DFIG wind farm dynamic equivalent method, system and device based on protection action state identification

Abstract

The application provides a DFIG wind farm dynamic equivalence method, system and equipment based on protection action state recognition. By setting different fault degrees and wind speeds, the protection action of the wind turbine under different operation scenarios is simulated to obtain data for learning of a LightGBM classifier; a wind turbine operation state criterion is proposed to quickly judge the protection action state of the wind turbine under different faults; an improved LightGBM wind turbine protection action recognition model is constructed, and a misjudgment penalty factor matrix is proposed to improve the classification accuracy; when applied, the state quantity of the wind turbine is input into the improved LightGBM classification model to quickly and accurately obtain the grouping result, and a wind farm equivalence model is obtained through parameter calculation; finally, the accuracy and universality of the method are verified based on Simulink. The application considers the operation condition of the partial off-grid of the wind farm under the fault, and the recognition and classification are based on data driving, so the speed is fast, the accuracy is high, and the universality is the highest.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of doubly-fed wind farm equivalence, in particular to a DFIG wind farm dynamic equivalence method, system and equipment based on protection action state identification. BACKGROUND

[0002] The installed capacity of wind power in China is increasing year by year, and the impact of large-scale wind turbine grid connection on the stable operation of the power system cannot be ignored, so it is urgent to establish a model that can accurately describe the dynamic characteristics of wind farm operation. Since a wind farm is often composed of dozens or even hundreds of wind turbines, in order to reduce the complexity and computational load of the simulation model and at the same time improve the simulation accuracy, it is of great significance to establish a reliable and accurate wind farm equivalence model.

[0003] The current doubly-fed wind turbine (DFIG) wind farm equivalence method mainly includes single-machine equivalence method and multi-machine equivalence method. The single-machine equivalence method equates the entire wind farm to one, which reduces the computational load but has lower accuracy; the multi-machine equivalence method groups similar units by studying different operating state quantities of different wind turbines during steady state and fault, such as active power, wind speed, pitch angle, etc., thereby obtaining a multi-machine equivalence model, which has higher accuracy than single-machine equivalence, but the selection of state quantities is crucial.

[0004] Although scholars at home and abroad have carried out a lot of research on the equivalence of DFIG wind farms, the existing equivalence methods do not consider the situation of partial wind turbine disconnection in the entire wind farm under severe faults, and partial wind turbine disconnection will change the power shown by the wind farm to the outside, so it is very important to identify and equivalent the disconnected units; the action characteristics of Crowbar are often the first grouping index to distinguish the differences between units during faults, but the existing Crowbar identification method has low accuracy, which cannot guarantee the equivalence accuracy. SUMMARY

[0005] The present application provides a DFIG wind farm dynamic equivalence method, system and equipment based on protection action state identification, which identifies the action state of Crowbar based on a data-driven classification model, and can guarantee high identification accuracy; considering the scenario of partial wind turbine disconnection in the wind farm, the disconnected units are identified, so that the equivalence model can adapt to more actual scenarios and has the highest universality.

[0006] To achieve the above purpose, the technical scheme of the method of the present application is as follows:

[0007] A DFIG wind farm dynamic equivalence method based on protection action state identification, the method comprising the following steps:

[0008] Data acquisition step: a wind farm system containing multiple DFIGs is constructed, each wind turbine in the wind farm system is equipped with LVRT control, Crowbar protection and low-voltage off-grid machine protection; different fault levels and different wind speeds are set to simulate the protection action of the wind turbine under different operating scenarios, and the power, voltage and wind speed of each wind turbine are collected to construct a sample data set, which is normalized and processed to divide the training set and the test set;

[0009] Running state judgment step: set the wind turbine running state criterion and label the wind turbine unit based on the wind turbine running state criterion;

[0010] Model training step: the data in the training set constructed in the data acquisition step is used as the model input, and the wind turbine running state label in the running state judgment step is used as the output, a wind turbine protection action recognition classification model based on the improved light gradient boosting machine is constructed, and the improved wind turbine protection state classification model is obtained based on the misjudgment penalty factor matrix. The improved wind turbine protection state classification model is trained, and the data in the test set obtained in the data acquisition step is used to evaluate the classification effect;

[0011] Wind farm equivalence step: the power, voltage and wind speed of each wind turbine in the actual wind farm are obtained and input into the improved wind turbine protection state classification model, and wind turbines with the same action state are divided into a group; a two-stage grouping method is proposed, and the Crowbar non-action unit is further divided into two machines according to the wind speed interval; the power and capacity of the equivalent unit are calculated based on the capacity weighting method, the line impedance is calculated based on the equal voltage loss method, and the equivalent wind speed is calculated based on the relationship curve of power and wind speed, so as to obtain a wind farm equivalence model with no more than 4 machines;

[0012] Simulation verification step: a 24-machine DFIG wind farm detailed model is built in MATLAB / Simulink for simulation verification.

[0013] Further, the data acquisition step is as follows:

[0014] By setting different short-circuit resistance values, different fault levels are simulated, so that the voltage at the wind farm grid connection point drops to 0.1p.u to 0.5p.u, and the protection action of the wind turbine under different operating scenarios is simulated by setting different wind speeds; five characteristic quantities of each wind turbine, including steady-state power, steady-state voltage, fault minimum point voltage, fault steady-state voltage and wind speed, are collected to construct a sample data set, which is normalized and processed to divide the training set and the test set.

[0015] Further, the running state judgment step is as follows:

[0016] Wind turbine running state criterion:

[0017]

[0018] where i r represents rotor current, U min represents minimum voltage, NAND represents Crowbar inaction unit; YAND represents Crowbar action but not off-grid unit; YAYD represents Crowbar action and off-grid unit;

[0019] According to Crowbar action and inaction, it is divided into two groups, Crowbar inaction unit NAND is labeled as 0; Crowbar action unit includes off-grid unit and non-off-grid unit, non-off-grid unit YAND is labeled as 1, and off-grid unit is labeled as YAYD.

[0020] Further, the model training step is specifically as follows:

[0021] Based on the machine learning framework of LightGBM, the histogram algorithm and the Leaf-wise decision tree growth strategy with depth limit are introduced, and the fan protection action recognition classification model based on improved light gradient boosting machine is constructed;

[0022] The misjudgment penalty factor matrix is introduced to improve the model to obtain the improved fan protection state classification model;

[0023] The steady-state power, steady-state voltage, fault minimum point voltage, fault steady-state voltage and wind speed five feature vectors are taken as the input of the improved fan protection state classification model, and the label of the fan unit according to the fan operation state criterion is taken as the output corresponding to each group of input;

[0024] The improved fan protection state classification model is trained and the classification effect is evaluated by using the test set;

[0025] Further, the misjudgment penalty factor matrix is as follows:

[0026]

[0027] where m 01 represents MPF when the actual value is 0 but predicted as 1, m 10 represents MPF when the actual value is 1 but predicted as 0, m 02 represents MPF when the actual value is 0 but predicted as 2, m 20 represents MPF when the actual value is 2 but predicted as 0, m 12 represents MPF when the actual value is 1 but predicted as 2, m 21 represents MPF when the actual value is 2 but predicted as 1,

[0028] Set m 01 , m 02, m 10 , m 12 is 1.5; set m 20 in the range of 2.5-3, m 21 in the range of 2-2.5.

[0029] The optimal parameters are selected by grid search to minimize the loss function and the misjudgment rate.

[0030] Further, a misjudgment penalty factor matrix is introduced to improve the model, and the improved fan protection state classification model comprises:

[0031] The loss function of LightGBM is:

[0032]

[0033] In the formula, N represents the number of decision trees, F m-1 (x i ; A m-1 ) is the predicted value of x i of the first m-1 trees, A m-1 is a parameter, y i is the actual value, F m-1 (x i ; A m-1 ) is the predicted value, and L(y i , F m-1 (x i ; A m-1 )) is the error function of the true value and the predicted value, and the logarithmic function is selected here;

[0034] The loss function expression after adding MPFM is as follows:

[0035]

[0036] In the formula, L modified represents the improved loss function.

[0037] Further, the equivalent step of the wind farm is specifically as follows:

[0038] The fault pre-power, fault pre-steady voltage, fault process minimum voltage, fault steady voltage and wind speed data of each fan of the actual wind farm are obtained and input into the improved fan protection state classification model;

[0039] The improved fan protection state classification model divides the fans with the same action state into a group, and further divides the Crowbar inaction unit into two machines according to the wind speed interval.

[0040] The power and capacity of the equivalent unit are calculated based on the capacity weighting method:

[0041]

[0042] In the formula, S equ , P equ represent the capacity and active power of the equivalent unit, S i , P i represent the capacity and active power of each unit, and M represents the number of the action group belonging to this type;

[0043] Meanwhile, the equivalent line impedance is calculated based on the equal voltage loss method, and the equivalent wind speed is calculated based on the relationship curve between power and wind speed.

[0044] Based on the above steps, the equivalent model of the wind farm with not more than 4 units is obtained.

[0045] Further, the specific steps of the simulation verification based on the MATLAB / Simulink simulation platform are as follows:

[0046] Based on MATLAB / Simulink, a detailed wind farm is established, a large amount of training data is obtained through a large number of running scene settings, the improved LightGBM model is trained, and part of the scene is selected for dynamic protection equivalence of the wind farm, and the rationality of the equivalence result is verified.

[0047] On the other hand, the application provides a DFIG wind farm dynamic equivalence system based on protection action state recognition, comprising:

[0048] The data acquisition module is used to construct a wind farm system containing multiple DFIGs, each wind turbine in the wind farm system is equipped with low voltage ride through (LVRT) control, Crowbar protection and low voltage off-grid machine protection; different fault degrees and different wind speeds are set to simulate the protection action of the wind turbine under different running scenes, and the power, voltage and wind speed of each wind turbine are collected to construct a sample data set, which is normalized and processed to divide the training set and the test set.

[0049] The running state judgment module is used to set the wind turbine running state criterion and label the wind turbine unit based on the wind turbine running state criterion.

[0050] The model training module is used to input the data in the training set constructed in the data acquisition step as the model input, and the wind turbine running state label in the running state judgment step as the output, to construct a wind turbine protection action recognition classification model based on the improved light gradient boosting machine, and to improve the model based on the misjudgment penalty factor matrix to obtain the improved wind turbine protection state classification model, train the improved wind turbine protection state classification model, and evaluate the classification effect with the data in the test set obtained in the data acquisition step.

[0051] The wind farm equivalent module is used for obtaining the power, voltage and wind speed of each wind turbine of an actual wind farm, inputting the improved wind turbine protection state classification model, and dividing wind turbines with the same action state into a group; a two-stage grouping method is proposed, and the Crowbar inaction units are further divided into two units according to the wind speed interval; the power and capacity of the equivalent unit are calculated based on the capacity weighting method, the line impedance is calculated based on the equal voltage loss method, and the equivalent wind speed is calculated based on the relationship curve of power and wind speed, so that the equivalent model of the wind farm with not more than 4 units is obtained.

[0052] The simulation verification module is used for building a 24-machine DFIG wind farm detailed model in MATLAB / Simulink for simulation verification.

[0053] In another aspect, the application provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to realize the steps of the DFIG wind farm dynamic equivalent method based on protection action state identification.

[0054] Compared with the prior art, the application has the following beneficial effects:

[0055] 1. The application fully considers more scenes in the operation process of the wind farm, that is, the scene of part of wind turbines of the wind farm being off the grid after a fault, and the proposed method can accurately identify the off-grid units, which is considered by the existing method; at the same time, the proposed equivalent model can also be applicable to the scene without wind turbine off-grid, and therefore has the highest universality.

[0056] 2. The application is based on the classification function of data driving, and proposes an MPFM matrix to improve the classification accuracy and reduce the misjudgment rate, thereby improving the identification accuracy of the Crowbar action state and further improving the equivalent effect of the wind farm equivalent model. DETAILED DESCRIPTION

[0057] In order to more clearly illustrate the technical solutions in the application or prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0058] Figure 1 The flow chart of the DFIG wind farm dynamic equivalent based on the improved LightGBM protection action state recognition proposed in the embodiment of the application.

[0059] Figure 2 It is a 24-machine DFIG wind farm detailed model based on MATLAB / Simulink.

[0060] Figure 3 is an improved LightGBM classifier classification effect confusion matrix.

[0061] Figure 4 is an equivalent model effect diagram under the scenario of partial wind turbine off-grid.

[0062] Figure 5 is an equivalent model effect diagram under the scenario of no wind turbine off-grid.

[0063] Figure 6 is a DFIG wind farm dynamic equivalent system based on improved LightGBM protection action state recognition. DETAILED DESCRIPTION

[0064] In order to make the objects, technical solutions and advantages of the present application clearer, the technical solutions in the present application will be described clearly and completely below in combination with the drawings in the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0065] Embodiment 1

[0066] The present application will be further described in detail below in combination with the drawings and specific embodiments:

[0067] As shown in the figure, it is a DFIG wind farm dynamic equivalent method based on protection action state recognition provided by the embodiment of the present application, which comprises the following steps: Figure 1

[0068] Step S101, data acquisition:

[0069] A wind farm system containing multiple DFIGs is constructed, and each wind turbine is equipped with low voltage ride through (LVRT) control, Crowbar protection and low voltage off-grid machine protection. Among them, the LVRT control refers to the control strategy that when the wind turbine detects voltage drop, the fault steady state control is switched to the LVRT control, the reactive power is preferentially increased, and the active power is appropriately reduced. Since the reactive power Q is only controlled by i qr , the active power P is only controlled by i dr , i qr and i dr are determined by the following formula respectively:

[0070]

[0071] In the formula, i qr represents the rotor q-axis current, i dr represents the rotor d-axis current, and the superscripts bf, df and af represent the state quantities before, during and after the fault respectively, u​ds ω represents the d-axis component of the stator voltage. s L represents the synchronous angular velocity. m L represents magnetizing inductance. s This represents the stator flux linkage, and k1 represents the dynamic reactive current proportional coefficient, which is set to 0.2. I represents the stator voltage before the fault. N This is the rated current of the fan. I represents the active power before the fault. rmax t1 represents the maximum rotor current, which is set to 1.2; k2 represents the active power recovery coefficient, which is set to 0.2; t1 represents the fault clearing time; and t represents time.

[0072] Crowbar is a resistor connected to the rotor side. When the rotor current is detected to be greater than the threshold of 1.5, Crowbar is activated, shorting the rotor side to prevent damage to the power electronic equipment.

[0073] The fan terminal is equipped with low-pressure trip protection as required by national standards, based on the following criteria:

[0074]

[0075] In the formula, u i The voltage at the turbine terminal is represented by t, which is time. When any phase voltage is lower than 0.2, the turbine is immediately shut down.

[0076] By setting different short-circuit resistance values, different fault levels are simulated, causing the grid connection voltage of the wind farm to drop to between 0.1 pu and 0.5 pu. By setting different wind speeds, including wind speeds with different wind directions based on the wake effect and random wind speed scenarios, the protection actions inside the wind turbines are simulated under different operating scenarios. Five features of each wind turbine are collected: steady-state power, steady-state voltage, voltage at the lowest fault point, steady-state voltage at the fault, and wind speed. A sample dataset is constructed, and after normalization and processing, it is divided into training and test sets.

[0077] Step S102, Running status determination:

[0078] The transient response of the crowbar differs significantly from that of a non-operating unit, and the crowbar operation depends on the magnitude of the rotor-side current of the wind turbine; the crowbar will operate when the rotor-side current is greater than 1.5.

[0079]

[0080] In the formula, i r P represents the rotor current. bf U represents the steady-state power before the fault. bf This is the steady-state voltage before the fault. ν is the steady-state voltage during the fault, and v is the wind speed.

[0081] At the same time, the embodiment proposes to consider the equivalence under the scenario of fan disconnection, and the fan disconnection is related to the minimum voltage of the grid-connected point during the fault.

[0082] Therefore, the following fan operation state criterion is proposed:

[0083]

[0084] In the formula, NAND represents a Crowbar inaction unit, and is labeled as 0; YAND represents a Crowbar action but not disconnection unit, and is labeled as 1; YAYD represents a Crowbar action and disconnection unit, and is labeled as 2, thereby obtaining sample output.

[0085] Step S103, model training:

[0086] LightGBM is a machine learning framework based on Gradient Boosting Decision Tree (GBDT), which can be used to solve classification problems. GBDT builds multiple decision trees N and increases the decision trees in an iterative manner until the accuracy improvement is less than a threshold; on the basis of GBDT, LightGBM introduces a histogram algorithm and a Leaf-wise decision tree growth strategy with depth limitation, thereby enhancing the robustness to noise, while ensuring good evaluation accuracy and training speed.

[0087] The loss function of LightGBM is:

[0088]

[0089] In the formula, N represents the number of decision trees, F m-1 (x i ; A m-1 ) is the prediction value of the first m-1 trees for x i , A m-1 is a parameter, y i is the actual value, F m-1 (x i ; A m-1 ) is the prediction value, and L(y i , F m-1 (x i ; A m-1 )) is an error function of the true value and the prediction value, and a logarithmic function is selected here.

[0090] It should be noted that the fan part is less off-grid in the actual scene, so the YAYD data volume will be much smaller than the NAND and YAND data volumes. However, the off-grid of the fan has the most important influence on the equivalence effect of the wind farm, and if the off-grid fan is misjudged, the accuracy of the equivalence effect will be significantly reduced. Based on this, the present application proposes a misjudgment penalty factor matrix (Misjudgment penalty factor matrix, MPFM):

[0091]

[0092] Wherein, m 01 represents the MPF when the actual value is 0 but is predicted to be 1, m 10 represents the MPF when the actual value is 1 but is predicted to be 0, and the remaining letters have similar meanings. Since the NAND and YAND data volumes are larger, the influence of misjudgment on the equivalence result is less than the misjudgment of YAYD, so m 01 , m 02 , m 10 , m 12 are set to 1.5; in order to improve the accuracy of the model in judging YAYD, a higher penalty is set for m 20 , m 21 , especially considering that the influence of judging YAYD as NAND is the most obvious, so the range of m 20 is 2.5-3, and the range of m 21 is 2-2.5. The optimal parameters are selected by grid search to minimize the loss function and minimize the misjudgment rate.

[0093] The loss function expression after adding MPFM is as follows:

[0094]

[0095] The input of the improved LightGBM model is the steady-state power, steady-state voltage, fault minimum point voltage, fault steady-state voltage and wind speed obtained in step S101, and the output corresponding to each group of input is 0, 1 and 2 marked in step S102. By introducing MPFM, the classification performance is further improved. The wind turbine protection state classification model is trained and the classification effect is evaluated with the test set;

[0096] Step S104, wind farm equivalence:

[0097] In practical applications, the pre-fault power, pre-fault steady-state voltage, fault process minimum voltage, fault steady-state voltage and wind speed data of each wind turbine in the actual wind farm are obtained, and the improved LightGBM classification model is input;

[0098] The classification model will divide the fans with the same action state into a group. In order to improve the accuracy of the equivalent model, a two-stage grouping method is proposed, and the Crowbar non-action unit is further divided into two machines according to the wind speed interval;

[0099] The power and capacity of the equivalent unit are calculated based on the capacity weighting method:

[0100]

[0101] In the formula, S equ , P equ represent the capacity and active power of the equivalent unit, S i , P i represent the capacity and active power of each unit, and M represents the number of action groups belonging to this type.

[0102] At the same time, the equivalent line impedance is calculated based on the equal voltage loss method, and the equivalent wind speed is calculated based on the relationship curve of power and wind speed.

[0103] Based on the above steps, the equivalent model of the wind farm with no more than 4 machines is obtained.

[0104] Step S105, simulation verification:

[0105] A detailed model of a 24-machine DFIG wind farm is built in MATLAB / Simulink, and the feasibility of the improved LightGBM classification model is verified based on the improved LightGBM classification model.

[0106] The detailed topology of the established wind farm is shown in Figure 2 The value of the short-circuit resistance is increased from 1Ω to 3Ω in steps of 0.1Ω, simulating 21 groups of different severity faults; 10 different wind speed conditions are set, so as to obtain 21*10*24=5040 groups of data, and the labels are marked according to the proposed operating state criterion; the model is trained with the training set, and the model training effect is judged with the test set, and the confusion matrix is drawn as shown in Figure 3 The overall accuracy of the model classification reaches 99.58%, which can significantly distinguish the different operating states of the fans; at the same time, based on the MPFM matrix proposed in the application, YAYD has a classification accuracy of 100% under the condition of the smallest sample size, which proves that the improved LightGBM proposed in the application can greatly reduce the misjudgment rate and improve the classification accuracy.

[0107] Select a group of scenarios to verify the accuracy of the equivalent model: set a three-phase short-circuit fault at the grid connection point of the wind farm at 10s, and the grid connection point voltage drops to below 0.2, collect the fault before power, fault before steady voltage, fault process minimum voltage, fault during steady voltage and wind speed data of 24 units, input the improved LightGBM, and get the penalty factor m 20 3, m21 is 2.45, the clustering results are: 1, 2, 3, 9, 10, 11, 17, 18 are YAYD, 4, 5, 6, 12, 13, 14, 19, 20, 21 are YAND, 7, 8, 15, 16, 22, 23, 24 are NAND, and the obtained classification results are completely consistent with the actual action conditions, and a three-machine equivalent model is obtained.

[0108] Figure 4 The comparison of the active power, the reactive power, the voltage and the current at the grid connection point between the equivalent model and the actual wind farm is shown, and the closer the curves are, the better the effect of the equivalent model is.

[0109] The method disclosed by the present application is compared with the single-machine equivalent method and the popular two-machine equivalent method with or without Crowbar action, and through the comparison, it can be obtained that under the operation condition that part of the wind turbines in the wind farm are off-grid, the existing methods cannot accurately identify the off-grid units, and therefore a large error will be generated; and the method disclosed by the present application is based on the classification function of the improved LightGBM and the wind turbine operation state criterion, and can accurately equivalent the units in different protection action conditions into a group, so that the highest precision is achieved.

[0110] In order to verify the universality of the method disclosed by the present application, an operation condition without off-grid units is selected for verification: it is set that a three-phase short-circuit fault occurs at the grid connection point of the wind farm at 10s, the voltage at the grid connection point drops to about 0.25, and no wind turbine is off-grid in the wind farm; the fault pre-power, the fault pre-steady voltage, the fault process minimum voltage, the fault steady voltage and the wind speed data of 24 units are collected, and are input into the improved LightGBM model, and the clustering results are: 1, 2, 3, 4, 5, 9, 10, 11, 12, 13, 17, 18 are YAND, 6, 7, 8, 14, 15, 16, 19, 20, 21, 22, 23, 24 are NAND, and the obtained classification results are completely consistent with the actual action conditions; the NAND group is further divided into two groups according to the wind speed interval, a three-machine equivalent model is obtained, and the comparison is shown as Figure 5 It can be seen that the method disclosed by the present application still has the highest precision.

[0111] Embodiment 2

[0112] As shown in Figure 6 The DFIG wind farm dynamic equivalent system based on the protection action state recognition includes:

[0113] The data acquisition module is configured to build a wind farm system with multiple DFIGs, each wind turbine in the wind farm system being equipped with low voltage ride through (LVRT) control, Crowbar protection, and low voltage off-grid cutting protection; different fault levels and different wind speeds are set to simulate the protection action of the wind turbine under different operating scenarios, and sample data sets of power, voltage, and wind speed of each wind turbine are collected, normalized, and processed to divide the training set and the test set;

[0114] The running state judgment module is configured to set wind turbine running state criteria and label wind turbine units based on the wind turbine running state criteria;

[0115] The model training module is configured to use the data in the training set constructed in the data acquisition step as model input and the wind turbine running state label in the running state judgment step as output, to build a wind turbine protection action recognition classification model based on the improved light gradient boosting machine, to improve the model based on the misjudgment penalty factor matrix to obtain an improved wind turbine protection state classification model, to train the improved wind turbine protection state classification model, and to evaluate the classification effect using the data in the test set obtained in the data acquisition step;

[0116] The wind farm equivalence module is configured to obtain the power, voltage, and wind speed of each wind turbine in the actual wind farm and input the improved wind turbine protection state classification model, to divide wind turbines with the same action state into a group; a two-stage grouping method is proposed, and Crowbar non-action units are further divided into two units according to the wind speed interval; the power and capacity of the equivalent unit are calculated based on the capacity weighting method, the line impedance is calculated based on the equal voltage loss method, and the equivalent wind speed is calculated based on the relationship curve of power and wind speed, to obtain a wind farm equivalence model with no more than 4 units;

[0117] The simulation verification module is configured to build a 24-machine DFIG wind farm detailed model in MATLAB / Simulink for simulation verification.

[0118] The DFIG wind farm data acquisition module 21, the wind turbine running state judgment module 22, the model training module 23, the DFIG wind farm equivalence module 24, and the simulation verification module 25.

[0119] The DFIG wind farm data acquisition module 21 is configured to acquire wind turbine running data, and is configured to:

[0120] Simulate the protection action of the wind turbine under different operating scenarios, collect the power, voltage, and wind speed of each wind turbine, and input the sample data set;

[0121] The wind turbine running state judgment module 22 is configured to judge the state of wind turbines under different operating states, and is configured to:

[0122] Set the fan operating state criterion and label the fan unit based on the fan operating state criterion, and divide it into two groups according to the Crowbar action and inaction, and label the Crowbar inaction unit NAND as 0; the Crowbar action unit includes the off-grid unit and the non-off-grid unit, label the non-off-grid unit YAND as 1, and label the off-grid unit YAYD as 2;

[0123] The model training module 23 is configured to train the fan operating state recognition model, and is configured to:

[0124] The power, voltage and wind speed of the fan are taken as model inputs, and different action conditions of the fan are taken as outputs, an improved LightGBM-based fan protection action recognition classifier is constructed, and the MPFM model is improved based on the MPFM model to improve the classification accuracy.

[0125] The DFIG wind farm equivalent module 24 is configured to equivalent the DFIG wind farm, and is configured to:

[0126] After obtaining the power, voltage and wind speed of each fan of the actual wind farm, the fans with the same action state are divided into a group based on the improved LightGBM classification model; the equivalent machine parameters are calculated, so as to obtain a wind farm equivalent model with no more than 4 machines.

[0127] The simulation verification module 25 is configured to simulate and verify, and is configured to:

[0128] A detailed model of a 24-machine DFIG wind farm is built in MATLAB / Simulink for simulation verification.

[0129] Although the preferred embodiments of the present application have been described, those skilled in the art can make further changes and modifications to the embodiments once they know the basic inventive concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0130] Obviously, those skilled in the art can make various modifications and variations to the embodiments of the present application without departing from the spirit and scope of the embodiments of the present application. Thus, if these modifications and variations of the embodiments of the present application fall within the scope of the claims of the present application and their equivalents, the present application also intends to include these modifications and variations.

[0131] Other parts not described in detail are prior art.

Claims

1. A DFIG wind farm dynamic equivalence method based on protection action state recognition, characterized in that, The method comprises the following steps: Data acquisition step: construct a wind farm system containing multiple DFIGs, each wind turbine in the wind farm system is equipped with LVRT control, Crowbar protection and low-voltage off-grid machine protection; set different fault levels and different wind speeds to simulate the protection action of the wind turbine under different operating conditions, collect the power, voltage and wind speed of each wind turbine to construct a sample data set, normalize and process the sample data set, and divide the training set and the test set; Running state judgment step: set the wind turbine running state criterion and label the wind turbine unit based on the wind turbine running state criterion; Model training step: use the data in the training set constructed in the data acquisition step as the model input, and use the wind turbine running state label in the running state judgment step as the output, construct a wind turbine protection action recognition classification model based on the improved light gradient boosting machine, improve the model based on the misjudgment penalty factor matrix to obtain an improved wind turbine protection state classification model, train the improved wind turbine protection state classification model, and evaluate the classification effect by using the data in the test set obtained in the data acquisition step; Wind farm equivalence step: input the power, voltage and wind speed of each wind turbine in the actual wind farm into the improved wind turbine protection state classification model, and divide the wind turbines with the same action state into a group; propose a two-stage grouping method, and further divide the Crowbar non-action unit into two units according to the wind speed interval; calculate the power and capacity of the equivalent unit based on the capacity weighting method, calculate the line impedance based on the equal voltage loss method, and calculate the equivalent wind speed based on the relationship curve of power and wind speed to obtain a wind farm equivalence model with no more than 4 units; Simulation verification step: build a 24-machine DFIG wind farm detailed model in MATLAB / Simulink for simulation verification.

2. The DFIG wind farm dynamic equivalence method based on protection action state recognition of claim 1, wherein, The data acquisition step is as follows: By setting different short-circuit resistance values, different fault levels are simulated, so that the voltage at the wind farm grid connection point drops to 0.1p.u to 0.5p.u, and the protection action of the wind turbine under different operating conditions is simulated by setting different wind speeds; five characteristic quantities of each wind turbine, including steady-state power, steady-state voltage, fault minimum voltage, fault steady-state voltage and wind speed, are collected to construct a sample data set, which is normalized and processed to divide the training set and the test set.

3. The DFIG wind farm dynamic equivalence method based on protection action state recognition of claim 1, wherein, The running state judgment step is as follows: Wind turbine running state criterion: wherein i r represents the rotor current, U min represents the minimum voltage, NAND represents Crowbar inactive units; YAND represents Crowbar active but not off-grid units; YAYD represents Crowbar active and off-grid units; According to the Crowbar action and non-action, the wind turbines are divided into two groups, and the Crowbar non-action unit NAND is labeled as 0; the Crowbar action unit includes off-grid units and non-off-grid units, the non-off-grid unit YAND is labeled as 1, and the off-grid unit is labeled YAYD as 2.

4. The DFIG wind farm dynamic equivalence method based on protection action status recognition of claim 2, wherein, The model training step is as follows: Based on the machine learning framework of LightGBM, the histogram algorithm and the Leaf-wise decision tree growth strategy with depth limitation are introduced to construct a wind turbine protection action recognition classification model based on the improved light gradient boosting machine; An improved wind turbine protection state classification model is obtained by introducing a misjudgment penalty factor matrix to improve the model. The five characteristic vectors of steady-state power, steady-state voltage, fault minimum point voltage, fault steady-state voltage and wind speed are taken as the inputs of the improved wind turbine protection state classification model, and the label of the wind turbine unit according to the wind turbine operation state criterion is taken as the corresponding output of each group of inputs; The improved wind turbine protection state classification model is trained and the classification effect is evaluated with the test set.

5. The DFIG wind farm dynamic equivalence method based on protection action status recognition of claim 4, wherein, The misjudgment penalty factor matrix is as follows: wherein, m 01 MPF representing actual value of 0 but predicted as 1, m 10 MPF representing actual value of 1 but predicted as 0, m 02 MPF representing actual value of 0 but predicted as 2, m 20 MPF representing actual value of 2 but predicted as 0, m 12 MPF representing actual value of 1 but predicted as 2, m 21 MPF representing actual value of 2 but predicted as 1, Set m 01 , m 02 , m 10 , m 12 1.5; set m 20 2.5-3, m 21 2-2.5; The optimal parameters are selected by grid search to minimize the loss function and the misjudgment rate.

6. The DFIG wind farm dynamic equivalence method based on protection action status recognition of claim 5, wherein, The improved wind turbine protection state classification model is obtained by introducing the misjudgment penalty factor matrix and improving the model, including: The loss function of LightGBM is: N represents the number of decision trees, F m-1 x i ；A m-1 is the predicted value of the previous m-1 tree, x i is the actual value, A m-1 is a parameter, y i is the actual value, F m-1 x i ；A m-1 is the predicted value, L y i ，F m-1 x i ；A m-1 is an error function of the true value and the predicted value, and a logarithmic function is selected herein;​​​​ The loss function expression after adding MPFM is as follows: In the formula, L modified represents the improved loss function.

7. The DFIG wind farm dynamic equivalence method based on protection action status recognition of claim 1, wherein, The equivalent steps of the wind farm are as follows: The fault pre-power, steady-state voltage before fault, minimum voltage during fault, steady-state voltage during fault and wind speed data of each wind turbine in the actual wind farm are obtained and input into the improved wind turbine protection state classification model; The improved wind turbine protection state classification model divides the wind turbines with the same action state into a group, and further divides the Crowbar non-action units into two machines according to the wind speed interval; The power and capacity of the equivalent unit are calculated based on the capacity weighting method: In the formula, S equ , P equ represent the capacity and active power of the equivalent unit, S i , P i represent the capacity and active power of each unit, M represent the number of fans belonging to the same action group with the same action state; At the same time, the equivalent line impedance is calculated based on the loss method of equal voltage, and the equivalent wind speed is calculated based on the relationship curve of power and wind speed; Based on the above steps, the equivalent model of the wind farm with not more than 4 machines is obtained.

8. The DFIG wind farm dynamic equivalence method based on protection action status recognition of claim 1, wherein, The specific steps of simulation verification based on MATLAB / Simulink simulation platform are as follows: Based on MATLAB / Simulink, a detailed wind farm is established, a large amount of training data is obtained through a large number of running scene settings, and the improved LightGBM model is trained; part of the scene is selected for dynamic protection equivalence of the wind farm, and the rationality of the equivalence result is verified.

9. A DFIG wind farm dynamic equivalence system based on protection action state recognition, characterized in that, It includes: The data acquisition module is used to construct a wind farm system containing multiple DFIGs, each wind turbine in the wind farm system is equipped with LVRT control, Crowbar protection and low-voltage off-grid machine protection; different fault degrees and different wind speeds are set to simulate the protection action of the wind turbine under different operating conditions, and the power, voltage and wind speed of each wind turbine are collected to construct a sample data set; the sample data set is normalized and processed, and then divided into training set and test set; The running state judgment module is used to set the wind turbine running state criterion and label the wind turbine unit with the wind turbine running state label based on the wind turbine running state criterion; The model training module is used to take the data in the training set constructed in the data acquisition step as the model input, and take the wind turbine running state label in the running state judgment step as the output, to construct a wind turbine protection action recognition classification model based on the improved Light Gradient Boosting Machine, and to improve the model by introducing the misjudgment penalty factor matrix to obtain the improved wind turbine protection state classification model, train the improved wind turbine protection state classification model, and evaluate the classification effect with the data in the test set obtained in the data acquisition step. The wind farm equivalent module is used to obtain the power, voltage and wind speed of each wind turbine of an actual wind farm, and input the improved wind turbine protection state classification model, and wind turbines with the same action state are classified into a group; a two-stage grouping method is proposed, and the Crowbar non-action units are further divided into two units according to the wind speed interval; the power and capacity of the equivalent units are calculated based on the capacity weighting method, the line impedance is calculated based on the equal voltage loss method, and the equivalent wind speed is calculated based on the relationship curve of power and wind speed, so as to obtain a wind farm equivalent model with not more than 4 units. The simulation verification module is used to build a 24-machine DFIG wind farm detailed model in MATLAB / Simulink for simulation verification. The DFIG wind farm dynamic equivalent system based on protection action state recognition is used to execute the steps in the DFIG wind farm dynamic equivalent method based on protection action state recognition in any one of claims 1-8.

10. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 9. The computer program is executed by a processor to realize the steps in the DFIG wind farm dynamic equivalent method based on protection action state recognition in any one of claims 1-8.

Images