Double-fed fan subsynchronous oscillation suppression method

Abstract

The invention discloses a double-fed fan subsynchronous oscillation suppression method. The method comprises the following steps: acquiring structure parameters and operation data of a double-fed fan grid-connected system by adopting a Storm cluster data processing architecture and a Spark memory batch processing technology; establishing a system linearization model according to the structure parameters of the doubly-fed fan grid-connected system; according to active transfer learning, carrying out subsynchronous oscillation source positioning on the system linearization model with subsynchronous oscillation; obtaining a subsynchronous oscillation current signal based on improved variational mode decomposition and a second-order blind identification algorithm, and extracting a signal oscillation frequency through a Prony identification model; and suppressing the subsynchronous oscillation current in the doubly-fed fan rotor by using the improved active disturbance rejection control. According to the invention, fault positioning and suppression of double-fed fan subsynchronous oscillation can be realized, and the stability and reliability of a wind power grid-connected system are improved.

Description

Technical Field

[0001] The present invention relates to a method for suppressing oscillations of a doubly-fed wind turbine, and particularly to a method for suppressing subsynchronous oscillations of a doubly-fed wind turbine. Background Art

[0002] With the continuous increase in the penetration rate of wind power, more and more wind power is transmitted to the large power grid and widely used in industrial production and residential life. However, there are significant differences between the distribution of electricity consumption loads and the geographical distribution of wind energy resources in China. In order to transmit wind power to load-intensive areas, long-distance power transmission is usually required. Without taking appropriate measures, long-distance power transmission will bring a large amount of power loss, reduce the power quality, and thus limit the transmission efficiency and capacity. Therefore, series compensation capacitors are usually installed on the transmission line to shorten the electrical distance and improve the transmission capacity. However, this approach may trigger subsynchronous oscillations in the wind power system, have an adverse impact on the normal operation of wind turbines, shorten the service life of wind turbines, and may even cause damage to the unit shafting, resulting in the unit tripping off the grid, thereby reducing the stability of the power grid.

[0003] Currently, the research on subsynchronous oscillation source localization mainly falls into two categories: one is the numerical simulation method based on mechanism analysis, and the other is the data-driven method. The method based on mechanism analysis mainly relies on the energy function method and its derivative methods to locate the low-frequency oscillation source. Although these methods have achieved certain success in low-frequency oscillation source localization, due to the wide frequency range and complex induction mechanism of subsynchronous oscillations, the universality and effectiveness of existing mechanism analysis-based methods in subsynchronous oscillation source localization still need to be further verified. The data-driven method has gradually become the mainstream in recent years with the application of big data and artificial intelligence technologies. This type of method is more in line with the actual engineering needs, especially suitable for parameterless modeling. However, existing data-driven methods still have certain limitations in dealing with large-scale and high-dimensional data, especially in terms of computational efficiency and model generalization ability.

[0004] Based on this, comprehensively adopting efficient data processing, system modeling, subsynchronous oscillation source localization, and improved active disturbance rejection control technology can effectively improve the stability of the wind power system, reduce power loss, ensure the safe and reliable operation of the power grid, and optimize the overall performance and efficiency of the system. Summary of the Invention

[0005] Aiming at the problems existing in the prior art, the present invention proposes a method for suppressing subsynchronous oscillations of a doubly-fed wind turbine, which can accurately locate and effectively suppress subsynchronous oscillations and improve the stability of the wind power system.

[0006] To achieve the above object, the technical solution adopted by the present invention is: a method for suppressing subsynchronous oscillations of a doubly-fed wind turbine, comprising the following steps:

[0007] Step 1: Obtain the structural parameters and operation data of the doubly-fed induction generator (DFIG) grid-connected system by using the Storm cluster data processing architecture and the Spark in-memory batch processing technology;

[0008] Step 2: Establish a system linearization model according to the structural parameters of the DFIG grid-connected system;

[0009] Step 3: Divide the structural parameters and operation data of the DFIG grid-connected system into a source domain and a target domain. After training the system linearization model in the source domain by using the active transfer learning method, migrate it to the target domain, and calculate and output the state variables through the linearization model adjusted in the target domain, so as to determine the part where subsynchronous oscillation occurs in the wind turbine and the power system;

[0010] Step 4: Obtain the current signal of the line on the output side of the DFIG where subsynchronous oscillation occurs; extract the oscillation frequency of the subsynchronous oscillation current signal from the current signal;

[0011] Step 5: According to the structural parameters and operation data of the wind turbine located in Step 3 and the oscillation frequency of the subsynchronous oscillation current signal, use the active disturbance rejection control (ADRC) to suppress the subsynchronous oscillation current in the rotor of the DFIG; the state observer in the ADRC adopts a linear extended state observer; during the control process, the ADRC estimates and compensates all disturbances except the output power of the turbine as external disturbances.

[0012] Further, the step of obtaining the structural parameters and operation data of the DFIG grid-connected system by using the Storm cluster data processing architecture and the Spark in-memory batch processing technology in Step 1 includes the following steps:

[0013] Step 101: Use sensors and monitoring devices to collect the real-time data of the DFIG grid-connected system, including wind speed, wind direction, temperature, humidity, generator speed, voltages and currents on the wind turbine side and the grid side; format the real-time data and convert it into a standard format;

[0014] Step 102: Transmit the real-time data converted into the standard format to the nodes in the Storm cluster for cleaning;

[0015] Step 103: Store the cleaned data as historical data in a distributed file system, a relational database or a NoSQL database;

[0016] Step 104: Load the historical data in the distributed file system, the relational database or the NoSQL database into the Spark cluster for batch processing to obtain the structural parameters and operation data of the DFIG grid-connected system, where the structural parameters include: generator parameters, converter parameters, electrical connections and configurations, and the operation data include: wind speed, speed, output power, voltage, current.

[0017] Furthermore, according to the structural parameters of the doubly-fed wind turbine grid-connected system in step 2, establishing a system linearization model includes the following steps:

[0018] Step 201: Establish the state space equation of the wind turbine subsystem:

[0019]

[0020] In the formula, X p is the column vector composed of all state variables of the wind turbine, X p = [X p1 X p2 …X pm T ; V w is the amplitude of the machine-side input voltage at the grid connection point of the wind turbine, V w = [V wd V wq T ; P p , Q p are the output variables, which are the active power and reactive power injected by the wind turbine into the power system respectively; A p is the state matrix of the wind turbine subsystem; b pd is the input matrix related to V w ; d p , d q are coefficients.

[0021] Step 202: Establish the state space equation of the remaining subsystem:

[0022]

[0023] In the formula, X s is the column vector composed of all state variables of the remaining subsystem; A s is the state matrix of the remaining subsystem; b sd , b sq are the responses of the state variables in the remaining subsystem to the changes in the d-axis and q-axis input variables respectively; d sp , d sq are coefficients.

[0024] Step 203: According to the state space equations of the wind turbine subsystem and the remaining subsystem, obtain the state space equation for describing the linearized system model:

[0025]

[0026] Furthermore, step 3 includes the following steps:

[0027] Step 301: In the source domain, use the features of the system linearization model to train the initial model. The expression of the loss function is as follows:

[0028]

[0029] In the formula, L s (θ) is the cross-entropy loss function, and N s represents the total number of samples in the source domain dataset; i is the index of the sample; X s i represents the i-th eigenvector of the coefficient A in the state space equation of the system linearization model; y s i represents the i-th sample label; f(X s i ; θ) is the model output; θ is the model parameter.

[0030] Step 302: In the target domain, initialize using the parameters of the source domain model and perform fine-tuning.

[0031] Use the source domain model to perform model prediction on the target domain data to generate pseudo-labels which is:

[0032]

[0033] In the formula: X t i represents the feature vector of the i-th sample in the target domain; c represents the category.

[0034] Combine the pseudo-labels for fine-tuning and use the weighted loss function:

[0035]

[0036] In the formula, L t (θ) is the weighted loss function, and N t represents the total number of samples in the target domain dataset; f(X t i ; θ) is the model output.

[0037] When performing fine-tuning, introduce the weight coefficient λ to balance the losses of the source domain and the target domain:

[0038] L(θ) = L s (θ) + λL t (θ)

[0039] In the formula, L(θ) is the loss function, and λ is the weight coefficient.

[0040] Step 303: Select samples with high uncertainty in the target domain through uncertainty measurement for manual annotation to improve the performance of the model.

[0041] Use entropy to measure the uncertainty of samples:

[0042]

[0043] Select the top k samples with the highest entropy values for manual annotation and update the training set;

[0044] Step 304. For each possible oscillation source position j, calculate the weighted average score S j :

[0045]

[0046] In the formula: ω i represents the weight assigned by each model; Y ij represents the score of the i-th model for the j-th position; n represents the number of models;

[0047] Normalize the comprehensive score into a probability distribution P j :

[0048]

[0049] In the formula: m represents the total number of different oscillation positions; k represents the index variable;

[0050] Select the position with the highest probability as the position of the oscillation source: For example:

[0051] Furthermore, in step 4, extracting the oscillation frequency of the subsynchronous oscillation current signal from the current signal includes: decomposing the current signal into intrinsic mode function components of different frequency bands by the variational mode decomposition algorithm; selecting the dominant intrinsic mode function component from the intrinsic mode function components of different frequency bands as the input of the second-order blind identification algorithm to obtain the subsynchronous oscillation current signal; and extracting the oscillation frequency of the subsynchronous oscillation current signal according to the subsynchronous oscillation current signal by the Prony algorithm.

[0052] Furthermore, in step 4, when the current signal is decomposed into intrinsic mode function components of different frequency bands by the variational mode decomposition algorithm, the variational mode decomposition algorithm is improved for the Lagrange operator, and the expression of the obtained iterative operator is as follows:

[0053]

[0054] In the formula: t n is the operator of the n-th iteration, and t n+1 is the operator of the (n + 1)-th iteration;

[0055] The quadratic iterative update of the system eigenvalue λ is realized by introducing the iterative operator, and the expression is:

[0056]

[0057] Wherein, is the system eigenvalue of the (n + 1)-th iteration, is the system eigenvalue of the n-th iteration;

[0058] The second-order blind identification algorithm processes the input signal using a linear mixing model:

[0059] X(t) = A 1 S(t) + N(t)

[0060] Wherein, X(t) is the dominant intrinsic mode function component; S(t) is the desired signal; N(t) is the interference signal; A 1 is an m×n dimensional mixed signal matrix;

[0061] The calculation formula of the Prony algorithm is:

[0062]

[0063] Wherein, y(n) is the signal value at the n-th sampling point in the time domain; n = 0, 1, 2, …, N - 1, N is the number of data points, Δt is the equal time interval of the N data points for sampling; Z is the pole, B is the corresponding residue; A i is the amplitude of the i-th exponential function; θ i is the initial phase of the i-th exponential function; σ i is the decay factor of the i-th exponential function; f i is the frequency of the i-th exponential function; p is the number of exponential functions in the linear combination.

[0064] Furthermore, the improved active disturbance rejection controller in step 5 includes an improved linear extended state observer and a linear state error feedback law, and the improved active disturbance rejection controller estimates and compensates all disturbances other than the turbine output power as external disturbances during the control process. The improved active disturbance rejection controller is connected to the rotor current loop of the doubly-fed wind turbine to suppress the subsynchronous oscillation current in the rotor.

[0065] Furthermore, the design of the improved linear extended state observer in step 5 is as follows:

[0066]

[0067] Wherein: ω m is the turbine speed; n p is the number of pole pairs of the motor; ψ f is the magnetic flux per pole; J is the moment of inertia; i q is the q-axis component current in the stator winding of the motor; B is the damping coefficient; Tm is mechanical torque.

[0068] Let x 1 = ω m , x 2 = (b - b 0 )i q * - (Bω m - T m ) / J, b 0 = -1.5(n p ψ f ). Among them, x 2 is the extended state variable, representing the total disturbance of the system; b and b 0 are control gains. Denote Then the above formula can be transformed into:

[0069]

[0070] Based on the deviation control principle, the mathematical model of the traditional linear extended state observer is rewritten. The derivative of the observed value is adjusted by using the error between each state variable and its observed value, and a linear extended state observer is introduced. The linear extended state observer constructed for the above formula is:

[0071]

[0072] In the formula: z 1 , z 2 are the observed values of the state variables x 1 , x 2 ; x 1 is the true state value of the system; e 1 is the observation error; a 1 , a 2 are the observer gains related to x 1 , x 2 ; b 0 is the control gain; i q is the q-axis component current in the stator winding of the motor; represents the change rate of the observed values z 1 , z 2 ; represents the change rate of the error.

[0073] The role of the linear state error feedback law is to form a control quantity through a suitable linear combination of the signals generated in the nonlinear tracking differentiator and the linear extended state observer, and then subtract the total disturbance of the system except the turbine output power from the total control quantity to obtain a pure control quantity without disturbance, making the control signal of the system more reasonable. The linear state error feedback law is designed as:

[0074]

[0075] In the formula: u 0 is the given rotational speed value ω m * and the observed output z 1 After taking the difference and inputting it into the linear state error feedback law, the obtained output quantity; k p is the error feedback rate gain; ω m * is the system desired signal; z 1 、z 2 are the observed output values; i q * is the compensated control quantity; b 0 is the control gain.

[0076] Beneficial effects: Compared with the prior art, the present invention has the following advantages: Based on the structural parameters and operation data of the doubly-fed wind turbine grid-connected system, the established system linearization model can accurately describe the dynamic behavior of the doubly-fed wind turbine grid-connected system; combined with transfer learning, a powerful method for locating the subsynchronous oscillation source is provided. The improved variational mode decomposition and second-order blind identification algorithms not only improve the accuracy of signal processing but also enhance the anti-noise ability of the algorithms. The signal oscillation frequency extracted by the Prony model provides key information for deeply understanding the system dynamic characteristics; the active disturbance rejection control technology has unique advantages in power tracking and frequency regulation of strong nonlinear complex power systems. The improved linear active disturbance rejection control technology is used to adjust the oscillation frequency, estimate and compensate the total system disturbance, and effectively suppress the subsynchronous oscillation current in the rotor of the doubly-fed wind turbine. In addition, the present invention adopts the Storm cluster data processing architecture and Spark in-memory batch processing technology, which can efficiently collect the structural parameters and operation data of the doubly-fed wind turbine grid-connected system. The Storm cluster is suitable for real-time stream data processing with its high throughput and low latency characteristics, while the in-memory computing ability of Spark significantly improves the data processing speed. Brief Description of the Drawings

[0077] Figure 1 is a flow chart of a method for suppressing subsynchronous oscillation of a doubly-fed wind turbine according to the present invention;

[0078] Figure 2 is a flow chart for identifying subsynchronous oscillation current signals in the present invention. Detailed Embodiment

[0079] The present invention will be further described below with reference to the drawings. The following specific embodiments are only used to more clearly illustrate the technical solution of the present invention and cannot be used to limit the protection scope of the present invention.

[0080] As Figure 1This is a flowchart of a method for suppressing subsynchronous oscillation of a doubly-fed wind turbine according to the present invention, including the following steps:

[0081] Step 1: The present invention uses a Storm cluster data processing architecture and Spark in-memory batch processing technology to obtain the structural parameters and operating data of the doubly-fed wind turbine grid-connected system, specifically including the following sub-steps:

[0082] Step 101: Use sensors and monitoring devices to collect real-time data of the doubly-fed wind turbine grid-connected system, including key parameters such as wind speed, wind direction, temperature, humidity, generator speed, grid voltage and current. Format the collected data and convert it into a standard format (such as JSON or CSV) for subsequent processing;

[0083] Step 102: Transmit the real-time data converted into the standard format to the nodes in the Storm cluster for cleaning processing;

[0084] Step 103: Store the cleaned data in a distributed file system (such as HDFS), a relational database (such as MySQL), or a NoSQL database (such as HBase, Cassandra);

[0085] Step 104: Load the historical data in the distributed file system, relational database, or NoSQL database into the Spark cluster for batch processing to obtain the structural parameters and operating data of the doubly-fed wind turbine grid-connected system, where the structural parameters include: generator parameters, converter parameters, electrical connections and configurations, and the operating data includes: wind speed, speed, output power, voltage, current.

[0086] The advantages of real-time monitoring of wind turbine data based on big data processing technology are as follows:

[0087] 1) By using a Storm cluster data processing architecture, it is possible to make full use of a large-scale cluster to quickly process a large amount of data and timely detect wind turbine faults, thereby improving the accuracy of diagnosis and early warning. In addition, the Storm cluster data processing architecture also has the following advantages: ① By increasing the number of nodes and improving load balancing, it is easy to horizontally expand the Storm cluster data processing architecture; ② The cluster data processing architecture ensures that data will not be lost during the processing by providing reliability guarantees; ③ It is possible to implement more complex data processing logics, such as calculations based on multiple metrics, data aggregation and filtering, etc.

[0088] 2) Spark is a distributed in-memory computing system that can perform batch processing on large-scale data. The advantage of using Spark's in-memory batch processing technology is that data processing can be carried out in memory, improving the speed and efficiency of data processing. In addition, Spark also supports multiple languages, facilitating secondary development and customization by developers. By adopting Spark's in-memory batch processing technology, the data processing speed in the real-time monitoring model of wind turbines can be accelerated, the efficiency of fault diagnosis and early warning can be improved, and the operation safety and stability of wind turbines can be better guaranteed. By leveraging the powerful functions of Spark's in-memory batch processing technology, wind turbine operators can accelerate the data processing speed in the real-time monitoring model, improve the efficiency of fault diagnosis and warning, and better ensure the safety and stability of wind turbines.

[0089] Step 2: In the present invention, establishing a system linearization model from the structural parameters of the doubly-fed wind turbine grid-connected system specifically includes the following sub-steps:

[0090] Step 201: Establish the state-space equation of the wind turbine subsystem:

[0091]

[0092] In the formula, X p is a column vector composed of all state variables of the wind turbine, X p = [X p1 X p2 …X pm T ; V w is the amplitude of the terminal input voltage at the grid connection point of the wind turbine, V w = [V wd V wq T ; P p and Q p are output variables, which are the active power and reactive power injected by the wind turbine into the power system respectively; A p is the state matrix of the wind turbine subsystem; b pd is the input matrix related to V w ; d p and d q are coefficients.

[0093] Step 202: Establish the state-space equation of the remaining subsystem:

[0094]

[0095] In the formula, X s is a column vector composed of all state variables of the remaining subsystem; A s is the state matrix of the remaining subsystem; bsd , b sq are the responses of the state variables in the remaining subsystems to the changes in the d-axis and q-axis input variables, respectively; d sp , d sq are coefficients.

[0096] Step 203: According to the state space equations of the fan subsystem and the remaining subsystems, obtain the state space equation for describing the linearized system model:

[0097]

[0098] Step 3: The present invention divides the structural parameters and operation data of the doubly-fed wind turbine grid-connected system into a source domain and a target domain. After training the system linearization model in the source domain through an active transfer learning method, it is transferred to the target domain. The state quantity is calculated and output through the linearization model adjusted in the target domain, so as to determine the part where subsynchronous oscillation occurs in the wind turbine unit and the power system. Specifically, it includes the following sub-steps:

[0099] Step 301: In the source domain, use the features of the system linearization model to train the initial model; the expression of the loss function is as follows:

[0100]

[0101] In the formula, L s (θ) is the cross-entropy loss function, N s represents the total number of samples in the source domain dataset; i is the index of the sample; X s i represents the i-th eigenvector of the coefficient A in the state space equation of the system linearization model; y s i represents the i-th sample label; f(X s i ; θ) is the model output; θ is the model parameter;

[0102] Step 302: In the target domain, use the parameters of the source domain model for initialization and fine-tuning;

[0103] Use the source domain model to perform model prediction on the target domain data and generate pseudo-labels which is:

[0104]

[0105] In the formula: X t i represents the feature vector of the i-th sample in the target domain; c represents the category;

[0106] Combine the pseudo-labels for fine-tuning and use the weighted loss function:

[0107]

[0108] Wherein, L t (θ) is the weighted loss function, and N t represents the total number of samples in the target domain dataset; f(X t i ; θ) is the model output;

[0109] When fine-tuning, a weight coefficient λ is introduced to balance the losses of the source domain and the target domain:

[0110] L(θ) = L s (θ) + λL t (θ)

[0111] Wherein, L(θ) is the loss function, and λ is the weight coefficient;

[0112] Step 303: Select samples with high uncertainty in the target domain through uncertainty measurement for manual annotation to improve the performance of the model;

[0113] Use entropy to measure the uncertainty of samples:

[0114]

[0115] Select the top k samples with the highest entropy values for manual annotation and update the training set;

[0116] Step 304: For each possible oscillation source position j, calculate the weighted average score S j :

[0117]

[0118] Wherein: ω i represents the weight assigned by each model; Y ij represents the score of the i-th model for the j-th position; n represents the number of models;

[0119] Normalize the comprehensive score into a probability distribution P j :

[0120]

[0121] Wherein: m represents the total number of different oscillation positions; k represents the index variable;

[0122] Select the position with the highest probability as the position of the oscillation source:

[0123]

[0124] Wherein, j * represents the position of the oscillation source.

[0125] Step 4: Based on the improved variational mode decomposition and second-order blind identification algorithms, the present invention obtains the subsynchronous oscillation current signal and extracts the signal oscillation frequency through the Prony identification model, as Figure 2 shown, which specifically includes the following sub-steps:

[0126] Step 401: Through simulation analysis and research, a group of subsynchronous oscillation current signals are extracted, and through the improvement of the variational mode decomposition algorithm, it is successfully decomposed into several intrinsic mode function components, thus obtaining more accurate results;

[0127] Step 402: By calculating the correlation between the intrinsic mode function components and the observed subsynchronous oscillation current signal, and through data preprocessing, the optimal multi-channel signal structure is determined;

[0128] Step 403: By integrating multiple channels into the second-order blind identification algorithm, not only can noise interference be effectively suppressed, but also the subsynchronous oscillation current signal can be completely captured;

[0129] Step 404: Through the Prony algorithm, the amplitude, initial phase, attenuation factor, and frequency of the denoised subsynchronous oscillation current signal can be accurately identified.

[0130] In the present invention, the variational mode decomposition algorithm aims to construct a variational model to determine the optimal bandwidth in the subsynchronous state and regularly adjust the core frequency of the circuit, thereby obtaining the optimal mode. To further improve the performance of the algorithm, a fast iterative idea is adopted for the Lagrange operator to shorten the operation time and more accurately obtain the iterative result. The expression of the iterative operator is as follows:

[0131]

[0132] where: t n is the operator of the nth iteration, and t 0 = 1 is the initial value.

[0133] By introducing the iterative operator t, without affecting the original calculation results, the system eigenvalue λ can be updated iteratively twice, thus converting the expression to:

[0134]

[0135] The second-order blind identification algorithm can effectively reduce the error caused by the delay of the subsynchronous oscillation current signal by adopting the joint approximate diagonalization technique. To achieve the blind source separation of the signal, a linear mixing model can be used to achieve this goal, as shown in the following formula:

[0136] X(t) = AS(t) + N(t)

[0137] Where: X(t) is the observed signal of subsynchronous oscillation current; S(t) is the desired signal; N(t) is the interference signal; A is an m×n dimensional mixed signal matrix.

[0138] Find the unbiased covariance matrix R X (0) and the covariance matrix R X (τ) with time delay τ, which are expressed as follows:

[0139] R X (0) = E[X(t)X(t) T = AR S (0)A T +R v (0)

[0140] R X (τ) = E[X(t)X(t + τ) T = AR S (τ)A T

[0141] Where: R S (0) is the signal covariance matrix; R v (0) is the noise covariance matrix; R S (τ) is the signal covariance matrix with time delay τ.

[0142] In the second-order blind identification algorithm, the correlation function of the source information is zero, and the autocorrelation function is also zero. Therefore, R X (0) and R X (τ) are diagonal matrices with non-zero diagonal elements.

[0143] Perform whitening processing on X(t), and we can get:

[0144] Z(t) = QX(t)

[0145] Where: Q is an m×n dimensional whitening matrix; Z(t) = [z 1 (t), z 2 (t), …, z n (t)] T .

[0146] After obtaining Z(t), solve its covariance matrix to get the formula:

[0147] R Z (τ) = QR X (τ)Q T = QAR S (τ)A T Q T

[0148] Where: RZ R(τ) is the time-delay covariance matrix of the whitened signal.

[0149] To eliminate the influence of time delay on positive definiteness, the joint approximate diagonalization of R Z (τ) is used to obtain the following formula:

[0150] R Z (τ i ) = UD i U T

[0151] In the formula: R Z (τ i ) is the covariance matrix with a time delay of τ i (i = 1, 2,..., k); D i is a diagonal matrix; U is an orthogonal matrix.

[0152] To estimate the problem of U, the objective function is constructed as follows:

[0153]

[0154] In the formula: off{Y T R Z (τ i )Y} represents the non-diagonal elements, that is, to find the minimum value R T R Z (τ i ) of Y Z (τ i ). Through the analysis of the orthogonal matrix, a series of estimated source signals can be deduced, including:

[0155] S(t) = U T QX

[0156] A = QU

[0157] In the formula: S(t) is the desired signal; A is an m×n-dimensional mixed signal matrix; U is an orthogonal matrix; Q is an m×n-dimensional whitening matrix; X is the observed signal.

[0158] The Prony algorithm is an algorithm used to evaluate different numbers of digital signals. By using the linear combination of exponential functions, an algorithm that can estimate the frequency, attenuation factor, amplitude, and initial phase of a given signal is formed. Assuming N data points sampled at equal time intervals Δt, it can be simulated by the linear combination of p exponential functions, that is

[0159]

[0160] In the formula: n = 0, 1, 2,..., N - 1; y(n) is the signal value of the nth sampling point; Z = [z 1 , z 2,…,z p is the pole; B = [b 1 , b 2 ,…, b p is the corresponding residue; A i is the amplitude; θ i is the initial phase; σ i is the attenuation factor; f i is the frequency.

[0161] The operating principle of the Prony identification algorithm is as follows:

[0162] By analyzing the sampled data points, establish the matrix Y to solve:

[0163]

[0164] Let YA T = 0, and solve the polynomial composed of the coefficients a 1 , a 2 ,…, a p :

[0165] Z p + a 1 Z p-1 + … + a p-1 Z + a p = 0

[0166] By solving the system of equations using the least squares method, the remaining value B can be calculated:

[0167]

[0168] From this, the values of the amplitude, initial phase, attenuation factor, and frequency of the simulated input signal are derived:

[0169]

[0170] Step 5: The improved active disturbance rejection controller in the present invention includes an improved linear extended state observer and a linear state error feedback law, and the improved active disturbance rejection controller estimates and compensates all disturbances other than the turbine output power as external disturbances during the control process. The improved linear active disturbance rejection control overcomes the contradiction between the response speed and overshoot under proportional-integral control while effectively improving the active power output during the system frequency response process. The improved active disturbance rejection controller is connected to the rotor current loop of the doubly-fed wind turbine to suppress the subsynchronous oscillation current in the rotor.

[0171] The design of the improved linear extended state observer in the present invention is as follows:

[0172]

[0173] Where: ωm is the turbine speed; n p is the number of pole pairs of the motor; ψ f is the magnetic flux per pole; J is the moment of inertia; i q is the q-axis component current in the stator winding of the motor; B is the damping coefficient; T m is the mechanical torque.

[0174] Let x 1 = ω m , x 2 = (b - b 0 )i q * - (Bω m - T m ) / J, b 0 = -1.5(n p ψ f ) / J. Wherein, x 2 is the extended state variable, representing the total system disturbance; b, b 0 are the control gains. Denote Then the above equation can be transformed into:

[0175]

[0176] Based on the deviation control principle, the mathematical model of the traditional linear extended state observer is rewritten. The derivative of the observed value is adjusted by using the error between each state variable and its observed value, and a linear extended state observer is introduced. The linear extended state observer constructed for the above equation is:

[0177]

[0178] In the formula: z 1 , z 2 are the observed values of the state variables x 1 , x 2 ; x 1 is the true state value of the system; e 1 is the observation error; a 1 , a 2 are the observer gains related to x 1 , x 2 ; b 0 is the control gain; i q is the q-axis component current in the stator winding of the motor; represents the change rate of the observed values z 1 , z 2 ; represents the change rate of the error.

[0179] The role of the linear state error feedback law is to form a control quantity through a suitable linear combination of the signals generated in the nonlinear tracking differentiator and the linear extended state observer, and then subtract the total disturbance of the system except the turbine output power from the total control quantity to obtain a pure control quantity without disturbance, making the control signal of the system more reasonable. The linear state error feedback law is designed as follows:

[0180]

[0181] where: u 0 is the output quantity obtained by subtracting the observed output z m * from the given rotational speed value ω 1 and inputting the result into the linear state error feedback law; k p is the error feedback rate gain; ω m * is the desired signal of the system; z 1 , z 2 are the observed output values; i q * is the compensated control quantity; b 0 is the control gain.

Claims

1. A method for suppressing subsynchronous oscillation of a doubly-fed wind turbine, characterized in that: The following steps are involved: Step 1, obtaining structural parameters and operating data of the double-fed wind turbine grid-connected system; Step 2: Establish a system linearization model based on the structural parameters of the double-fed wind turbine grid-connected system; Step 3: The structural parameters and operation data of the doubly-fed wind turbine grid-connected system are divided into a source domain and a target domain. The system linearization model is trained in the source domain through an active transfer learning method and then transferred to the target domain. The state quantity is calculated and output through the system linearization model adjusted in the target domain, so as to determine the part of the wind turbine and the power system where subsynchronous oscillation occurs, and locate the wind turbine involved in the subsynchronous oscillation. Step 4, obtaining a current signal of an output side line of a doubly-fed wind turbine where subsynchronous oscillation occurs; extracting the oscillation frequency of the subsynchronous oscillating current signal from the current signal; Step 5: According to the structural parameters and operating data of the wind turbine located in step 3, and the oscillation frequency of the subsynchronous oscillating current signal, the subsynchronous oscillating current in the doubly fed wind turbine rotor is suppressed by using active disturbance rejection control; the state observer in the active disturbance rejection control adopts a linear extended state observer; during the control process, the active disturbance rejection control estimates and compensates for all disturbances except the turbine output power as external disturbances.

2. The method for suppressing subsynchronous oscillation of a doubly-fed wind turbine according to claim 1, characterized in that: In step 1, the Storm cluster data processing architecture and Spark memory batch processing technology are used to obtain the structural parameters and operating data of the doubly fed wind turbine grid-connected system.

3. The method for suppressing subsynchronous oscillation of a doubly-fed wind turbine according to claim 2, characterized in that: The method of using Storm cluster data processing architecture and Spark memory batch processing technology to obtain structural parameters and operating data of a double-fed wind turbine grid-connected system includes the following steps: Step 101: Use sensors and monitoring equipment to collect real-time data of the doubly-fed wind turbine grid-connected system, including wind speed, wind direction, temperature, humidity, generator speed, and voltage and current on the wind turbine side and grid side; format the real-time data and convert it into a standard format; Step 102: Transmit the real-time data converted into a standard format to the nodes in the Storm cluster for cleaning; Step 103: storing the cleaned data as historical data in a distributed file system, a relational database, or a NoSQL database; Step 104, load the historical data in the distributed file system, relational database or NoSQL database into the Spark cluster for batch processing to obtain the structural parameters and operating data of the doubly fed wind turbine grid-connected system, wherein the structural parameters include: generator parameters, converter parameters, electrical connection and configuration, and the operating data include: wind speed, rotation speed, output power, voltage, and current.

4. The method for suppressing subsynchronous oscillation of a doubly-fed wind turbine according to claim 1, characterized in that: The state space equation of the linearized system model described in step 2 is: In the formula, t is time, X p is the column vector composed of all state variables of the wind turbine; s is the column vector of all state variables of the remaining subsystems; A is the coefficient; V w is the input voltage amplitude of the wind turbine grid connection point, V w =[V wd V wq ] T , V wd and V wq Respectively represent the amplitudes of the two orthogonal components of the voltage in the dq axis coordinate system; A p A is the state matrix of the fan subsystem; s is the state matrix of the remaining subsystem; b pd For V w The relevant input matrix is ​​used to describe the response of the wind turbine state variables to the change of the input voltage amplitude at the machine end; b sd 、b sq are the responses of the state variables in the remaining subsystems to the changes of the input variables on the d and q axes, respectively; d p , d q are the coefficients of the state space equation of the fan subsystem; d sp d sq are the coefficients of the state-space equations of the remaining subsystems.

5. The method for suppressing subsynchronous oscillation of a doubly-fed wind turbine according to claim 1, characterized in that: Step 3 includes the following steps: Step 301: In the source domain, the features of the system linearization model are used to train the initial model; the loss function expression is as follows: Where, L s (θ) is the cross entropy loss function, N s represents the total number of samples in the source domain dataset; i is the index of the sample; X s i represents the i-th eigenvector of coefficient A in the state space equation of the system linearization model; y s i represents the i-th sample label; f(X s i ; θ) is the source domain model output; θ is the model parameter; Step 302: In the target domain, the parameters of the source domain model are used for initialization and fine-tuning; Use the source domain model to predict the target domain data and generate pseudo labels for: Where: represents the pseudo label, X t i represents the feature vector of the i-th sample in the target domain; c represents the category; f(X t i ; θ) is the model output; Combine pseudo labels for fine-tuning and use a weighted loss function: Where, L t (θ) is the weighted loss function, N t Represents the total number of samples in the target domain dataset; During fine-tuning, a weight coefficient λ is introduced to balance the loss of the source domain and the target domain: L(θ)=L s (θ)+λL t (i) In the formula, L(θ) is the loss function, λ is the weight coefficient; Step 303: Use entropy to measure the uncertainty of the sample: In the formula, Indicates the uncertainty of the sample, selects the first k samples with the highest entropy value for manual labeling, and updates the training set; Step 304: For each possible oscillation source position j, calculate the weighted average score S j : Where: S j represents the weighted average score, ω i represents the weight assigned to each model; Y ij represents the score of the i-th model for the j-th position; n represents the number of models; Normalize the comprehensive score to the probability distribution P j : Where: P j represents the probability distribution, m represents the total number of different oscillation positions; k represents the index variable; The position with the highest probability is selected as the location of the oscillation source.

6. The method for suppressing subsynchronous oscillation of a doubly-fed wind turbine according to claim 1, characterized in that: The oscillation frequency of a subsynchronous oscillating current signal is extracted from the current signal, comprising: decomposing the current signal into intrinsic modal function components of different frequency bands by a variational modal decomposition algorithm; selecting a dominant intrinsic modal function component from the intrinsic modal function components of different frequency bands as an input of a second-order blind identification algorithm to obtain a subsynchronous oscillating current signal; and extracting the oscillation frequency of the subsynchronous oscillating current signal by a Prony algorithm according to the subsynchronous oscillating current signal.

7. The method for suppressing subsynchronous oscillation of a doubly-fed wind turbine according to claim 6, characterized in that: The current signal is decomposed into intrinsic mode function components of different frequency bands by a variational mode decomposition algorithm. The variational mode decomposition algorithm is improved for the Lagrange operator. The expression of the improved iterative operator is as follows: Where: t n is the operator of the nth iteration, t n+1 is the operator of the n+1th iteration; By introducing the iterative operator, the system eigenvalue λ is updated twice iteratively, and the expression is: In the formula, is the system eigenvalue of the n+1th iteration, is the system eigenvalue of the nth iteration; The second-order blind identification algorithm uses a linear mixture model to process the input signal: X(t)=A1S(t)+N(t) Where X(t) is the dominant intrinsic mode function component; S(t) is the desired signal; N(t) is the interference signal; A1 is the m×n dimensional mixed signal matrix; The calculation formula of Prony algorithm is: Where y(n) is the signal value of the nth sampling point in the time domain; n = 0, 1, 2, ..., N-1, N is the number of data points, Δt is the equal time interval of the N data points for sampling; Z is the pole, B is the corresponding residue; A i is the amplitude of the i-th exponential function; θ i is the initial phase of the i-th exponential function; σ i is the attenuation factor of the i-th exponential function; f i is the frequency of the i-th exponential function; p is the number of exponential functions in the linear combination.

8. The method for suppressing subsynchronous oscillation of a doubly-fed wind turbine according to claim 1, characterized in that: The active disturbance rejection control adopts improved active disturbance rejection control, which is based on the active disturbance rejection control and is obtained by improving the state observer.

9. The method for suppressing subsynchronous oscillation of a doubly-fed wind turbine according to claim 8, characterized in that: The improved state observer adopts a linear extended state observer, which is: Where: z1 is the observed value of the state variable x1; z2 is the observed value of the expanded state variable x2; x1 is the true state value of the system; e1 is the observation error between z1 and x1; a1 is the observer gain associated with the state variable x1; a2 is the observer gain associated with the state variable x2; b0 is the control gain; i q is the q-axis component current in the motor stator winding; Indicates the rate of change of observations z1 and z2; Indicates the rate of change of error.

10. The method for suppressing subsynchronous oscillation of a doubly-fed wind turbine according to claim 8, characterized in that: The improved ADRC also includes an improved linear state error feedback law, which is calculated as follows: Where: u0 is the given speed value ω m * The output obtained by subtracting the observed output z1 through the linear state error feedback law; k p is the error feedback rate gain; ω m * is the expected signal of the system; z1 and z2 are the observed output values; i q * is the control quantity after compensation; b0 is the control gain.

Images