METHOD AND SYSTEM FOR CONTROLLING A QUANTITY OF A WIND TURBINE BY CHOOSING THE CONTROLLER VIA MACHINE LEARNING

Abstract

The present invention relates to a method of controlling a wind turbine by automatic online selection of a controller that minimizes the wind turbine fatigue. The method therefore relies on an (offline constructed) database (BDD) of simulations of a list (LIST) of controllers, and on an online machine learning step for determining the optimal controller in terms of wind turbine (EOL) fatigue. Thus, the method allows automatic selection of controllers online, based on a fatigue criterion, and switching between the controllers according to the measured evolution of wind condition.

Claims

1.-10. (canceled)

11. A method of controlling a quantity of a wind turbine for which a list of plural controllers of the quantity of the wind turbine is available, comprising steps of: a) constructing a database offline by simulating, for each controller of the list and for plural wind data, a cost function representative of fatigue of the wind turbine; b) measuring wind data online; c) determining online a controller from the list that minimizes fatigue of wind turbine for the measured wind data by machine learning from the database; and d) controlling online the quantity of the wind turbine by use of the determined controller.

12. A control method as claimed in claim 11, wherein the plural controllers of the list are selected from among proportional integral PI controllers, at least one of Hoc regulators with different weighting functions, and linear quadratic regulators with different weightings, and model predictive controls with different weightings and LiDAR-based predictive controls.

13. A control method as claimed in claim 11, wherein the machine learning is implemented by a regression method predicting the fatigue of the wind turbine for each controller of the list or by use of a method of classifying the controllers of the list that minimizes cost criterion according to the measured wind data.

14. A control method as claimed in claim 12, wherein the machine learning is implemented by a regression method predicting the fatigue of the wind turbine for each controller of the list or by use of a method of classifying the controllers of the list that minimizes cost criterion according to the measured wind data.

15. A control method as claimed in claim 13, wherein the machine learning is implemented by use of a regression method carrying out steps of: i) the measured wind data; ii) performing a polynomial increase in the measured wind data; and iii) performing a linear regression of the polynomially increased wind data by use of a change in space of a target value.

16. A control method as claimed in claim 14, wherein the machine learning is implemented by use of a regression method carrying out steps of: i) the measured wind data; ii) performing a polynomial increase in the measured wind data; and iii) performing a linear regression of the polynomially increased wind data by use of a change in space of a target value.

17. A control method as claimed in claim 13, wherein the machine learning is implemented by use of a regression method based on a random forest method, a neural network method, a support vector machine method or a Gaussian process method.

18. A control method as claimed in claim 14, wherein the machine learning is implemented by use of a regression method based on a random forest method, a neural network method, a support vector machine method or a Gaussian process method.

19. A control method as claimed in claim 11, wherein an individual angle or the individual pitch of at least one blade of the wind turbine is controlled.

20. A control method as claimed in claim 12, wherein an individual angle or the individual pitch of at least one blade of the wind turbine is controlled.

21. A control method as claimed in claim 13, wherein an individual angle or the individual pitch of at least one blade of the wind turbine is controlled.

22. A control method as claimed in claim 14, wherein an individual angle or the individual pitch of at least one blade of the wind turbine is controlled.

23. A control method as claimed in claim 11, wherein the controllers of the list further account for a regulation error between a setpoint for regulating the quantity of the wind turbine and a measurement of the quantity of the wind turbine.

24. A control method as claimed in claim 11, wherein the wind data used for constructing the database results from measurements on the site of the wind turbine.

25. A control method as claimed in claim 11, wherein the wind data used for constructing the database is provided by a wind simulator.

26. A control method as claimed in claim 12, wherein the wind data used for constructing the database is provided by a wind simulator.

27. A control method as claimed in claim 13, wherein the wind data used for constructing the database is provided by a wind simulator.

28. A control method as claimed in claim 14, wherein the wind data used for constructing the database is provided by a wind simulator.

29. A control method as claimed in claim 15, wherein the wind data used for constructing the database is provided by a wind simulator.

30. A system for controlling a quantity of a wind turbine using the control method as claimed in claim 11, the control system comprising means for storing the controller list and the database constructed by simulating for plural wind data a cost function representative of fatigue of the wind turbine for each controller of the list, means for measuring wind data, means for determining a controller of the list that minimizes fatigue of the wind turbine for the measured wind data by machine learning from the database, and for controlling the quantity of the wind turbine.

Description

BRIEF DESCRIPTION OF THE FIGURES

[0048] Other features and advantages of the method according to the invention will be clear from reading the description hereafter of embodiments given by way of non-imitative example, with reference to the accompanying figures wherein:

[0049] FIG. 1 illustrates the steps of the control method according to an embodiment of the invention;

[0050] FIG. 2 illustrates the regression machine learning step according to a variant embodiment of the invention; and

[0051] FIG. 3 is a curve of the real and the estimated fatigue values for an example embodiment of the invention.

DETAILED DESCRIPTION OF THE INVENTION

[0052] The invention relates to a method of controlling a quantity of a wind turbine in order to minimize the fatigue of the wind turbine or of at least a part of the wind turbine (that is a wind turbine component) according to measured wind data. The method according to the invention is based on the selection of the optimal controller (in terms of fatigue) by machine learning. The principle develops a learning algorithm allowing construction of a map relating measured wind conditions to a mechanical fatigue quantity. One of the goals of the invention can be to create a substitution model for estimating the service life of the wind turbine in a wind farm with an almost instantly given wind distribution.

[0053] In the rest of the description, the expression “wind turbine fatigue” also designates the fatigue of at least one turbine component.

[0054] A wind turbine quantity is understood to be any parameter of the wind turbine that can be controlled. According to a preferred embodiment, the quantity can be the individual inclination angle or the individual pitch of the blades used in the individual pitch control TPC.

[0055] Wind data is understood to be information relative to the incoming wind. This wind data can be measured notably by a LiDAR (laser imaging, detection and ranging) sensor, an anemometer or a SODAR (sonic detection and ranging) sensor, etc. By way of non-limitative example, wind data can notably comprise the following information: mean and standard deviation of the rotor averaged wind speed, horizontal and vertical gradients of the rotor averaged wind speed, pitch and yaw misalignments, rotor averaged wind turbulence intensity.

[0056] In order to select the optimal turbine quantity control, the method is based on the use of a predetermined list of a plural controllers (at least two controllers). Using plural controllers provides control adaptability to different wind conditions and it therefore enables optimal control whatever the wind conditions. The plural controllers of the list can be selected from among proportional integral PI controllers, and/or H∞ D regulators with different weighting functions, and at least one of linear quadratic regulators LQR, and model predictive controls MPC and/or LiDAR-based predictive controls with different weightings. The controller list can comprise controllers of the same type, that is several differently parametrized controllers.

[0057] The method according to the invention can then combine three aspects: control of a quantity of the wind turbine (individual blade control for example), wind characteristics that evolve slowly and turbine fatigue prediction. Substitution model techniques for fatigue can be used to predict a cost for the wind turbine subjected to the current wind for different controllers of a discrete set. This enables automatic online selection of the controllers, based on a fatigue criterion, and control of the wind turbine quantity by switching between controllers according to the wind condition evolutions.

[0058] The method according to the invention comprises the following steps:

[0059] 1) database construction;

[0060] 2) wind data measurement;

[0061] 3) determining the controller by machine learning; and

[0062] 4) controlling the wind turbine quantity.

[0063] Step 1) is carried out offline beforehand to limit the duration of the online control process. Moreover, the highest calculation cost of the method according to the invention is thus related to a step carried out offline.

[0064] Steps 2) to 4) are carried out online during operation of the wind turbine for real-time selection of the controller.

[0065] FIG. 1 schematically illustrates, by way of non-limitative example, the steps of the method according to one embodiment of the invention. A list LIST of controllers of the wind turbine quantity is determined beforehand. From this controller list LIST, and by use of simulations, database BDD representative of the turbine fatigue is constructed offline for the controllers of list LIST and for wind data. Wind data X is measured online. This measured wind data X, controller list LIST and database BDD are then used by a high-level controller CHN to determine, online, a controller K* of list LIST that minimizes the turbine fatigue for the measured wind data X. This determination of controller K* is performed by machine learning.

[0066] The determined controller K* is then used for online control CONT of the wind turbine quantity. In the embodiment illustrated, control CONT is carried out by considering a regulation error E corresponding to the difference between the regulation setpoint r of the turbine quantity and a measurement y of the turbine quantity y. Control CONT then generates a control signal u (an individual blade pitch for example) for wind turbine EOL. According to an embodiment option, regulation setpoint r can be, in most cases, zero so that the control, notably control IPC, can regulate the loads that unbalance the wind turbine to 0. In a variant, notably in the case of floating wind turbines, setpoint r can be given by an external controller for stabilizing and/or balancing the turbine.

[0067] The steps of the control method are detailed in the rest of the description.

[0068] 1. Database Construction

[0069] in this step, a database is constructed offline by simulating, for each controller of the predetermined list and for a plural wind data, a cost function representative of the wind turbine fatigue.

[0070] According to one embodiment of the invention, the plural wind data used for this step can be obtained by use of preliminary measurements on the wind turbine site. Thus, the database will be as representative as possible.

[0071] Alternatively, the plural wind data used for this step can be obtained by a wind simulator, for example the TurbSim™ software (NREL, National Renewable Energy Laboratory), which is a stochastic full-field turbulence simulator.

[0072] Simulation of the turbine behavior can be performed by a numerical simulator, for example an aeroelastic wind turbine simulator such as the FAST™ software (NREL, National Renewable Energy Laboratory).

[0073] A cost criterion J, whose complexity is not a limitation since it is assessed offline, is then designed. Typically, it is possible to use complex fatigue models of the mechanical elements of the turbine in order to have a cost criterion true to the damage undergone by the turbine. These models are most often not usable online because a rather long time series is necessary to evaluate the fatigue with these models. Thus, one advantage of the method according to the invention is that it can use complex turbine fatigue cost models that cannot be used directly online.

[0074] According to an implementation of the invention, the fatigue model can be a Palmgren-Miner model that counts the number of loading and unloading hysteresis loops. This counting may be discontinuous. Preferably, the counting method can be the rainflow counting method RFC. These counting methods do not allow fatigue to be expressed as the integral of an algebraic loading function, which is conventionally used in optimal control (cost of the integrals of quadratic functions). The integral of a quadratic cost function does not enable evaluation of the number of fatigue cycles, which is a problem when a compromise is to be assessed between the fatigues of various elements. One of the main advantages of the method according to the invention is to make it possible to integrate the fatigue calculation in the global control strategy.

[0075] Each simulation is evaluated with the previously designed cost criterion J. Thus, the simulation of the wind turbine subjected to the wind i, denoted by in a closed loop with controller K.sub.j belonging to the list K.sub.list, has a cost y.sub.ij=J(w.sub.i,K.sub.j). On the other hand, in order to reduce the number of variables and to simplify the problem, it is possible to extract from the wind measurements characteristics capable of unequivocally characterizing wind woo and which could be correlated with the value of the cost criterion. Function g giving, from wind the wind characteristic vector X.sub.i=g(w.sub.i) can then be defined.

[0076] 2. Wind Data Measurement

[0077] In this step, the wind data is measured online to know the incoming wind in real time.

[0078] According to one embodiment, these measurements can be carried out by a LiDAR sensor.

[0079] 3. Determining the Controller by Machine Learning

[0080] This step determines online the optimal controller in terms of wind turbine fatigue for the wind data measured in the previous step. The controller is determined from among the controller list by machine teaming using the database constructed in step 1) and the wind data measurements of step 2), as well as the controller list.

[0081] According to one implementation of the invention, two ways of combining the data and machine learning for selecting the controllers can be considered: cost prediction via regression techniques (one regression per controller in the list) and classification of the controllers that minimize the cost criterion according to the current wind (measured wind data).

[0082] According to a first embodiment, the regression can reconstruct the map J(w.sub.i,K.sub.j)=(g(w.sub.i), K.sub.j)=(X.sub.i, K.sub.j) with a function ƒ.sub.reg such that:

[00001]$f_{\mathrm{reg}}=\underset{\tilde{f}}{\mathrm{argmin}}\underset{i}{.Math.}\underset{j}{.Math.}.Math.\tilde{f}\left(X_i,K_j\right)-y\left(X_i,K_j\right).Math.$

where Y is a map associating wind characteristic vector X.sub.i and controller K.sub.j with the corresponding cost, ƒ defines a class of functions whose parameters are to be optimized so as to minimize the difference between the predictions of the model and the map. Function ƒ.sub.reg predicts the value of the cost criterion for the wind turbine in a closed loop with each controller of the list under the current wind (measured wind data). It is thus possible to select the controller K* that is best suited for the current wind conditions X (measured wind data), by taking the controller that minimizes the cost criterion according to function ƒ.sub.reg:

[00002]$K^{*}\left(X,f_{reg},K_{\mathrm{list}}\right)=\underset{K_j\in K_{\mathrm{list}}}{\arg \min }{f}_{\mathrm{reg}}\left(X,K_j\right)$

[0083] According to a second embodiment, the regression can comprise the following steps:

[0084] i) standardizing the measured wind data;

[0085] ii) performing a polynomial increase in the measured wind data; and

[0086] iii) performing a linear regression of the polynomially increased wind data by use of a change in space of the target value.

[0087] Standardization of the wind data allows the measured wind data to be brought to a centered normal law.

[0088] The polynomial increase corresponds to multiplying together the coordinates of the wind data vector up to a certain predefined degree. For example, data (x1, x2, x3) can be converted to (1, x1, x2, x3, x1x2, x1x3, x2x3, x12, x22, x32) for a polynomial increase of degree 2.

[0089] The space change of the target value can be a Box-Cox transformation allow adding a non-linearity at the output. In statistics, the Box-Cox transformation is a family of functions applied to create a monotonic transformation of data using power functions. Such a transformation allows stabilizing the variance, to make the data closer to a normal type distribution and to improve the measurement validity.

[0090] According to a third embodiment, the regression can be based on a random forest method, a neural network method, a support vector machine (SVM) method or a Gaussian process method.

[0091] According to one aspect of the invention, classification of the controllers can directly synthesize a function ƒ.sub.cl predicting the controller best suited for the current wind condition X, denoted by K.sup.+=ƒ.sub.cl(X). Function ƒ.sub.cl can be defined as follows:

[00003]$f_{cl}=\underset{\tilde{f}}{\arg \min }\underset{i}{.Math.}h\left(\tilde{f}\left(X_i\right),\underset{K_j\in K_{\mathrm{list}}}{\arg \min }y\left(X_i,K_j\right).Math.\right)$

where function h provide a good classification of the controller:

[00004]$h\left(K_a,K_b\right)=\{\begin{array}{c}1\mathrm{if}{K}_a\ne K_b\\ \mathrm{otherwise}\end{array}$

[0092] According to the initial results, the two methods (regression and classification) seem to be equivalent. It is noted that, according to the classification technique used, regression of a pseudo cost function (fatigue) can be performed. This cost function is the probability that a controller K is the most suitable controller under a wind condition X, denoted by p(X,K). Finally, the result of ƒ.sub.cl is the controller that maximizes this probability under a wind condition.

[0093] Regression has the advantage of predicting the (fatigue) cost directly. It is therefore possible to determine a threshold for controller switch and to limit switching from one controller to another only to the switches providing a net gain. Classification has the advantage of directly minimizing the classification error, and thus limiting risk of taking the wrong controller when selecting the most suitable controller.

[0094] 4. Controlling the Wind Turbine Quantity

[0095] This step controls online the wind turbine quantity by applying the controller determined in step 3).

[0096] According to an embodiment corresponding to FIG. 1, the controller is applied in the feedback loop. The controller accounts for the regulation error between a regulation setpoint and the turbine quantity measurement. In this case, the method can comprise a step of measuring the controlled wind turbine quantity.

[0097] Furthermore, the present invention relates to a system of controlling a wind turbine quantity, capable of implementing the method according to any one of the variant combinations described.

[0098] The control system comprises at least: [0099] means for storing the controller list and the database constructed by simulation; [0100] means for wind data measurement; [0101] means for determining a controller, which uses the controller list and the database of the means for storage and the wind data measurements of the means for measuring; and [0102] control means for applying the determined controller to the wind turbine.

[0103] According to one embodiment of the invention, the means for deter mining a controller and the means for storing can be a computer.

[0104] Moreover, the control system may comprise a numerical simulation computer for constructing the database.

[0105] The advantage of using the method according to the invention rather than conventional optimal control methods also intended to minimize a cost criterion is that significant latitude is provided to the cost criterion. Indeed, the method according to the invention allows any cost criterion to be used. It is therefore possible to use precise mechanical fatigue models that can only be used offline, unlike the conventional MPC (Model Predictive Control) models that require that the cost criterion can be continuously re-evaluated online.

[0106] The second advantage is that the method according to the example can allow optimizing the control over a very complex cost function using relatively simple control techniques, thereby having a very low online calculation cost. Furthermore, the control method according to the invention is intrinsically designed to adapt to various wind conditions, unlike most other control techniques based on linear models, which require an additional work of generalization to the different cases encountered by the wind turbine.

Example

[0107] Other features and advantages of the control method according to the invention will be clear from reading the description of the example hereafter.

[0108] In order to validate the control method according to the invention, the method was first tested with a wind data set generated by the TurbSim™ wind generator and simulated in closed loops on the FAST™ aeroelastic wind turbine simulator, with 4 controllers. The controllers considered are proportional integral (PI) IPC controllers corresponding to the one described in Bossanyi et al. (Bossanyi, 2003). For this example, a CPC controller mentioned in Jonkman et al. (Jonkman, 2007) provides good regulation of the rotor speed and power. A PI controller gives, from the regulation error between the measurement and the desired value ε(t), defined as the difference between the measured quantity to be regulated and the regulation setpoint, the input for the system to be regulated u(t) as follows:

u(t)=∫.sub.t.sub.0.sup.tK.sub.Iε(τ)dτ+K.sub.pε(t)

where K.sub.p and K.sub.I are the proportional and integrator coefficients that define the controller. The parameters of the 4 PI controllers considered in the example are:

TABLE-US-00001 TABLE 1 Controller K.sub.p K.sub.I 1 4 .Math. 10.sup.−5 3.2889 .Math. 10.sup.−5 2 4 .Math. 10.sup.−5 5.1556 .Math. 10.sup.−5 3 0.086 0.0031 4 0.0186 0.0066

[0109] The winds used to create the database (learning data) are non-uniform three-dimensional wind fields with coherent turbulences. For the learning data, 588 winds were generated with 147 combinations of parameters (average speed, direction, vertical speed gradient, turbulence intensity).

[0110] To be able to predict fatigue as a function of wind, the characteristics allowing to explain the fatigue that could be obtained from wind reconstruction algorithms need to be extracted from the wind.

[0111] From the TurbSim™ wind fields, the wind vector {right arrow over (V)}(t, y, z)=[u(t, y, z), v(t, y, z), w(t, y, z)].sup.T is obtained at the time t in the rotor plane where y and z are the horizontal and vertical coordinates of the field respectively. Let V be the norm L.sub.2 of vector {right arrow over (V)}(t, y, z).

[0112] The wind characteristics considered are the average and the standard deviation over the simulation time (300 seconds), starting at t0 and ending at tf, of the rotor averaged wind speed RAMS, of the horizontal and vertical gradients denoted by δy and δz, and of the pitch and yaw misalignments denoted by θ.sub.y and θ.sub.z. Finally, the rotor averaged turbulence intensity RATI is calculated for each simulation. The instantaneous values of RAWS, δy, δz, θ.sub.y and θ.sub.z, as well as the value for the entire simulation of RATI are mathematically expressed as follows:

[00005]$\mathrm{RAWS}=\frac{1}{S}{\int }_{S}Vds$${\delta }_y\left(t\right)=\frac{1}{S}{\int }_s\frac{\partial V}{\partial y}ds$${\delta }_z\left(t\right)=\frac{1}{S}{\int }_S\frac{\partial V}{\partial z}ds$${\theta }_y\left(t\right)=\frac{1}{S}{\int }_S{\tan }^{-1}\frac{w}{u}ds$${\theta }_z\left(t\right)=\frac{1}{S}{\int }_S{\tan }^{-1}\frac{v}{u}ds$$\mathrm{RATI}=\frac{1}{S}{\int }_S\frac{{\int }_{t_0}^{t_f}{V}^{2}dt-{\left({\int }_{t_0}^{t_f}\mathrm{Vdt}\right)}^2}{{\int }_{t_0}^{t_f}Vdt}\mathrm{ds}$

with S the rotor area and ds=dydz an infinitesimal surface of the rotor.

[0113] In this example, the machine learning function ƒ has the following structure, illustrated in FIG. 2 (by way of non-(imitative example): [0114] standardization STD of data X (distribution brought to a centered normal law) polynomial increase (multiplying together the coordinates of the vector up to a certain degree) [0115] (e.g. (x.sub.1, x.sub.2, x.sub.3).fwdarw.(1, x.sub.1x.sub.2, x.sub.1x.sub.3, x.sub.2x.sub.3, x.sub.1.sup.2, x.sub.2.sup.2, x.sub.3.sup.2) for a polynomial increase of degree 2). In the present case, the x are the wind characteristics. We then obtain data X.sub.poly [0116] linear regression REG from data X.sub.poly [0117] Box-Cox transformation TBC of the target value (Box & Cox, 1964), which allows a non-linearity to be added at the output

[00006]$y_{\mathrm{BoxCox}}=\frac{1}{\lambda }\left(y^{\lambda }-1\right)$ [0118] reverse Box-Cox transformation TBC-INV allowing to calculate the fatigue prediction.

[0119] w* is a vector which results from the optimization of these coefficients so as to minimize the difference between the predictions and the map in the Box-Cox space. The equation of shows how w* is used to predict the cost in the Box-Cox space from X.sub.poly.

[0120] This regression scheme is performed for each controller K.sub.j, and all these regressions give function ƒ. We can therefore write: ƒ(X, K.sub.j)=(X, K.sub.j).

[0121] The first tests show that the substitution model of the cost evaluation procedure actually allows to predict the cost correctly on test data not used during learning (database construction). The regression algorithm has learned on a randomly drawn set without redelivery of 294 winds, 4 regressions were obtained, one for each controller.

[0122] The algorithm is tested on 294 randomly drawn wind samples, without redelivery, not used for learning (database). FIG. 3 shows the real values VR and the estimated values VE obtained with the method according to the invention. It is possible to see in FIG. 3 that the algorithm predicts the cost of each simulation correctly, and the predictions are represented here for the four simulations.

[0123] To evaluate the quality of the method according to the invention, two indicators can be used:

[00007]$R^2\left(K\right)=1-\frac{\nu \mathrm{ar}\left(y\left(K\right)-\hat{y}\left(K\right)\right)}{\mathrm{var}\left(y\left(K\right)\right)}$$R_{\mathrm{dec}}\left(K\right)=1-\frac{\underset{i}{.Math.}y\left(K^{*},X_i\right)}{\underset{K}{\min }\underset{i}{.Math.}\left(K,X_i\right)}$

[0124] R.sup.2 gives an indication of the regression algorithm quality, the closer it is to 1, the higher the quality of the regression. R.sub.dec gives an approximation of the fatigue decrease that could be obtained using the best controllers K* determined by the regression, without accounting for the cost that could be added by switching from one controller to another.

[0125] Table 2 gives the values of the indicators. Scores R.sup.2 are above 0.9 for each regression. Therefore, the regression method is of good quality. According to scores R.sub.dec, the algorithm could indeed allow reduction of the wind turbine cost by at least 20% in relation to the best controller of the set of candidates alone.

TABLE-US-00002 TABLE 2 Controller R.sup.2 R.sub.dec 1 0.93 23% 2 0.96 35% 3 0.93 36% 4 0.92 26%