A method for identifying control parameters of a doubly-fed wind turbine based on LSTM-IPSO

Abstract

The application discloses a kind of based on LSTM-IPSO doubly-fed fan control parameter identification method, including the following steps: using RT-LAB semi-physical simulation platform obtains the hardware-in-the-loop experiment data of doubly-fed fan derived from real controller, and the isomorphic identification model of real controller of doubly-fed fan is built on Matlab / simulink platform;Increase input feature set dimension and remove irrelevant features, select the correlation higher characteristic value as the input feature set of the neural network model;Control parameter-input feature set is formed by input feature set and corresponding control parameter set;Using LSTM neural network, control parameter-input feature set is trained and predicted, and the predicted initial value and optimization range are obtained;Using IPSO algorithm as the secondary optimization method of accurate identification, the purpose of accurate optimization is achieved;Judge the reliability of identification model;The application solves the technical problem that traditional identification method is difficult to accurately identify the control parameter of doubly-fed fan electromagnetic model under low voltage ride through condition.

Description

Technical Field

[0001] This invention relates to the field of new energy power generation parameter identification technology, specifically to a method for identifying control parameters of doubly fed wind turbines based on LSTM-IPSO. Background Technology

[0002] In recent years, with the large-scale grid connection of wind turbines, the modeling issues of wind turbines have become increasingly prominent. Currently, power grid analysis mainly relies on model simulation or digital simulation, and the model parameters are mostly given by the manufacturer or set as empirical values. Among them, grid-side control parameters are difficult to collect from nameplates and technical manuals, but their accuracy directly affects the stability and safety of power grid operation. Therefore, from the perspective of wind turbine characteristic research and power system stability analysis, it is necessary to build a refined simulation model of wind turbines based on measured data and identify grid-side control parameters.

[0003] Therefore, to address the problem that traditional identification methods are unable to accurately identify the electromagnetic model control parameters of doubly-fed induction generators (DFIGs) under low-voltage ride-through conditions, it is necessary to propose a method for identifying DFIG control parameters based on a long short-term memory (LSTM) neural network combined with an improved particle swarm optimization (IPSO) algorithm. Summary of the Invention

[0004] The purpose of this invention is to overcome the above-mentioned shortcomings and provide a method for identifying control parameters of doubly fed wind turbines based on LSTM-IPSO, so as to solve the technical problem that traditional identification methods are difficult to identify the control parameters of the electromagnetic model of doubly fed wind turbines with high accuracy under low voltage ride-through conditions.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by this invention is: a method for identifying control parameters of a doubly fed wind turbine based on LSTM-IPSO, which includes the following steps:

[0006] Step 1: Obtain hardware-in-the-loop experimental data of the doubly-fed wind turbine from the real controller using the RT-LAB hardware-in-the-loop simulation platform, and build an isomorphic identification model of the real controller of the doubly-fed wind turbine on the Matlab / simulink platform.

[0007] Step 2: Increase the dimension of the input feature set and remove irrelevant features, selecting highly relevant feature values ​​as the input feature set of the neural network model; the input feature set and the corresponding control parameter set constitute the control parameter-input feature set;

[0008] Step 3: Use an LSTM neural network to train and predict the control parameters-input feature set to obtain the initial prediction values ​​and the optimization range;

[0009] Step 4: Based on the predicted initial value and optimization range obtained in Step 4, the IPSO algorithm is used as a secondary optimization method for accurate identification to achieve the goal of accurate optimization.

[0010] Step 5: Compare and verify the identified model with the hardware-in-the-loop experimental data to determine the reliability of the identified model.

[0011] Preferably, in step one, the hardware-in-the-loop experimental data of the doubly-fed wind turbine derived from the real controller is obtained using the RT-LAB hardware-in-the-loop simulation platform, including the DC bus voltage u. dc The dq component of the output current i dg and i qg Where the d-axis reference current i dg It is obtained by DC voltage control, while the q-axis reference current i qg Set it to zero, and build an isomorphic identification model of the real controller of the doubly fed wind turbine on the Matlab / simulink platform.

[0012] Preferably, the doubly fed wind turbine includes a wind turbine, a gearbox, an induction asynchronous generator, a back-to-back converter, and its controller;

[0013] The control equations involved in building the isomorphic identification model of a real controller for a doubly-fed wind turbine on the Matlab / Simulink platform are as follows:

[0014] According to the law of conservation of energy, the KCL equation for the DC capacitor C is:

[0015]

[0016] In the formula, P r P is the active power output from the machine-side converter to C. rc Active power injected into the grid by the grid-side converter;

[0017] The dynamic model of the filter reactance is as follows:

[0018]

[0019] In the formula, u cd u cq with u sd u sq X represents the d-axis and q-axis components of the voltage at the output of the grid-side converter and the generator-side converter, respectively. r This is the filter reactance value.

[0020] The control equations for the grid-side controller can be written as:

[0021]

[0022] In the formula, udcref i is the reference value for DC voltage. dgref and i qgref These are the actual values ​​of the d-axis and q-axis currents on the grid side, respectively, K p1 For the proportional coefficient of the voltage outer loop PI controller, K p2 K is the proportionality coefficient of the inner loop of the q-axis current. p3 Let x1, x2, and x3 be the proportional coefficient of the inner loop of the q-axis current control. Then, the dynamic equation of the controller is:

[0023]

[0024] In the formula, K i1 For the integral coefficient of the voltage outer loop PI controller, K i2 K represents the integral coefficient of the inner loop of the q-axis current. i3 i is the integral coefficient of the inner loop of the q-axis current control. qgref It is usually 0, but under fault conditions, i qgref The value is directly given by the low-profile control module:

[0025]

[0026]

[0027] In the formula, k is the reactive current support coefficient, and U N I is the rated voltage at the grid connection point of the wind turbine. N The rated current of GSC is ; when low voltage ride-through occurs, the grid-side converter will generate a certain amount of reactive power. The value of the reactive current can be obtained according to equation (5). At this time, the inner loop of the q-axis current switches to the low voltage ride-through control mode, and the q-axis reference value is given by the low voltage ride-through control module.

[0028] Preferably, step two specifically includes the following steps:

[0029] Step A1: Collect the isomorphic identification model control parameters - input feature set, where the input feature values ​​are parameters related to the wind turbine grid-side controller: u dc i dg and i qg The error value F between the control parameter set and the hardware-in-the-loop test data, and the control parameter set is the control parameter dataset of the grid-side converter corresponding to the input feature value;

[0030] Step A2: Remove sample data with outlier data points to avoid affecting the accuracy of identification;

[0031] Step A3: According to the "Guidelines for Modeling Electrical Simulation Models of Wind Turbine Units", the simulation time is divided into 5 intervals: steady-state intervals ① and ⑤, low-voltage ride-through intervals ②, ③ and ④, and the i in the input feature value is...dg and i qg This method also involves dividing the data into intervals and using them as feature sets, where u dc The eigenvalue V is obtained by taking values ​​in intervals. a V b V c V d and V e i dg The eigenvalue I is obtained by dividing the intervals. da I db I dc I dd and I qe i qg The eigenvalue I is obtained by dividing the intervals. qa I qb I qc I qd and I qe ;

[0032] Step A4: Use the maximum information coefficient (MIC) to analyze the correlation between the input control parameter and the output feature variable. First, the scatter plot of the control parameter (K) and the feature variable (T) is gridded with column a and row b to find the maximum mutual information value. Then, the maximum mutual information value is normalized. Finally, the maximum mutual information value at different scales is selected as the MIC value.

[0033] Mutual information can be viewed as the amount of information contained in one random variable about another. The mutual information I(K,T) between K and T is defined as:

[0034]

[0035] In the formula, K is the sample sum of variable k, T is the sample sum of variable t, P(k,t) is the joint probability between variables k and t, and P(k) and P(t) are the marginal probabilities of k and t, respectively.

[0036]

[0037] In the formula, n is the number of samples;

[0038] The MIC value obtained from equation (8) is between 0 and 1. The larger the MIC value, the higher the correlation between the two variables. After obtaining the MIC value of the output signal, the key output signal index is selected as the input feature of the LSTM neural network.

[0039] Step A5: Normalize the dataset. Dimensionless data can speed up the gradient descent process to find the optimal solution.

[0040] Preferably, in step three, an LSTM neural network is used to make preliminary predictions of the initial values ​​and optimization range of the doubly fed wind turbine control parameters based on the processed dataset, making it approximate the minimum objective function of the identification model. After extensive debugging, the structural parameters and network training parameters of the LSTM neural network are obtained, and the LSTM neural network model is trained and tested. Its core cell state c is controlled by the forget gate, input gate, and output gate, where: σ is the sigmoid function, x t For the input of the t-th unit, c t Let h be the cell state of the t-th unit. t Let ⊕ be the hidden state of the t-th unit. For summing and multiplying vector elements, c t-1 h represents the cell state of the (t-1)th unit. t-1 This represents the hidden state of the (t-1)th unit;

[0041] The output parameters of an LSTM neural network change under the influence of the input values, and this change can be expressed by the following formula:

[0042]

[0043] In the formula, f t W is the output value for the forget gate. i W f and W o b represents the network layer weights for the input gate, forget gate, and output gate. i b f and b o b is the bias term for the input gate, forget gate, and output gate. c It is the network layer bias. o represents the information input to the neuron at time t. t ☉ represents the output value of the output gate, and ☉ represents the Hadamard product operator.

[0044] Preferably, in step three, the specific steps for training and prediction are as follows:

[0045] Step B1, the proper selection and allocation of datasets and parameter tuning, has a significant impact on the accuracy of the prediction results. Randomly select 90% of the dataset as the training set and 10% of the dataset as the validation set to verify the model's accuracy.

[0046] Step B2: Set the number of training iterations to 500, the loss function to MSE, the optimizer to Adam, the input dimension to 12, the output dimension to 6, the number of hidden layers to 2, the number of hidden layer neurons to 11, and the learning rate to 0.015.

[0047] Step B3: Use an LSTM neural network to train and predict the collected sequence, and obtain the initial values ​​of the predicted parameters and the optimization range.

[0048] Preferably, in step four, the PSO algorithm speed and position update formulas are as follows:

[0049] V u+1 =ωV u +c1r1(pbest u -x u )+c2r2(gbest u -x u (10)

[0050] x u+1 =x u +V u+1 (11)

[0051] In the formula, V represents the particle update speed, x represents the particle, u represents the current time, u+1 represents the next update time, c1 and c2 are learning factors, r1 and r2 are random values ​​between 0 and 1, pbest is the individual optimal value, gbest is the global optimal value, and ω is the adaptive inertia weight.

[0052] PSO uses the mean squared difference between the output value at the actual position and the output value at the optimal position of each particle as the objective function of the algorithm, and obtains the fitness value of each particle at the current iteration number. The objective function is expressed as:

[0053]

[0054] In the formula, U i,error I di,error and I qi,error For u dc i dg and i qg The error, U i I di and I qi For u dc i dg and i qg The measured data value, U ti I dti and I qti For u dc i dg and i qg The output response value obtained from the identification results;

[0055] Step C1, Improve the objective function: Based on the adaptive weighting method, when constructing the objective function for multiple output values, the weight allocation is dynamically adjusted non-linearly according to the MIC value and actual value of each output value. This allows the objective function for multiple output values ​​to adaptively adjust for output values ​​with high correlation or small actual values. The improved formula is as follows:

[0056]

[0057] In the formula, k U k ID and k IQ For U i I di and I qi Adaptive weighting coefficients.

[0058] Where k U As shown in equation (14), the adaptive weighting coefficients for the remaining output values ​​can be calculated according to equation (14):

[0059]

[0060] Among them mic U For u dc The sum of MIC values ​​for each interval is used during iteration. For each particle, the algorithm updates pbest and gbest, and then the objective function F for each sample is calculated. I Save the optimal value and the corresponding parameter value;

[0061] Step C2, Improve the inertia weight ω: Adopt a non-linear decreasing method for the inertia weight so that ω is adaptively adjusted with the number of particle iterations, ensuring the global search capability of the particle in the early stage while improving the local optimization capability in the later stage. The improvement is shown in formula (15):

[0062]

[0063] In the formula, ω min and ω max Let ω be the minimum and maximum values, and e and G be the current and total number of iterations, respectively.

[0064] Step C3, Improved learning factors c1, c2: The learning factors c1 and c2 in the particle velocity update formula reflect the optimization ability of individual particles and the population. Appropriate learning factors can improve the identification accuracy, shorten the convergence time and reduce the probability of getting trapped in local optima. The learning factors decrease nonlinearly with the increase of the number of iterations during the optimization process. Improved formula (16):

[0065]

[0066] Preferably, step five specifically includes the following steps:

[0067] Step D1: To further verify the effectiveness and practicality of the proposed model, the LSTM-IPSO algorithm is compared with the PSO algorithm and the LSTM-PSO algorithm. The same dataset is used to train the above model, and the three sets of identification curves are compared with the measured curves at voltage drop levels of 20%, 40%, 60%, and 80%.

[0068] Step D2: Calculate the average deviation of the three model identification results in intervals ①, ②, ③, ④ and ⑤ for voltage drop of 20%, 40%, 60% and 80% respectively according to formula (17), and then take the average of the average deviations under the four working conditions to obtain the error of the PSO, LSTM-PSO and LSTM-IPSO model output results in each interval.

[0069]

[0070] In the formula, F Vi For interval V i The average deviation, u M To measure the DC capacitor voltage, u i To identify the DC capacitor voltage of the model, K start and K end These are the sequence numbers of the first and last simulation data in the interval.

[0071] The beneficial effects of this invention are:

[0072] 1. This invention obtains the initial value and optimization range of IPSO based on the prediction results of control parameters of LSTM neural network, and uses IPSO algorithm as a secondary optimization method for accurate identification to achieve the purpose of accurate optimization. It solves the technical problem that traditional identification methods are difficult to accurately identify the control parameters of the electromagnetic model of doubly fed wind turbine under low voltage ride-through conditions.

[0073] 2. In order to remove irrelevant features, the maximum information coefficient is used to analyze the correlation between the input control parameters and the output feature variables, and the key output signal indicators are selected as the input features of the LSTM neural network.

[0074] 3. Compared with previous parameter identification methods, the parameter identification method for doubly fed wind turbine controllers based on LSTM-LSTM in this invention simulates the input-output characteristics of the wind turbine grid-side control system through an LSTM training dataset. Without running the wind turbine model, the measured data is input into the LSTM neural network to obtain the control parameter prediction results.

[0075] 4. This method, based on the adaptive weighting approach, dynamically adjusts the weight allocation non-linearly according to the MIC value and actual value of each output value when constructing the objective function for multiple output values. This allows the objective function for multiple output values ​​to adaptively adjust for output values ​​with high correlation or small actual values, effectively solving the problem of poor algorithm optimization performance caused by improper objective function construction. It also applies adaptive inertia weights and adaptive learning factors to reduce the optimization range in the later stages of the identification algorithm and improve parameter identification accuracy. Attached Figure Description

[0076] Figure 1 This is a schematic diagram of the doubly fed fan control parameter identification method of the present invention;

[0077] Figure 2 The MIC value of each control parameter is the input characteristic value of each interval of the DC capacitor voltage in this invention.

[0078] Figure 3 The MIC values ​​of each control parameter are the input characteristic values ​​of each DC interval of the d-axis in this invention.

[0079] Figure 4 The MIC values ​​of the input characteristic values ​​of each interval of the q-axis DC region in this invention are relative to the control parameters.

[0080] Figure 5 This is an iterative graph of the loss function during the training of the LSTM model in this invention;

[0081] Figure 6 The mean square error iterative diagrams of the LSTM-IPSO identification algorithm, PSO identification algorithm and LSTM-PSO identification algorithm of the present invention are shown.

[0082] Figure 7 This is a comparison of the DC capacitor voltage of the LSTM-IPSO identification curve, PSO identification curve, and LSTM-PSO identification curve of the present invention with the measured curve under a voltage drop of 20%.

[0083] Figure 8 This is a DC comparison graph of the d-axis of the LSTM-IPSO identification curve, PSO identification curve, and LSTM-PSO identification curve of the present invention with the measured curve under a voltage drop of 20%.

[0084] Figure 9 This is a DC comparison of the q-axis of the LSTM-IPSO identification curve, PSO identification curve, and LSTM-PSO identification curve of the present invention with the measured curve under a voltage drop of 20%.

[0085] Figure 10This is a comparison of the DC capacitor voltage of the LSTM-IPSO identification curve, PSO identification curve, and LSTM-PSO identification curve of the present invention with the measured curve under a voltage drop of 80%.

[0086] Figure 11 This is a DC comparison graph of the d-axis of the LSTM-IPSO identification curve, PSO identification curve, and LSTM-PSO identification curve of the present invention with the measured curve under a voltage drop of 80%.

[0087] Figure 12 This is a DC comparison of the q-axis of the LSTM-IPSO identification curve, PSO identification curve, and LSTM-PSO identification curve of the present invention with the measured curve under a voltage drop of 80%.

[0088] Figure 13 Error diagrams showing the DC capacitor voltage identification results of the LSTM-IPSO algorithm, PSO algorithm, and LSTM-PSO algorithm of this invention;

[0089] Figure 14 Error diagrams of the d-axis DC identification results of the LSTM-IPSO algorithm, PSO algorithm and LSTM-PSO algorithm of this invention;

[0090] Figure 15 This is an error graph showing the q-axis DC identification results of the LSTM-IPSO algorithm, PSO algorithm, and LSTM-PSO algorithm of this invention. Detailed Implementation

[0091] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0092] like Figure 1-15 As shown, a method for identifying control parameters of a doubly fed wind turbine based on LSTM-IPSO includes the following steps:

[0093] Step 1: Obtain hardware-in-the-loop experimental data of the doubly-fed wind turbine from the real controller using the RT-LAB hardware-in-the-loop simulation platform, and build an isomorphic identification model of the real controller of the doubly-fed wind turbine on the Matlab / simulink platform.

[0094] Step 2: Increase the dimension of the input feature set and remove irrelevant features, selecting highly relevant feature values ​​as the input feature set of the neural network model; the input feature set and the corresponding control parameter set constitute the control parameter-input feature set;

[0095] Step 3: Use an LSTM neural network to train and predict the control parameters-input feature set to obtain the initial prediction values ​​and the optimization range;

[0096] Step 4: Based on the predicted initial value and optimization range obtained in Step 4, the IPSO algorithm is used as a secondary optimization method for accurate identification to achieve the goal of accurate optimization.

[0097] Step 5: Compare and verify the identified model with the hardware-in-the-loop experimental data to determine the reliability of the identified model.

[0098] Preferably, in step one, the hardware-in-the-loop experimental data of the doubly-fed wind turbine derived from the real controller is obtained using the RT-LAB hardware-in-the-loop simulation platform, including the DC bus voltage u. dc The dq component of the output current i dg and i qg Where the d-axis reference current i dg It is obtained by DC voltage control, while the q-axis reference current i qg Set it to zero, and build an isomorphic identification model of the real controller of the doubly fed wind turbine on the Matlab / simulink platform.

[0099] Preferably, the doubly-fed wind turbine includes a wind turbine, a gearbox, an induction asynchronous generator, a back-to-back converter, and its controller; for the grid-side converter, its purpose is to maintain a stable DC connection voltage, ensure a good sinusoidal output current, and control the power factor. Therefore, the d-axis reference current i dg It is obtained by DC voltage control, while the q-axis reference current i dq It is usually set to zero.

[0100] The control equations involved in building the isomorphic identification model of a real controller for a doubly-fed wind turbine on the Matlab / Simulink platform are as follows:

[0101] According to the law of conservation of energy, the KCL equation for the DC capacitor C is:

[0102]

[0103] In the formula, P r P is the active power output from the machine-side converter to C. rc Active power injected into the grid by the grid-side converter;

[0104] The dynamic model of the filter reactance is as follows:

[0105]

[0106] In the formula, u cd u cq with u sd u sq X represents the d-axis and q-axis components of the voltage at the output of the grid-side converter and the generator-side converter, respectively. r This is the filter reactance value.

[0107] The control equations for the grid-side controller can be written as:

[0108]

[0109] In the formula, u dcref i is the reference value for DC voltage. dgref and i qgref These are the actual values ​​of the d-axis and q-axis currents on the grid side, respectively, K p1 For the proportional coefficient of the voltage outer loop PI controller, K p2 K is the proportionality coefficient of the inner loop of the q-axis current. p3 Let x1, x2, and x3 be the proportional coefficient of the inner loop of the q-axis current control. Then, the dynamic equation of the controller is:

[0110]

[0111] In the formula, K i1 For the integral coefficient of the voltage outer loop PI controller, K i2 K represents the integral coefficient of the inner loop of the q-axis current. i3 i is the integral coefficient of the inner loop of the q-axis current control. qgref It is usually 0, but under fault conditions, i qgref The value is directly given by the low-profile control module:

[0112]

[0113]

[0114] In the formula, k is the reactive current support coefficient, and U N I is the rated voltage at the grid connection point of the wind turbine. N The rated current of GSC is ; when low voltage ride-through occurs, the grid-side converter will generate a certain amount of reactive power. The value of the reactive current can be obtained according to equation (5). At this time, the inner loop of the q-axis current switches to the low voltage ride-through control mode, and the q-axis reference value is given by the low voltage ride-through control module.

[0115] Preferably, step two specifically includes the following steps:

[0116] Step A1: Collect the isomorphic identification model control parameters - input feature set, where the input feature values ​​are parameters related to the wind turbine grid-side controller: u dc i dg and i qg The error value F between the control parameter set and the hardware-in-the-loop test data, and the control parameter set is the control parameter dataset of the grid-side converter corresponding to the input feature value;

[0117] Step A2: Remove sample data with outlier data points to avoid affecting the accuracy of identification;

[0118] Step A3: According to the "Guidelines for Modeling Electrical Simulation Models of Wind Turbine Units", the simulation time is divided into 5 intervals: steady-state intervals ① and ⑤, low-voltage ride-through intervals ②, ③ and ④, and the i in the input feature value is... dg and i qg This method also involves dividing the data into intervals and using them as feature sets, where u dc The eigenvalue V is obtained by dividing the intervals and taking values. a V b V c V d and V e i dg The eigenvalue I is obtained by dividing the interval and taking values. da I db I dc I dd and I qe i qg The eigenvalue I is obtained by dividing the interval and taking values. qa I qb I qc I qd and I qe ;

[0119] Step A4: Use the maximum information coefficient (MIC) to analyze the correlation between the input control parameter and the output feature variable. First, the scatter plot of the control parameter (K) and the feature variable (T) is gridded with column a and row b to find the maximum mutual information value. Then, the maximum mutual information value is normalized. Finally, the maximum mutual information value at different scales is selected as the MIC value.

[0120] Mutual information can be viewed as the amount of information contained in one random variable about another. The mutual information I(K,T) between K and T is defined as:

[0121]

[0122] In the formula, K is the sample sum of variable k, T is the sample sum of variable t, P(k,t) is the joint probability between variables k and t, and P(k) and P(t) are the marginal probabilities of k and t, respectively.

[0123]

[0124] In the formula, n is the number of samples;

[0125] The MIC value obtained from equation (8) is between 0 and 1. The larger the MIC value, the higher the correlation between the two variables. After obtaining the MIC value of the output signal, the key output signal index is selected as the input feature of the LSTM neural network.

[0126] Step A5: Normalize the dataset. Dimensionless data can speed up the gradient descent process to find the optimal solution.

[0127] Preferably, in step three, an LSTM neural network is used to make preliminary predictions of the initial values ​​and optimization range of the doubly fed wind turbine control parameters based on the processed dataset, making it approximate the minimum objective function of the identification model. After extensive debugging, the structural parameters and network training parameters of the LSTM neural network are obtained, and the LSTM neural network model is trained and tested. Its core cell state c is controlled by the forget gate, input gate, and output gate, where: σ is the sigmoid function, x t For the input of the t-th unit, c t Let h be the cell state of the t-th unit. t Let ⊕ be the hidden state of the t-th unit. For summing and multiplying vector elements, c t-1 h represents the cell state of the (t-1)th unit. t-1 This represents the hidden state of the (t-1)th unit;

[0128] The output parameters of an LSTM neural network change under the influence of the input values, and this change can be expressed by the following formula:

[0129]

[0130] In the formula, f t W is the output value for the forget gate. i W f and W o b represents the network layer weights for the input gate, forget gate, and output gate. i b f and b o b is the bias term for the input gate, forget gate, and output gate. c It is the network layer bias. o represents the information input to the neuron at time t. t ☉ represents the output value of the output gate, and ☉ represents the Hadamard product operator.

[0131] Preferably, in step three, the specific steps for training and prediction are as follows:

[0132] Step B1, the reasonable selection and allocation of datasets and parameter tuning, has a great impact on the accuracy of prediction results. Randomly select 90% of the dataset as the training set and 10% of the dataset as the validation set to verify the accuracy of the model.

[0133] Step B2: Set the number of training iterations to 500, the loss function to MSE, the optimizer to Adam, the input dimension to 12, the output dimension to 6, the number of hidden layers to 2, the number of hidden layer neurons to 11, and the learning rate to 0.015.

[0134] Step B3: Use an LSTM neural network to train and predict the collected sequence, and obtain the initial values ​​of the predicted parameters and the optimization range.

[0135] Preferably, in step four, the PSO algorithm speed and position update formulas are as follows:

[0136]

[0137]

[0138] In the formula, V represents the particle update speed, x represents the particle, u represents the current time, u+1 represents the next update time, c1 and c2 are learning factors, r1 and r2 are random values ​​between 0 and 1, pbest is the individual optimal value, gbest is the global optimal value, and ω is the adaptive inertia weight.

[0139] PSO uses the mean squared difference between the output value at the actual position and the output value at the optimal position of each particle as the objective function of the algorithm, and obtains the fitness value of each particle at the current iteration number. The objective function is expressed as:

[0140]

[0141] In the formula, U i,error I di,error and I qi,error For u dc i dg and i qg The error, U i I di and I qi For u dc i dg and i qg The measured data value, U ti I dti and I qti For u dc i dg and i qg The output response value obtained from the identification results;

[0142] Step C1, Improve the objective function: Based on the adaptive weighting method, when constructing the objective function for multiple output values, the weight allocation is dynamically adjusted non-linearly according to the MIC value and actual value of each output value. This allows the objective function for multiple output values ​​to adaptively adjust for output values ​​with high correlation or small actual values. The improved formula is as follows:

[0143]

[0144] In the formula, k U k ID and k IQ For U i I di and I qi Adaptive weighting coefficients.

[0145] Where k U As shown in equation (14), the adaptive weighting coefficients for the remaining output values ​​can be calculated according to equation (14):

[0146]

[0147] mic U For u dc The sum of MIC values ​​for each interval is used during iteration. For each particle, the algorithm updates pbest and gbest, and then the objective function F for each sample is calculated. I Save the optimal value and the corresponding parameter value;

[0148] Step C2, Improve the inertia weight ω: Adopt a non-linear decreasing method for the inertia weight so that ω is adaptively adjusted with the number of particle iterations, ensuring the global search capability of the particle in the early stage while improving the local optimization capability in the later stage. The improvement is shown in formula (15):

[0149]

[0150] In the formula, ω min and ω max Let ω be the minimum and maximum values, and e and G be the current and total number of iterations, respectively.

[0151] Step C3, Improved learning factors c1, c2: The learning factors c1 and c2 in the particle velocity update formula reflect the optimization ability of individual particles and the population. Appropriate learning factors can improve the identification accuracy, shorten the convergence time and reduce the probability of getting trapped in local optima. The learning factors decrease nonlinearly with the increase of the number of iterations during the optimization process. Improved formula (16):

[0152]

[0153] Preferably, step five specifically includes the following steps:

[0154] Step D1: To further verify the effectiveness and practicality of the proposed model, the LSTM-IPSO algorithm is compared with the PSO algorithm and the LSTM-PSO algorithm. The same dataset is used to train the above model, and the three sets of identification curves are compared with the measured curves at voltage drop levels of 20%, 40%, 60%, and 80%.

[0155] Step D2: Calculate the average deviation of the three model identification results in intervals ①, ②, ③, ④ and ⑤ for voltage drop of 20%, 40%, 60% and 80% respectively according to formula (17), and then take the average of the average deviations under the four working conditions to obtain the error of the PSO, LSTM-PSO and LSTM-IPSO model output results in each interval.

[0156]

[0157] In the formula, F Vi For interval V i The average deviation, u M To measure the DC capacitor voltage, u i To identify the DC capacitor voltage of the model, K start and K end These are the sequence numbers of the first and last simulation data in the interval.

[0158] Figure 6 The convergence curves of the mean squared error corresponding to the minimum fitness value obtained by each algorithm are presented. The PSO algorithm has a larger initial value and changes less in the later stages, but the error update speed is significantly slower. The LSTM-PSO algorithm has a smaller initial value, but changes less during the iteration process. The LSTM-IPSO algorithm has the smallest initial value and the strongest local search ability. Due to its adaptive adjustment, it finds the smallest target value in the subsequent extreme value search optimization process, and has advantages in terms of optimization accuracy and convergence speed.

[0159] like Figure 13 , 14 As shown in Figure 15, the output results of the PSO algorithm, LSTM-PSO algorithm, and LSTM-IPSO algorithm models are u. dc i dg and i qg The average deviation in intervals ①, ②, ③, ④, and ⑤. Different algorithms affect the final control parameter K. p1 K i1 K p2 K i K i3 and K i3The identification effect is greatly affected by the PSO algorithm. The average deviation of the output signal results of the control parameters identified by the PSO algorithm is relatively large in intervals ①, ②, ③, ④, and ⑤. Adding the LSTM algorithm to perform preliminary optimization of the objective function, obtaining the predicted control parameters and the optimization range, and then using the PSO algorithm to iteratively optimize the objective function, results in more accurate identification results. The sum of the average deviations of the identified output signal results in intervals ①, ②, ③, ④, and ⑤ is smaller than that of the PSO algorithm. After improving the PSO algorithm, the identification results become even more accurate. The LSTM-IPSO algorithm has smaller identification errors in intervals ①, ②, ③, ④, and ⑤, with average deviations all less than 5%, demonstrating feasibility. Through comparative analysis, the algorithm proposed in this paper has advantages over the PSO algorithm and the LSTM-PSO algorithm in terms of computational accuracy and convergence speed, and is suitable for the identification of control parameters in wind turbine transient models.

[0160] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The embodiments and features described in these embodiments can be arbitrarily combined without conflict. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A method for identifying control parameters of a doubly fed wind turbine based on LSTM-IPSO, characterized in that: It includes the following steps: Step 1: Obtain hardware-in-the-loop experimental data of the doubly-fed wind turbine from the real controller using the RT-LAB hardware-in-the-loop simulation platform, and build an isomorphic identification model of the real controller of the doubly-fed wind turbine on the Matlab / simulink platform. Step 2: Increase the dimension of the input feature set and remove irrelevant features, selecting highly relevant feature values ​​as the input feature set of the neural network model; the input feature set and the corresponding control parameter set constitute the control parameter-input feature set; Step 3: Use an LSTM neural network to train and predict the control parameters-input feature set to obtain the initial prediction values ​​and the optimization range; Step 4: Based on the predicted initial value and optimization range obtained in Step 4, the IPSO algorithm is used as a secondary optimization method for accurate identification to achieve the goal of accurate optimization. Step 5: Compare and verify the identified model with the hardware-in-the-loop experimental data to determine the reliability of the identified model; Step one describes using the RT-LAB hardware-in-the-loop simulation platform to obtain hardware-in-the-loop experimental data of the doubly-fed wind turbine from the actual controller, including the DC bus voltage. u dc dq component of output current i dg and i qg d-axis reference current i dg It is obtained by DC voltage control, while the q-axis reference current is... i qg Set to zero and build an isomorphic identification model of the real controller of the doubly fed wind turbine on the Matlab / simulink platform; The doubly fed wind turbine includes a wind turbine, a gearbox, an induction asynchronous generator, a back-to-back converter, and its controller. The control equations involved in building the isomorphic identification model of a real controller for a doubly-fed wind turbine on the Matlab / Simulink platform are as follows: According to the law of conservation of energy, the DC capacitor can be obtained. C The KCL equation is: (1)； In the formula, P r For the output of the machine-side converter to C active power, P rc Active power injected into the grid by the grid-side converter; The dynamic model of the filter reactance is as follows: (2)； In the formula, u cd , u cq and u sd , u sq These represent the d-axis and q-axis components of the voltage at the output of the grid-side converter and the generator-side converter, respectively. X r This is the filter reactance value; The control equations for the grid-side controller are written as follows: (3)； In the formula, u dcref This is a reference value for DC voltage. i dgref and i qgref These are the actual values ​​of the d-axis and q-axis currents on the grid side, respectively. K p1 For the proportional coefficient of the voltage outer loop PI controller, K p2 This is the proportionality coefficient of the inner loop of the q-axis current. K p3 Let the proportional coefficient of the inner loop of the q-axis current control be denoted as , and let the intermediate variables output by the integral element of the PI controller be denoted as . x 1. x 2 and x 3. The dynamic equation of the controller is: (4)； In the formula, K i1 The integral coefficient of the voltage outer loop PI controller, K i2 The integral coefficient of the inner loop of the q-axis current. K i3 The integral coefficient of the inner loop for q-axis current control. i qgref Normally it is 0, but under fault conditions, i qgref The value is directly given by the low-profile control module: (5)； (6)； In the formula, k This is the reactive current support factor. U N The rated voltage at the grid connection point of the wind turbine. I N The rated current of GSC is ; when low voltage ride-through occurs, the grid-side converter will generate a certain amount of reactive power. The value of the reactive current can be obtained according to equation (5). At this time, the inner loop of the q-axis current switches to the low-voltage ride-through control mode, and the q-axis reference value is given by the low-voltage ride-through control module. Step two specifically includes the following steps: Step A1: Collect the isomorphic identification model control parameters - input feature set, where the input feature values ​​are parameters related to the wind turbine grid-side controller: u dc , i dg and i qg and the error value between it and the hardware-in-the-loop test data F The control parameter set is the control parameter dataset of the grid-side converter corresponding to the input feature values; Step A2: Remove sample data with outlier data points to avoid affecting the accuracy of identification; Step A3: According to the "Guidelines for Modeling Electrical Simulation Models of Wind Turbine Units", the simulation time is divided into 5 intervals: steady-state intervals ① and ⑤, low-voltage ride-through intervals ②, ③ and ④, and the input feature values ​​are... i dg and i qg This method also involves selecting values ​​from intervals as the feature set, where... u dc The eigenvalues ​​are obtained by dividing the values ​​into intervals. V a , V b , V c , V d and V e , i dg The eigenvalues ​​are obtained by dividing the values ​​into intervals. I da , I db , I dc , I dd and I qe , i qg The eigenvalues ​​are obtained by dividing the values ​​into intervals. I qa , I qb , I qc , I qd and I qe ; Step A4: Use the Maximum Information Coefficient (MIC) to analyze the correlation between the input control parameters and the output characteristic variables. First, analyze the correlation between the control parameters ( K ) and characteristic variables ( T The scatter plot is gridded with column a and row b to find the maximum mutual information value. Then the maximum mutual information value is normalized, and finally the maximum mutual information value at different scales is selected as the MIC value. Mutual information is the amount of information contained in one random variable about another random variable. K and T mutual information I ( K,T ) is defined as: (7)； In the formula, K For variables k The sample and, T For variables t The sample and, P ( k,t ) is a variable k and t The joint probability between them P ( k )and P ( t ) are respectively k and t The marginal probability; (8)； In the formula, n The number of samples; The MIC value obtained from equation (8) is between 0 and 1. The larger the MIC value, the higher the correlation between the two variables. After obtaining the MIC value of the output signal, the key output signal index is selected as the input feature of the LSTM neural network. Step A5: Normalize the dataset. Dimensionless data can speed up the gradient descent process to find the optimal solution. In step four, the PSO algorithm's speed and position update formulas are as follows: (10)； (11)； In the formula, V Indicates the particle update rate, x Represents particles, u Indicates the current moment. u +1 indicates the time of the next update. c 1. c 2 is the learning factor. r 1. r 2 is a random value between 0 and 1. pbest For the individual optimal value, gbest The global optimal value. ω For adaptive inertia weights; PSO uses the mean squared difference between the output value at the actual position and the output value at the optimal position of each particle as the objective function of the algorithm, and obtains the fitness value of each particle at the current iteration number. The objective function is expressed as: (12)； In the formula, U i,error , I di,error and I qi,error for u dc , i dg and i qg The error, U i , I di and I qi for u dc , i dg and i qg The measured data values, U ti , I dti and I qti for u dc , i dg and i qg The output response value obtained from the identification results; Step C1, Improve the objective function: Based on the adaptive weighting method, when constructing the objective function for multiple output values, the weight allocation is dynamically adjusted non-linearly according to the MIC value and actual value of each output value. This allows the objective function for multiple output values ​​to adaptively adjust for output values ​​with high correlation or small actual values. The improved formula is as follows: (13)； In the formula, k U , k ID and k IQ for U i , I di and I qi Adaptive weighting coefficients; in k U As shown in equation (14), the adaptive weighting coefficients for the remaining output values ​​are calculated according to equation (14): (14)； in mic U for u dc The sum of the MIC values ​​in each interval is updated using an algorithm for each particle during the iteration process. pbest and gbest Then calculate the objective function for each sample. F I Save the optimal value and the corresponding parameter value; Step C2, Improve inertia weight ω : By employing a non-linear decreasing method for inertial weights, ω The algorithm adaptively adjusts with the number of particle iterations, ensuring the global search capability of the particle in the early stage while improving the local optimization capability in the later stage. The improvement is shown in formula (15): (15)； In the formula, ω min and ω max for ω Minimum and maximum values e G represents the current and total number of iterations; Step C3: Improve learning factors c 1, c 2: The learning factor in the particle velocity update formula c 1 and c 2 reflects the optimization ability of individual particles and the population. A suitable learning factor can improve the identification accuracy, shorten the convergence time and reduce the probability of getting trapped in local optima. The learning factor decreases nonlinearly with the number of iterations during the optimization process. Improved formula (16): (16)。 2. The method for identifying control parameters of a doubly fed wind turbine based on LSTM-IPSO according to claim 1, characterized in that: In step three, an LSTM neural network is used to make preliminary predictions on the initial values ​​and optimization range of the doubly fed wind turbine control parameters based on the processed dataset, so that it approximates the minimum objective function of the identification model. After a lot of debugging work, the structural parameters and network training parameters of the LSTM neural network are obtained, and the LSTM neural network model is trained and tested. Its core cell state c Controlled by the forget gate, input gate, and output gate, among which: σ For the sigmoid function, x t For the first t The input of each unit, c t For the first t The cellular state of each unit, h t For the first t The hidden states of each unit, and Sum and multiply vector elements. c t-1 For the first t- The cell state of a single unit. h t-1 For the first t- The hidden state of one unit; The output parameters of an LSTM neural network change under the influence of the input values, as expressed by the following formula: (9)； In the formula, f t Output the value for the forget gate. W i 、W f and W o These are the network layer weights for the input gate, forget gate, and output gate. b i 、b f and b o These are the bias terms for the input gate, forget gate, and output gate. b c It is the network layer bias. for t Information constantly input into neurons o t ☉ represents the output value of the output gate, and ☉ represents the Hadamard product operator.

3. The method for identifying control parameters of a doubly fed wind turbine based on LSTM-IPSO according to claim 2, characterized in that: In step three, the specific steps for training and prediction are as follows: Step B1, the reasonable selection and allocation of datasets and parameter tuning, has a great impact on the accuracy of prediction results. Randomly select 90% of the dataset as the training set and 10% of the dataset as the validation set to verify the accuracy of the model. Step B2: Set the number of training iterations to 500, the loss function to MSE, the optimizer to Adam, the input dimension to 12, the output dimension to 6, the number of hidden layers to 2, the number of hidden layer neurons to 11, and the learning rate to 0.

015. Step B3: Use an LSTM neural network to train and predict the collected sequence, and obtain the initial values ​​of the predicted parameters and the optimization range.

4. The method for identifying control parameters of a doubly fed wind turbine based on LSTM-IPSO according to claim 1, characterized in that: Step five specifically includes the following steps: Step D1: To further verify the effectiveness and practicality of the proposed model, the LSTM-IPSO algorithm is compared with the PSO algorithm and the LSTM-PSO algorithm. The same dataset is used to train the above model, and the three sets of identification curves are compared with the measured curves at voltage drop levels of 20%, 40%, 60% and 80%. Step D2: Calculate the average deviation of the three model identification results in intervals ①, ②, ③, ④ and ⑤ for voltage drop of 20%, 40%, 60% and 80% respectively according to formula (17), and then take the average of the average deviations under the four working conditions to obtain the error of the PSO, LSTM-PSO and LSTM-IPSO model output results in each interval. (17)； In the formula, F Vi For interval V i The average deviation, u M To measure the DC capacitor voltage, u i To identify the DC capacitor voltage of the model, K start and K end These are the sequence numbers of the first and last simulation data in the interval.

Images