Model Reference Based Adaptive Decoupling Control Method of RBF Neural Network for Cabin Suspension System

Abstract

The present invention discloses an RBF neural network adaptive decoupling control method for a nacelle suspension system based on model reference. A suspension model at both ends of the nacelle with two degrees of freedom of axial and pitch motions is constructed. By combining model reference adaptive control with the RBF neural network, relying on the strictly linear and non-coupled characteristics of the reference model and the infinite approximation ability of the RBF neural network, an RBF neural network adaptive controller is designed based on the model deviation, so that the suspension system model completely approximates the reference model, thereby achieving complete decoupling. The linear tracking controller takes the decoupled suspension system as a reference to complete the nacelle suspension tracking control. The present invention will greatly improve the nacelle suspension tracking, disturbance rejection and synchronization performance of both ends of the suspension, which is beneficial to the yaw alignment of the wind turbine nacelle under suspension, and at the same time has strong guiding significance for the multi-point suspension control of heavier suspended objects.

Description

Technical Field

[0001] The present invention relates to a model-reference-based RBF neural network adaptive decoupling control method for a nacelle suspension system, and in particular to a method for yaw alignment after the nacelle of a horizontal-axis wind power system is stably suspended, which solves the problem that the difference in the windward areas on the blade side and the tail fin side of the nacelle easily causes the nacelle to pitch, belonging to the field of electromagnetic suspension for wind power generation. Background Art

[0002] The horizontal-axis wind power system is a popular type in the wind power system. The traditional wind yaw device adopts a mechanical coupling yaw structure, which has problems such as large frictional power consumption, poor wind alignment accuracy, and high failure rate. Therefore, the New Energy Research Institute of Qufu Normal University proposed a wind magnetic suspension yaw system, which greatly reduces the yaw power consumption of the nacelle. Due to the harsh suspension conditions of the nacelle, the wind speed and direction are time-varying, and the masses on the blade side and the tail fin side of the nacelle are not completely the same, resulting in the nacelle being prone to pitch, seriously affecting the operation safety of the wind turbine. How to improve the axial suspension stability of the nacelle, effectively suppress the nacelle pitch, and improve the synchronization performance of the suspension system is the key to the stable suspension of the wind nacelle. Although the synchronous control method adopted in Patent 202010552436 can reduce the synchronous error at both ends of the nacelle and make the wind nacelle have a certain anti-interference ability, it does not completely solve the coupling problem between the blade side and the tail fin side of the nacelle. For the decoupling control of the suspension system, traditional decentralized PID plus cross-coupling control and linear decoupling methods require that the controlled system must be described by an accurate mathematical model, which makes it difficult for most decoupling control methods to achieve the expected suspension control effect when applied to the suspension system at both ends of the wind nacelle, seriously restricting the suspension stability of the wind nacelle and the accuracy of yaw alignment. Summary of the Invention

[0003] The object of the present invention is to overcome the deficiencies of the above-mentioned prior art, and provide an RBF neural network adaptive decoupling control method for the nacelle suspension system based on model reference, which transforms the coupled suspension system at both ends of the nacelle into single-end suspension independent control, constructs a single-end suspension linear decoupling model, and with the help of the infinite approximation ability of the RBF neural network, makes the single-end suspension system of the nacelle infinitely approximate the single-end suspension linear decoupling model, realizes the decoupling and interference suppression of the suspension system at both ends of the nacelle, and at the same time provides a suspension current reference for the suspension converter; the single-end suspension linear decoupling model adopts a third-order linear non-coupled stable system model; the single-end suspension independent control introduces an RBF neural network on the basis of model reference adaptive control, and designs an RBF neural network adaptive controller and a linear tracking controller based on model reference; the RBF neural network adaptive controller based on model reference adopts a structure with 5 hidden layer neurons, designs the adaptive law of the RBF neural network weights based on the model deviation, the first derivative of the model deviation, and the second derivative of the model deviation between the suspension system at both ends of the nacelle and the linear decoupling model, and online optimizes and adjusts the network weights; the effective reference input of the linear tracking controller is jointly composed of the suspension air gap reference and the output of the RBF neural network adaptive controller, and is easy to feedback the suspension air gap, generates the suspension air gap tracking error, the first derivative of the error, and the second derivative of the tracking error as the state feedback control input, completes the nacelle suspension tracking control, and realizes the decoupling of the suspension at both ends and the synchronous control of the suspension at both ends. It includes the following steps:

[0004] Step 1: Construct the motion equations with two degrees of freedom in the axial direction and pitch

[0005]

[0006] In the formula, ω is the pitch angular velocity, is the pitch angle, F A 、F B are the independent suspension suction forces on both sides respectively, J is the pitch moment of inertia of the nacelle, m is the mass of the wind turbine nacelle, g is the acceleration due to gravity, δ is the axial suspension air gap, f d is the axial interference of the nacelle, T s is the overturning moment of the nacelle, and r is the rotation radius of the nacelle.

[0007] Step 2: Construct the suspension force equations at both ends of the nacelle

[0008]

[0009] In the formula, μ 0 is the vacuum permeability, N is the number of turns of the suspension winding on both sides, S is the pole area, δ A 、i A are the suspension air gap and suspension current on the blade side, δ B 、i BThey are the fin side suspension air gap and suspension current.

[0010] Step 3 Conversion of the suspension dynamic model at both ends of the fan nacelle

[0011] First step, use coordinate transformation to convert the two-degree-of-freedom motion equation in Equation (1) into the front and rear side air gap motion equation as

[0012]

[0013] Second step, based on (δ 0 , i 0 ) convert Equation (3) into the linearized dynamic model at both ends of the nacelle:

[0014]

[0015] In the formula, δ 0 is the air gap between the suspension winding and the nacelle at the equilibrium point, i 0 is the suspension current flowing through the suspension winding at the equilibrium point, Δf is the high-order term after linearization.

[0016] Third step, taking the derivative of Equation (4) gives

[0017]

[0018] Fourth step, since the inner loop suspension current is controlled by the suspension converter, for the convenience of research, the suspension winding coil is modeled, that is, the suspension winding coil is replaced by a resistor and an inductor in series. According to the electromagnetic induction law and Kirchhoff's law of the circuit, the suspension winding voltage equation of a single-side nacelle is u(t) = Ri(t) + dψ(t) / dt, and the air gap magnetic field ψ can be expressed as ψ = Li = Nφ m , so the dynamic model of the suspension converter can be expressed as:

[0019]

[0020] In the formula, R and L are the equivalent resistance and equivalent inductance in the suspension converter respectively.

[0021] Fifth step, assuming that the parameters such as the resistance and inductance in the suspension converter do not change during the suspension process of the nacelle, then it can be expressed by Equation (6) as:

[0022]

[0023] Sixth step, when the suspended nacelle is in the equilibrium state, its acceleration is zero, that is then it can be obtained from Equation (4):

[0024]

[0025] Step 7. By combining Equations (7) and (8), Equation (5) can be transformed into:

[0026]

[0027] Step 8. The cross-coupling term, axial perturbation term, and pitch perturbation term in the above equation are classified as system uncertainties, which are respectively expressed as

[0028] Then Equation (9) can be simplified into the following form:

[0029]

[0030] Step 4 Selection of the single-ended suspension linear decoupling model

[0031] Step 1. Construct a linear system model as the desired model of the suspension system at both ends of the nacelle, which is expressed as:

[0032]

[0033] Step 2. It can be seen from Equation (11) that this desired model is a completely linear and non-coupled model, and its differential equation can be described as:

[0034]

[0035] where A m , B m are expected constants; r is the reference air gap input, and the state variables of the desired model are the same as those of the suspension system model, that is, X m = X.

[0036] Step 3. To ensure good tracking performance, take ξ = 0.8, ω n = 70, then the coefficient matrix in Equation (12) is:

[0037]

[0038] At the same time, the dominant poles s 0 = -60 of this desired model can be obtained, and there are also poles s 1 = -70 + 2.48×10 -8 i, s 2 = -70 - 2.48×10 -8 i. Obviously, the three poles of this desired model are all distributed in the left half-plane and there is no overshoot, which proves that the selected linear system is asymptotically stable.

[0039] Step 5 Design of the model-reference RBF neural network adaptive controller and linear tracking controller

[0040] In the first step, when designing the controller, taking side A as an example, let the state variable be u as the control input. Then, the state-space equation of single-ended suspension independent control can be written as:

[0041]

[0042] In the formula, K is the linear controller parameter matrix, which can be obtained from the adaptive decoupling matching condition of the ideal model reference RBF neural network.

[0043] In the second step, it can be seen from the selection of the single-ended suspension linear decoupling model in step 4 that the differential equation of the desired model of the nacelle single-ended suspension system is:

[0044]

[0045] In the third step, an RBF neural network adaptive controller based on model reference is used to approximate the nacelle single-ended suspension system model to the desired model, so that the output of the RBF neural network can adjust the reference air gap and the feedback air gap in real time. At this time, the RBF neural network approximates the composite uncertain disturbance term of the suspension system to Φ * , there exists an ideal neural network weight vector θ * , such that

[0046] Φ * = θ *T h(x) + ε (16)

[0047] In the formula, h(x) is the radial basis function vector, and ε is the neural network approximation error, satisfying |ε| ≤ ε 0 .

[0048] In the fourth step, combining formula (16), the state-space description of the unilateral suspension system model is transformed into:

[0049]

[0050] In the fifth step, taking the control objective requires designing a control law:

[0051] u = K(X ref - X + Φ * ) (18)

[0052] In the formula, K is the linear controller feedback gain.

[0053] In the sixth step, substituting formula (18) into formula (17) gives:

[0054]

[0055] Step 7: Compare Equation (19) with the expected reference dynamic Equation (15). For the controller in the form of Equation (18) to exist, the ideal control gain must satisfy the matching condition

[0056] A - BK = A m

[0057] BKX ref = B m r (20)

[0058] Step 8: Assume that these matching conditions hold. Using Equation (20), a closed-loop system identical to the reference model can be obtained. Therefore, for any bounded reference input signal, the fixed-gain controller Equation (20) guarantees global uniform asymptotic tracking performance. The value of the parameter matrix K of the linear tracking controller in this chapter can be obtained from Equation (20) as BK = A - A m . Where A and B are defined by Equation (14), and A m is defined by Equation (15), then:

[0059]

[0060] In the formula,

[0061] Step 9: The parameter matrix K of the linear tracking controller can be obtained from Equation (21) as:

[0062]

[0063] Step 10: Define the difference between the desired output air gap and the output air gap of the two-point suspension system as the state tracking error. Then the state tracking error is E(t) = X m (t) - X(t). This state tracking error E(t) is used as the input of the RBF neural network adaptive controller, and the control objective is to make the state tracking error E(t) → 0 when t → ∞. Let be the estimate of the neural network weight θ * , then the output of the RBF neural network is:

[0064]

[0065] Step 11: The control law of the suspension system can be written as:

[0066]

[0067] Step 12: Combining Equation (15), Equation (17), Equation (20), and Equation (24), the closed-loop dynamics of E(t) = X m (t) - X(t) can be obtained:

[0068]

[0069] Step 13: Take Then

[0070]

[0071] Step 14: Construct the Lyapunov function of the closed-loop system as:

[0072]

[0073] where α is a positive constant, matrix P is a symmetric positive definite matrix and satisfies A m T P + PA m = -Q.

[0074] Step 15: Differentiate Equation (27) to obtain:

[0075]

[0076] Step 16: Take the weight adaptation law as

[0077]

[0078] Step 17: Combining Equation (31), Equation (30) is transformed into:

[0079]

[0080] Since the RBF neural network can be designed to make its approximation error ε small enough, so that

[0081] Based on the model reference RBF neural network adaptive decoupling control method for the cabin suspension system, the above five working steps are carried out. The coordinate transformation equation in Step 3 is:

[0082]

[0083] where δ A is the suspension air gap on the blade side, δ B is the suspension air gap on the tail wing side, and r is the radius of the suspended cabin.

[0084] The conversion method is to take the second derivative of the coordinate transformation equation (20) as

[0085]

[0086] The beneficial effects of the present invention are:

[0087] 1) The proposed RBF neural network adaptive controller serves as the adaptive mechanism for the suspension model at both ends of the nacelle and the reference model. The design of this controller does not rely on the exact mathematical model of the suspension system. By leveraging the strictly linear uncoupled characteristics of the reference model and the infinite approximation ability of the RBF neural network, the suspension system model is made to fully approximate the reference model, thereby achieving complete decoupling, effectively suppressing the pitching moment, and greatly reducing the synchronization error at both ends of the nacelle.

[0088] 2) The proposed linear tracking controller takes the decoupled suspension system as a reference to complete the suspension tracking control, greatly enhancing the suspension stability of the wind turbine nacelle. Brief Description of the Drawings

[0089] Figure 1 It is a schematic diagram of the nacelle suspension structure of the horizontal axis wind turbine yaw system for the model-reference-based RBF neural network adaptive decoupling control method of the nacelle suspension system of the present invention.

[0090] Figure 2 It is a control structure diagram of the nacelle suspension of the horizontal axis wind turbine yaw system for the model-reference-based RBF neural network adaptive decoupling control method of the nacelle suspension system of the present invention.

[0091] Figure 3 It is an experimental diagram of the variable air-gap tracking of the nacelle air gap under the control of the present invention's model-reference-based RBF neural network adaptive decoupling control method of the nacelle suspension system and PID control.

[0092] Figure 4 It is an experimental diagram of applying an axial interference force to the nacelle under PID control.

[0093] Figure 5 It is an experimental diagram of applying an axial interference force to the nacelle under the control of the present invention's model-reference-based RBF neural network adaptive decoupling control method of the nacelle suspension system.

[0094] Figure 6 It is an experimental diagram of applying a pitching interference force to the nacelle under PID control.

[0095] Figure 7 It is an experimental diagram of applying a pitching interference force to the nacelle under the control of the present invention's model-reference-based RBF neural network adaptive decoupling control method of the nacelle suspension system.

[0096] In the figure: 1 - wind turbine blade, 2 - wind turbine nacelle, 3 - yaw stator, 4 - front side winding, 5 - rear side winding, 6 - front side air gap sensor, 7 - rear side air gap sensor, 8 - tower, 9, 10 - linear decoupling model, 11 - blade side RBF neural network adaptive controller, 12 - blade side neural network weight adaptive law, 13, 16 - Riccati equation, 14 - tail side RBFNMN adaptive controller, 15 - tail side neural network weight adaptive law, 17 - blade side linear tracking controller, 18 - tail side linear tracking controller, 19 - blade side current tracking controller, 20 - blade side suspension converter, 21 - tail side current tracking controller, 22 - tail side suspension converter, 23 - nacelle two - end suspension model. Detailed implementation manner

[0097] A model - reference - based RBF neural network adaptive decoupling control method for the nacelle suspension system transforms the coupled suspension system at both ends of the nacelle into a single - end suspension independent control, constructs a single - end suspension linear decoupling model (9, 10), and by virtue of the infinite approximation ability of the RBF neural network, makes the single - end suspension system of the nacelle infinitely approximate the single - end suspension linear decoupling model, realizing the decoupling and interference suppression of the suspension system at both ends of the nacelle, and at the same time providing a suspension current reference for the suspension converters (19, 20, 21, 22); the single - end suspension linear decoupling model (9, 10) adopts a third - order linear uncoupled stable system model; the single - end suspension independent control (9, 11, 12, 13, 17 or 10, 14, 15, 16, 18) introduces an RBF neural network on the basis of model - reference adaptive control, and designs a model - reference - based RBF neural network adaptive controller and a linear tracking controller; the model - reference - based RBF neural network adaptive controller (11, 12, 13, 14, 15, 16) adopts a structure with 5 hidden - layer neurons, designs an adaptive law for the weights of the RBF neural network based on the model deviation, the first - order derivative of the model deviation, and the second - order derivative of the model deviation between the suspension system at both ends of the nacelle and the linear decoupling model, and online optimizes and adjusts the network weights; the effective reference input of the linear tracking controller (17, 18) is jointly composed of the suspension air - gap reference and the output of the RBF neural network adaptive controller, and is easy to feedback the suspension air - gap, generating the suspension air - gap tracking error, the first - order derivative of the error, and the second - order derivative of the tracking error as the state - feedback control input to complete the nacelle suspension tracking control, realizing the decoupling of the two - end suspension and the synchronous control of the two - end suspension. It includes the following steps.

[0098] Step 1: Construct the motion equations with two degrees of freedom in the axial and pitch directions

[0099]

[0100] In the formula, ω is the pitch angular velocity, is the pitch angle, FA , F B are respectively the independent suspension suction forces on both sides, J is the pitching moment of inertia of the nacelle, m is the mass of the wind turbine nacelle, g is the acceleration due to gravity, δ is the axial suspension air gap, f d is the axial interference of the nacelle, T s is the overturning moment of the nacelle, and r is the rotation radius of the nacelle.

[0101] Step 2: Construct the suspension force equations at both ends of the nacelle

[0102]

[0103] In the formula, μ 0 is the vacuum permeability, N is the number of turns of the suspension winding on both sides, S is the pole area, δ A , i A are the suspension air gap and suspension current on the blade side, δ B , i B are the suspension air gap and suspension current on the tail wing side.

[0104] Step 3: Transformation of the suspension dynamic model at both ends of the wind turbine nacelle

[0105] First step, use coordinate transformation to transform the two-degree-of-freedom motion equation in Equation (1) into the front and rear side air gap motion equations as

[0106]

[0107] Second step, based on (δ 0 , i 0 ) transform Equation (3) into the linearized dynamic model at both ends of the nacelle:

[0108]

[0109] In the formula, δ 0 is the air gap between the suspension winding and the nacelle at the equilibrium point, i 0 is the suspension current flowing through the suspension winding at the equilibrium point, Δf is the high-order term after linearization.

[0110] Third step, take the derivative of Equation (4) to obtain

[0111]

[0112] Fourth step, since the inner loop suspension current is controlled by the suspension converter, for the convenience of research, the suspension winding coil is modeled, that is, the suspension winding coil is replaced by a resistor and an inductor in series. According to the electromagnetic induction law and Kirchhoff's law of the circuit, the suspension winding voltage equation of a single-sided nacelle is u(t) = Ri(t) + dψ(t) / dt, and the air gap magnetic field ψ can be expressed as ψ = Li = Nφ m, so the dynamic model of the suspension converter can be expressed as:

[0113]

[0114] In the formula, R and L are the equivalent resistance and equivalent inductance in the suspension converter respectively.

[0115] Step 5: Assume that the parameters such as resistance and inductance in the suspension converter do not change during the suspension process of the nacelle. Then, i can be expressed by Equation (6) as:

[0116]

[0117] Step 6: When the suspended nacelle is in a balanced state, its acceleration is zero, that is Then it can be obtained from Equation (4):

[0118]

[0119] Step 7: Combining Equations (7) and (8), Equation (5) can be transformed into:

[0120]

[0121] Step 8: The cross-coupling term, axial disturbance term, and pitch disturbance term in the above formula are classified as system uncertainty terms, which are respectively expressed as

[0122]

[0123] Then Equation (9) can be simplified into the following form:

[0124]

[0125] Step 4 Selection of the single-ended suspension linear decoupling model

[0126] Step 1: Construct a linear system model as the desired model of the suspension system at both ends of the nacelle, which is expressed as:

[0127]

[0128] Step 2: It can be seen from Equation (11) that the desired model is a completely linear and non-coupled model, and its differential equation can be described as:

[0129]

[0130] In the formula, A m , B m are expected constants; r is the reference air gap input, and the state variables of the desired model are the same as those of the suspension system model, that is, X m = X.

[0131] In the third step, to ensure good tracking performance, take ξ = 0.8 and ω n = 70, then the coefficient matrix in Equation (12) is:

[0132]

[0133] Meanwhile, the dominant pole s 0 of the desired model can be obtained as s = -60, and there are also poles s 1 = -70 + 2.48×10 -8 i, s 2 = -70 - 2.48×10 -8 i. Obviously, the three poles of the desired model are all distributed in the left half-plane and there is no overshoot, which proves that the selected linear system is asymptotically stable.

[0134] Step 5 Design of the model-reference RBF neural network adaptive controller and the linear tracking controller

[0135] In the first step, when designing the controller, take side A as an example. Let the state variable u be the control input, then the state-space equation of the single-ended suspension independent control can be written as:

[0136]

[0137] where K is the linear controller parameter matrix, which can be obtained from the ideal model-reference RBF neural network adaptive decoupling matching condition.

[0138] In the second step, from the selection of the single-ended suspension linear decoupling model in Step 4, the differential equation of the desired model of the nacelle single-ended suspension system is:

[0139]

[0140] In the third step, adopt the model-reference RBF neural network adaptive controller to approximate the nacelle single-ended suspension system model to the desired model, so that the RBF neural network output can adjust the reference air gap and the feedback air gap in real time. At this time, the RBF neural network approximates the composite uncertain disturbance term of the suspension system to Φ * , there exists an ideal neural network weight vector θ * , such that

[0141] Φ * = θ *T h(x) + ε (16)

[0142] where h(x) is the radial basis function vector and ε is the neural network approximation error, satisfying |ε| ≤ ε 0 .

[0143] Step 4. Combining with Equation (16), the state - space description of the single - side suspension system model is transformed into:

[0144]

[0145] Step 5. Taking The control objective requires designing a control law:

[0146] u = K(X ref - X + Φ * ) (18)

[0147] where K is the feedback gain of the linear controller.

[0148] Step 6. Substituting Equation (18) into Equation (17), we can get:

[0149]

[0150] Step 7. Comparing Equation (19) with the expected reference dynamics Equation (15), for the controller in the form of Equation (18) to exist, the ideal control gain must satisfy the matching condition

[0151]

[0152] Step 8. Assuming that these matching conditions hold, using Equation (20) we can obtain a closed - loop system identical to the reference model. Therefore, for any bounded reference input signal, the fixed - gain controller Equation (20) guarantees global uniform asymptotic tracking performance. From Equation (20), the value of the parameter matrix K of the linear tracking controller in this chapter can be obtained. There is BK = A - A m . Where A and B are defined by Equation (14), and A m is defined by Equation (15), then:

[0153]

[0154] where

[0155] Step 9. From Equation (21), the parameter matrix K of the linear tracking controller is:

[0156]

[0157] Step 10. Define the difference between the desired output air gap and the output air gap of the two - point suspension system as the state tracking error. Then the state tracking error is E(t)=X m (t)-X(t). This state tracking error E(t) is used as the input of the RBF neural network adaptive controller. The control objective is that when t→∞, the state tracking error E(t)→0. Let be the neural network weight θ *For the estimation, the output of the RBF neural network is:

[0158]

[0159] Step 11, the suspension system control law can be written as:

[0160]

[0161] Step 12, combining equations (15), (17), (20) and (24), we can obtain the closed-loop dynamics of E(t) = X m (t) - X(t):

[0162]

[0163] Step 13, take Then

[0164]

[0165] Step 14, construct the Lyapunov function of the closed-loop system as:

[0166]

[0167] where α is a positive constant, the matrix P is a symmetric positive definite matrix and satisfies A m T P + PA m = -Q.

[0168] Step 15, taking the derivative of equation (27) gives:

[0169]

[0170] Step 16, the weight adaptation law takes

[0171]

[0172] Step 17, combining equation (31), equation (30) is transformed into:

[0173]

[0174] Since the RBF neural network can be designed to make its approximation error ε small enough, so that

[0175] Based on the RBF neural network adaptive decoupling control method of the engine room suspension system with model reference, the above five working steps are carried out. The coordinate transformation equation in step 3 is:

[0176]

[0177] where δ A is the suspension air gap on the blade side, and δ B is the suspension air gap on the fin side, and r is the radius of the suspended nacelle.

[0178] The conversion method is to take the second derivative of the coordinate conversion equation (20) to obtain

[0179]

[0180] The present invention will be further described in detail below with reference to the accompanying drawings and examples.

[0181] The nacelle suspension parameters of the wind power magnetic suspension yaw system are shown in Table 1. The nacelle suspension weight is 484 kg, the total number of turns of the suspension winding is 930 turns, the number of turns of the front and rear side windings is 465 turns, the nacelle rotation radius is 360 mm, the power of each of the two suspension converters is 1 kW, and the suspension air gap sensor uses an eddy current displacement sensor with an accuracy of 0.27 v / mm. The following three examples are carried out respectively, namely the variable air gap tracking experiment, the axial interference force application experiment, and the anti-pitching moment experiment, to illustrate the effective effect of the present invention.

[0182] Table 1 Nacelle suspension system parameters of the wind power magnetic suspension yaw system

[0183]

[0184] Example 1 Variable air gap tracking experiment, as Figure 3 shown. At t = 0 s, the nacelle starts to suspend, and the initial suspension reference height is set to 13 mm. At t = 4 s, the suspension height reference value is switched to 15 mm. At t = 15 s, the suspension height reference value is switched back to the initial suspension reference value again. The comparison table of the variable air gap tracking performance is shown in Table 2. When 0 s ≤ t < 4 s, the nacelle is in the suspension start-up stage. At this time, the suspension start-up time and the steady-state fluctuation value of the nacelle are mainly examined. The start-up time controlled by the present invention is 0.2 s, and the steady-state fluctuation value is 0.0099 mm, which is smaller than the steady-state suspension height fluctuation of the PID controller; when 4 s ≤ t < 15 s, in the nacelle suspension height switching stage, the suspension switching time and the overshoot of the nacelle are mainly examined. The reference switching stable time controlled by the present invention is 0.2 s, and there is no overshoot. It can be seen that the control of the present invention not only significantly improves the dynamic performance of the suspension system, but also greatly reduces the overshoot and the steady-state error, achieving a good decoupling control effect on the suspension systems at both ends of the nacelle.

[0185] Table 2 Variable air gap tracking performance

[0186]

[0187] Example 2 Axial interference force application experiment, as Figure 4 and Figure 5As shown in the figure, the reference value of the initial suspension height of the nacelle is set to 13 mm. At t = 4 s, an axial downward pressure disturbance of 1000 N is applied to one side of the suspension system to simulate the axial interference of the external wind on the nacelle. At t = 15 s, the axial downward pressure disturbance is withdrawn, and the maximum drop value, drop recovery time, maximum recovery value after the disturbance is withdrawn, and return to stable time of the nacelle after being disturbed are observed. The comparison of the anti-axial disturbance performance of one-side nacelle is shown in Table 3. It can be seen that when the suspended nacelle is subjected to one-side disturbance at t = 4 s and t = 15 s, the maximum drop value controlled by the present invention is 0.0194 mm, and the suspended nacelle can return to the initial suspension height after 0.1 s, which is smaller than the maximum drop value of the PID control and has a shorter drop return time, effectively improving the response speed of the suspension system and enabling the suspended nacelle to have better anti-axial disturbance ability.

[0188] Table 3 Comparison of Axial Interference Force Application Performance

[0189]

[0190] Example 3 Anti-Pitching Moment Experiment, as Figure 6 and Figure 7 shown in the figure, analyze the synchronous performance of the suspension control on both sides of the nacelle; set the reference value of the initial suspension height of the nacelle to 13 mm, apply a pitching moment disturbance of 1000 N to one side of the suspension system at t = 4 s to simulate the external side wind interference, and withdraw the pitching moment disturbance at t = 15 s, and observe the maximum drop value, drop recovery time, maximum recovery value after the disturbance is withdrawn, and return to stable time of the nacelle after being disturbed. The comparison of the anti-pitching disturbance performance of one-side nacelle is shown in Table 4. It can be seen that when the suspended nacelle is subjected to one-side disturbance at t = 4 s and t = 15 s, the traditional controller directly loses control, and at this time, the suspended nacelle tilts and cannot return to the initial suspension height smoothly, while the maximum drop controlled by the present invention is 0.044 mm, and the suspended nacelle can return to the initial suspension height after 0.2 s. It can be seen that when the present invention is used for control, the drop value is smaller and the drop return time is shorter, effectively improving the response speed of the suspension system and enabling the suspended nacelle to have better anti-disturbance performance, and can quickly suppress the difference in the air gaps on both sides of the suspended nacelle.

[0191] Table 4 Comparison Table of One-Side Disturbance Performance

[0192]

Claims

1. Model-reference based RBF neural network adaptive decoupling control method for nacelle suspension system, characterized in that: The coupled suspension system at both ends of the nacelle is transformed into single-end suspension independent control, a single-end suspension linear decoupling model is constructed, and with the infinite approximation ability of the RBF neural network, the single-end suspension system of the nacelle is infinitely approximated to the single-end suspension linear decoupling model to achieve decoupling and interference suppression of the suspension system at both ends of the nacelle, and at the same time provide a suspension current reference for the suspension converter; The single-end suspension linear decoupling model adopts a third-order linear non-coupled stable system model; the single-end suspension independent control introduces an RBF neural network on the basis of model-reference adaptive control, and designs an RBF neural network adaptive controller and a linear tracking controller based on model reference; the RBF neural network adaptive controller based on model reference adopts a structure with 5 hidden layer neurons, and designs an adaptive law for the weights of the RBF neural network based on the model deviation, the first derivative of the model deviation, and the second derivative of the model deviation between the suspension system at both ends of the nacelle and the linear decoupling model, and online optimizes and adjusts the network weights; The effective reference input of the linear tracking controller consists of the suspension air gap reference and the output of the RBF neural network adaptive controller, and is easy to feedback the suspension air gap, generate the suspension air gap tracking error, the first derivative of the error, and the second derivative of the tracking error as the state feedback control input to complete the nacelle suspension tracking control and achieve decoupling of the suspension at both ends and synchronous control of the suspension at both ends.

2. The model-reference based RBF neural network adaptive decoupling control method for nacelle suspension system according to claim 1, characterized in that, it includes the following steps: Step 1: Construct the motion equations with axial and pitch two degrees of freedom where ω is the pitch angular velocity, is the pitch angle, F A and F B are the independent suspension suction forces on both sides, J is the pitch moment of inertia of the nacelle, m is the mass of the wind turbine nacelle, g is the acceleration due to gravity, δ is the axial suspension air gap, f d is the axial interference of the nacelle, T s is the tipping moment of the nacelle, and r is the rotation radius of the nacelle; Step 2: Construct the suspension force equations at both ends of the nacelle where μ 0 is the permeability of vacuum, N is the number of turns of the suspension windings on both sides, S is the pole area, δ A , i A are the suspension air gaps and suspension currents on the blade side, and δ B , i B are the suspension air gaps and suspension currents on the fin side; Step 3: Transformation of the suspension dynamic model at both ends of the fan nacelle First step, use coordinate transformation to transform the two-degree-of-freedom motion equation in Equation (1) into the front and rear side air gap motion equations Step 2: Based on (δ 0 , i 0 ), transform Equation (3) into the linearized dynamic model at both ends of the engine room: where, δ 0 is the air gap between the suspended winding and the nacelle at the equilibrium point, i 0 is the suspension current flowing through the suspended winding at the equilibrium point, Δf is the high-order term after linearization; Third step, taking the derivative of Equation (4) gives Fourth step, since the inner loop suspension current is controlled by the suspension converter, for the convenience of research, the suspension winding coil is modeled, that is, the suspension winding coil is replaced by a resistor and an inductor in series. According to the electromagnetic induction law and Kirchhoff's law of the circuit, the suspension winding voltage equation of a single-side nacelle is u(t)=Ri(t)+dψ(t) / dt, and since the air gap magnetic field ψ can be expressed as ψ = Li, the dynamic model of the suspension converter can be expressed as: In the formula, R and L are the equivalent resistance and equivalent inductance in the suspension converter respectively; Fifth step, assuming that the parameters such as the resistance and inductance in the suspension converter do not change during the nacelle suspension process, then i can be expressed from Equation (6) as: Step 6: When the hovering cabin is in a balanced state, its acceleration is zero, i.e., It can be obtained from Equation (4): Seventh step, combining Equation (7) and (8), Equation (5) can be transformed into: In the eighth step, the cross-coupling term, axial perturbation term, and pitch perturbation term in the above equation are classified as system uncertainties and are respectively expressed as Then equation (9) can be simplified to the following form: Step 4: Selection of the single-end suspension linear decoupling model First step, construct a linear system model as the desired model of the suspension system at both ends of the nacelle, expressed as: Second step, it can be seen from Equation (11) that this desired model is a completely linear non-coupled model, and its differential equation can be described as: where A m , B m are expected constants, r is the reference air-gap input, and it is expected that the model state variables are consistent with those of the suspension system, i.e., X m = X; Step 3: To ensure good tracking performance, take ξ = 0.8 and ω n = 70. Then the coefficient matrix in Equation (12) is as follows: Meanwhile, the dominant pole s of the desired model can be obtained 0 = -60, and there are also poles s 1 = -70 + 2.48×10 -8 i, s 2 = -70 - 2.48×10 -8 i. Obviously, all three poles of the desired model are distributed in the left half-plane and there is no overshoot, proving that the selected linear system is asymptotically stable; Step 5: Design of the RBF neural network adaptive controller and linear tracking controller based on model reference In the first step, when designing the controller, taking side A as an example, let the state variable u be the control input, then the state-space equation of single-ended floating independent control can be written as: where \(K\) is the parameter matrix of the linear controller, which can be obtained from the adaptive decoupling matching conditions of the ideal model reference RBF neural network; Step 2. As can be seen from the selection of the single-ended suspension linear decoupling model in Step 4, the differential equation of the desired model of the nacelle single-ended suspension system is: In the third step, a model-reference-based RBF neural network adaptive controller is used to approximate the single-ended suspension system model of the nacelle to the desired model, so that the RBF neural network output can adjust the reference air gap and the feedback air gap in real time. At this time, the RBF neural network approximates the composite uncertain disturbance term of the suspension system to Φ * , there exists an ideal neural network weight vector θ * , such that Φ * = θ *T h(x) + ε (16) where \(h(x)\) is the radial basis function vector, and \(\varepsilon\) is the neural network approximation error, satisfying \(|\varepsilon|\leq\varepsilon\) 0 ; Step 4. Combining Equation (16), the state-space description of the unilateral suspension system model is transformed into: Step 5, take The control objective requires designing a control law: u = K(X ref - X + Φ * ) (18) where \(K\) is the feedback gain of the linear controller; Step 6. Substituting Equation (18) into Equation (17) gives: Step 7. Comparing Equation (19) with the expected reference dynamic Equation (15), for the controller in the form of Equation (18) to exist, the ideal control gain must satisfy the matching condition Step 8. Assume that these matching conditions hold. Using Equation (20), the same closed-loop system as the reference model can be obtained. Therefore, for any bounded reference input signal, the fixed-gain controller Equation (20) guarantees the globally uniformly asymptotically tracking performance. The value of the parameter matrix K of the linear tracking controller in this chapter can be obtained from Equation (20), where BK = A - A m , where A and B are defined by Equation (14), and A m is defined by Equation (15). Then: In the formula, Step 9. From Equation (21), the parameter matrix \(K\) of the linear tracking controller is: Step 10: Define the difference between the expected output air gap and the output air gap of the two-point suspension system as the state tracking error. Then the state tracking error is E(t) = X m (t) - X(t). This state tracking error E(t) is used as the input of the RBF neural network adaptive controller. The control objective is to make the state tracking error E(t) → 0 when t → ∞. Let be the estimate of the neural network weight θ * . Then the output of the RBF neural network is: Step 11. The control law of the suspension system can be written as: Step 12, combining Equation (15), Equation (17), Equation (20) and Equation (24), we can obtain E(t) = X m (t) - Closed-loop dynamics of X(t): Step 13, take Then Step 14. The Lyapunov function of the closed-loop system is constructed as: where α is a positive constant, the matrix P is a symmetric positive definite matrix and satisfies A m T P + PA m = -Q; Step 15. Differentiating Equation (27) gives: Step 16. The weight adaptive law is taken as Step 17. Combining Equation (29), Equation (28) is transformed into: Since By designing an RBF neural network, its approximation error ε can be made small enough so that 3. The RBF neural network adaptive decoupling control method for the nacelle suspension system based on model reference according to claim 2, characterized in that: the coordinate transformation equation in Step 3 is where δ A is the suspension air gap on the blade side, and δ B is the suspension air gap on the fin side, and r is the radius of the suspended nacelle; The transformation method is to take the second derivative of the coordinate transformation equation (31) as

Images