Flywheel energy storage hybrid magnetic bearing lm neural network inverse decoupling controller

Abstract

The application discloses a flywheel energy storage hybrid magnetic bearing LM neural network inverse decoupling controller, output ends of two dynamic prediction modules are connected in series with input ends of LM neural network inverse systems, the first dynamic prediction module is composed of two series branches, the second dynamic prediction module is composed of three series branches and a composite signal calculation module, each series branch is composed of a feedback correction module, a control increment calculation module, a prediction value calculation module and an optimal value calculation module in series, input of the first series branch is weight and reference weight, output is optimal initial weight, input of the second series branch is threshold and reference threshold, output is optimal initial threshold, combining respective advantages of dynamic prediction, LM neural network control and inverse system, initial weight and threshold of the LM neural network are optimized by the dynamic prediction module, independent and accurate control among three degrees of freedom of the flywheel energy storage hybrid magnetic bearing is realized.

Description

Technical Field

[0001] This invention belongs to the technical field of electric drive control equipment, and relates to the decoupling control technology of hybrid magnetic bearings for flywheel energy storage. Specifically, it is an LM neural network inverse decoupling controller for hybrid magnetic bearings for flywheel energy storage, which is suitable for decoupling control of multivariable, nonlinear, and strongly coupled hybrid magnetic bearings for flywheel energy storage. Background Technology

[0002] Currently, the main energy storage methods include chemical energy storage, superconducting energy storage, and physical energy storage. Among them, flywheel energy storage, a physical energy storage method, has become a focus of new energy due to its high energy density, long service life, and high energy conversion efficiency, and has been successfully applied in fields such as uninterruptible power supplies (UPS) and wind power plants. Traditional flywheel energy storage systems use mechanical bearings as support devices, resulting in significant frictional losses, which greatly limits the rotor's mass and speed, and requires periodic bearing maintenance. Magnetic bearings, with their advantages of being frictionless, requiring no lubrication, having high rotor displacement accuracy, and long service life, are used in flywheel energy storage. Flywheel energy storage using hybrid magnetic bearings is a multi-input, multi-output, strongly coupled, nonlinear, multivariable system, thus requiring linear decoupling control to ensure stable and accurate operation. Common linear decoupling control methods include: approximate linearization, differential geometry, analytical inverse system methods, and neural network inverse methods. However, approximate linearization can only perform static decoupling; changes in parameters, critical speeds, disturbances, and loads during system operation can lead to deterioration in control performance. Differential geometry uses abstract mathematical tools, is computationally complex, and is not easily applied in a wide range of applications. Neural network inverse methods utilize the ability of neural networks to approximate nonlinear systems with arbitrary precision to establish inverse models. However, neural networks themselves suffer from slow convergence, susceptibility to local minima, and the significant influence of training samples on weight adjustments.

[0003] To address these issues, Chinese patent application number 201710511782.9 discloses a "fuzzy neural network decoupling controller for a five-degree-of-freedom bearingless permanent magnet synchronous motor." This controller uses a fuzzy neural network combined with dynamic prediction to decouple the five-degree-of-freedom bearingless permanent magnet synchronous motor. However, the fuzzy neural network method requires high accuracy of the system model, and too many model parameters can easily lead to overfitting. Furthermore, the inference and learning speed is slow, and the initial values ​​of the network are randomly selected, which has a significant impact on the decoupling effect. Therefore, the control effect is not ideal. Summary of the Invention

[0004] The purpose of this invention is to solve the problems existing in the control technology of hybrid magnetic bearings for flywheel energy storage, and to propose a robust and fast learning speed LM neural network inverse decoupling controller for hybrid magnetic bearings for flywheel energy storage.

[0005] The technical solution adopted by the hybrid magnetic bearing LM neural network inverse decoupling controller for flywheel energy storage of this invention is as follows: the output end of the LM neural network inverse system is connected to a composite controlled object including a six-pole radial-axial hybrid magnetic bearing for flywheel energy storage; the output ends of two dynamic prediction modules are connected in series to the input end of the LM neural network inverse system; the first dynamic prediction module consists of the first and second series branches; the second dynamic prediction module consists of the third, fourth, and fifth series branches and a composite signal calculation module; each of the series branches consists of a feedback correction module, a control increment calculation module, a prediction value calculation module, and an optimal value calculation module connected in series; the input of the first series branch is the weight ω and the reference weight ω. * The output is the optimal initial weight ω. c The inputs to the second serial branch are the threshold θ and the reference threshold θ. * The output is the optimal initial threshold θ. c The inputs to the third series branch are the radial displacement x and the radial displacement setpoint x. * The output is the final radial displacement prediction value x. ac (k+1); The inputs to the fourth series branch are the radial displacement y and the radial displacement setpoint yk+1. * The output is the radial displacement prediction value y. ac (k+1); The input axial displacement z and the given axial displacement z of the fifth series branch. * The output is the predicted axial displacement value z. ac (k+1); Final radial displacement prediction value x ac (k+1), y ac (k+1) and the final predicted axial displacement z ac (k+1) is used as the input to the composite signal calculation module, which calculates the composite control signal j. a The inverse system of the LM neural network uses the aforementioned optimal initial weights ω. c Optimal initial threshold θ c and composite control signal j a As input, the expected value of the output equivalent control current i z * i x * i y * Give it to the composite controlled object.

[0006] Furthermore, in the first series branch, the feedback correction module will use the weight ω and the reference weight ω *The weight error e1(k) at the current time k is obtained by comparison. The weight at time k+1 is corrected using the weight error e1(k) to obtain the corrected weight ω'(k+1) at time k+1. The initial predicted weight at time k+1 is set based on the weight ω'(k+1) at time k+1. Initial predicted value of weights The control increment calculation module and the prediction value calculation module are respectively input, and the control increment calculation module calculates the weighted control increment Δu. ω (k), the prediction value calculation module calculates the prediction value based on the initial prediction value of the weights. and weight control increment Δu ω (k) Calculate the final predicted value ω(k+1), input the final predicted value ω(k+1) into the optimal value calculation module, and the optimal value calculation module outputs the optimal solution ω with weights. c The principles of the other four series branches are similar to those of the first series branch.

[0007] Furthermore, the correction weight ω'(k+1) at time k+1 is ω * +σe1(k), the initial predicted value of the weights The weight control increment The final predicted value The optimal solution of the weights ω c =α|ω(k+1)|+(1-α)ω 2 (k+1); α is the optimization factor, ranging from 0 to 1; σ is the feedback correction coefficient, taken as 0.55; s is the shift coefficient; i = 0, 1, ..., k; d i is the error coefficient; a is the increment coefficient, with a value ranging from 0 to 1.

[0008] The advantages of this invention using the above technical solution are as follows:

[0009] 1. This invention combines the advantages of dynamic prediction, LM neural network control and inverse system to achieve independent and precise control of the three degrees of freedom of the hybrid magnetic bearing for flywheel energy storage. It overcomes the problem of control performance degradation caused by system parameter perturbation, disturbance and load change, and makes the system have excellent dynamic and static characteristics.

[0010] 2. The initial weights and thresholds of the LM neural network are optimized by using a dynamic prediction module to reduce the impact of the initial weight and threshold selection on the training of the neural network. The weight matrix of the neural network system is adjusted in real time. Combining the advantages of the LM neural network's good learning ability and fast convergence speed, the training accuracy and generalization ability of the neural network are improved, the accuracy of the online neural network inverse module is improved, and dynamic decoupling control between the outputs is realized.

[0011] 3. The LM algorithm is an improved form of the Gauss-Newton method, possessing both the local characteristics of the Gauss-Newton method and the global characteristics of the gradient method. The introduction of regularization reduces the number of parameters in the neural network model, significantly improving the generalization ability, approximation accuracy, and convergence speed of the LM neural network. Employing a rolling optimization strategy instead of a single global optimization, the LM algorithm can promptly compensate for uncertainties caused by model distortion, interference, and other factors, improving the algorithm's real-time performance. The neural network was trained to an accuracy of 0.002, demonstrating good dynamic performance.

[0012] 4. Compared to the standard BP neural network inverse decoupling controller, this invention does not require calculating the Hessian matrix of the objective function, greatly reducing computational complexity, improving convergence speed, and solving the problem of the non-invertibility of the Hessian matrix during the solution process. It can effectively improve the real-time performance of the controller while ensuring control performance. By decoupling the nonlinear, strongly coupled system of the flywheel energy storage hybrid magnetic bearing into a second-order position pseudolinear system, the mature linear control theory can be used to achieve precise and high-performance control of the flywheel energy storage hybrid magnetic bearing. Attached Figure Description

[0013] Figure 1 This is a block diagram of the hybrid magnetic bearing LM neural network inverse decoupling controller for flywheel energy storage of the present invention;

[0014] Figure 2 yes Figure 1 A schematic diagram illustrating the equivalent effects of a composite controlled object;

[0015] Figure 3 yes Figure 1 Block diagram of the inverse system of the LM neural network;

[0016] Figure 4 yes Figure 1 The structural block diagram of the first dynamic prediction module in the program;

[0017] Figure 5 yes Figure 1 The structural block diagram of the second dynamic prediction module;

[0018] In the diagram: 1. Composite controlled object; 2. Six-pole radial-axial hybrid magnetic bearing for flywheel energy storage; 3. DC power amplifier; 4. Extended current hysteresis three-phase power inverter; 5. LM neural network inverse system; 41. Clark transform; 42. Current hysteresis three-phase power inverter; 51. LM neural network model; 52. Second-order differential processor; 61. Dynamic prediction module; 62. Dynamic prediction module; 611, 615, 621, 624, 627. Feedback correction modules; 612, 616, 622, 625, 628. Control increment calculation modules; 613, 617, 623, 626, 629. Predicted value calculation modules; 614, 618. Optimal value calculation modules; 620. Composite signal calculation module. Detailed Implementation

[0019] like Figure 1 As shown, the flywheel energy storage hybrid magnetic bearing LM neural network inverse decoupling controller of the present invention consists of two dynamic prediction modules 61, 61 and an LM neural network inverse system 5. The output terminals of the two dynamic prediction modules 61, 61 are connected in series with the input terminal of the LM neural network inverse system 5. The output terminal of the LM neural network inverse system 5 is connected to the composite controlled object 1 containing the flywheel energy storage six-pole radial-axial hybrid magnetic bearing 2.

[0020] The output of the composite controlled object 1 is the radial displacement x, y, and axial displacement z of the six-pole radial-axial hybrid magnetic bearing used for flywheel energy storage. The inputs to the first dynamic prediction module 61 are random weights ω and a threshold θ, as well as a reference weight ω. * and reference threshold θ * The output is the optimal initial weight ω. c and the optimal initial threshold θ c The inputs to the second dynamic prediction module 62 are radial displacements x and y, axial displacement z, and the given radial displacement value x. * y * axial displacement given value z * The output is a composite control signal j a The input to the inverse system 51 of the LM neural network is the optimal initial weight ω. c Optimal initial threshold θ c and composite control signal j a The output is the expected value of the equivalent control current i. z * i x * i y * .

[0021] Combination Figure 2The composite controlled object 1 shown consists of an axial DC power amplifier 3, a radially extended current-hysteresis three-phase power inverter 4, and a six-pole radial-axial hybrid magnetic bearing 2 for flywheel energy storage. The outputs of the DC power amplifier 3 and the extended current-hysteresis three-phase power inverter 4 are connected in series with the input of the six-pole radial-axial hybrid magnetic bearing 2. The extended current-hysteresis three-phase power inverter 4 is composed of a radial Clark inverse converter 41 connected in series with the current-hysteresis three-phase power inverter 42. The desired radial equivalent control current i of the six-pole radial-axial hybrid magnetic bearing 2 for flywheel energy storage is... x * i y * The three-phase current is transformed into the expected value i by Clark inverse transform 41. A * i B * i C * The current hysteresis three-phase power inverter 42 tracks the desired value of the three-phase current and outputs the radial control current i of the six-pole radial-axial hybrid magnetic bearing 2 for flywheel energy storage. A i B i C DC power amplifier 3 is based on the desired value of the axial control current i. z * The axial control current i of the six-pole radial-axial hybrid magnetic bearing 2 used for output flywheel energy storage z Therefore, the input to the composite controlled object 1 is the expected value of the equivalent control current i. x * i y * i z * The output is the axial and radial displacements z, x, and y.

[0022] The magnetic circuit of the six-pole radial-axial hybrid magnetic bearing 2 for flywheel energy storage is analyzed, and a mathematical model of the six-pole radial-axial hybrid magnetic bearing 2 for flywheel energy storage under ideal conditions is established. The mathematical model of the six-pole radial-axial hybrid magnetic bearing 2 for flywheel energy storage is a 6th-order differential matrix equation with a relative vector order of {2,2,2}. It can be verified by the Interactor algorithm. Therefore, the six-pole radial-axial hybrid magnetic bearing 2 is invertible, that is, a right-hand inverse system exists.

[0023] like Figure 3As shown, based on the order of the composite controlled object 1, the inverse model of the composite controlled object 1 can be established by adding a second-order difference processor 52 to the LM neural network model 51. That is, the LM neural network inverse system 5 can be composed of a second-order difference processor 52 and an LM neural network model 51 connected in series. Using a random signal as the excitation signal of the composite controlled object 1, input and output data are collected to fully obtain the dynamic and static characteristics of the composite controlled object 1. The output displacements z, x, y and their first and second derivatives are predicted and calculated. Then, the data is normalized to obtain the training samples of the LM neural network model 51. 3000 sets of training samples that can fully reflect the dynamic and static characteristics of the composite controlled object 1 are selected. The composite control signals z, x, y and their first and second derivatives are used as the inputs of the LM neural network model 51, and the input i of the composite controlled object 1 is used as the inputs of the LM neural network model 51. z * i x * i y * As the expected output based on the LM neural network model 51, the first dynamic prediction module 61 is used to optimize the initial weights and thresholds of the LM neural network model 51. The LM neural network model 51 is optimized using the LM algorithm in a rolling optimization manner, and the objective function of the rolling optimization is... Among them, i r * Let i be the reference output value at time k+1. p Let u(k) be the predicted output value at time k+1, u(k) be the control quantity at time k, u(k-1) be the control quantity at time k-1, and ||·|| 2 All are in L2 regularization norm form, where λ is the regularization parameter, ranging from 0 to 1. Regularization is used to limit the weights in the network. After training until the accuracy reaches 0.002, the LM neural network model is obtained. The LM neural network topology includes an input layer, a hidden layer, and an output layer. The hyperbolic tangent sigmoid function is selected as the hidden layer node transfer function, the linear purelin function is set as the output layer node transfer function, and the trainlm function of Levenbrg-Marquardt (LM) is set as the training function of the LM neural network algorithm.

[0024] Connect the input of the second-order difference processor 52 to the output of the second dynamic prediction module 62. The input of the second-order difference processor 52 is the composite control signal j. a The output is a composite control signal j a and its first-order and second-order differential signals The output of the first dynamic prediction module 61 is connected to the input of the LM neural network model 51, and the optimal initial weight ω output by the first dynamic prediction module 61 is used. c Threshold θ c and composite control signal ja and its first-order and second-order differential signals The inputs are fed into the LM neural network model 51 for inverse system regression training, and the expected value of the equivalent control current i is output. x * i y * i z * To the composite controlled object 1.

[0025] The second-order differential processor 52 pairs the input composite control signal j a The processing is performed as follows: the composite control signal j at time k. a First-order difference signal The composite control signal j at time k-4 a The composite control signal j at times (k-4) and (k-3) a The composite control signal j at times (k-3) and (k-1) a The composite control signal j at time (k-1) and time k a (k) is calculated using the following formula: The composite control signal j at time k a Second-order difference signal The composite control signal j at time k-4 a The composite control signal j at times (k-4) and (k-3) a The composite control signal j at times (k-3) and (k-2) a The composite control signal j at times (k-2) and (k-1) a The composite control signal j at time (k-1) and time k a (k) is calculated using the following formula:

[0026] like Figure 4 As shown, the first dynamic prediction module 61 consists of two feedback correction modules 611 and 615, two control increment calculation modules 612 and 616, two prediction value calculation modules 613 and 617, and two optimal value calculation modules 614 and 618. Each of these modules is connected in series to form two parallel branches, designated as the first and second parallel branches. The outputs of both optimal value calculation modules 614 and 618 are connected to the input of the LM neural network inverse system 5, meaning the outputs of the first and second parallel branches are both connected to the input of the LM neural network inverse system 5. The inputs of the first parallel branch are the weight ω and the reference weight ω. * The output is the optimal initial weight ω.c The inputs to the second serial branch are the threshold θ and the reference threshold θ. * The output is the optimal initial threshold θ. c .

[0027] Taking the first series branch as an example: the first feedback correction module 611 uses weight ω and reference weight ω * As input, the weights ω and reference weights ω * By comparison, the weight error e1(k) = ω at the current time k is obtained. * -ω, using the weight error e1(k) to correct the weight at the next time step, i.e., time k+1, the corrected weight ω'(k+1) at time k+1 is calculated using the following formula:

[0028] ω'(k+1)=ω * +σe1(k),

[0029] In the formula, σ is the feedback correction coefficient, which is taken as 0.55 based on the actual control effect.

[0030] The initial predicted value of the weight at time k+1 is set by shifting the weight ω'(k+1) at time k+1.

[0031]

[0032] In the formula, s is the shift coefficient, which is determined based on the actual control effect. When predicting at time k+1, k+2 is the first predicted value output, but we need to obtain the value starting at time k+1, so we need to multiply it by a shift coefficient to shift it.

[0033] The corrected and shifted initial predicted weight values The initial predicted weights are respectively input into the first control increment calculation module 612 and the first predicted value calculation module 613. As the first input to the first prediction calculation module 613, the first control increment calculation module 612 uses the weighted initial prediction value. To calculate the weighted control increment Δu ω (k), weight control increment Δu ω The magnitude of (k) is the sum of the historical errors at time k:

[0034]

[0035] Where i = 0, 1, ..., k, d i This is the error coefficient, determined based on the actual control effect.

[0036] This invention introduces a control increment Δu ω(k) enables dynamic predictive control to not only utilize current and past deviation values, but also predict future deviation values, so that the deviation between the controlled variable and the expected value in online optimization is minimized.

[0037] The first control increment calculation module 612 will control increment Δu ω (k) Input into the first prediction value calculation module 613, as the second input to the first prediction value calculation module 613, the first prediction value calculation module 613 calculates the final prediction value ω(k+1):

[0038]

[0039] Where 'a' is the incremental coefficient, which ranges from 0 to 1 and can be determined based on the actual control effect.

[0040] The final predicted value ω(k+1) is input into the first optimal value calculation module 614 to obtain the optimal solution ω with weights. c :

[0041] ω c =α|ω(k+1)|+(1-α)ω 2 (k+1),

[0042] α is an optimization factor, ranging from 0 to 1. The larger α is, the stronger the system robustness, but the slower the control speed becomes. The specific value needs to be determined based on the actual control effect.

[0043] The second series branch operates on a similar principle to the first series branch, differing only in its input and output values. The second feedback correction module 615 is similar to the second feedback correction module 611, using a threshold θ and a reference threshold θ... * As input, the threshold θ'(k+1) at time k+1 is output. The threshold θ'(k+1) is input into the second control increment calculation module 616 and the second prediction value calculation module 617, respectively. The second control increment calculation module 616 calculates the threshold control increment Δu. θ (k), threshold control increment Δu θ (k) Input the second predicted value calculation module 617. The second predicted value calculation module 617 calculates the final threshold predicted value θ(k+1). The final threshold predicted value θ(k+1) is input into the second optimal value calculation module 618 to obtain the optimal threshold solution θ. c .

[0044] The optimal solution of weight ω c and threshold optimal solution θ c The input is fed into the inverse system 5 of the LM neural network for training.

[0045] like Figure 5As shown, the second dynamic prediction module 62 consists of three feedback correction modules 621, 624, and 627, three control increment calculation modules 622, 625, and 628, three prediction value calculation modules 623, 626, and 629, and a composite signal calculation module 620. Each of the feedback correction modules 621, 624, and 627, the control increment calculation modules 622, 625, and 628, and the prediction value calculation modules 623, 626, and 629 is connected in series to form a series branch, resulting in three series branches: the third, fourth, and fifth series branches. The outputs of the three prediction value calculation modules 623, 626, and 629 are all connected to the input of the composite signal calculation module 620. In other words, the outputs of the third, fourth, and fifth series branches are all connected to the input of the composite signal calculation module 620. The output of the composite signal calculation module 620 is connected to the inverse neural network system 5.

[0046] The principles of the third, fourth, and fifth series branches are similar to those of the first series branch. The third series branch uses radial displacement x and a given radial displacement value x. * As input, the final radial displacement prediction value x ac (k+1) represents the output. Radial displacement x and the radial displacement setpoint x0 * The input is fed into the third feedback correction module 621, where the initial predicted value of the radial displacement at the next time step k+1 is obtained after correction and shifting. The radial displacement control increment Δu is then calculated by the third control increment calculation module 622. x (k), radial displacement control increment Δu x (k) and initial predicted value of radial displacement In the common input third prediction value calculation module 623, the third prediction value calculation module 623 calculates the final radial displacement prediction value x. ac (k+1). Similarly, the fourth feedback correction module 624 uses the radial displacement y and the radial displacement setpoint y * As input, the fifth feedback correction module 627 uses the axial displacement z and the axial displacement setpoint z as inputs. * The fourth feedback correction module 624 outputs the initial predicted value of the radial displacement at time k+1. The fifth feedback correction module 627 outputs the initial predicted value of the axial displacement at time k+1. The fourth control increment calculation module 625 outputs the radial displacement control increment Δu. y (k), the fifth control increment calculation module 628 outputs the axial displacement control increment Δu. z (k); The fourth prediction value calculation module 626 calculates the final radial displacement prediction value y. ac(k+1), the fifth prediction value calculation module 629 calculates the final axial displacement prediction value z. ac (k+1).

[0047] Final radial displacement prediction value x ac (k+1), y ac (k+1) and the final predicted axial displacement z ac (k+1) is used as the input to the composite signal calculation module 620. The composite signal calculation module 620 processes the three predicted values ​​to obtain the composite control signal j. a :

[0048]

[0049] α1 and α2 are optimization coefficients, and their values ​​range from 0 to 1, which can be determined based on the actual control effect.

[0050] The composite signal calculation module 620 outputs a composite control signal j. a The signal is fed into the inverse system of the LM neural network.

[0051] This invention first establishes Figure 2 The composite controlled object 1 shown is then established. Figure 3 The LM neural network inverse system 5 shown is trained by learning the LM neural network model 51 in the LM neural network inverse system 5. Sinusoidal signals of different phases and frequencies are selected as inputs to fully excite the LM neural network inverse system 5. 3000 sets of inputs (i) of the six-pole radial-axial hybrid magnetic bearing 2 used for flywheel energy storage are sampled in real time. A i B i C i z The training process is as follows: Two radial displacements {x,y} and one axial displacement z of the composite controlled object 1 are acquired through sensors. Then, the composite control signal ja of the radial displacements {x,y} and axial displacement z is obtained by online calculation of the control increment Δu(k). Simultaneously, the first-order and second-order difference signals of the composite control signal ja are obtained by numerical difference calculation. The signals are then normalized to form the input signal training samples for the inverse system 5 of the LM neural network, using the current step signal (i... z * i x * i y *The output training samples of the inverse system 5 of the LM neural network were used as the basis for optimization. A combination of gradient descent and Gauss-Newton methods was used to adjust the weights and thresholds of the LM neural network model 51. The objective was to minimize the mean square error between the actual value and the network's expected value. When the error objective function value decreased, the damping coefficient was reduced to make the LM algorithm closer to the Gauss-Newton method; when the error objective function value increased, the damping coefficient was increased to make the LM algorithm closer to the gradient descent method. Online rolling optimization of the LM algorithm was used; each step was a static optimization, but globally it was a dynamic optimization, thereby improving training accuracy and accelerating convergence. After 265 training steps, the error was 0.0015, meeting the set accuracy requirements. Then, a system was established... Figure 4 and Figure 5 The two dynamic prediction modules 61 and 62 shown correct the actual output and predicted output at each sampling time, and then perform new optimizations. This online optimization, changing with time, is repeated to ensure that the model predictions nearly accurately match the actual controlled process, thereby improving the training accuracy of the LM neural network model 51. Finally, the two dynamic prediction modules 61 and 62 are connected in series with the LM neural network inverse system 5, and the LM neural network inverse system 5 is connected in series with the composite controlled object 1. This constructs the LM neural network inverse decoupling controller for the hybrid magnetic bearing used in flywheel energy storage, achieving decoupling control of the six-pole radial-axial hybrid magnetic bearing 2 used in flywheel energy storage. Figure 1 As shown.

Claims

1. A hybrid magnetic bearing LM neural network inverse decoupling controller for flywheel energy storage, wherein the output of the LM neural network inverse system (5) is connected to a composite controlled object containing a six-pole radial-axial hybrid magnetic bearing for flywheel energy storage, characterized in that: The output of the two dynamic prediction modules (61, 62) is connected in series with the input of the LM neural network inverse system (5), the first dynamic prediction module (61) is composed of the first and second series branches, the second dynamic prediction module (62) is composed of the third, fourth and fifth series branches and a composite signal calculation module (620), the first and second series branches are both composed of a feedback correction module, a control increment calculation module, a prediction value calculation module and an optimal value calculation module connected in series, the third, fourth and fifth series branches are all composed of a feedback correction module, a control increment calculation module and a prediction value calculation module; the input of the first series branch is the weight ω and the reference weight ω * , and the output is the optimal initial weight ω c ; the input of the second series branch is the threshold θ and the reference threshold θ * , and the output is the optimal initial threshold θ c ; the input of the third series branch is the radial displacement x and the given value of the radial displacement x * , and the output is the final radial displacement prediction value x ac (k+1); the input of the fourth series branch is the radial displacement y and the given value of the radial displacement y * , and the output is the radial displacement prediction value y ac (k+1); the input of the fifth series branch is the axial displacement z and the given value of the axial displacement z * , and the output is the axial displacement prediction value z ac (k+1); the final radial displacement prediction value x ac (k+1), y ac (k+1) and the final axial displacement prediction value z ac (k+1) are the inputs of the composite signal calculation module (620), the composite signal calculation module (620) calculates the composite control signal j a ; the LM neural network inverse system (5) takes the optimal initial weight ω c , the optimal initial threshold θ c and the composite control signal j a as the inputs, and outputs the equivalent control current expected value i z * , i x * , i y * to the composite controlled object; The LM neural network inverse system (5) is composed of a second-order difference processor (52) and an LM neural network model (51) in series, the input of the second-order difference processor (52) is the complex control signal j a , the output is the complex control signal j a and the first-order and second-order difference signals The optimal initial weight ω c , the optimal initial threshold θ c and the complex control signal j a and the first-order and second-order difference signals are input into the LM neural network model (51), and the LM neural network model (51) outputs the equivalent control current expected value i x * , i y * , i z * ; The first series branch, wherein the feedback correction module corrects the weight ω and the reference weight ω * The weight error e1(k) is obtained by comparison, the weight error e1(k) is used to correct the weight of k+1 time to obtain the corrected weight ω ’ (k+1) of k+1 time, the weight ω ’ (k+1) of k+1 time is shifted according to the weight ω (k+1) of k+1 time is shifted according to the weight ω The control increment calculation module and the prediction value calculation module are respectively input, the control increment calculation module calculates the weight control increment Δu ω (k), the prediction value calculation module calculates the final prediction value ω(k+1) according to the weight initial prediction value (k+1) and the weight control increment Δu ω (k), the final prediction value ω(k+1) is input into the optimal value calculation module, and the optimal value calculation module outputs the weight optimal solution ω c The principles of the remaining four series branches are the same as those of the first series branch. The composite controlled object is composed of an axial direct current power amplifier (3), a radial extended current hysteresis three-phase power inverter (4) and a six-pole radial-axial hybrid magnetic bearing for flywheel energy storage, the output ends of the direct current power amplifier (3) and the extended current hysteresis three-phase power inverter (4) are connected with the input ends of the six-pole radial-axial hybrid magnetic bearing in series; The extended current hysteresis three-phase power inverter (4) is composed of a radial Clark inverter (41) and a current hysteresis three-phase power inverter (42) in series, the radial equivalent control current expected value i x * 、 y * The Clark inverter (41) is transformed into a three-phase current expected value i A * 、 B * 、 C * The current hysteresis three-phase power inverter (42) tracks the three-phase current expected value, and outputs the radial control current i A 、 B 、 C .

2. The LM neural network inverse decoupling controller of claim 1, wherein: The correction weight ω of k+1 moment ’ (k+1)=ω * +σe1(k), the weight initial prediction value The weight control increment The final prediction value The weight optimal solution ω c =α|ω(k+1)|+(1-α)ω 2 (k+1); α is an optimization factor, the value range is 0~1; σ is a feedback correction coefficient, taking 0.55; s is a shift coefficient; i=0, 1, …, k; d i is an error coefficient; a is an increment coefficient, the value range is 0~1.

3. The LM neural network inverse decoupling controller of claim 1, wherein: The composite signal calculation module (620) calculates the composite signal according to the formula The composite control signal j is calculated a ; a1, a2 are optimization coefficients, and the value range is 0-1.

4. The LM neural network inverse decoupling controller of claim 1, wherein: The LM neural network model (51) adopts LM algorithm rolling optimization, and the objective function of rolling optimization i r * is the reference output value at k+1 moment, i p is the predicted output value at k+1 moment, u(k) is the control amount at k moment, u(k-1) is the control amount at k-1 moment, and λ is a regularization parameter, and the value range is 0-1.

5. The flywheel energy storage hybrid magnetic bearing (LM) neural network inverse decoupling controller of claim 4, wherein: The LM neural network topology structure comprises an input layer, a hidden layer and an output layer, adopts a hyperbolic tangent S-shaped tansig function as a hidden layer node transfer function, and adopts a linear purelin function as an output layer node transfer function.

Images