A method and system for enhancing stability of a converter of a doubly-fed wind power series compensation grid-connected system

Abstract

The present application relates to the field of wind power grid connection stability, in particular to a method and system for enhancing stability of a converter of a doubly-fed wind power series compensation grid connection system. The method comprises: collecting rotor current signals and initializing a reference frequency; extracting subsynchronous oscillation components through recursive least squares; performing sliding window spectrum transformation to identify a dominant oscillation frequency, and updating the reference frequency when the stability condition is met; inputting historical spectrum features into a long short-term memory network to predict the oscillation frequency, and updating the reference frequency according to the prediction deviation; calculating a compensation voltage instruction based on the reference frequency and the oscillation components, and superimposing the compensation voltage instruction on a converter control loop to establish a local control model; training the local model by taking multiple units as federated learning clients, and weighting and aggregating model parameters by a central server and then distributing updates. The present application realizes adaptive oscillation suppression and multi-station collaborative optimization, and guarantees data privacy and security.

Description

Technical Field

[0001] This invention relates to the field of wind power grid connection stability technology, specifically to a method and system for enhancing the stability of a converter in a doubly fed wind power series compensation grid-connected system. Background Technology

[0002] Wind power has become a strategic energy choice due to its abundant reserves, wide distribution, and clean and environmentally friendly characteristics. To improve the efficiency of long-distance power transmission, series capacitor-compensated transmission lines are widely used for wind farm grid connection.

[0003] However, when doubly-fed induction generator (DFIG) wind farms are connected to the grid via series compensation lines, the characteristics of the induction generator and the converter control mechanism interact with the line compensation, leading to subsynchronous oscillation stability issues. In existing technologies, damping control methods based on fixed parameters are highly dependent on system operating conditions; when disturbances such as series compensation degree or wind speed change, the oscillation suppression effect significantly decreases. While adaptive methods based on converter control have some robustness, their frequency tracking response speed under dynamic operating conditions is insufficient, making it difficult to adapt to rapidly changing oscillation characteristics in real time. Furthermore, with the large-scale distributed deployment of wind farms, there is a lack of collaborative optimization mechanisms between stations, and centralized data processing faces security risks such as measurement data privacy leaks and the theft of network topology parameters. Summary of the Invention

[0004] This invention provides a method and system for enhancing the stability of a converter in a doubly-fed wind power series-compensated grid-connected system, aiming to solve the problems of lack of adaptability in converter control and data security risks in doubly-fed wind power series-compensated grid-connected systems, and improve system stability.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] This invention discloses a method for enhancing the stability of a converter in a doubly-fed wind power series-compensated grid-connected system, comprising:

[0007] S100: Collects the rotor current signal of the rotor-side converter of the doubly fed wind turbine, and initializes the reference frequency to the preset reference frequency;

[0008] S200: Construct a harmonic basis function based on the reference frequency, fit the rotor current signal using the recursive least squares method, and extract the subsynchronous oscillation component corresponding to the reference frequency;

[0009] S300: Perform a sliding window spectrum transformation on the rotor current signal, obtain the spectrum characteristics of each window and identify the current dominant oscillation frequency. When the dominant oscillation frequency meets the preset stability condition in multiple consecutive windows, update the reference frequency to the dominant oscillation frequency.

[0010] S400: Input the spectral characteristics of multiple historical windows into the long short-term memory network to obtain the predicted oscillation frequency output by the network, and update the reference frequency according to the deviation between the predicted oscillation frequency and the reference frequency;

[0011] S500: Calculate the compensation voltage command based on the reference frequency and the subsynchronous oscillation component, superimpose the compensation voltage command onto the voltage control loop of the rotor-side converter, and establish a local control model including the weight parameters of the long short-term memory network.

[0012] S600: Multiple doubly fed wind turbine units are used as federated learning clients. Each client trains the local control model and uploads the model weight parameters to the central server. The central server performs weighted aggregation on the received model weight parameters to obtain global weight parameters, and then distributes the global weight parameters to each client to update the local control model.

[0013] As a preferred embodiment of the present invention, the extraction step of the subsynchronous oscillation component specifically includes:

[0014] Construct a harmonic basis function composed of cosine and sine components based on the reference frequency;

[0015] The rotor current signal is iteratively fitted using the recursive least squares method, and the covariance matrix and gain vector are updated by the forgetting factor to obtain the coefficients of the cosine component and the coefficients of the sine component.

[0016] The subsynchronous oscillation component is reconstructed based on the coefficients of the cosine component and the coefficients of the sine component.

[0017] As a preferred embodiment of the present invention, S300 specifically includes:

[0018] Set the window length and window sliding step size, wherein the window length is determined based on the sampling frequency and the target frequency resolution;

[0019] The rotor current signal is subjected to a fast Fourier transform according to the window sliding step size to obtain the amplitude spectrum of multiple windows.

[0020] The frequency point with the largest amplitude within the subsynchronous frequency range of the amplitude spectrum is identified as the current dominant oscillation frequency, and the current dominant oscillation frequency and its corresponding amplitude are recorded as the spectral feature.

[0021] As a preferred embodiment of the present invention, the preset stability conditions include:

[0022] Within a consecutive preset number of windows, the variation amplitude of the dominant oscillation frequency does not exceed a preset frequency threshold, and the amplitude corresponding to the dominant oscillation frequency is higher than a preset amplitude threshold.

[0023] The duration for which the dominant oscillation frequency remains stable within a predetermined number of consecutive windows is not less than a predetermined time threshold.

[0024] As a preferred embodiment of the present invention, the step of obtaining the predicted oscillation frequency specifically includes:

[0025] The spectral features of the multiple historical windows are normalized, and the spectral features include the dominant oscillation frequency and corresponding amplitude identified by the multiple historical windows.

[0026] Obtain the disturbance parameters at the corresponding time, including wind speed parameters and series complement parameters;

[0027] The normalized spectral features are combined with the perturbation parameters to form a feature vector, and then arranged in time series to form an input sequence.

[0028] The input sequence is input into the input layer of the long short-term memory network. The trend of the spectral characteristics changing with the perturbation parameters within the historical time period is preserved through the forget gate, input gate and output gate of the hidden layer. The predicted oscillation frequency is output by the output layer.

[0029] As a preferred embodiment of the present invention, the training process of the Long Short-Term Memory network includes:

[0030] Using the spectral characteristics and perturbation parameters of the historical window as input, the dominant oscillation frequency actually identified at the next moment is used as the label to construct training samples;

[0031] The mean squared error is used as the loss function to calculate the error between the predicted oscillation frequency and the actual identified dominant oscillation frequency;

[0032] The gradient is calculated using the backpropagation algorithm, and the weight parameters of the Long Short-Term Memory network are updated. The weight parameters include the forget gate weight, input gate weight, output gate weight, and weights from the hidden layer to the output layer.

[0033] As a preferred embodiment of the present invention, updating the reference frequency specifically includes:

[0034] Calculate the absolute value of the difference between the predicted oscillation frequency and the reference frequency;

[0035] When the absolute value of the difference is less than a preset deviation threshold, the reference frequency is updated to the predicted oscillation frequency;

[0036] When the absolute value of the difference is greater than or equal to the preset deviation threshold, the reference frequency remains unchanged.

[0037] As a preferred embodiment of the present invention, S500 specifically includes:

[0038] The gain coefficient is calculated based on the reference frequency, and the gain coefficient changes nonlinearly with the reference frequency.

[0039] The initial compensation voltage is obtained by multiplying the gain coefficient by the subsynchronous oscillation component;

[0040] The amplitude of the initial compensation voltage is limited to obtain the compensation voltage command;

[0041] The compensation voltage command is directly superimposed onto the reference voltage of the dq-axis voltage control circuit of the rotor-side converter.

[0042] As a preferred embodiment of the present invention, the updating of the local control model specifically includes:

[0043] Initialize global weight parameters and distribute them to all clients;

[0044] All clients use locally acquired rotor current signals, spectral characteristics, and disturbance parameters to train the local control model and calculate the update amount of the local model weight parameters relative to the initial parameters.

[0045] All clients upload the update data to the central server;

[0046] The central server obtains the amount of training data from each client and calculates the proportion of each client's training data to the total training data as the aggregation weight of the corresponding client.

[0047] The central server calculates a weighted average of the update volume uploaded by all clients according to their respective aggregation weights to obtain the global update volume.

[0048] The central server adds the global update amount to the current global weight parameter to obtain the updated global weight parameter, and then sends it to all clients to update their local control models.

[0049] This invention also proposes a converter stability enhancement system for a doubly-fed wind power series compensation grid-connected system, comprising:

[0050] The signal acquisition module is used to acquire the rotor current signal of the rotor-side converter of the doubly fed wind turbine, and initializes the reference frequency to the preset reference frequency.

[0051] The component extraction module is used to construct a harmonic basis function based on the reference frequency, fit the rotor current signal by recursive least squares method, and extract the subsynchronous oscillation component corresponding to the reference frequency.

[0052] The frequency identification module is used to perform a sliding window spectrum transformation on the rotor current signal, obtain the spectrum characteristics of each window and identify the current dominant oscillation frequency. When the dominant oscillation frequency meets the preset stability condition in multiple consecutive windows, the reference frequency is updated to the dominant oscillation frequency.

[0053] The frequency prediction module is used to input the spectral characteristics of multiple historical windows into a long short-term memory network to obtain the predicted oscillation frequency output by the network, and update the reference frequency according to the deviation between the predicted oscillation frequency and the reference frequency.

[0054] The compensation control module is used to calculate the compensation voltage command based on the reference frequency and the subsynchronous oscillation component, superimpose the compensation voltage command onto the voltage control loop of the rotor-side converter, and establish a local control model including the weight parameters of the long short-term memory network.

[0055] The federated learning module is used to treat multiple doubly-fed wind turbine units as federated learning clients. Each client trains the local control model and uploads the model weight parameters to the central server. The central server performs weighted aggregation on the received model weight parameters to obtain global weight parameters, and then distributes the global weight parameters to each client to update the local control model.

[0056] The beneficial effects of this invention are:

[0057] 1. In view of the problem of insufficient frequency tracking response speed in the prior art, this invention proposes a two-layer frequency tracking mechanism that combines sliding window spectrum recognition and long short-term memory network prediction. The disturbance parameters are integrated into the neural network to learn the oscillation evolution law, which realizes accurate capture and advance prediction of the dominant oscillation frequency, and significantly improves the adaptability and response speed of converter control.

[0058] 2. This invention is based on a dual-layer frequency tracking mechanism to adaptively update the reference frequency in real time and dynamically adjust the compensation voltage gain coefficient accordingly. This overcomes the problem of the traditional fixed parameter method being highly dependent on the system operating conditions. It achieves rapid response to changes in disturbance conditions such as series compensation degree and wind speed, and precise suppression of oscillations at different frequencies, effectively shortening the oscillation duration and improving system stability.

[0059] 3. To address the data privacy and network parameter security risks of large-scale distributed wind farms, this invention adopts a federated learning framework to achieve collaborative optimization among multiple wind farms. Each client only uploads the model parameter update, and the central server weights and aggregates the data to form a global model. Under the premise of ensuring data privacy and network parameter security, the control strategy has the ability to generalize across multiple disturbance conditions at different wind farms. Attached Figure Description

[0060] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0061] Figure 1 This is a topology diagram of a doubly fed wind farm series compensation grid-connected system;

[0062] Figure 2 This is a flowchart illustrating a method for enhancing the stability of a converter in a doubly fed wind power series compensation grid-connected system according to the present invention.

[0063] Figure 3 This is a schematic diagram of the rotor-side converter control of the present invention;

[0064] Figure 4 This is a schematic diagram of the structure of a converter stability enhancement system for a doubly fed wind power series compensation grid-connected system according to the present invention;

[0065] Figure 5(a) shows the DC side voltage variation curve under series compensation disturbance;

[0066] Figure 5(b) shows the electromagnetic torque variation curve under series compensation disturbance;

[0067] Figure 5(c) shows the DC side voltage variation curve under wind speed disturbance;

[0068] Figure 5(d) shows the electromagnetic torque variation curve under wind speed disturbance. Detailed Implementation

[0069] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0070] See Figure 1 This invention is applied to a doubly-fed wind farm series compensation grid-connected system. For example... Figure 1 As shown, the system includes N doubly-fed wind turbine generators (DFWG). Each generator is connected to a series-compensated transmission line via step-up transformers T1 and T2. The line includes line resistance. Line reactance and series compensation capacitor Ultimately, the system is connected to the grid via transformer T3. Each doubly-fed wind turbine includes a rotor-side converter and a grid-side converter. The rotor-side converter enables maximum power point tracking (MPPT) and stator reactive power control.

[0071] Example 1: As Figure 2 As shown, the present invention provides a method for enhancing the stability of a converter in a doubly-fed wind power series-compensated grid-connected system, comprising:

[0072] S100: Collects the rotor current signal of the rotor-side converter of the doubly fed wind turbine, and initializes the reference frequency to the preset reference frequency;

[0073] Specifically, the rotor current signal of the rotor-side converter of the doubly-fed wind turbine is acquired through a current sensor. The rotor current signal includes a d-axis component in a synchronous rotating coordinate system. and q-axis components .

[0074] Set sampling frequency Subsequently, the rotor-side current signal is sampled and collected in real time. Sampling frequency... The settings need to ensure that they can capture signal characteristics within the subsynchronous oscillation frequency range. Preferably, Set to 5kHz to 10kHz.

[0075] Initialize reference frequency Reference frequency The setting range is 5Hz to 45Hz, which covers the subsynchronous oscillation frequency range commonly found in series compensation systems.

[0076] The acquired rotor current signals are stored in time series, and the time series number corresponding to each sampling time is denoted as . The sampling time point is recorded as .

[0077] S200: Construct a harmonic basis function based on the reference frequency, fit the rotor current signal using the recursive least squares method, and extract the subsynchronous oscillation component corresponding to the reference frequency;

[0078] Furthermore, the extraction step of the subsynchronous oscillation component specifically includes:

[0079] Construct a harmonic basis function composed of cosine and sine components based on the reference frequency;

[0080] The rotor current signal is iteratively fitted using the recursive least squares method, and the covariance matrix and gain vector are updated by the forgetting factor to obtain the coefficients of the cosine component and the coefficients of the sine component.

[0081] The subsynchronous oscillation component is reconstructed based on the coefficients of the cosine component and the coefficients of the sine component.

[0082] Specifically, based on the reference frequency A harmonic basis function is constructed as the assumed oscillation baseline. The expression for the constructed harmonic basis function is:

[0083] ;

[0084] In the formula, For each sampling time point The harmonic basis functions constructed below, and These are the coefficients of the cosine and sine functions, respectively. The current reference frequency is set. The time series number corresponding to the sampling time point. The sampling frequency.

[0085] The collected rotor current signal is iteratively fitted using the recursive least squares method to estimate the coefficients. and The value of the value. Recursive least squares is an online parameter estimation algorithm that recursively updates parameter estimates by processing sampled data points one by one. This method introduces a forgetting factor to adjust the weight of historical and new data. The forgetting factor ranges from 0 to 1, with a preferred value of 0.95 to 0.99. A larger forgetting factor allows the algorithm to retain more historical information, while a smaller forgetting factor allows the algorithm to adapt to new data more quickly. The specific value of the forgetting factor is dynamically determined based on the system operating conditions: when disturbance parameters such as series complement or wind speed change rapidly, and the subsynchronous oscillation frequency changes at a high rate, a smaller forgetting factor, such as 0.95 to 0.96, is chosen to allow the algorithm to adapt to new data more quickly and improve frequency tracking speed; when the system operating conditions are relatively stable and the oscillation frequency changes slowly, a larger forgetting factor, such as 0.98 to 0.99, is chosen to allow the algorithm to retain more historical information, improve parameter estimation accuracy, and reduce noise interference.

[0086] The recursive least squares method involves initializing the covariance matrix and parameter vector, then calculating a gain vector for each new sampling point. This gain vector determines the magnitude and direction of the parameter update. The parameter estimates are updated using the gain vector and the error between the current measurement and the predicted value, while the covariance matrix is ​​updated to reflect the changes in estimation uncertainty. Through this recursive process, the optimal estimates of the cosine and sine component coefficients are gradually obtained. and .

[0087] Based on the estimated coefficients and Reconstructing the dominant frequency component of the subsynchronous oscillation (SSO):

[0088] ;

[0089] In the formula, This is the dominant frequency component of the subsynchronous oscillation (SSO). and These are the coefficients of the cosine and sine functions estimated using the recursive least squares method, respectively. This synchronous oscillation component represents the rotor current signal relative to the reference frequency. The corresponding oscillation component.

[0090] S300: Perform a sliding window spectrum transformation on the rotor current signal, obtain the spectrum characteristics of each window and identify the current dominant oscillation frequency. When the dominant oscillation frequency meets the preset stability condition in multiple consecutive windows, update the reference frequency to the dominant oscillation frequency.

[0091] Furthermore, S300 specifically includes:

[0092] Set the window length and window sliding step size, wherein the window length is determined based on the sampling frequency and the target frequency resolution;

[0093] The rotor current signal is subjected to a fast Fourier transform according to the window sliding step size to obtain the amplitude spectrum of multiple windows.

[0094] The frequency point with the largest amplitude within the subsynchronous frequency range of the amplitude spectrum is identified as the current dominant oscillation frequency, and the current dominant oscillation frequency and its corresponding amplitude are recorded as the spectral feature.

[0095] Furthermore, the preset stability conditions include:

[0096] Within a consecutive preset number of windows, the variation amplitude of the dominant oscillation frequency does not exceed a preset frequency threshold, and the amplitude corresponding to the dominant oscillation frequency is higher than a preset amplitude threshold.

[0097] The duration for which the dominant oscillation frequency remains stable within a predetermined number of consecutive windows is not less than a predetermined time threshold.

[0098] Specifically, a sliding window fast Fourier transform is performed on the acquired rotor-side current signal for spectral analysis. First, the sliding window length is set. and window sliding step size The window length N is determined based on the sampling frequency and the target frequency resolution. The frequency resolution is equal to the sampling frequency divided by the window length, i.e. To obtain sufficient frequency resolution to distinguish different subsynchronous oscillation frequencies, the window length... The preferred setting is 512 to 2048 sampling points. Window sliding step size. The time interval for spectrum updates is determined and is preferably set to one-quarter to one-half of the window length to balance computational burden and spectrum update speed.

[0099] Slide the window according to the set step size A sliding window fast Fourier transform is performed on the rotor current signal. Specifically, a segment of length [missing information] is extracted from the rotor current signal sequence. The data segment is processed, and a Fast Fourier Transform is performed on it to obtain the frequency domain representation of the window. Then the window is slid backward. For each sampling point, the next data segment is extracted and a Fast Fourier Transform is performed. This process is repeated to obtain the amplitude spectrum of multiple windows.

[0100] In the amplitude spectrum obtained in each window, the frequency point with the largest amplitude within the subsynchronous frequency range is identified as the current dominant oscillation frequency. The subsynchronous frequency range is defined as the frequency interval between 0Hz and the grid rated frequency, typically 0 to 50Hz or 0 to 60Hz. During the identification process, in addition to finding the maximum amplitude, it is also necessary to determine whether the amplitude at that frequency point is significantly higher than other frequency bands and conforms to oscillation characteristics, in order to avoid misidentifying noise or non-oscillating components as the dominant frequency. Record the identified dominant oscillation frequency. The amplitude and its corresponding value are used as the spectral features of this window.

[0101] Determine the dominant oscillation frequency Does it meet the preset stability conditions? The preset stability conditions include two aspects: first, within a preset number of consecutive windows, the dominant oscillation frequency... The variation range does not exceed the preset frequency threshold. And the dominant oscillation frequency The corresponding amplitudes are all higher than the preset amplitude threshold; secondly, the dominant oscillation frequency. The duration of stability within a consecutive preset number of windows is not less than a preset time threshold. .

[0102] Specifically, set the number of consecutive windows. This number is preferably set to 3 to 10 windows. Preset frequency threshold. Set to 1Hz to 3Hz to determine frequency stability. The preset amplitude threshold is determined based on the system noise level to ensure the identified dominant frequency has sufficient energy. Preset time threshold... Set it to 0.2 seconds to 1 second to ensure the frequency remains stable for a sufficient duration.

[0103] When the dominant oscillation frequency is identified In continuous The magnitude of change within each window is less than The corresponding amplitudes are all higher than the preset amplitude threshold, and the duration is not less than [a certain value]. When the frequency is determined to meet the stability condition, the reference frequency is then set. Updated to the dominant oscillation frequency ,Right now If the stability condition is not met, then maintain the current reference frequency. The reference frequency remains unchanged. This step locks the reference frequency to the dominant subsynchronous oscillation frequency of the actual system, ensuring correct identification and subsequent compensation of that frequency.

[0104] S400: Input the spectral characteristics of multiple historical windows into the long short-term memory network to obtain the predicted oscillation frequency output by the network, and update the reference frequency according to the deviation between the predicted oscillation frequency and the reference frequency;

[0105] Furthermore, the step of obtaining the predicted oscillation frequency specifically includes:

[0106] The spectral features of the multiple historical windows are normalized, and the spectral features include the dominant oscillation frequency and corresponding amplitude identified by the multiple historical windows.

[0107] Obtain the disturbance parameters at the corresponding time, including wind speed parameters and series complement parameters;

[0108] The normalized spectral features are combined with the perturbation parameters to form a feature vector, and then arranged in time series to form an input sequence.

[0109] The input sequence is input into the input layer of the long short-term memory network. The trend of the spectral characteristics changing with the perturbation parameters within the historical time period is preserved through the forget gate, input gate and output gate of the hidden layer. The predicted oscillation frequency is output by the output layer.

[0110] Specifically, the first step is data preprocessing. The collected data... The spectral characteristics of each historical window are normalized. The spectral characteristics include the dominant oscillation frequency identified in each window. And their corresponding amplitudes. Normalization processing uses min-max normalization or standardization methods to map frequency and amplitude data to the same numerical range, usually 0 to 1 or -1 to 1, to facilitate neural network training.

[0111] Obtain the disturbance parameters at the corresponding time moment. These parameters include wind speed and series complement parameters. The wind speed parameter, obtained from the wind speed disturbance module, represents the current input wind speed of the wind turbine. The series complement parameter, obtained from the series-compensated transmission line, represents the ratio of the capacitive reactance of the series compensation capacitor to the inductive reactance of the line. These disturbance parameters directly affect the subsynchronous oscillation characteristics of the system. Combine the normalized spectral characteristics with the disturbance parameters at the corresponding time moment to construct a feature vector. Each feature vector contains four elements: dominant frequency, corresponding amplitude, wind speed, and series complement. The feature vectors of the historical windows are arranged in a time sequence to form the input sequence, which has a dimension of . .

[0112] The input sequence is fed into the input layer of a Long Short-Term Memory (LSTM) network. The hidden layers of the LSTM network employ three gating mechanisms: a forget gate, an input gate, and an output gate. The forget gate determines which information is discarded from the cell state, the input gate determines which new information is added to the cell state, and the output gate determines which information is output. Through these three gating mechanisms, the hidden layers can retain the trend of spectral characteristics changing with perturbation parameters over historical time periods, learning the dynamic relationship between oscillation frequency and perturbation conditions. The number of hidden layer units in the LSTM network is preferably set to 32 to 128, which can be adjusted according to system complexity and data volume. After retaining some of the information memorized by the hidden layers, the output layer outputs the dominant oscillation frequency for the next time step, i.e., the predicted dominant oscillation frequency. .

[0113] Furthermore, the training process of the Long Short-Term Memory network includes:

[0114] Using the spectral characteristics and perturbation parameters of the historical window as input, the dominant oscillation frequency actually identified at the next moment is used as the label to construct training samples;

[0115] The mean squared error is used as the loss function to calculate the error between the predicted oscillation frequency and the actual identified dominant oscillation frequency;

[0116] The gradient is calculated using the backpropagation algorithm, and the weight parameters of the Long Short-Term Memory network are updated. The weight parameters include the forget gate weight, input gate weight, output gate weight, and weights from the hidden layer to the output layer.

[0117] Specifically, training samples are constructed using the spectral characteristics and perturbation parameters of historical windows as input, and the dominant oscillation frequency actually identified at the next moment is used. As a label. (Previous) The feature vector corresponding to the nth window is used as input, and the nth window... The dominant frequency identified by each window is used as a label to form a training sample. A large number of training samples are generated from historical data using a sliding window approach.

[0118] The mean square error (MSE) is used as the loss function to calculate the predicted oscillation frequency. The dominant oscillation frequency actually identified The error between them. The formula for calculating the mean square error is:

[0119] ;

[0120] In the formula, The dominant oscillation frequency predicted by LSTM. The dominant oscillation frequency is identified using the sliding window fast Fourier transform algorithm. This represents the number of training samples.

[0121] The gradient of the loss function with respect to the network parameters is calculated using the backpropagation algorithm, and the weight parameters of the Long Short-Term Memory (LSTM) network are updated. These weight parameters include the forget gate weights, input gate weights, output gate weights, and connection weights from hidden layers to the output layer. An optimizer such as Adam or SGD is used for parameter updates, with the learning rate preferably set to 0.001 to 0.01. Training samples are fed into the LSTM network to obtain predicted frequencies. The mean squared error (MSE) of the error between the predicted and true dominant frequencies is calculated, and the gradient is calculated to update the network weights and biases. This training process is repeated multiple times until the MSE is sufficiently small to satisfy the early stopping condition, or until a preset number of training epochs is reached. The early stopping condition can be set to ensure that the loss function on the validation set no longer decreases within a certain number of consecutive epochs.

[0122] Furthermore, updating the reference frequency specifically includes:

[0123] Calculate the absolute value of the difference between the predicted oscillation frequency and the reference frequency;

[0124] When the absolute value of the difference is less than a preset deviation threshold, the reference frequency is updated to the predicted oscillation frequency;

[0125] When the absolute value of the difference is greater than or equal to the preset deviation threshold, the reference frequency remains unchanged.

[0126] Specifically, after obtaining the trained LSTM network, the network is used to predict the current system state and obtain the predicted dominant oscillation frequency. Based on the predicted oscillation frequency Compared with the current reference frequency The deviation is used to update the reference frequency. The specific steps are: calculate the predicted oscillation frequency. With reference frequency The absolute value of the difference, i.e. The difference is compared with a preset deviation threshold. A comparison is made. When the absolute value of the difference is less than a preset deviation threshold, i.e. To determine if the predicted frequency is reliable, the reference frequency will be used. Updated to predict oscillation frequency ,Right now When the absolute value of the difference is greater than or equal to the preset deviation threshold, the predicted frequency is determined to be too large or unreliable, and the reference frequency is maintained. Unchanged. Preset deviation threshold. The preferred setting is 2Hz to 5Hz.

[0127] S500: Calculate the compensation voltage command based on the reference frequency and the subsynchronous oscillation component, superimpose the compensation voltage command onto the voltage control loop of the rotor-side converter, and establish a local control model including the weight parameters of the long short-term memory network.

[0128] Furthermore, the S500 specifically includes:

[0129] The gain coefficient is calculated based on the reference frequency, and the gain coefficient changes nonlinearly with the reference frequency.

[0130] The initial compensation voltage is obtained by multiplying the gain coefficient by the subsynchronous oscillation component;

[0131] The amplitude of the initial compensation voltage is limited to obtain the compensation voltage command;

[0132] The compensation voltage command is directly superimposed onto the reference voltage of the dq-axis voltage control circuit of the rotor-side converter.

[0133] Specifically, see Figure 3 , Figure 3 A schematic diagram of rotor-side converter control is shown. Based on the current or updated reference frequency. Compared with the estimated subsynchronous oscillation SSO main frequency component Calculate the compensation voltage command .

[0134] First, based on the reference frequency Calculate the gain coefficient Gain coefficient The gain coefficient is a function of the reference frequency and varies nonlinearly with it. The design of the gain coefficient needs to consider the differences in system impedance characteristics at different frequencies to effectively suppress oscillations at different frequencies. The gain function can be implemented using a lookup table, a piecewise linear function, or a polynomial function. Preferably, the gain coefficient increases with frequency within the subsynchronous oscillation frequency range to compensate for the slower attenuation of high-frequency oscillations. Specifically, the gain coefficient is designed based on the system impedance-frequency characteristic curve, setting a larger gain value near the system resonant frequency and a smaller gain value in regions far from the resonant frequency. The specific numerical range of the gain coefficient is determined based on system parameters and control requirements, preferably set between 0.5 and 5, ensuring that the compensation voltage effectively suppresses oscillations without introducing new instability factors due to excessive gain.

[0135] Gain coefficient With subsynchronous oscillation component Multiplying these yields the initial compensation voltage. Since the subsynchronous oscillation component contains two components, one along the d-axis and the other along the q-axis, respectively... and Therefore, the initial compensation voltage also includes two components: the d-axis and the q-axis, which are respectively... and .

[0136] The initial compensation voltage is amplitude-limited to prevent excessive compensation voltage from affecting normal system operation or exceeding the converter's output capacity. Amplitude limiting can be achieved using a saturation function. When the amplitude of the initial compensation voltage exceeds a preset upper limit, it is limited to the upper limit; when it falls below a preset lower limit, it is limited to the lower limit. The preset upper and lower limits are determined based on the rated voltage and safety margin of the rotor-side converter, preferably set to 10% to 30% of the rated voltage. After amplitude limiting, the final compensation voltage command is obtained. Its expression is:

[0137] ;

[0138] In the formula, It is a gain function. This is the dominant frequency component of the subsynchronous oscillation (SSO). To compensate for the control input superimposed on the inner loop control of the RSC control of the rotor-side converter, This is the reference frequency.

[0139] like Figure 3 As shown, the compensation voltage command will be executed. The voltage control circuit superimposed on the rotor-side converter. Figure 3 The diagram illustrates the workflow of the SSO dominant component identification module and the SSO dominant frequency update module. The input signal is the subsynchronous oscillation SSO dominant frequency component identified through the rotor current. These components, along with the reference frequency continuously updated through the long short-term memory network, generate compensating control quantities according to the control law. The compensation voltage command is superimposed onto the reference voltage of the rotor-side converter voltage control loop in the synchronous rotating coordinate system dq-axis. In the rotor-side converter's inner current loop control, the original d-axis voltage reference is... The q-axis voltage reference is The new reference voltages after superimposing the compensation voltage are respectively and . Figure 3 The text also shows that it includes damping terms. and filtering stage The control structure, in which The synchronization angular frequency, It is the electric angular frequency. Leakage inductance coefficient, The stator inductance is used. The generated dq-axis voltage signal is transformed by Parker transformation and PWM modulation to obtain the control signal of the rotor-side converter, thereby suppressing subsynchronous oscillation.

[0140] A local control model incorporating the weight parameters of the Long Short-Term Memory (LSTM) network is established. This local control model includes all weight parameters of the LSTM network, specifically the weight matrices from the input layer to the hidden layers, the forget gate weight matrix of the hidden layers, the input gate weight matrix, the output gate weight matrix, the weight matrix from the hidden layers to the output layers, and the bias vectors for each layer. These weight parameters are continuously updated during local training and are used to predict the dominant oscillation frequency. The local control model also includes a gain function. The parameters include control parameters such as amplitude limits and upper and lower limits. These parameters are stored uniformly as a local LSTM-SSO control parameter model, serving as the basis for subsequent federated learning aggregation and updates.

[0141] S600: Multiple doubly fed wind turbine units are used as federated learning clients. Each client trains the local control model and uploads the model weight parameters to the central server. The central server performs weighted aggregation on the received model weight parameters to obtain global weight parameters, and then distributes the global weight parameters to each client to update the local control model.

[0142] Furthermore, the updating of the local control model specifically includes:

[0143] Initialize global weight parameters and distribute them to all clients;

[0144] All clients use locally acquired rotor current signals, spectral characteristics, and disturbance parameters to train the local control model and calculate the update amount of the local model weight parameters relative to the initial parameters.

[0145] All clients upload the update data to the central server;

[0146] The central server obtains the amount of training data from each client and calculates the proportion of each client's training data to the total training data as the aggregation weight of the corresponding client.

[0147] The central server calculates a weighted average of the update volume uploaded by all clients according to their respective aggregation weights to obtain the global update volume.

[0148] The central server adds the global update amount to the current global weight parameter to obtain the updated global weight parameter, and then sends it to all clients to update their local control models.

[0149] Specifically, initialization begins. The global weight parameters are initialized and distributed to all clients. The global weight parameters include all weight matrices and bias vectors of the Long Short-Term Memory (LSTM) network; initial values ​​can be obtained through random initialization or pre-trained model parameters. The initialized global weight parameters are denoted as... It is distributed to each client through a central server.

[0150] Define the roles of the client and server. Each doubly-fed induction generator (DFIG) wind turbine is treated as a federated client. Each client retains local data, including actual operating data, spectral characteristics, disturbance parameters, and its local LSTM-SSO control parameter model. The server acts as the central control center, responsible for global model initialization, aggregating client model updates, generating updated local LSTM-SSO control models for each client, and distributing them to the clients. Clients do not communicate directly with each other; they only interact with the central server.

[0151] Set the number of training rounds. The period for local training and server-wide aggregation updates should be set according to the changes in perturbation parameters. The number of training rounds is determined based on the system perturbation frequency and model convergence speed, preferably between 10 and 100 rounds. Each training round includes two phases: local training on the client and global aggregation on the server.

[0152] All clients train their local control models using locally acquired rotor current signals, spectral characteristics, and disturbance parameters. Each client trains its model based on its current local LSTM-SSO control parameter model, employing the training method described in the S400 section above. The loss function is calculated using local data, and model parameters are updated via backpropagation. After training, the update amount of the local model weight parameters relative to the initial parameters is calculated. Specifically, let the client... The initial weight parameters are The weight parameters after training are The update quantity is The client only updates the amount of data. Data privacy is protected by uploading data to a central server instead of the original data or complete model parameters.

[0153] All clients upload their calculated updates to the central server. The server receives updates from all clients participating in the training.

[0154] The central server obtains the amount of training data from each client. Let the client... The amount of training data is This indicates the number of samples used by the client in this training round. The server calculates the proportion of each client's training data to the total training data, which is used as the aggregation weight for that client. The total training data volume is... ,in Total number of clients. The aggregate weight is .

[0155] The central server calculates a weighted average of all client-uploaded updates based on their respective aggregation weights to obtain the global update volume. The formula for calculating the global update volume is:

[0156] ;

[0157] This weighted averaging method, known as the FedAvg algorithm, comprehensively considers the amount of data from each client, allowing clients with larger data volumes to contribute more to the global model.

[0158] The central server adds the global update value to the current global weight parameter to obtain the updated global weight parameter. The calculation formula is:

[0159] ;

[0160] It can also be expressed as:

[0161] ;

[0162] In the formula, The weights of the global LSTM-SSO control parameter model. For the client The parameters of the local LSTM-SSO control parameter model. Total data volume for all clients For the client The amount of training data.

[0163] The central server will update the global weight parameters. The global weight parameters are distributed to all clients to update their local control models. Each client receives the global weight parameters and uses them as the new weight parameters for its local LSTM-SSO control parameter model, replacing the original local weight parameters. The updated local model is then used for the next round of local training or practical control applications.

[0164] The process of local training, uploading and updating data, server aggregation, and distributing global parameters is repeated until the preset number of training rounds is completed or the model performance meets the requirements. Through multiple iterations, the local models of each client gradually become consistent, forming a general control model that can adapt to various disturbance conditions. At the same time, the original data of each client is always kept locally, ensuring data privacy and security.

[0165] Example 2: To verify the effectiveness of the method of the present invention, a simulation model of a doubly fed wind farm series compensation system was built as a test case.

[0166] This invention employs a converter stability enhancement system for a doubly-fed wind power series compensation grid-connected system, such as... Figure 4 As shown, it includes: a signal acquisition module, a component extraction module, a frequency identification module, a frequency prediction module, a compensation control module, and a federated learning module.

[0167] Specifically, the system comprises three doubly-fed wind farms, each acting as a client in the federated learning process. Initial operating conditions are set to a 15% compensation degree for the series-compensated transmission lines and a wind speed of 9 m / s.

[0168] This embodiment sets up two disturbance conditions for verification. The first condition is series complement disturbance: at time 2s, the system increases the series complement from 15% to 20%, that is, increases the series complement disturbance by 5%. The second condition is wind speed disturbance: at time 2s, the system increases the wind speed from 9m / s to 12m / s.

[0169] The simulation results for the first operating condition are shown in Figures 5(a) and 5(b). Figure 5(a) shows the change of the DC-side voltage V(pu) of one of the wind farms over time. As can be seen from the figure, in the initial 2 seconds, when the series complement is 15%, the voltage fluctuates significantly over time before gradually stabilizing. At the 2nd second, when the series complement is increased by 5%, the voltage waveform oscillates and then quickly stabilizes. This indicates that the method of this invention can quickly identify the new dominant oscillation frequency under series complement disturbance conditions and effectively suppress oscillations through compensation control.

[0170] Figure 5(b) shows the variation of the electromagnetic torque Te(pu) of the wind farm over time. As can be seen from the figure, the fluctuation of the electromagnetic torque is more severe in the first 2 seconds, then gradually stabilizes. When a 5% series complement disturbance is added at the 2nd second, the waveform of the electromagnetic torque oscillates for 0.5 seconds before quickly stabilizing. This rapid stabilization of the electromagnetic torque demonstrates that the present invention, through its frequency prediction module, anticipates frequency changes caused by disturbances and updates the compensation control quantity in a timely manner, significantly shortening the oscillation duration of the system.

[0171] The simulation results for the second operating condition are shown in Figures 5(c) and 5(d). Figures 5(c) and 5(d) illustrate the changes in DC-side voltage V(pu) and electromagnetic torque Te(pu) over time under a step wind speed condition. As can be seen from the figures, the voltage waveform changes and electromagnetic torque changes are basically consistent with those under the series complement disturbance condition. The difference is that this method responds faster and improves stability more significantly under wind speed disturbances. This is because the frequency prediction module inputs wind speed as a disturbance parameter into the long short-term memory network, enabling it to learn the correlation between wind speed changes and oscillation frequency, thus achieving more accurate frequency prediction and faster control response.

[0172] Through the federated learning module, the three wind farm clients collaboratively learn under different operating conditions and disturbances, sharing experience without sharing the original operating data. After multiple rounds of federated learning training, the local models of each client have the ability to generalize to various disturbance conditions and maintain good oscillation suppression under complex operating conditions where series complement and wind speed change simultaneously.

[0173] This embodiment verifies the following advantages of the method of the present invention: First, through the dual-layer frequency tracking mechanism of the frequency identification module and the frequency prediction module, it can accurately identify and predict the changes in the dominant oscillation frequency and adapt to disturbance conditions; second, through the adaptive compensation voltage command generated by the compensation control module, it can effectively suppress subsynchronous oscillations, shorten the oscillation duration, and improve system stability; third, through the federated learning module, it can realize the collaborative optimization of models among multiple wind farms, improve the robustness and generalization ability of the control strategy, and at the same time protect the data privacy and security of each wind farm.

[0174] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for enhancing the stability of a converter in a doubly-fed wind power series-compensated grid-connected system, characterized in that, include: S100: Collects the rotor current signal of the rotor-side converter of the doubly fed wind turbine, and initializes the reference frequency to the preset reference frequency; S200: Construct a harmonic basis function based on the reference frequency, fit the rotor current signal using the recursive least squares method, and extract the subsynchronous oscillation component corresponding to the reference frequency; S300: Perform a sliding window spectrum transformation on the rotor current signal, obtain the spectrum characteristics of each window and identify the current dominant oscillation frequency. When the dominant oscillation frequency meets the preset stability condition in multiple consecutive windows, update the reference frequency to the dominant oscillation frequency. The preset stability conditions include: Within a consecutive preset number of windows, the variation amplitude of the dominant oscillation frequency does not exceed a preset frequency threshold, and the amplitude corresponding to the dominant oscillation frequency is higher than a preset amplitude threshold. The duration for which the dominant oscillation frequency remains stable within a predetermined number of consecutive windows is not less than a predetermined time threshold. S400: Input the spectral characteristics of multiple historical windows into the long short-term memory network to obtain the predicted oscillation frequency output by the network, and update the reference frequency according to the deviation between the predicted oscillation frequency and the reference frequency; The steps for obtaining the predicted oscillation frequency specifically include: The spectral features of the multiple historical windows are normalized, and the spectral features include the dominant oscillation frequency and corresponding amplitude identified by the multiple historical windows. Obtain the disturbance parameters at the corresponding time, including wind speed parameters and series complement parameters; The normalized spectral features are combined with the perturbation parameters to form a feature vector, and then arranged in time series to form an input sequence. The input sequence is input into the input layer of the long short-term memory network. The trend of the spectral characteristics changing with the perturbation parameters within the historical time period is preserved through the forget gate, input gate and output gate of the hidden layer. The predicted oscillation frequency is output by the output layer. S500: Calculate the compensation voltage command based on the reference frequency and the subsynchronous oscillation component, superimpose the compensation voltage command onto the voltage control loop of the rotor-side converter, and establish a local control model including the weight parameters of the long short-term memory network. S600: Multiple doubly fed wind turbine units are used as federated learning clients. Each client trains the local control model and uploads the model weight parameters to the central server. The central server performs weighted aggregation on the received model weight parameters to obtain global weight parameters, and then distributes the global weight parameters to each client to update the local control model.

2. The method for enhancing the stability of a converter in a doubly-fed wind power series-compensated grid-connected system according to claim 1, characterized in that, The extraction steps for the subsynchronous oscillation component specifically include: Construct a harmonic basis function composed of cosine and sine components based on the reference frequency; The rotor current signal is iteratively fitted using the recursive least squares method, and the covariance matrix and gain vector are updated by the forgetting factor to obtain the coefficients of the cosine component and the coefficients of the sine component. The subsynchronous oscillation component is reconstructed based on the coefficients of the cosine component and the coefficients of the sine component.

3. The method for enhancing the stability of a converter in a doubly-fed wind power series-compensated grid-connected system according to claim 1, characterized in that, Specifically, S300 includes: Set the window length and window sliding step size, wherein the window length is determined based on the sampling frequency and the target frequency resolution; The rotor current signal is subjected to a fast Fourier transform according to the window sliding step size to obtain the amplitude spectrum of multiple windows. The frequency point with the largest amplitude within the subsynchronous frequency range of the amplitude spectrum is identified as the current dominant oscillation frequency, and the current dominant oscillation frequency and its corresponding amplitude are recorded as the spectral feature.

4. The method for enhancing the stability of a converter in a doubly-fed wind power series-compensated grid-connected system according to claim 1, characterized in that, The training process of the Long Short-Term Memory network includes: Using the spectral characteristics and perturbation parameters of the historical window as input, the dominant oscillation frequency actually identified at the next moment is used as the label to construct training samples; The mean squared error is used as the loss function to calculate the error between the predicted oscillation frequency and the actual identified dominant oscillation frequency; The gradient is calculated using the backpropagation algorithm, and the weight parameters of the Long Short-Term Memory network are updated. The weight parameters include the forget gate weight, input gate weight, output gate weight, and weights from the hidden layer to the output layer.

5. The method for enhancing the stability of a converter in a doubly-fed wind power series-compensated grid-connected system according to claim 1, characterized in that, Updating the reference frequency specifically includes: Calculate the absolute value of the difference between the predicted oscillation frequency and the reference frequency; When the absolute value of the difference is less than a preset deviation threshold, the reference frequency is updated to the predicted oscillation frequency; When the absolute value of the difference is greater than or equal to the preset deviation threshold, the reference frequency remains unchanged.

6. The method for enhancing the stability of a converter in a doubly-fed wind power series-compensated grid-connected system according to claim 1, characterized in that, The S500 specifically includes: The gain coefficient is calculated based on the reference frequency, and the gain coefficient changes nonlinearly with the reference frequency. The initial compensation voltage is obtained by multiplying the gain coefficient by the subsynchronous oscillation component; The amplitude of the initial compensation voltage is limited to obtain the compensation voltage command; The compensation voltage command is directly superimposed onto the reference voltage of the dq-axis voltage control circuit of the rotor-side converter.

7. The method for enhancing the stability of a converter in a doubly-fed wind power series-compensated grid-connected system according to claim 1, characterized in that, The updated local control model specifically includes: Initialize global weight parameters and distribute them to all clients; All clients use locally acquired rotor current signals, spectral characteristics, and disturbance parameters to train the local control model and calculate the update amount of the local model weight parameters relative to the initial parameters. All clients upload the update data to the central server; The central server obtains the amount of training data from each client and calculates the proportion of each client's training data to the total training data as the aggregation weight of the corresponding client. The central server calculates a weighted average of the update volume uploaded by all clients according to their respective aggregation weights to obtain the global update volume. The central server adds the global update amount to the current global weight parameter to obtain the updated global weight parameter, and then sends it to all clients to update their local control models.

8. A system for enhancing the stability of a converter in a doubly-fed wind power series-compensated grid-connected system, characterized in that, The aforementioned doubly-fed wind power series-compensated grid-connected system converter stability enhancement system is used to execute the doubly-fed wind power series-compensated grid-connected system converter stability enhancement method according to any one of claims 1-7, the system comprising: The signal acquisition module is used to acquire the rotor current signal of the rotor-side converter of the doubly fed wind turbine, and initializes the reference frequency to the preset reference frequency. The component extraction module is used to construct a harmonic basis function based on the reference frequency, fit the rotor current signal by recursive least squares method, and extract the subsynchronous oscillation component corresponding to the reference frequency. The frequency identification module is used to perform a sliding window spectrum transformation on the rotor current signal, obtain the spectrum characteristics of each window and identify the current dominant oscillation frequency. When the dominant oscillation frequency meets the preset stability condition in multiple consecutive windows, the reference frequency is updated to the dominant oscillation frequency. The frequency prediction module is used to input the spectral characteristics of multiple historical windows into a long short-term memory network to obtain the predicted oscillation frequency output by the network, and update the reference frequency according to the deviation between the predicted oscillation frequency and the reference frequency. The compensation control module is used to calculate the compensation voltage command based on the reference frequency and the subsynchronous oscillation component, superimpose the compensation voltage command onto the voltage control loop of the rotor-side converter, and establish a local control model including the weight parameters of the long short-term memory network. The federated learning module is used to treat multiple doubly-fed wind turbine units as federated learning clients. Each client trains the local control model and uploads the model weight parameters to the central server. The central server performs weighted aggregation on the received model weight parameters to obtain global weight parameters, and then distributes the global weight parameters to each client to update the local control model.

Images