A decoupling controller for PWM rectification system of bearingless permanent magnet synchronous generator

Abstract

The invention discloses a decoupling controller for a PWM rectifier system of a bearingless permanent magnet synchronous generator, comprising three internal model controllers, a particle swarm algorithm optimization module and a least squares support vector machine. The optimal kernel function parameters and penalty factors output by the particle swarm algorithm optimization module are used as the first and second inputs of the least squares support vector machine, and the voltage U output by the rectangular window algorithm module is used as the first input and the second input of the least squares support vector machine. DC , U DC The eight training samples, namely the first-order derivative, horizontal radial displacement x, the first-order derivative of x, the second-order derivative of x, the vertical radial displacement y, the first-order derivative of y and the second-order derivative of y, are used as the third to tenth inputs of the least squares support vector machine. The kernel function, symmetric positive definite matrix, parameter matrix, bias and Lagrange multiplier matrix are obtained based on the optimal kernel function parameters. The particle swarm algorithm is used to select the kernel function parameters and penalty factor, and the adaptive inertia weight and dynamic learning factor are introduced to improve the particle swarm algorithm.

Description

Technical Field

[0001] The present invention relates to a generator PWM rectifier system composed of a bearingless permanent magnet synchronous generator and a PWM rectifier, and specifically to a decoupling controller for controlling the generator PWM rectifier system. The present invention is applied to power generation occasions where bearings are difficult to replace, such as wind turbine generators, gas turbine generators, aviation power supplies, hybrid vehicles, flywheel energy storage systems, etc. Background Art

[0002] Compared with the traditional permanent magnet synchronous generator, the bearingless permanent magnet synchronous generator has many advantages: it combines bearingless technology with permanent magnet synchronous generator, embeds suspension force winding in the stator, and uses magnetic field imbalance to suspend the rotor, so that the motor has the advantages of no friction and wear and high-speed and high-precision operation. The generator PWM rectifier system composed of a bearingless permanent magnet synchronous generator and a PWM rectifier in series is a complex nonlinear system with multiple variables and strong coupling. The coupling relationship between radial suspension force and between suspension force and output voltage has a great influence on the control performance of the system. In order to obtain excellent dynamic control performance, it is necessary to dynamically decouple the radial two-degree-of-freedom suspension force and output voltage. At present, the commonly used control methods include vector control and direct torque control. Vector control can only perform static decoupling control on the output voltage and radial suspension force, and requires complex coordinate transformation units and accurate rotor position angle information. Its controller parameters are difficult to debug, the dynamic performance is poor, and it is sensitive to the parameter changes of the generator. Although direct torque control has the advantages of simple implementation, no need for coordinate transformation, strong system robustness, and fast dynamic response, it has the disadvantages of unstable switching frequency and high sampling frequency. The generalized inverse of the least squares support vector machine can be used to achieve dynamic decoupling of the system, but its Gaussian kernel function parameter σ 2 The penalty factor γ is difficult to adjust, and the traditional least squares support vector machine is an offline algorithm, and the nonlinear system model cannot be corrected as subsequent samples are continuously updated.

[0003] The control method disclosed in the document with Chinese patent application number 201410217088.2 and titled “A Direct Torque Control Method for Permanent Magnet Synchronous Generator” solves the inverse system expression of the original system of the permanent magnet synchronous generator, constructs a pseudo-linear subsystem, and realizes the linear decoupling control of the electromagnetic torque and the stator flux amplitude in the direct torque control of the permanent magnet synchronous generator. The control object is the permanent magnet synchronous generator; about the Gaussian kernel function parameter σ of the support vector machine 2 The penalty factor γ is adjusted by trial and error, which makes the search for parameters blind and the controller lacks wide applicability. Summary of the invention

[0004] The purpose of the present invention is to provide a decoupling controller for a PWM rectifier system of a bearingless permanent magnet synchronous generator, which can achieve decoupling between radial suspension forces and between radial suspension forces and output voltage of the bearingless permanent magnet synchronous generator, while enabling the system to obtain good dynamic and static performance.

[0005] The decoupling controller of the bearingless permanent magnet synchronous generator PWM rectifier system of the present invention adopts a technical solution: it includes three internal model controllers, a particle swarm algorithm optimization module and a least squares support vector machine, the output end of the least squares support vector machine is connected to the composite controlled object having a PWM rectifier and a bearingless permanent magnet synchronous generator connected in series, and a given voltage is and the PWM rectifier output voltage U DC The voltage difference obtained by the difference is used as the input of the first internal model controller, and the horizontal radial displacement x is given. * The horizontal radial displacement difference obtained by subtracting the horizontal radial displacement x output by the bearingless permanent magnet synchronous generator is used as the input of the second internal model controller, and the vertical radial displacement y is set. * The vertical radial displacement difference obtained by subtracting the vertical radial displacement y output by the bearingless permanent magnet synchronous generator is used as the input of the third internal model controller. The first internal model controller is successively transferred through the first-order transfer function The rectangular window algorithm module is connected to the least squares support vector machine, and the second internal model controller is sequentially connected to the second-order standard transfer function The first integrator s -1 The least squares support vector machine is connected to the rectangular window algorithm module, and the third internal model controller is sequentially connected through the second-order transfer function The second integrator s -1 The least squares support vector machine is connected to the rectangular window algorithm module, and the particle swarm algorithm optimization module outputs the optimal kernel function parameter σ 2 and penalty factor γ, optimal kernel function parameter σ 2 And the penalty factor γ is used as the first and second inputs of the least squares support vector machine. The output of the rectangular window algorithm module is the voltage U DC , U DC The eight training samples are the first-order derivative, horizontal radial displacement x, first-order derivative of x, second-order derivative of x, vertical radial displacement y, first-order derivative of y, and second-order derivative of y. These eight training samples are used as the third to tenth inputs of the least squares support vector machine; the least squares support vector machine is based on the optimal kernel function parameter σ 2 Get the kernel function at time k X k and X' k Generally refers to two different training samples input into the support vector machine at time k, according to the kernel function K at time k kand the penalty factor γ to obtain the symmetric positive definite matrix at time k Parameter Matrix Bias θ k and the Lagrange multiplier α k The matrix And the output of the least squares support vector machine ψ=∑α k K k +b k , ψ is the dq axis given generating control current of the input composite controlled object and dq axis given suspension force control current M k-1 is the parameter matrix at time k-1, I is the identity matrix, β is the forgetting factor, and 0<β<1, b k-1 , b k are the biases at time k-1 and k respectively, α k-1 , α k are the Lagrange multipliers at time k-1 and k respectively, z k Output matrix for k time instant.

[0006] The advantages of the present invention are:

[0007] 1. Through the online least squares support vector machine generalized inverse, the multivariable, strongly coupled nonlinear time-varying system of the bearingless permanent magnet synchronous generator PWM rectifier system is transformed into the control of two displacement second-order integral pseudo-linear subsystems and one voltage first-order integral pseudo-linear subsystem. An additional closed-loop controller is used to achieve dynamic decoupling between the radial two-degree-of-freedom suspension force and between the radial suspension force and the output voltage, thereby achieving high-performance independent control of the displacement system and output voltage of the bearingless permanent magnet synchronous generator PWM rectifier system.

[0008] 2. The rectangular window algorithm is combined with the least squares support vector machine to form an online least squares support vector machine based on the rectangular window algorithm. The estimated value of the rectangular window algorithm only depends on a limited number of past data. The old data before this is completely deleted, so that the samples of the least squares support vector machine change over time, and the forgetting factor β is introduced. The forgetting factor is added to the historical data, which takes into account both the influence of historical data and the role of new data. Therefore, the inverse model can be continuously updated in real time to achieve correction and shorten the training time.

[0009] 3. Particle swarm optimization is a global optimization algorithm that is used to select the kernel function parameter σ of the online least squares support vector machine. 2and penalty factor γ, avoiding the blindness of the trial and error method to determine the parameters; in view of the problem that the particle swarm algorithm is prone to fall into the local optimum, the adaptive inertia weight and dynamic learning factor are introduced to improve the particle swarm algorithm, so that the particle swarm algorithm focuses on global exploration ability in the initial stage and local search in the later iteration, which improves the search efficiency of the algorithm. Compared with the traditional empirical parameter setting method, this parameter optimization method improves the robustness and wide applicability of the online least squares support vector machine algorithm, can solve the problem of poor adaptability of the general inverse system and inability to self-adjust the mutation amount, reduce the risk brought by empirical errors, and effectively reduce the complexity of the distribution of sample data in high-dimensional space. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 It is a structural block diagram of a decoupling controller of a bearingless permanent magnet synchronous generator PWM rectification system of the present invention;

[0011] Figure 2 yes Figure 1 A magnified block diagram of the structure of the online least squares support vector machine generalized inverse 35 optimized by particle swarm algorithm;

[0012] Figure 3 yes Figure 1 The structural block diagram and equivalent diagram of the pseudo-linear system 44 composed of the online least squares support vector machine generalized inverse 35 optimized by the particle swarm algorithm and the composite controlled object 117;

[0013] Figure 4 yes Figure 1 A control block diagram of a first internal model controller 51;

[0014] Figure 5 yes Figure 1 4 is a control block diagram of the second internal model controller 52. DETAILED DESCRIPTION

[0015] like Figure 1 As shown, a decoupling controller for a bearingless permanent magnet synchronous generator PWM rectification system of the present invention includes an additional closed-loop controller 54 and an online least squares support vector machine generalized inverse 35 optimized by a particle swarm algorithm. The additional closed-loop controller 54 is connected in series before the online least squares support vector machine generalized inverse 35 optimized by a particle swarm algorithm. The online least squares support vector machine generalized inverse 35 optimized by a particle swarm algorithm is connected in series before the composite controlled object 117. The three are connected in series in sequence. The composite controlled object 117 has a PWM rectifier 19 and a bearingless permanent magnet synchronous generator 115 connected in series. The output end of the bearingless permanent magnet synchronous generator 115 is connected in series with the input end of the PWM rectifier 19. The PWM rectifier 19 and the bearingless permanent magnet synchronous generator 115 form a generator PWM rectification system 116. The input of the decoupling controller is a given voltage Given radial displacement x in horizontal direction * and vertical direction given radial displacement y * The output of the decoupling controller is the three-phase suspension force control current i Ba 、i Bb 、i Bc and the power generation control switch signal S G , using three-phase suspension force to control current i Ba 、i Bb 、i Bc Control the suspension of the bearingless permanent magnet synchronous generator 115, the bearingless permanent magnet synchronous generator 115 outputs a horizontal radial displacement x and a vertical radial displacement y, and the power generation control switch signal S G Input PWM rectifier 19, PWM rectifier 19 output voltage U DC .

[0016] The additional closed-loop controller 54 is composed of three internal model controllers 51, 52, and 53 in parallel. And the output voltage U DC The voltage difference is obtained by subtracting the voltage difference, which is used as the input of the first internal model controller 51. The first internal model controller 51 outputs the first-order derivative of the voltage The horizontal radial displacement x * The horizontal radial displacement difference is obtained by subtracting the horizontal radial displacement x, and the horizontal radial displacement difference is used as the input of the second internal model controller 52. The second internal model controller 52 outputs the second derivative of the horizontal radial displacement The vertical direction is given by the radial displacement y * The vertical radial displacement difference is obtained by subtracting the vertical radial displacement y, and the vertical radial displacement difference is used as the input of the third internal model controller 53, and the third internal model controller 53 outputs the second derivative of the vertical radial displacement

[0017] The online least squares support vector machine generalized inverse 35 optimized by particle swarm algorithm consists of a particle swarm algorithm optimization module 31, a rectangular window algorithm module 32, a least squares support vector machine 33, a first-order transfer function Second-order standard transfer function Second-order standard transfer function and two integrators s -1 Composition, where a 31 、a 32 is the first-order standard transfer function coefficient, a 10 、a 11 、a 12 、a 20 、a 21 、a 22is the second-order standard transfer function coefficient, and s is a complex space variable.

[0018] Since the inertia weight and learning factor of the traditional particle swarm algorithm are fixed, it is easy to fall into local extreme values. The particle swarm algorithm optimization module 31 uses the improved particle swarm algorithm to optimize the kernel function parameter σ of the online least squares support vector machine. 2 The optimization is performed with the penalty factor γ, and the following adaptive inertia weight ω and dynamic learning factors c1 and c2 are used:

[0019]

[0020]

[0021] Among them, ω max is the maximum value of the inertia weight, ω min is the minimum value of the inertia weight, t is the current iteration number, T max is the total number of iterations, constant α≥0, c 1,max , c 1,min , c 2,max , c 2,min is the initial constant in the interval [0,4], e is a natural constant, constants β1≥0, β2≥0. Compared with the inertia weight and learning factor in the traditional particle swarm algorithm, the adaptive inertia weight ω will decrease nonlinearly as the number of iterations increases, and the dynamic learning factors c1 and c2 will change nonlinearly in the range of [0,4]. Therefore, with the change of the adaptive inertia weight ω and the dynamic learning factors c1 and c2, the improved particle swarm algorithm has a good performance in the number of iterations. When the inertia weight factor is large, the particle will perform a global search with a larger step size. The inertia weight factor is small, and the particles tend to perform fine local search, which can prevent the algorithm from falling into the local optimum and improve the search efficiency of the algorithm.

[0022] The initial kernel function parameter σ of the online least squares support vector machine 2 and penalty factor γ as the two-dimensional particles of the particle swarm algorithm, where the kernel function parameter σ 2 The value range of is 0-1, the value range of the penalty factor γ is 0-1000, the particle swarm is initialized, and the kernel function parameter σ is randomly assigned as the initial particle 2 and the initial position and speed of the penalty factor γ; the regression accuracy of the least squares support vector machine is used as the measure of the fitness value of the particle swarm algorithm, according to the fitness evaluation formula Calculate the fitness value F of each particle, where h is the kernel function parameter σ of the test set input in each iteration 2 and the regression result of the penalty factor γ, h tis the test set output. Initially, the current particle is taken as the individual extreme value, and the particle with the largest fitness value F is selected as the global extreme value through comparison; the particle speed and position are updated based on the adaptive inertia weight ω and the dynamic learning factors c1 and c2. The update rules of the particle speed and position are:

[0023]

[0024] in is the velocity of the nth generation particle, is the velocity of the n+1th generation particle, is the position of the nth generation particle, is the position of the n+1th generation particle, r1 and r2 are random numbers in the interval [0,1], is the individual extreme value of the nth generation, is the global extreme value of the nth generation; recalculate the fitness value of the updated particle, compare the fitness value of the updated particle with the previous individual extreme value, if it is greater than the original individual extreme value, replace the original individual extreme value, compare the fitness value of the updated particle with the previous global extreme value, if it is greater than the original global extreme value, replace the original global extreme value; update the inertia weight and learning factor according to the adaptive inertia weight and dynamic learning factor update formula to participate in the next update of particle speed and position; determine whether the number of iterations or accuracy requirements are met, if not, continue to iterate, otherwise the particle swarm algorithm optimization module 31 outputs the optimal kernel function parameter σ 2 And the penalty factor γ. The optimal kernel function parameter σ 2 And the penalty factor γ is used as the first and second inputs of the least squares support vector machine 33.

[0025] The first internal model controller 51 is sequentially connected through a first-order transfer function The least squares support vector machine 33 is connected to the rectangular window algorithm module 32. The first voltage first-order derivative output by the first internal model controller 51 Through the first-order transfer function Get voltage U DC , voltage U DC As the first input of the rectangular window algorithm module 32, the voltage U DC After passing through the rectangular window algorithm module 32, the voltage training sample of m duration at the current time k is output. It is used as the third input of the least squares support vector machine 33, where is the voltage at km-1, is the voltage at time k. The first-order derivative of voltage As the second input of the rectangular window algorithm module 32, the voltage first-order derivative training sample of m duration at the current time k is obtained after passing through the rectangular window algorithm module 32. And use it as the fourth input of the least squares support vector machine 33, where is the first-order derivative of voltage at km-1, is the first-order derivative of voltage at time k.

[0026] The second internal model controller 51 is sequentially connected through the second-order transfer function The first integrator s -1 The least squares support vector machine 33 is connected to the rectangular window algorithm module 32. The second internal model controller 51 outputs the second derivative of the horizontal radial displacement Through the second-order transfer function Get the first-order derivative of the horizontal radial displacement First-order derivative of horizontal radial displacement As the third input of the rectangular window algorithm module 32, the first-order derivative of the horizontal radial displacement After the rectangular window algorithm module 32, the first-order derivative training sample of the horizontal radial displacement of m time length at the current k moment is obtained. and uses it as the fifth input of the least squares support vector machine 33, is the first-order derivative of the horizontal radial displacement at km-1, is the first-order derivative of the horizontal radial displacement at time k. After the first integrator s -1 The horizontal radial displacement x is obtained, and the horizontal radial displacement x is used as the fourth input of the rectangular window algorithm module 32. After the horizontal radial displacement x passes through the rectangular window algorithm module 32, the horizontal radial displacement training sample x(k)=[x k-m+1 ,...,x k ] and uses it as the sixth input of the least squares support vector machine 33, x k-m-1 is the horizontal radial displacement at km-1, and x is the horizontal radial displacement at k. The second-order derivative of radial displacement As the fifth input to the rectangular window algorithm module 32, the radial displacement second derivative After the rectangular window algorithm module 32, the second-order derivative training sample of the radial displacement with a duration of m at the current time k is obtained. and use it as the seventh input of the least squares support vector machine 33, is the second-order derivative of the horizontal radial displacement at km-1, is the second-order derivative of the horizontal radial displacement at time k.

[0027] The third internal model controller 53 is sequentially connected through the second-order transfer function The second integrator s -1The least squares support vector machine 33 is connected to the rectangular window algorithm module 32. The vertical radial displacement second-order derivative output by the third internal model controller 53 is Through the second-order transfer function Get the first-order derivative of the vertical radial displacement First derivative of radial displacement in vertical direction As the sixth input of the rectangular window algorithm module 32, the first-order derivative of the radial displacement in the vertical direction After the rectangular window algorithm module 32, the vertical radial displacement first-order derivative training sample of the current k time and m duration is obtained. and use it as the eighth input of the least squares support vector machine 33, is the first-order derivative of the vertical radial displacement at the time km-1, is the first-order derivative of the vertical radial displacement at time k. After the second integrator s -1 The vertical radial displacement y is obtained, and the vertical radial displacement y is used as the seventh input of the rectangular window algorithm module 32. After the vertical radial displacement y passes through the rectangular window algorithm, the vertical radial displacement training sample y(k) of the current time k and the length m is obtained. k-m+1 ,...,y k ] and uses it as the ninth input of the least squares support vector machine 33, y k-m+1 is the vertical radial displacement at km-1, y k is the vertical radial displacement at time k. The second-order derivative of the vertical radial displacement is As the eighth input of the rectangular window algorithm module 32, after the rectangular window algorithm, the vertical radial displacement second-order derivative training sample of the current k time and m time length is obtained. and use it as the tenth input of the least squares support vector machine 33, is the second-order derivative of the vertical radial displacement at the time km-1, is the second-order derivative of the vertical radial displacement at time k.

[0028] The output of the least squares support vector machine 33 is the given generation control current of the dq axis With dq axis given suspension force control current dq axis given generation control current With dq axis given suspension force control current Both are input into the composite controlled object 117 to control the PWM rectifier 19 and the bearingless permanent magnet synchronous generator 115 respectively.

[0029] The mathematical model of the composite controlled object 117 is established, that is, a fifth-order differential equation, whose vector relative order is {2,2,1}. It can be proved by derivation that the fifth-order differential equation is reversible, that is, the inverse system exists, and it can be determined that the input of the inverse system is the second-order derivative of the radial displacement in the horizontal and vertical directions and the first-order derivative of voltage The four outputs are respectively the dq axis given generation control current With dq axis given suspension force control current like Figure 2 As shown, the rectangular window algorithm module 32 is combined with the least squares support vector machine 33 to form an online least squares support vector machine 34, and the step excitation signal dq axis is given to the power generation control current With dq axis given suspension force control current Added to the input end of the composite controlled object 117, collect the horizontal and vertical radial displacements x, y of the bearingless permanent magnet synchronous generator 115 and the output voltage U DC , calculate the first-order derivative of radial displacement in horizontal and vertical directions Second Derivative and the first-order derivative of voltage The signal is normalized to obtain the training sample set of the online least squares support vector machine 34

[0030] The window width in the rectangular window algorithm module 32 is measured in time, and the window length is set to m. Training is performed only based on data within the current m time length, and all old data before these data are completely eliminated. Therefore, the training samples are rolled over time, and when a new sample comes in, an old sample will be deleted, thereby realizing real-time identification of the system inverse model.

[0031] Since the input parameters are updated in real time after being processed by the rectangular window algorithm module 32, the least squares support vector machine 33 must also be adjusted in real time to ensure the accuracy of the trained inverse system. The least squares support vector machine 33 is based on the optimal kernel function parameter σ 2 , calculate the kernel function at time k according to the formula X k and X' k Generally refers to two different training samples input into the support vector machine at time k. Then according to the kernel function K at time k k and penalty factor γ to calculate the symmetric positive definite matrix at time k According to the symmetric positive definite matrix Φ at time k k Get the parameter matrix at time k Among them, M k-1 is the parameter matrix at time k-1, I is the identity matrix, β is the forgetting factor, and 0<β<1.

[0032] Therefore, as the training samples are continuously input into the least squares support vector machine 33, different kernel functions K are obtained. k , the symmetric positive definite matrix Φ at time k k With the parameter matrix M k Follow the iterative update. According to the formula θ k =θ k-1 +M k-1 Φ k (βI+Φ k M k-1 Φ k ) -1 (z k -Φ k θ k-1 ) to obtain the bias θ at time k k and the Lagrange multiplier α at time k k The matrix of but Among them, b k-1 , b k are the biases at time k-1 and time k respectively, α k-1 , α k are the Lagrange multipliers at time k-1 and time k, respectively, z k is the output matrix at time k, that is, the inverse system outputs the matrix at time k. Therefore, as the training samples are continuously input, Φ k With M k By continuously iterating and updating, we can finally obtain the bias b of the online least squares support vector machine 34 at time k. k and the Lagrange multiplier α at time k k Based on the bias b k and the Lagrange multiplier α k The output of the least squares support vector machine 33 is calculated as ψ = ∑α k K k +b k , The dq axis given generation control current is obtained With dq axis given suspension force control current

[0033] like Figure 3 As shown, the online least squares support vector machine generalized inverse 35 optimized by the particle swarm algorithm is connected in series with the composite controlled object 117 to form a pseudo-linear system 44, which is composed of two displacement second-order integral pseudo-linear subsystems 42 and 43 and a voltage first-order integral pseudo-linear subsystem 41, thereby realizing the nonlinear dynamic decoupling between the radial two-degree-of-freedom suspension force and between the radial suspension force and the output voltage, and converting the complex nonlinear system control into a simple three-variable linear system control.

[0034] The first internal model controller 51 is designed for the voltage first-order integral pseudo-linear subsystem 41, such as Figure 4 As shown, the first internal model controller 51 is composed of a first built-in controller 21, a first internal model 22 and a first low-pass filter 23. The output end of the first built-in controller 21 is connected to the voltage first-order integral type pseudo-linear subsystem 41 and the input end of the first internal model 22 respectively. With voltage U DC The first built-in controller 21 outputs the first-order derivative of the voltage Voltage first derivative The first internal model 22 is input into the voltage first-order integral pseudo-linear subsystem 41 and the first internal model 22 respectively. The output terminals of the voltage first-order integral pseudo-linear subsystem 41 and the first internal model 22 are connected in parallel. The voltage output by the voltage first-order integral pseudo-linear subsystem 41 is subtracted from the voltage output by the first internal model 22. The voltage difference is then fed back to the first internal controller 21 after passing through the first low-pass filter 23. The first internal model 22 adopts the transfer function The transfer function of the first low-pass filter 23 is Where τ1 is the parameter of the first low-pass filter 23, s is a complex space variable, and the transfer function of the first built-in controller 21 is G c1 (s)=F1(s)G1(s) -1 , G1(s) -1 Represents the inverse of the transfer function G1(s) of the first internal model 22.

[0035] Similarly, the second internal model controller 52 is designed for the displacement second-order integral pseudo-linear subsystem 42, such as Figure 5 As shown, the second internal model controller 52 is composed of a second built-in controller 24, a second internal model 25 and a second low-pass filter 26. * The difference between the horizontal radial displacement x and the feedback is input into the second built-in controller 24, and the output of the second built-in controller 24 is the second-order derivative of the horizontal radial displacement The output end of the second built-in controller 24 is connected to the displacement second-order integral pseudo-linear subsystem 42 and the second internal model 25 respectively. The second internal model 25 is connected in parallel with the output end of the displacement second-order integral pseudo-linear subsystem 42. The horizontal radial displacement outputted by the second internal model 25 is subtracted. The displacement difference is fed back to the second built-in controller 24 through the second low-pass filter 26. The transfer function of the second internal model 25 is: The transfer function of the second low pass filter 26 is set to Where τ2 is the parameter of the second low-pass filter 26, s is a complex space variable, then the transfer function of the second built-in controller 24 is G c2 (s)=F2(s)G2(s) -1 , where F2(s) is the transfer function of the second low-pass filter 26, G2(s) -1 represents the inverse of the transfer function G2(s) of the second internal model 25.

[0036] Similarly, the third internal model controller 53 is designed for the displacement second-order integral pseudo-linear subsystem 43, and its structure is similar to Figure 5 The second inner model controller 52 shown is similar.

[0037] like Figure 1 As shown, the composite controlled object 117 is composed of a PWM rectifier 19, a bearingless permanent magnet synchronous generator 115, four PI controllers 11, 12, 13, 14, two Park inverse transforms 15, 16, two Park transforms 113, 114, two Clark transforms 111, 112, two SVPWMs 17, 18 and a voltage source inverter 110. The three-phase AC current i output by the bearingless permanent magnet synchronous generator 115 is Ga 、i Gb 、i Gc The three-phase AC current i Ga 、i Gb 、i Gc The first Clark transformation 111 and the first Park transformation 113 are sequentially transformed into the dq axis power generation control current i Gd 、i Gq , the dq axis given generation control current output by the least squares support vector machine 33 Respectively with the corresponding dq axis generation control current i Gd 、i Gq The difference is respectively passed through the corresponding first PI controller 11 and the second PI controller 12, and then passed through the first Park inverse transform 15 and the first SVPWM 17 in sequence to output the power generation control switch signal S G , controls the PWM rectifier 19. The voltage source inverter 110 outputs a three-phase suspension force control current i Ba 、i Bb 、i Bc Control the suspension of the bearingless permanent magnet synchronous generator 115, the three-phase suspension force controls the current i Ba 、i Bb 、i BcAfter the second Clark transformation 112 and the second Park transformation 114, it is transformed into the dq axis suspension force control current i Bd 、i Bq The dq axis given suspension force control current output by the least squares support vector machine 33 The corresponding dq axis suspension force control current i Bd 、i Bq The difference is respectively passed through the corresponding third PI controller 13 and the fourth PI controller 14, and then passed through the first Park inverse transform 16 and the second SVPWM 18 in sequence to output the suspension force control switch signal S B To the voltage source inverter 110.

Claims

1. A decoupling controller for a PWM rectifier system of a bearingless permanent magnet synchronous generator, comprising three internal model controllers (51, 52, 53), a particle swarm algorithm optimization module (31) and a least squares support vector machine (33), wherein the output end of the least squares support vector machine (33) is connected to a composite controlled object, wherein the composite controlled object comprises a PWM rectifier and a bearingless permanent magnet synchronous generator connected in series, wherein a given voltage is and the PWM rectifier output voltage U DC The voltage difference obtained by the differential operation is used as the input of the first internal model controller (51) to set the horizontal radial displacement x * The horizontal radial displacement difference obtained by subtracting the horizontal radial displacement x output by the bearingless permanent magnet synchronous generator is used as the input of the second internal model controller (52). The vertical radial displacement y is set to * The vertical radial displacement difference obtained by subtracting the vertical radial displacement y output by the bearingless permanent magnet synchronous generator is used as the input of the third internal model controller (53), which is characterized by: The first internal model controller (51) is sequentially connected via a first-order transfer function The least square support vector machine (33) is connected to the rectangular window algorithm module (32), and the second internal model controller (51) is sequentially connected to the second order standard transfer function The first integrator s -1 The least square support vector machine (33) is connected to the rectangular window algorithm module (32), and the third internal model controller (53) is sequentially connected through the second-order transfer function The second integrator s -1 The least squares support vector machine (33) is connected to the rectangular window algorithm module (32), and the particle swarm algorithm optimization module (31) outputs the optimal kernel function parameter σ 2 and penalty factor γ, optimal kernel function parameter σ 2 and the penalty factor γ as the first and second inputs of the least squares support vector machine (33), and the output of the rectangular window algorithm module (32) is the voltage U DC , U DC The eight training samples are the first-order derivative, horizontal radial displacement x, first-order derivative of x, second-order derivative of x, vertical radial displacement y, first-order derivative of y, and second-order derivative of y. The eight training samples are used as the third to tenth inputs of the least squares support vector machine (33); the least squares support vector machine 33 is based on the optimal kernel function parameter σ 2 Get the kernel function at time k X k and X′ k Generally refers to two different training samples input into the support vector machine at time k, according to the kernel function K at time k k and the penalty factor γ to obtain the symmetric positive definite matrix at time k Parameter Matrix Bias θ k and the Lagrange multiplier α k The matrix And the output of the least squares support vector machine (33) ψ = ∑α k K k +b k , The dq axis given generation control current of the input composite controlled object and dq axis given suspension force control current M k-1 is the parameter matrix at time k-1, I is the identity matrix, β is the forgetting factor, and 0<β<1, b k-1 、b k are the biases at time k-1 and k respectively, α k-1 , α k are the Lagrange multipliers at time k-1 and k respectively, z k Output matrix for k time instant.

2. A bearingless permanent magnet synchronous generator PWM rectifier system decoupling controller according to claim 1, characterized in that: The particle swarm optimization module (31) sets the initial kernel function parameter σ 2 and penalty factor γ as two-dimensional particles, based on adaptive inertia weight And dynamic learning factors c1, c2: To update the particle's velocity and position, ω max is the maximum value of the inertia weight, ω min is the minimum value of the inertia weight, t is the current iteration number, T max is the total number of iterations, constant α≥0, c 1,max , c 1,min , c 2,max , c 2,min is the initial constant in the interval [0,4], e is a natural constant, and the constants β1≥0, β2≥0.

3. A bearingless permanent magnet synchronous generator PWM rectifier system decoupling controller according to claim 2, characterized in that: The update rules for particle velocity and position are: is the velocity of the nth generation particle, is the velocity of the n+1th generation particle, is the position of the nth generation particle, is the position of the n+1th generation particle, r1 and r2 are random numbers in the interval [0,1], is the individual extreme value of the nth generation, is the global extreme value of the nth generation.

4. A bearingless permanent magnet synchronous generator PWM rectifier system decoupling controller according to claim 1, characterized in that: The first internal model controller (51) outputs the first-order derivative of the voltage Voltage first derivative Through the first-order transfer function Get voltage U DC , voltage U DC As the first input of the rectangular window algorithm module (32), the voltage U DC After passing through the rectangular window algorithm module (32), a voltage training sample of m duration at the current time k is output. As the third input to the least squares support vector machine (33), is the voltage at km-1, is the voltage at time k, the first-order derivative of the voltage As the second input of the rectangular window algorithm module (32), after passing through the rectangular window algorithm module (32), a voltage first-order derivative training sample of m duration at the current time k is obtained. As the fourth input of the least squares support vector machine 33, is the first-order derivative of voltage at km-1, is the first-order derivative of voltage at time k.

5. The decoupling controller of a bearingless permanent magnet synchronous generator PWM rectification system according to claim 1 is characterized in that: The second derivative of the horizontal radial displacement output by the second internal model controller (51) is Second order derivative of displacement Through the second-order transfer function Get the first-order derivative of displacement First-order derivative of displacement As the third input to the rectangular window algorithm module (32), the first-order derivative of the displacement After passing through the rectangular window algorithm module (32), the first-order derivative training sample of the horizontal radial displacement of the current time k and the length m is obtained. It is used as the fifth input of the least squares support vector machine (33). is the first-order derivative of the horizontal radial displacement at km-1, is the first-order derivative of the horizontal radial displacement at time k, and the first-order derivative of the displacement After the first integrator s -1 Then the horizontal radial displacement x is obtained, and the displacement x is used as the fourth input of the rectangular window algorithm module (32). After the displacement x passes through the rectangular window algorithm module (32), the horizontal radial displacement training sample x(k)=[x k-m+1 ,...,x k ] and uses it as the sixth input of the least squares support vector machine (33), x k-m+1 is the horizontal radial displacement at km-1, x k is the horizontal radial displacement at time k; the second-order derivative of radial displacement As the fifth input of the rectangular window algorithm module (32), the second-order derivative of the radial displacement After passing through the rectangular window algorithm module (32), the second-order derivative training sample of the radial displacement of the current time k and the length m is obtained. and used as the seventh input of the least squares support vector machine (33). is the second-order derivative of the horizontal radial displacement at km-1, is the second-order derivative of the horizontal radial displacement at time k.

6. A bearingless permanent magnet synchronous generator PWM rectifier system decoupling controller according to claim 1, characterized in that: The third internal model controller (53) outputs the second-order derivative of the vertical radial displacement Second order derivative of displacement Through the second-order transfer function Get the first-order derivative of displacement First-order derivative of displacement As the sixth input to the rectangular window algorithm module (32), the first-order derivative of the displacement After passing through the rectangular window algorithm module (32), a training sample of the first-order derivative of the vertical radial displacement at the current time k and the length m is obtained. and used as the eighth input of the least squares support vector machine (33). is the first-order derivative of the vertical radial displacement at the time km-1, is the first-order derivative of the vertical radial displacement at time k, and the first-order derivative of the displacement After the second integrator s -1 The vertical radial displacement y is obtained, and the radial displacement y is used as the seventh input of the rectangular window algorithm module (32). After the radial displacement y passes through the rectangular window algorithm, the vertical radial displacement training sample y(k) of the current time k and the length m is obtained. k-m+1 ,...,y k ] and uses it as the ninth input of the least squares support vector machine (33), y k-m+1 is the vertical radial displacement at km-1, y k is the vertical radial displacement at time k, and the second-order derivative of the displacement As the eighth input of the rectangular window algorithm module (32), after the rectangular window algorithm, the vertical radial displacement second-order derivative training sample of the current time k and the length m is obtained. and used as the tenth input of the least squares support vector machine (33). is the second-order derivative of the vertical radial displacement at the time km-1, is the second-order derivative of the vertical radial displacement at time k.

7. A bearingless permanent magnet synchronous generator PWM rectifier system decoupling controller according to claim 1, characterized in that: The window width in the rectangular window algorithm module (32) is measured in time. Based on the data within the current m time length, all old data before these data are eliminated.

8. The decoupling controller of a bearingless permanent magnet synchronous generator PWM rectification system according to claim 1, characterized in that: The first internal model controller (51) is composed of a first built-in controller (21), a first internal model (22) and a first low-pass filter (23). The output end of the first built-in controller (21) is connected to the voltage first-order integral type pseudo-linear subsystem (41) and the input end of the first internal model (22) respectively. With voltage U DC The difference is input to the first built-in controller (21), and the first built-in controller (21) outputs the first-order derivative of the voltage Voltage first derivative The voltage is input to the voltage first-order integral type pseudo-linear subsystem (41) and the first internal model (22) respectively. The output ends of the voltage first-order integral type pseudo-linear subsystem (41) and the first internal model (22) are connected in parallel. The voltage output by the voltage first-order integral type pseudo-linear subsystem (41) is subtracted from the voltage output by the first internal model (22). The voltage difference is then fed back to the first built-in controller (21) after passing through the first low-pass filter (23).

9. A bearingless permanent magnet synchronous generator PWM rectifier system decoupling controller according to claim 8, characterized in that: The first internal model (22) uses the transfer function The transfer function of the first low-pass filter (23) is τ1 is the parameter of the first low-pass filter (23), s is a complex space variable, and the transfer function of the first built-in controller (21) is G c1 (s)=F1(s)G1(s) -1 , G1(s) -1 The inverse of the transfer function G1(s) of the first internal model 22.

Images