conventional MPC for the yaw control system, which aims at achieving structural loads minimization and power production maximization simultaneously [23]. Besides, wind direction is estimated by an inverted function of wind power and wind speed, and then is employed into the yaw control system with logic controls [24]. The controls with indirect measurements may be potential for improving performance yaw system, but the presented control solution ignoring prediction algorithms is incomplete and needs to be further investigated. • This paper proposes a novel control structure that consists of a wind • • predictive model and a novel control method. To do this, a hybrid autoregressive integrated moving average method-based Kalman filter (ARIMA-KF) model is used to predict the wind direction. Then, two novel yaw control strategies are proposed: one created by using the predicted wind direction as the tracking reference, and the other based on a model predictive control (MPC). This paper introduces the novel MPC with a finite control set for controlling the yaw system. The predicted wind directions are discrete data obtained at every sampling period, and thus the predictive model for yaw control system including predicted wind directions are internally discrete model. Thus, compared with the conventional MPC with a continuous control set, the proposed MPC strategy is more suitable for controlling yaw systems using predicted wind directions. On one hand, the yaw command sets during each sampling period can be categorized into a finite control set rather than a continuous control set. On the other hand, the algorithm solution is directly selected from available control sets, and thus reduces the computational burden. This paper discusses how wind power extraction of the WTs can benefit from the predicted wind direction-based control structure and introduces four performance indexes to evaluate the yaw control system performance, namely yaw error, yaw action time, yaw action count, and power reduction factor. The remainder of this work is organized as follows: the wind direction measurement, two industrial yaw control methods, and performance indexes of yaw system are discussed in Section 2; and Section 3 describes the novel yaw control solution. It is followed by simulation tests and result discussions in Section 4. Finally, conclusions are drawn in Section 5. 2. Yaw control system of industrial WTS 2.1. Wind direction measurement 2.1.1. Wind direction sensor For current industrial WTs, the wind direction measurement is normally provided by one or two wind direction sensors which are installed on the rear of the nacelle. A typical wind direction sensor is shown in Fig. 1, which is a product of Kriwan with number INT30 [33]. Its basic specification is given in Table 2. 2.1.2. Wind direction measurement Fig. 2 shows the principle of wind direction measurements. Since the wind direction sensor rotates along with the WT’s nacelle, it measures a yaw error rather than the wind direction. Besides the wind From the above, it is concluded that controls with advanced measurements and with indirect measurements may improve performance for the yaw system, because the wind direction information in the future can be utilized. By comparison to the advanced measurements, the indirect measurements are normally cost-effective. Until now, some developed forecasting approaches have been proposed, such as the wind-power prediction by Azimi et al. [25], Yesilbudak et al. [26], and Mohammadi et al. [27]; wind-speed prediction by Zameer et al. [28], Zhang et al. [29], and Noorollahi et al. [30]; and wind direction prediction by Ouyang et al. [31] and Song et al. [32]. These studies addressed the prediction issues relevant to the wind source, but none of them tried to employ the predicted wind data into the control application of the WTs. Motivated by the aforementioned observations, this study proposes a novel control solution by taking advantage of the predicted wind Fig. 1. The outlook of a typical wind direction sensor [33]. 589 Energy Conversion and Management 157 (2018) 587–599 D. Song et al. to reduce θye to zero. Since θye is normally disturbed by operation of the WT operation, it is usually filtered before being employed as the controller reference. Table 2 Specifications of a typical wind direction sensor [33]. Parameter Value Measuring range Resolution Accuracy Start-up wind speed Permitted ambient temperature Permissible relative humility 0–360° <1° ± 2.5° <0.4 m/s −40 to + 70° 0–100% . r . h. 2.2. Yaw control methods with normal measurements For megawatt WTs, the yaw speed is normally designed in a range of [0.2deg/s,0.8deg/s], because a fast movement of the yaw system may induce high loads to the WT. Meanwhile, to avoid over-usage of yaw actuator, the yaw system is always activated at discrete intervals. The yaw control methods with normal measurements for modern WTs normally employ the logic controls, where the yaw actuators are activated when the yaw error measured by wind vanes exceeds some thresholds. In this study, two yaw control algorithms with normal measurements are given and used as the baseline control algorithms. North Wi nd dir ect 2.2.1. Yaw control algorithm I The yaw control algorithm I is taken from the MY (Ming Yang) 1.5 MW WTs manufactured by China Ming Yang Wind Power, which is illustrated in Fig. 3 and includes following four steps [18]: ion Blade rotor (1) Yaw error filtering. In this step, the yaw error is averaged by three averaged units with different averaged periods: 10 s, 30 s and 60 s; 10s 30s 60s and thus, three averaged yaw errors (denoted as θye ,θye ,θye ) are obtained. 10s 30s 60s (2) Yaw error judgment. In this step, θye ,θye ,θye are compared to three predefined amplitude thresholds (denoted as Ah1,Ah2,Ah3) respectively. When any of the three comparisons is satisfied, and the sustaining time is longer than the corresponding predefined time thresholds (denoted as Th1,Th2,Th3), the control loop goes on; otherwise, the control loop is ended for this cycle. (3) Yaw time calculation. In this step, the yaw action time is calculated using the corresponding averaged yaw error divided by the yaw speed (Yawspeed ). (4) Yaw movement. In this step, yaw movement is activated during the activation time. As a consequence, the nacelle moves toward facing the wind. wd np ye Nacelle Wind wane Fig. 2. Schematic diagram of wind direction measurement [32]. direction sensor, there is another transducer used in yaw control system, namely, nacelle position transducer. The yaw error is the difference between the wind direction and nacelle position, which can be calculated by θye = θwd−θnp In Fig. 3, the parameters utilized are summarized in Table 3. 2.2.2. Yaw control algorithm II The yaw control algorithm II is taken from the NREL CART3 (Controls Advanced Research Turbine 3-Bladed) turbine, which is illustrated in Fig. 4. The control logic is comparably simple and detailed in [16,19,20]. The yaw error is filtered by two low-pass filters, one with a time constant of 1 s, and the other 60 s, producing a more quickly (1) Since the wind direction varies along with the time, the yaw control system is needed to adjust the nacelle position to track the wind direction. From Eq. (1), it is obtained that tracking the wind direction is Step 2 Step 1 10s ye T( 10 s ye Step 3 Ah1) Th1 10 s ye Yes Yaw _ speed No ye 30s ye T( 30 s ye Ah2) Th2 Yes 30 s ye Yaw _ speed No 60s ye T( 60 s ye Ah3) Th3 Fig. 3. Schematic of the yaw control algorithm for the MY 1.5 MW WTs [18]. 60 s ye Yes Yaw _ speed No Do nothing this cycle 590 Step 4 Start yaw movement Energy Conversion and Management 157 (2018) 587–599 D. Song et al. Table 3 Yaw control parameters used in Fig. 3. wd Parameter Ah1 Ah2 Ah3 Th1 Th2 Th3 Values 13° 10° 8° 10 s 5s 5s 2 changing measurement of error and a more slowly changing measurement. The quickly changing measurement error is integrated and monitored. When the integrated error (notated AccErr in Fig. 4) reaches a value such that it has been off by 10 degrees for 10 min, the yaw angle of the turbine is moved to the location given by the slowly changing measurement of the error. (cos(θye ))eq = N (2) ̇ (t )| > 0) (|θnp (3) The yaw action count (denoted as Cyaw ) is the activation count of the yaw actuators. When the current yaw speed is different from the last one, Cyaw is increased by one and it can be expressed as Cyaw (t ) = ̇ (t ) = θnp ̇ (t −1) ∃ θnp ⎧Cyaw (t −1), ̇ (t ) ≠ θnp ̇ (t −1) ⎨Cyaw (t −1) + 1, ∃ θnp ⎩ erfast Lowpass TC=1s N ∑ j=1 (cos(θyej))2 f j , θyej ∈ [−180°,180°] Wind direction prediction approaches have been recently addressed by some researchers, such as the data mining algorithm-based predicting approach by Ouyang et al. [31], and the ARIMA-KF based hybrid approach by Song et al. [32]. The ARIMA model is suitable for Fig. 4. Schematic of the CART3 yaw controller [16]. AccErr sign(erfast )* erfast Do nothing this cycle 2 erslow Lowpass TC=60s (7) 3.1. Wind direction prediction (4) Since θye has a direct influence on wind power extraction of the WT, the wind power Pa extracted by a horizontal axis WT should be also evaluated and it can be expressed by [5] ye (6) From the above, it is clear that the available yaw control algorithms depend entirely on the historical data of the measured yaw error. To handle the issue of the measured yaw error disturbed by operation of the WT, averaging or low-pass filters are normally utilized to provide the reference for the yawing movement, but they may bring about a time delay. Besides, the filtered measurements only reflect the past yaw error and may be different from the future wind directions. Therefore, it is reasonable to use prediction values, which may improve the performance of the yaw control system. Hence, this study proposes a novel solution. Fig. 5 shows its structure.Compared to the methods shown in Figs. 3 and 4, the proposed method comprises two parts: a wind direction prediction module (depicted as 1) and a novel yaw control module (depicted as 2). The module 1 is designed to estimate the future wind direction θwd (k + 1|k ) at the kth sampling period (Ts ), and the module 2 uses θwd (k + 1|k ) to make a decision for the yaw system action in the next control period. These two modules are explained as follows. |θye (t )| t=1 (5) 3. Novel yaw control solution N ∑ Yaw Command where (cos(θye ))eq is calculated by The yaw action time (denoted as tyaw ) is the activation time of the ̇ (t ) ), tyaw is yaw actuators. When there is a yaw speed (denoted as θnp cumulated and it can be expressed as tyaw = Novel yaw control ξ = Pred/ Pideal = 1−((cos(θye ))eq)2 t=1 ∑t = 1 (θye (t ))2 Yaw system To better check the influence of θye with different yaw control algorithm on Pa , we introduce the power reduction factor (denoted as ξ ) which is calculated based on the following formula N 1 N (k ) Pa = ρAr Cp V03 (cos(θye ))2 /2 To evaluate performance of the yaw control system, two aspects should be taken into account, one to measure the accuracy of the wind direction tracking, and the other to measure the usage of the yaw actuators. The accuracy of the wind direction tracking can be evaluated by checking the statistical value of yaw error, and the usage of the yaw actuators by calculating the yaw action time and action count. The yaw error is the difference between the wind direction and nacelle position. In this paper, mean absolute error (MAE) and root mean squared error (RMSE) of θye are calculated to evaluate the accuracy of the wind direction tracking. Their calculations are expressed as ∑ wd Fig. 5. Block diagram of the novel yaw control solution. 2.3. Performance indexes for evaluating yaw control system ⎧ 1 ⎪ MAE (θye ) = N ⎨ ⎪ RMSE (θye ) = ⎩ 1 (k 1| k ) Wind direction prediction Yes Yaw position AccErr yaw setpoint Yaw to yaw setpoint Yes Yaw setpoint in cone? Threshold No Stop the turbine 591 No Energy Conversion and Management 157 (2018) 587–599 D. Song et al. errors rather than the filtered measurements. These differences may change the performance of the yaw control system, which will be discussed in Section 4. capturing short-range correlations and has been widely in a variety of forecasting applications [34], but it has a difficulty of adjusting the model’s parameters when the time series contains new information. Thus, the ARIMA model in combination with a Kalman filter is used to solve this problem [35]. In this study, the ARIMA-KF approach is employed because of its capability of providing accurate short-term prediction results. Here, it is worthy noticing that the ARIMA-KF methods is not the only one that can be utilized here, and other machine learning-based prediction approaches may provide more accurate results [36,37]. To build an ARIMA-KF predictive model for predicting the wind direction, the relevant procedures refer to three steps, which are demonstrated in Fig. 6. Its details are given as follows: 3.2.2. Method II The MPC enables a novel solution of a control problem while honoring the constraints imposed upon by the designers of the WT [38]. Meanwhile, it has a potential advantage that it is able to predict the behavior of a plant in future using its model. There are two types of MPC [39]: the MPC with a continuous control set, which requires solving a quadratic programming problem on line and thus has a computation burden, and the MPC with a finite control set, which has recently been proposed to control pitch angle and generator torque for the WTs [40]. The control command for the yaw system can be categorized into a finite control set; therefore, the MPC with a finite control set is proposed and employed in this study. As suggested in [39,40], the MPC with a finite control set can be developed by following four steps: (1) Choosing the original wind direction series. Since the yaw controller uses the mean values of wind direction data, three types of wind direction data with different averaged periods: 10 s, 30 s and 60 s, are used to form the original data series. (2) Defining the ARIMA initial model. In this step, the three types of original data series are used to train the ARIMA predictive models, in which a three iterative step is included: model identification, parameter estimation, and diagnostic checking. (3) Designing the KF-based predictive model. The KF is generally described as an approach consisting of a state equation and a measurement equation. The system state equation is created by reformulating the ARIMA model, and the measurement equation is defined according to the relation of the measurement variables and system state variables. Finally, the newly predicted values are obtained by using the KF iteration algorithm. KF algorithm also includes three steps: system modeling, measurement update and time update filter gain. (4) Outputting one-step prediction values. The ARIMA-KF predictive models may provide multiple-step predictions, but multiple-step predictions are with the considerable prediction error. Thus, only the one-step prediction values are used in this study. (1) (2) (3) (4) Define a quality function (QF). Build a model of the target system and its possible control set. Build a model of the controlled variables for prediction. Evaluate the predicted value for each control set and select the one with minimal value of QF. The block diagram of the proposed MPC-based yaw control method is shown in Fig. 8, which mainly refers to the construction of the predictive model and the QF. In this study, the primary control objective is to track the predicted wind direction, which in other words is to decrease the yaw error. Thus, the predictive model is needed to predict the yaw error. And, the secondary control objective is to avoid the yaw actuator over-usage. In the QF, these two control objectives can be evaluated using a convex function. In this way, both objectives of wind direction tracking and actuator usage can be taken care by adjusting the corresponding weighting factors. The proposed method is detailed as follows. 3.2. Two novel yaw control methods a. Finite control set for yaw actuator In this section, the two novel yaw control methods are proposed and discussed: one is created by using wind direction predictions as the tracking references (Method I), and the other is based on the MPC using a finite control set (Method II). For modern WTs, the yaw system is driven by yaw motors. When the yaw system is activated, the nacelle will be moved by the yaw motors at a certain speed; otherwise, the nacelle position remains unchanged. The permissible actions for the yaw actuators comprise three elements and they can be categorized into a finite set as follows: 3.2.1. Method I The schematic diagram of Method I is illustrated in Fig. 7, where three types of predicted wind directions with different sampling periods: 10 s, 30 s, and 60 s, are used to provide the tracking references. As a consequence, the resulted controller has three control periods. In other words, the controller updates its output when the inputs are updated at each 10 s, 30 s, and 60 s. When comparing with the yaw control method shown in Fig. 3, the proposed method shown in Fig. 7 has a main distinction that the update period of its output depends on Ts of the predictive model. Besides, it differs in following details: in Step 1, its inputs uses three types of predicted yaw errors instead of filtered measurements; in Step 2, it employs only three amplitude thresholds ( Ah1,Ah2,Ah3), without time thresholds; and in Step 3, its tracking references use the prediction yaw 1 Original wind direction series 2 ARIMA initial model ⎧− Yawspeed, j = 0 ̇ (j ) = Yawspeed, j = 1 θnp ⎨ ⎩ 0, j = 2 (8) b. Predictive model Since the control objective is to decrease the yaw error, the yaw error is selected as the predictive variable, which can be expressed as follows: ̇ (j ) Tc θye (k + 1|k ) = θwd (k + 1|k )−θnp (k )−θnp (9) where the control period Tc is equal to the sampling period Ts . 4 3 One-step prediction KF model 592 Fig. 6. Development of the ARIMA-KF wind-direction prediction model. Energy Conversion and Management 157 (2018) 587–599 D. Song et al. Step 1 wd (k 1| k )Ts 10 s ye (k 1| k )Ts wd Step 2 10 s (k 1| k ) ye ye (k ) Step 3 (k 1| k )Ts 10 s Yes Ah1 ye (k 1| k )Ts 10 s Yaw _ speed Step 4 No wd (k 1| k )Ts 30 s ye (k 1| k )Ts wd 30 s (k 1| k ) ye (k 1| k )Ts ye (k ) 30 s Ah2 60 s Ah3 Yes ye (k 1| k )Ts Start yaw movement 30 s Yaw _ speed No wd (k 1| k )Ts 60 s ye (k 1| k )Ts 60 s wd ( k 1| k ) ye ye ( k ) (k 1| k )Ts Yes ye (k 1| k )Ts 60 s Yaw _ speed No Do nothing this cycle Fig. 7. Schematic diagram of method I. The yaw system w1 , w2 , w3 np ( Minization ye ( k 1| k ) of QF j) Predictive model Fig. 8. Block diagram of method II. wd ( k ) np ( k ) Wind direciton prediction 2 1 np ( k ) wd ( k model. The first two elements are linear functions which are the penalties on the increment of yaw error and yaw actuator action distance, whereas the third element is established to punish the action count, which uses a logical function where the value is w3 when the current yaw action is different from the last one. c. Quality function The quality function is a representation of the control objectives which are usually related to make the variables follow the references. Hence, the quadratic value of the error, or its absolute value, is commonly employed to find the minimum value of the quality function. In this study, minimization of yaw error and yaw actuator usage should be achieved simultaneously, so they will be combined in a form of sum. Since the yaw error has a direct influence on the power production of the WT in a cosine-squared fashion, the first part (QF (1) ) of the quality function can be chosen as follows: QF (1) = w1 (θye (k + 1|k ))2 Apply Obtaining (10) np np (k ), (op) wd (k 1| k ) QF (op) where w1 is the weighting factor to be adjusted. The yaw actuator usage can be evaluated by two indexes: actuator action distance and action count. These two indexes are combined to form the second part (QF (2) ) of the quality function, which is given as follows: ̇ (j )|Tc + w3 (θnp ̇ (k + 1) ≠ θnp ̇ (k )) QF (2) = w2 |θnp 1| k ) for ye end for(j=0:2) (k 1| k ) wd (k 1| k ) np (k ) np ( j)Tc (11) Combining Eq. (11) with Eq. (10), the final QF is obtained as follows: ̇ (j )|Tc + w3 (θnp ̇ (k + 1) ≠ θnp ̇ (k )) QF = w1 (θye (k + 1|k ))2 + w2 |θnp If(QF(j)<QF(op)) QF(op) = QF(j); op=j; End (12) As shown in Eq. (12), it is clear that one of the aspects where the MPC shows its flexibility is the inclusion of different elements in a Fig. 9. Flowchart of the implemented control algorithm. 593 Energy Conversion and Management 157 (2018) 587–599 D. Song et al. Wind direction [degree] 250 Fig. 10. Original wind data in a specific day: (a) wind direction; (b) wind speed. Original wind direction Averaged wind direction/10s Averaged wind direction/30s Averaged wind direction/60s 200 150 100 50 0 0 4 8 12 16 20 24 16 20 24 Time [h] (a) Wind speed [m/s] 20 15 10 5 0 0 4 8 12 Time [h] (b) The control algorithm is detailed in Fig. 9 as a flow diagram. As shown in Fig. 9, the minimization of the quality function is implemented as a “for” cycle predicting for three permissible control action, evaluating the quality function, and storing the minimum value and the index value of the corresponding control action. By the flowchart, the proposed control algorithm can be easily implemented in the wind turbine systems. data shown in Fig. 10(a), where the first 20 h are used to establish the models, and the leaving 4 h to check the model validity. Since the results have been detailed in [32], only the wind direction prediction results in the form of time series are chosen to show and illustrated in Fig. 11. Besides, three types of performance indexes (including MAE, RMSE, and MAPE) are computed and summarized in Table 4. These results show that the three types of the prediction results are in good agreements with the original data. Therefore, these results will be employed in the proposed novel yaw methods. 4. Method validations 4.2. Validation of novel yaw control methods The proposed methods are tested using wind data obtained from a wind farm located in Guangdong Province of China, which was saved in a SCADAS (Supervisory Control and Data Acquisition System) per second. In this study, the data used are referred to a specific day, including 86, 400 data points. Fig. 10(a) and (b) shows the original wind direction and wind speed series, respectively. Besides, the averaged wind direction series are also illustrated in Fig. 10(a), which refer to mean values averaged in 10 s, 30 s and 60 s periods, respectively. As shown in Fig. 10, both the original data of wind direction and wind speed are mixed with high-frequency noises. After averaging, the wind direction data become much smooth, and thus they can be utilized as inputs for the conventional yaw controllers. After obtaining the wind direction prediction data, two novel yaw controllers were developed using the Matlab software: the first one (denoted as C1) followed the schematic illustrated in Fig. 7 and took the parameters given in Table 3, whereas the second one (denoted as C2) employed the flow chart illustrated in Fig. 9. For the C2, the control period and the weighting factors in the QF have to be determined in advance. In this study, the control period is set to Tc = 10 s , which is chosen based on the fact that the predicted wind direction can provide data per 10 s. By following the trial and error procedures introduced in [41], the three weighting factors are chosen as: w1 = 1.0 , w2 = 14.0 and w3 = 50.0 , respectively. Besides, the yaw speed is set to 0.5 deg/s, which is the designing parameter of the yaw system for the MY 1.5 MW WT. 4.1. Wind direction prediction results 4.2.1. Results of time series By using the wind direction data shown in Fig. 10(a), the two proposed yaw controllers are simultaneously simulated in parallel with d. Control algorithm The ARIMA-KF predictive model is trained by using wind direction 594 Energy Conversion and Management 157 (2018) 587–599 D. Song et al. Fig. 11. Wind direction prediction results of (a) 10 s; (b) 30 s; (c) 60 s. the performance of a yaw controller can be evaluated in terms of three performance indexes: yaw error, yaw action time and yaw action count, these data were calculated during the simulation. The simulation results are shown in Fig. 12, which are obtained from the four controllers and refer to nacelle position, yaw error, yaw action time and yaw action count, respectively. In Fig. 12, all curves of nacelle position, yaw error, yaw action time and yaw action count under the four controllers presented similar trends, which prove that all the four controllers were in a normal work condition. However, there were obvious differences among yaw error, yaw action time and yaw action count under different controllers. Fig. 12(b) shows that the variations of yaw error from different Table 4 Indexes of the three prediction results. MAE RMSE MAPE Time series of 10 s Time series of 30 s Time series of 60 s 0.9218 1.3038 15.74% 0.9573 1.2839 12.91% 1.1954 1.6431 12.03% each other. Also, for the sake of comparison, the results of the two normal yaw controllers were collected together, where the ones from MY and NREL are denoted as C4_MY and C3_NREL, respectively. Since 595 Energy Conversion and Management 157 (2018) 587–599 D. Song et al. Nacelle position [degree] 200 Fig. 12. Simulation results of the four controllers for: (a) nacelle position; (b) yaw error; (c) yaw time; (d) yaw action count. C3_NREL C4_MY C1 C2 150 100 50 0 4 8 12 16 20 24 Time [h] (a) Yaw error [degree] 50 C3_NREL C4_MY C1 C2 25 0 -25 -50 0 4 8 12 16 20 24 Time [h] (b) 600 C3_NREL C4_MY C1 C2 Yaw time [min] 500 400 300 200 100 0 0 4 8 12 16 20 16 20 24 Time [h] (c) 1400 C3_NREL C4_MY C1 C2 Yaw count [-] 1200 1000 800 600 400 200 0 0 4 8 12 Time [h] 24 (d) controllers. It can be found out that among the four controllers, C3_NREL has the biggest variation of the yaw error, C4_MY comes secondly, and C1 and C2 have similar and smaller variations of the yaw error. Fig. 12(c) and (d) shows the yaw time and yaw count, respectively. It is very clear that the trends of yaw time for different controllers are similar as the ones of yaw count. Among the four 596 Energy Conversion and Management 157 (2018) 587–599 D. Song et al. 3.41%. By comparison to the ones of the former two controllers, C1 and C2 have less power reductions, being 2.43% and 2.24%, respectively. These numerical data well justify the developed novel controllers in the aspect of maximizing wind power extraction. The above observations prove that C2 noticeably outperforms the other three controllers in term of performance of wind power extraction with a moderate yaw actuator usage, whereas C1 outperforms the two industrial controllers in term of wind power extraction at the expense of yaw actuator usage. Table 5 Statistical data comparisons among the four controllers. C3_NREL C4_MY C1 C2 MAE (θ ye ) 9.69 7.96 6.73 6.23 RMSE (θ ye ) 12.88 10.57 8.87 8.18 tyaw [min] 85.4 359.4 523.3 328.9 Cyaw [−] 168 876 1377 622 ξ [%] 4.53 3.41 2.43 2.24 5. Conclusions and future work controllers, C1 and C4_MY take the first and second places of spending the most yaw time and yaw count, respectively, C2 comes next, and C3_NREL is with the last place. These observations indicate a fact that the four controllers have different performance: C3_NREL has the best and the worst performance in the aspect of yaw actuator usage and yaw error reduction, respectively; C1 and C2 outperform the other two controllers in terms of yaw error reduction, but C1 has a big payback of yaw actuator usage; overall, C2 has the most favorable performance. To operate the yaw system of the WTs effectively, a novel yaw control solution using predicted wind direction data has been proposed in this study. Using discrete data obtained at every sampling time for prediction, this feature of the predicted wind direction was fully taken into consideration for constructing the two novel methods. Three kinds of wind direction prediction data (10 s, 30 s and 60 s mean values which are provided by a hybrid ARIMA-KF model) were used by the first novel yaw control method as the tracking references. The second novel method employed the MPC with a finite control set, which utilized the predicted wind direction with 10 s mean value to estimate future yaw error and achieved multiple control objectives through explicitly including them into the quality function. The two novel methods are transformed to two novel controllers, whose performance is compared to the one of two industrial yaw controllers. Finally, the proposed two novel yaw controllers developed by the proposed methods demonstrated the superiority and the applicability through some simulation tests using wind direction data obtained from a wind farm located in Guangdong Province of China. That is, comparison results showed that the two novel controllers extracted 1% more wind power through an accurate wind direction tracking in comparison with the two industrial controllers. Also, C1 increased wind power extraction at the high expense of yaw actuator usage, while C2 showed its capability of maximizing power extraction and reducing yaw actuator usage at the same time, which owed to the quality function, constructed by adding penalties on the two interesting control objectives. Although encouraging results have been obtained by the proposed novel solution, there are some uncovered issues to be addressed in future studies, such as incorporating advanced machine-learning based prediction algorithms into the control solution, investigating the impacts of the prediction uncertainty on the control method, studying adaptive algorithms for parameter adjustment corresponding to different wind conditions. 4.2.2. Statistical data comparison and discussion To further evaluate the performance of the four controllers, the numerical results from Fig. 12(b–d) are collected to obtain statistical data, which are convenient for comparison and analysis. Accordingly, the four performance indexes discussed in Section 2.3 are calculated and summarized in Table 5.Table 5 shows the results of MAE (θye ) and RMSE (θye ) . Among the four controllers, C3_NREL has the biggest MAE (θye ) and RMSE (θye ) , which are 9.69 and 12.88, respectively. C4_MY has smaller results, where MAE (θye ) and RMSE (θye ) are decreased by 17.86% (being 7.96) and by 17.93% (being 10.57), respectively. C1 and C2 have closer results, whereas C2 has the smallest values, where MAE (θye ) and RMSE (θye ) are decreased by 35.71% (being 6.23) and by 36.49% (being 8.18), respectively. These numerical results are in good agreement with time series of yaw error shown in Fig. 12(b), and they reveal that the developed two novel controllers noticeably outweigh the other two controllers in reducing yaw error. Table 5 also shows the total yaw action time (tyaw ) and yaw action count (Cyaw ) of the four controllers during the day. Among the four controllers, C3_NREL has the least yaw action time and action count, which are 85.4 min and 168 times, respectively. C4_MY has a considerable yaw action time and action count, which are 359.4 min and 876 times, respectively. It means that the yaw actuators keep working for one fourth of the time on that day, and are activated per two minutes. The lengthy working time and frequent start may bring high loads, thus shorten the lifetime of the yaw actuators. However, C1 has a longer yaw time and frequenter action, which are increased by 45.6% (being 523.3 s) and 57.2% (being 1377 times), respectively. The payback of reducing yaw error for C1 is enormous but reasonable, because there is no time threshold for the C1. By comparison, C2 shows a satisfactory behavior: the yaw time is reduced by 8.6%, being 328.9 s, whereas the yaw action count is decreased by 54.8%, going down to 622 times. Finally, Table 5 shows the result of the reduction factor (ξ ) of different controller, the calculation of which is given in Appendix A. It can be seen that C3_NREL has the most power reduction caused by the yaw error, which is 4.53%, and C4_MY comes the second with the value of Acknowledgments This work is supported by the National Natural Science Foundation of China under Grant 51777217, the Project of Innovation-driven Plan in Central South University, the Australian Research Council under Grant DE160100675 and the Project funded by China Postdoctoral Science Foundation (2017M622605). This work is also financially supported by the Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education, NRF-2016R1A6A1A03013567. Appendix A The reduction factor ξ is calculated using Eqs. (6) and (7). To do this, θyej and their distribution probability ( f j ) are collected and calculated based on the results in Fig. 12(b). The results of θyej and f j are shown in the form of histogram in Fig. 13(a) and (b), respectively. Based on these distribution data, ξ of the four controllers can be calculated and obtained. 597 Energy Conversion and Management 157 (2018) 587–599 D. Song et al. Distribution probability [%] 5 Fig. 13. Distribution probability of yaw error under (a) C4_MY and C3_NREL; (b) C1 and C2. C3_NREL C4_MY 4 3 2 1 0 -30 -20 -10 0 Yaw error [degree] 10 20 30 (a) Distribution probability [%] 6 C1 C2 5 4 3 2 1 0 -30 -20 -10 0 Yaw error [degree] 10 20 30 (b) References [16] Kragh KA, Fleming PA. Rotor speed dependent yaw control of wind turbines based on empirical data. AIAA aerospace sciences meeting including the new horizons forum and aerospace exposition. 2012. [17] Bu FF, Huang WX, Hu YW, Shi K, Wang QS. Study and implementation of a control algorithm for wind turbine yaw control system. Proceedings of the 2009 world nongrid-connected wind power and energy conference, Nanjing, China, 24–26 September 2009. 2009. p. 1–5. [18] China Mingyang wind power. MY 1.5MW safety and control scheme; 2009. [19] Scholbrock AK, Fleming PA, Wright A, Slinger C, Medley J, Harris M. Field test results from lidar measured yaw control for improved power capture with the NREL controls advanced research turbine. In: 33rd Wind energy symposium (AIAA). Proceedings of SPIE - the international society for optical engineering, vol. 5685, issue no. 6; 2014. p. 1080–4. [20] Fleming PA, Scholbrock AK, Jehu A, Davoust S, Osler E, Wright AD, et al. Field-test results using a nacelle-mounted lidar for improving wind turbine power capture by reducing yaw misalignment. J Phys Conf Ser 2014;524(1):012002. [21] Kragh KA, Hansen MH. Potential of power gain with improved yaw alignment. Wind Energy 2015;18(6):979–89. [22] Spencer M, Stol K, Cater J. Predictive yaw control of a 5 MW wind turbine model. AIAA aerospace sciences meeting including the new horizons forum and aerospace exposition. 2013. [23] Hure N, Turnar R, Vasak M, Vasak M, Bencic G. Optimal wind turbine yaw control supported with very short-term wind predictions. IEEE international conference on industrial technology. 2015. p. 385–91. [24] Tsioumas E, Karakasis N, Jabbour N, Mademlis C. Indirect estimation of the YawAngle misalignment in a horizontal axis wind turbine. IEEE 11th international symposium on SDEMPED. 2017. p. 45–51. [25] Azimi R, Ghofrani M, Ghayekhloo M. A hybrid wind power forecasting model based on data mining and wavelets analysis. Energy Convers Manage 2016;127:208–25. [26] Yesilbudak M, Sagiroglu S, Colak I. A novel implementation of kNN classifier based on multi-tupled meteorological input data for wind power prediction. Energy Convers Manage 2017;135:434–44. [27] Mohammadi K, Alavi O, Mcgowan JG. Use of Birnbaum-Saunders distribution for estimating wind speed and wind power probability distributions: a review. Energy Convers Manage 2017;143:109–22. [28] Zameer A, Arshad J, Khan A, Raja MAZ. Intelligent and robust prediction of short term wind power using genetic programming based ensemble of neural networks. Energy Convers Manage 2017:361–72. [29] Zhang C, Zhou JZ, Li CS, Fu WL, Peng T. A compound structure of ELM based on [1] Chehouri A, Younes R, Ilinca A, Perron Jean. Review of performance optimization techniques applied to wind turbines. Appl Energy 2015;142(4):361–88. [2] Song DR, Yang J, Cai ZL, Dong M, Su M, Wang YH. Wind estimation with a nonstandard extended Kalman filter and its application on maximum power extraction for variable speed wind turbines. Appl Energy 2017;190:670–85. [3] Bontempo R, Manna M. The axial momentum theory as applied to wind turbines: some exact solutions of the flow through a rotor with radially variable load. Energy Convers Manage 2017;143:33–48. [4] Burton T, Sharpe D, Jenkins N, Bossanyi E. Wind Energy Handbook. Wiley; 2011. [5] Medici D. Experimental studies of wind turbine wakes: power optimisation and meandering. Ph.D. thesis, KTH, Mechanics. [6] Schepers JG. IEA annex XX: dynamic inflow effects at fast pitching steps on a wind turbine placed in the NASA-Ames wind tunnel. ECN, report; 2007 [7] Boorsma K. Power and loads for yawed flow conditions. ECN, report; 2012 [8] Kragh KA, Hansen MH. Load alleviation of wind turbines by yaw misalignment. Wind Energy 2014;17(7):971–82. [9] Pedersen TF, Paulsen US, Pedersen SM, Enevoldsen P. Operational experience and analysis of a spinner anemometer on a MW size wind turbine. Conference proceedings (online) European Wind Energy Association. 2008. [10] Ribrant J, Bertling LM. Survey of failures in wind power systems with focus on Swedish wind power plants during 1997–2005. IEEE Trans Energy Convers 2007;22(1):167–73. [11] Pérez JMP, Márquez FPG, Tobias A, Papaelias M. Wind turbine reliability analysis. Renew Sustain Energy Rev 2013;23(23):463–72. [12] Farret FA, Pfitscher LL, Bernardon DP. Sensorless active yaw control for wind turbines. Proceedings of the 27th annual conference of the IEEE Industrial Electronics Society, Denver, CO, USA, 29s November–2 December 2001, vol. 2. 2001. p. 1370–5. [13] Wu X, Liu Y, Teng W. Modified hill climbing method for active yaw control in wind turbine. IEEE control conference. 2012. p. 6677–80. [14] Karakasis N, Mesemanolis A, Nalmpantis T, Mademlis C. Active yaw control in a horizontal axis wind system without requiring wind direction measurement. IET Renew Power Gener 2017;10(9):1441–9. [15] Chen FQ, Yang JM. Fuzzy PID controller used in yaw system of Wind Turbine. IEEE international conference on power electronics systems and applications. 2009. p. 1–4. 598 Energy Conversion and Management 157 (2018) 587–599 D. Song et al. [30] [31] [32] [33] [34] [35] [36] Mostafavi ES, Mostafavi SI, Jaafari A, Jaafari A, Hosseinpour F. A novel machine learning approach for estimation of electricity demand: an empirical evidence from Thailand. Energy Convers Manage 2013;74(10):548–55. [37] Zhang YC, Liu KP, Qin L, An XL. Deterministic and probabilistic interval prediction for short-term wind power generation based on variational mode decomposition and machine learning methods. Energy Convers Manage 2016;112:208–19. [38] Gosk A. Model predictive control of a wind turbine Master thesis Lyngby (Denmark): Technical University of Denmark; 2011 [39] Rodriguez J, Kazmierkowski MP, Espinoza JR, Zanchetta PH, Abu-Rub, Young A, et al. State of the art of finite control set model predictive control in power electronics. IEEE Trans Ind Inf 2013;9(2):1003–16. [40] Song DR, Yang J, Dong M, Joo YH. Model predictive control with finite control set for variable-speed wind turbines. Energy 2017;126:564–72. [41] Cortés P, Kouro S, Rocca BL, Vargas R, Rodríguez J. Guidelines for weighting factors design in model predictive control of power converters and drives. IEEE international conference on industrial technology. 2009. p. 1–7. feature selection and parameter optimization using hybrid backtracking search algorithm for wind speed forecasting. Energy Convers Manage 2017;143:360–76. Noorollahi Y, Jokar MA, Kalhor A. Using artificial neural networks for temporal and spatial wind speed forecasting in Iran. Energy Convers Manage 2016;115:17–25. Ouyang T, Kusiak A, He Y. Predictive model of yaw error in a wind turbine. Energy 2017;123:119–30. Song DR, Yang J, Liu Y, Su M, Liu AF, Joo YH. Wind direction prediction for yaw control of wind turbines. Int J Control Autom Syst 2017;15(4):1720–8. KRIWAN industie-Elektronik GmbH. INT30 wind direction sensors product manual; 2016. < http://www.kriwan.com/en/search/?q=N234_N291_INT30_71000317/ > [accessed 16.09.17]. Bivona S, Bonanno G, Burlon R, Gurrera D, Leone C. Stochastic models for wind speed forecasting. Energy Convers Manage 2011;52:1157–65. Su ZY, Wang JZ, Lu HY, Zhao G. A new hybrid model optimized by an intelligent optimization algorithm for wind speed forecasting. Energy Convers Manage 2014;85(9):443–52. 599