Dynamic System Model Predictive Control Method Based on RBF Neural Network

Abstract

The invention relates to a dynamic system model predictive control method based on RBF neural network, which belongs to the field of control technology of dynamic system. The invention solves the problems of poor control accuracy and slow dynamic response speed in the existing dynamic system model predictive control method. In the invention, a dynamic equation between the control input and the state variable of the dynamic system is established according to the dynamic parameters of the dynamic system, and the dynamic equation is converted into a nominal form based on the virtual control variable by using a virtual control variable, and then converted into a state space form. The cost function of the model predictive control algorithm is designed by using the dynamic equation in the state space form, and the cost function is optimized and solved. The disturbance approximation model based on the neural network is adopted, and the unmodeled dynamics and external disturbances of the dynamic system are obtained by inverse solution. The unmodeled dynamics and external disturbances are used to obtain the system output after disturbance compensation, and then the control signal of the dynamic system is obtained. The invention is suitable for the dynamic system control of quadrotor aircraft.

Description

Technical Field

[0001] The invention belongs to the technical field of dynamic system control. Background Art

[0002] Dynamic systems such as aircraft, marine vehicles, and land-based mobile machines are widely used in both military and civilian applications, such as military reconnaissance, film and television production, agricultural monitoring, and express delivery. Rotary-wing drones (UAVs) are multifunctional flying robots capable of performing tasks such as target surveillance and urban aerial photography, enabling them to accomplish a variety of tasks at a fraction of the cost. Quadcopters, a typical example of rotary-wing UAVs, have been widely researched due to their simple airframe structure, excellent maneuverability, and efficient mission execution. While quadcopters are now widely used in the market, achieving stable flight under disturbances is technically challenging.

[0003] Model predictive control (MPC) is an advanced control strategy with the inherent ability to solve constraints, effectively addressing coupled nonlinearities in dynamic systems. Its optimization capabilities also enable effective control even when systems contain coupled nonlinearities, dead zones, time lags, saturation, and other nonlinearities. However, classic MPC designs rely on precise and detailed mathematical models of the controlled object. This means that control reliability depends largely on the accuracy of the established dynamic system model. When the system model contains unmodeled dynamics or significant external interference, making it difficult to obtain an accurate dynamic model, the control effectiveness of classic MPC is compromised, making the optimization process complex and difficult. Summary of the Invention

[0004] The purpose of the present invention is to solve the problems of poor control accuracy and slow dynamic response speed in the existing dynamic system model predictive control method, and proposes a dynamic system model predictive control method based on RBF neural network.

[0005] The dynamic system model predictive control method based on RBF neural network of the present invention comprises:

[0006] Step 1: Based on the dynamic parameters of the dynamic system, establish the dynamic equation between the control input and the state variable of the dynamic system; the dynamic parameters of the dynamic system include: moment of inertia and motor coefficient;

[0007] Step 2: Design virtual control variables according to the number of moments of inertia, and use the virtual control variables to transform the dynamic equation into a nominal form based on the virtual control variables to obtain the nominal form of the dynamic equation;

[0008] Step 3: Convert the nominal form of the dynamic equation into a state space form, and use the state space form of the dynamic equation to design a cost function of the model predictive control algorithm;

[0009] Step 4: Optimize and solve the cost function according to the control constraints of the dynamic system to obtain a control output, and use the first element of the control output as the input of the dynamic system to obtain an output signal of the dynamic system;

[0010] Step 5: Using the nominal form of the dynamic equations, the virtual control variables, and the output signal of the dynamic system obtained in step 4, a disturbance approximation model based on a neural network is used to inversely solve the unmodeled dynamics and external disturbances of the dynamic system;

[0011] Step 6: Use the unmodeled dynamics, external disturbances and virtual control variables obtained in step 5 as dynamic system inputs to obtain the system output signal after disturbance compensation. Use the control allocation matrix to transform the system output signal after disturbance compensation to obtain the control signal of the dynamic system.

[0012] Furthermore, in the present invention, in step 1, the dynamic system is a quadrotor dynamic system.

[0013] Furthermore, in the present invention, in step 1, the dynamic equation between the dynamic system control input and the state variable is:

[0014]

[0015]

[0016]

[0017] Among them, V f 、V b 、V r 、V l are the control inputs of the four motors up, down, right and left of the quadrotor aircraft, J p 、J r 、J y K is the moment of inertia of the quadrotor aircraft about the pitch axis, roll axis, and yaw axis; f is the lift coefficient, K t is the torque coefficient, d p d r d y They are the unmodeled dynamics and external disturbances of the quadrotor aircraft in three degrees of freedom: pitch angle, roll angle, and yaw angle. is θ p The second derivative of is θ r The second derivative of is θy The second derivative of θ p is the pitch angle. When the longitudinal axis of the quadrotor is above the horizontal plane, θ p is a positive value, θ r is the roll angle, looking forward from the tail, when the quadrotor tilts to the right, θ r The angle is positive, θ y is the yaw angle, θ y When viewed from the angular plane, if the projection from the OX axis to the longitudinal axis of the quadrotor on the horizontal plane rotates counterclockwise, then θ y The angle is positive.

[0018] Furthermore, in the present invention, in step 2, the kinetic equation in nominal form is:

[0019]

[0020] y=Cx+Du

[0021] in, is the derivative of the state variable x of the dynamic system, is θ y The first derivative of is θ p The first derivative of is θ y The first derivative of , y is the output of the dynamic system, y T =[θ y θ p θ r ]; L is the wheelbase between the four-rotor propellers; u is the virtual control variable, u=[v p v r v y ] T v p , v r , v y Corresponding to the three attitude channels θ of the quadrotor aircraft: pitch axis, roll axis, and yaw axis p ,θ r ,θ y dummy control variables for ;

[0022]

[0023]

[0024] Furthermore, in the present invention, in step 3, the cost function of the model predictive control algorithm is:

[0025]

[0026] x(i|k) represents the forward prediction of the i-th step at time k, u(i-1|k) represents the i-1-th step control executed at time k, Q and R are N×N weighted matrices respectively, N is the prediction step size of the system, and J(k) is the cost function at time k.

[0027] Furthermore, in the present invention, in step 4, the control constraints of the dynamic system include: system output condition constraints, system state constraints and control force constraints.

[0028] Furthermore, in the present invention, in step 4, the process of optimizing and solving the cost function is:

[0029]

[0030]

[0031] x(0|k)=x(k)

[0032]

[0033]

[0034] N is the prediction step size of the system, To control the constraint set, is the state constraint set, x(k) is the state of the system at the kth moment, and x(0|k) is the predicted value at the initial moment.

[0035] Furthermore, in the present invention, in step 5, the method for obtaining the system disturbance by inversely solving the disturbance approximation model based on a neural network using the nominal form of the dynamic equation, the virtual control variables and the dynamic system output obtained in step 4 is:

[0036] The dynamic equation between the control input and state variables of the dynamic system is rewritten as:

[0037]

[0038]

[0039]

[0040] Transformed into:

[0041]

[0042]

[0043]

[0044] Among them, d p d r and dy They are the unmodeled dynamics and external disturbances on the three degrees of freedom of pitch, roll and yaw.

[0045] Furthermore, in the present invention, in step 5, the perturbation approximation model based on the neural network is:

[0046]

[0047]

[0048]

[0049] Among them, E1 is the disturbance estimation error of the quadrotor roll channel, E2 is the disturbance estimation error of the quadrotor pitch channel, and E3 is the disturbance estimation error of the quadrotor yaw channel. is the neural network weight of the quadrotor roll channel, is the neural network weight of the quadrotor pitch channel, is the neural network weight of the quadrotor yaw angle channel, H1(X1) is the vector form of the Gaussian kernel function of the quadrotor roll channel, H2(X2) is the vector form of the Gaussian kernel function of the quadrotor pitch channel, and H3(X3) is the vector form of the Gaussian kernel function of the quadrotor yaw angle channel.

[0050] Furthermore, in the present invention, in step 6, the system output after disturbance compensation is:

[0051]

[0052] Among them, u is the total control output of the dynamic system, u * is the model predictive control output for the system nominal model.

[0053] Furthermore, in the present invention, in step 6, the control signal of the dynamic system obtained by converting the system output after the disturbance compensation through the control allocation matrix is:

[0054]

[0055] The present invention addresses the problem of difficulty in obtaining an accurate mathematical model for a dynamic system under a disturbed state and proposes a model predictive control method based on RBF disturbance compensation. External disturbances and unmodeled dynamics in the dynamic system are approximated by RBF and combined with the model predictive control method, so that the dynamic system controller under the influence of disturbances has good control accuracy and dynamic performance. By designing a model predictive control algorithm with RBF neural network disturbance compensation, the problem of dynamic system control under a disturbed state is solved. The system has good control accuracy and fast dynamic response capability. External disturbances and unmodeled dynamics of the system are estimated and compensated by neural network, solving the problem of model predictive controller design under the condition of inaccurate model. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 is a flow chart of the method of the present invention;

[0057] Figure 2 It is a schematic diagram of the quadrotor model;

[0058] Figure 3 is the quadrotor θ y Channel attitude output signal curve;

[0059] Figure 4 is the quadrotor θ p Channel attitude output signal curve;

[0060] Figure 5 is the quadrotor θ r Channel attitude output signal curve;

[0061] Figure 6 is the quadrotor θ y Channel rbf perturbation estimation output;

[0062] Figure 7 is the quadrotor θ p Channel rbf perturbation estimation output;

[0063] Figure 8 is the quadrotor θ r Channel rbf perturbation estimation output;

[0064] Figure 9 is the quadrotor θ y Control force output;

[0065] Figure 10 is the quadrotor θ p Control force output;

[0066] Figure 11 is the quadrotor θ r Control force output. DETAILED DESCRIPTION

[0067] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0068] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0069] Specific implementation method 1: Figure 1 This embodiment describes a dynamic system model predictive control method based on an RBF neural network, including:

[0070] Step 1: Based on the dynamic parameters of the dynamic system, establish the dynamic equation between the control input and the state variable of the dynamic system; the dynamic parameters of the dynamic system include: moment of inertia and motor coefficient;

[0071] Step 2: Design virtual control variables according to the number of moments of inertia, and use the virtual control variables to transform the dynamic equation into a nominal form based on the virtual control variables to obtain the nominal form of the dynamic equation;

[0072] Step 3: Convert the nominal form of the dynamic equation into a state space form, and use the state space form of the dynamic equation to design a cost function of the model predictive control algorithm;

[0073] Step 4: Optimize and solve the cost function according to the control constraints of the dynamic system to obtain a control output, and use the first element of the control output as the input of the dynamic system to obtain an output signal of the dynamic system;

[0074] Step 5: Using the nominal form of the dynamic equations, the virtual control variables, and the output signal of the dynamic system obtained in step 4, a neural network-based disturbance approximation model is used to inversely solve the unmodeled dynamics of the dynamic system and the external disturbances.

[0075] Step 6: Use the unmodeled dynamics, external disturbances, and virtual control variables obtained in step 5 as dynamic system inputs to obtain the system output signal after disturbance compensation. Use the control allocation matrix to transform the system output signal after disturbance compensation to obtain the control signal of the dynamic system.

[0076] Furthermore, in the present invention, in step 1, the dynamic system is a quadrotor dynamic system.

[0077] Furthermore, in the present invention, in step 1, the dynamic equation between the dynamic system control input and the state variable is:

[0078]

[0079]

[0080]

[0081] Among them, V f 、V b 、V r 、V l are the control inputs of the four motors up, down, right and left of the quadrotor aircraft, J p 、J r 、J y K is the moment of inertia of the quadrotor aircraft about the pitch axis, roll axis, and yaw axis; f is the lift coefficient, K t is the torque coefficient, d p d r d y They are the unmodeled dynamics and external disturbances of the quadrotor aircraft in three degrees of freedom: pitch angle, roll angle, and yaw angle. is θ p The second derivative of is θ r The second derivative of is θ y The second derivative of θ p is the pitch angle. When the longitudinal axis of the quadrotor is above the horizontal plane, θ p is a positive value, θ r is the roll angle, looking forward from the tail, when the quadrotor tilts to the right, θ r The angle is positive, θ y is the yaw angle, θ y When viewed from the angular plane, if the projection from the OX axis to the longitudinal axis of the quadrotor on the horizontal plane rotates counterclockwise, then θ y The angle is positive.

[0082] Furthermore, in the present invention, in step 2, the kinetic equation in nominal form is:

[0083]

[0084] y=Cx+Du

[0085] in, is the derivative of the state variable x of the dynamic system, is θ y The first derivative of is θ p The first derivative of is θ yThe first derivative of , L is the wheelbase between the four-rotor propellers; y is the output of the dynamic system, y T =[θ y θ p θ r ]; u is the virtual control variable of the control system, u=[v p v r v y ] T , v p , v r , v y Corresponding to the three attitude channels θ of the quadrotor aircraft: pitch axis, roll axis, and yaw axis p ,θ r ,θ y A virtual control variable is designed for each posture channel;

[0086]

[0087]

[0088] Furthermore, in the present invention, in step 3, the cost function of the model predictive control algorithm is:

[0089]

[0090] x(i|k) represents the forward prediction of the i-th step at time k, u(i-1|k) represents the i-1-th step control executed at time k, Q and R are N×N weighted matrices respectively, and J(k) is the cost function at time k.

[0091] Furthermore, in the present invention, in step 4, the control constraints of the dynamic system include: system output condition constraints, system state constraints and control force constraints.

[0092] Furthermore, in the present invention, in step 4, the process of optimizing and solving the cost function is:

[0093]

[0094]

[0095] x(0|k)=x(k)

[0096]

[0097]

[0098] N is the prediction step size of the system, To control the constraint set, is the state constraint set, and x(k) is the state of the system at the kth moment.

[0099] Furthermore, in the present invention, in step 5, the method for obtaining the system disturbance by inversely solving the disturbance approximation model based on a neural network using the nominal form of the dynamic equation, the virtual control variables and the dynamic system output obtained in step 4 is:

[0100] The dynamic equation between the control input and state variables of the dynamic system is rewritten as:

[0101]

[0102]

[0103]

[0104] Transformed into:

[0105]

[0106]

[0107]

[0108] Among them, d p d r d y They are the unmodeled dynamics and external disturbances on the three degrees of freedom of pitch, roll and yaw.

[0109] Furthermore, in the present invention, in step 5, the perturbation approximation model based on the neural network is:

[0110]

[0111]

[0112]

[0113] The neural network weight update model is:

[0114] ω γj (k+1)=ω γj (k)+η γ E γ s γj ,j=1,2,...,N

[0115]

[0116]

[0117] in, is the weight of the RBF neural network at time k+1, is the weight of the RBF neural network at time k, is the neural network weight update rate, E γ is the neural network estimation error, is the Gaussian kernel function, is the width of the Gaussian kernel function of the RBF neural network at time k+1, is the width of the Gaussian kernel function of the RBF neural network at time k, X γ is the input variable of the function to be estimated, is the vector form of the center of the Gaussian kernel function of the RBF neural network at time k+1, same is the center of the Gaussian kernel function of the RBF neural network at time k+1, c ij (k) is the center of the Gaussian kernel function of the RBF neural network at time k, η γ is the neural network weight update rate (no need to add (k)), same is an element of the center of the Gaussian kernel function of the RBF neural network, c ij Same as c ij (k), E1 is the disturbance estimation error of the quadrotor roll channel, E2 is the disturbance estimation error of the quadrotor pitch channel, and E3 is the disturbance estimation error of the quadrotor yaw channel. is the neural network weight of the quadrotor roll channel, is the neural network weight of the quadrotor pitch channel, is the neural network weight of the quadrotor yaw angle channel, H1(X1) is the vector form of the Gaussian kernel function of the quadrotor roll channel, H2(X2) is the vector form of the Gaussian kernel function of the quadrotor pitch channel, and H3(X3) is the vector form of the Gaussian kernel function of the quadrotor yaw angle channel.

[0118] Furthermore, in the present invention, in step 6, the system output after disturbance compensation is:

[0119]

[0120] Among them, u * is the model predictive control output for the system nominal model.

[0121] Furthermore, in the present invention, in step 6, the control signal of the dynamic system obtained by converting the system output after the disturbance compensation through the control allocation matrix is:

[0122]

[0123] Specific embodiment: The parameters of the quadrotor verification platform used are shown in the following table.

[0124] Table 1 Parameters of three-degree-of-freedom quadrotor

[0125]

[0126] Combine Figure 1 This specific embodiment describes a quadrotor attitude model predictive control method based on RBF neural network disturbance compensation. This method uses a neural network to estimate and compensate for external disturbances and unmodeled dynamics of the quadrotor, and designs a model predictive controller. Specifically, the method is implemented as follows:

[0127] Step 1: Establish a quadrotor dynamics model. Figure 1 As shown, the quadrotor is a cross configuration, where θ p is the pitch angle. When the longitudinal axis of the quadrotor is above the horizontal plane, θ p is positive, otherwise it is negative. r is the roll angle, looking forward from the tail, when the quadrotor tilts to the right, θ r The angle is positive, and the opposite is negative. y is the yaw angle, θ y When viewed from the angular plane, if the projection from the OX axis to the longitudinal axis of the quadrotor on the horizontal plane rotates counterclockwise, then θ y The angle is positive, otherwise it is negative. The dynamic equations of the quadrotor are three rotation equations, see formula (1). Among them, J p 、J r 、J y is the moment of inertia of the three axes of the quadrotor; V f 、V b 、V r 、V l is the control input of the four motors. f is the lift coefficient, K t is the torque coefficient, d p d r d y represent the unmodeled dynamics and external disturbances in three degrees of freedom, respectively.

[0128]

[0129] Step 2: Design virtual control variable u=[v p v r v y ] T , convert the dynamic model of step 1 into the form of nominal state space without unmodeled dynamics and external disturbances. The state of the dynamic system in equation (1) is recorded as The system output is denoted as y T =[θ y θ p θr ], ignoring the unmodeled dynamics and external disturbances of the system, and finally converting it into the form of nominal state space, see formula (2).

[0130]

[0131] in,

[0132]

[0133]

[0134] Step 3: Design a model predictive control algorithm based on the dynamic model of the nominal state space that does not contain the system's unmodeled dynamics and external disturbances. Define the cost function, see formula (3). Without loss of generality, assume that the prediction step size and control step size of the system are equal, denoted as N. x(i|k) represents the prediction of the i-th step at time k, u(i-1|k) represents the i-1-th step control executed at time k, and Q and R are N×N weighted matrices respectively. By optimizing and solving the cost function to minimize the cost function, the optimal control sequence is obtained, and the first element of the control sequence is applied to the controlled object. As shown in formula (4), the constraints that need to be met during the optimization solution are the system dynamics equation, the system output condition constraint, the system state constraint, and the control force constraint.

[0135]

[0136]

[0137] Step 4: Design a neural network perturbation approximation algorithm. Rewrite the system dynamics equation into the form of formula (5), and then convert it into the expected signal of the system's unmodeled dynamics and external disturbances, see formula (6). Give this expected signal to the RBF neural network, and finally obtain the approximate value of the system's unmodeled dynamics and external disturbances.

[0138] In specific implementation, an RBF neural network is designed for each channel, and the weights of the neural network are shown in formula (7). In this example, three RBF neural networks are designed for the three attitude channels of the quadrotor, that is, γ = 1, 2, and 3. The weights of the neural network are updated using the gradient descent method. The update rate of the neural network for each channel is related to the function to be approximated. In this example, the weight update rates of the three neural networks are shown in formula (8).

[0139]

[0140]

[0141]

[0142]

[0143] Step 5: Design of a model predictive controller based on neural network disturbance approximation. Based on the output of the model predictive controller, the neural network disturbance compensation is superimposed to obtain the final control output, as shown in Equation (9). The final output control force is converted into the final control voltage input for each motor using the control allocation matrix, as shown in Equation (10).

[0144]

[0145]

[0146] Although the present invention is described herein with reference to specific embodiments, it should be understood that these embodiments are merely illustrative of the principles and applications of the invention. It should be understood that many modifications may be made to the illustrative embodiments, and that other arrangements may be devised, without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that the various dependent claims and features described herein may be combined in ways other than those described in the original claims. It should also be understood that features described in conjunction with individual embodiments may be employed in conjunction with other described embodiments.

Claims

1. A dynamic system model predictive control method based on RBF neural network, characterized in that: It includes: Step 1: Based on the dynamic parameters of the dynamic system, establish the dynamic equation between the control input and the state variable of the dynamic system; the dynamic parameters of the dynamic system include: moment of inertia and motor coefficient; Step 2: Design virtual control variables according to the number of moments of inertia, and use the virtual control variables to transform the dynamic equation into a nominal form based on the virtual control variables to obtain the nominal form of the dynamic equation; Step 3: Convert the nominal form of the dynamic equation into a state space form, and use the state space form of the dynamic equation to design a cost function of the model predictive control algorithm; Step 4: Optimize and solve the cost function according to the control constraints of the dynamic system to obtain a control output, and use the first element of the control output as the input of the dynamic system to obtain an output signal of the dynamic system; Step 5: Using the nominal form of the dynamic equations, the virtual control variables, and the output signal of the dynamic system obtained in step 4, a neural network-based disturbance approximation model is used to inversely solve the unmodeled dynamics of the dynamic system and the external disturbances. Step 6: Use the unmodeled dynamics, external disturbances, and virtual control variables obtained in step 5 as dynamic system inputs to obtain the system output signal after disturbance compensation. Use the control allocation matrix to transform the system output signal after disturbance compensation to obtain the control signal of the dynamic system.

2. The dynamic system model predictive control method based on RBF neural network according to claim 1, characterized in that: In step 1, the dynamic system is a quadrotor dynamic system.

3. The dynamic system model predictive control method based on RBF neural network according to claim 2, characterized in that: In step 1, the dynamic equation between the control input and the state variable of the dynamic system is: Among them, V f 、V b 、V r 、V l are the control inputs of the four motors up, down, right and left of the quadrotor aircraft, J p 、J r 、J y K is the moment of inertia of the quadrotor aircraft about the pitch axis, roll axis, and yaw axis; f is the lift coefficient, K t is the torque coefficient, d p d r d y They are the unmodeled dynamics and external disturbances of the quadrotor aircraft in three degrees of freedom: pitch angle, roll angle, and yaw angle. is θ p The second derivative of is θ r The second derivative of is θ y The second derivative of θ p is the pitch angle. When the longitudinal axis of the quadrotor is above the horizontal plane, θ p is a positive value, θ r is the roll angle, looking forward from the tail, when the quadrotor tilts to the right, θ r The angle is positive, θ y is the yaw angle, θ y When viewed from the angular plane, if the projection from the OX axis to the longitudinal axis of the quadrotor on the horizontal plane rotates counterclockwise, then θ y The angle is positive.

4. The dynamic system model predictive control method based on RBF neural network according to claim 3 is characterized in that: In step 2, the kinetic equations in nominal form are: y=Cx+Du in, is the derivative of the state variable x of the dynamic system, is θ y The first derivative of is θ p The first derivative of is θ y The first derivative of , y is the output of the dynamic system, y T =[θ y θ p θ r ]; L is the wheelbase between the four-rotor propellers; u is the virtual control variable of the control system, u=[v p v r v y ] T , v p , v r , v y Corresponding to the three attitude channels θ of the quadrotor aircraft: pitch axis, roll axis, and yaw axis p ,θ r ,θ y dummy control variables for ; 5. The dynamic system model predictive control method based on RBF neural network according to claim 4 is characterized in that: In step 3, the cost function of the model predictive control algorithm is: x(i|k) represents the forward prediction of the i-th step at time k, u(i-1|k) represents the i-1-th step control executed at time k, Q and R are N×N weighted matrices respectively, N is the prediction step size of the system, and J(k) is the cost function at time k.

6. The dynamic system model predictive control method based on RBF neural network according to claim 5, characterized in that: In step 4, the control constraints of the dynamic system include: system output condition constraints, system state constraints and control force constraints.

7. The dynamic system model predictive control method based on RBF neural network according to claim 6, characterized in that: In step 4, the process of optimizing and solving the cost function is as follows: x(0|k)=x(k) N is the prediction step size of the system, To control the constraint set, is the state constraint set, x(k) is the state of the system at the kth moment, and x(0|k) is the predicted value at the initial moment.

8. The dynamic system model predictive control method based on RBF neural network according to claim 7, characterized in that: In step 5, the perturbation approximation model based on neural network is used to inversely solve the unmodeled dynamics and external disturbances of the dynamic system: The dynamic equation between the control input and state variables of the dynamic system is rewritten as: Transformed into: Among them, d p d r d y They are the unmodeled dynamics and external disturbances on the three degrees of freedom of pitch, roll and yaw.

9. The dynamic system model predictive control method based on RBF neural network according to claim 8, characterized in that: In step 5, the perturbation approximation model based on the neural network is: Among them, E1 is the disturbance estimation error of the quadrotor roll channel, E2 is the disturbance estimation error of the quadrotor pitch channel, E3 is the disturbance estimation error of the quadrotor yaw channel, W1 T is the neural network weight of the quadrotor roll channel, is the neural network weight of the quadrotor pitch channel, is the neural network weight of the quadrotor yaw angle channel, H1(X1) is the vector form of the Gaussian kernel function of the quadrotor roll channel, H2(X2) is the vector form of the Gaussian kernel function of the quadrotor pitch channel, and H3(X3) is the vector form of the Gaussian kernel function of the quadrotor yaw angle channel.

10. The dynamic system model predictive control method based on RBF neural network according to claim 9, characterized in that: In step 6, the system output after disturbance compensation is: Among them, u * is the output signal of the model predictive control for the nominal model of the system.

Images