An unmanned aerial vehicle gust disturbance suppression method and system based on model predictive control

Abstract

The application discloses a model predictive control-based unmanned aerial vehicle gust disturbance suppression method and system, which comprises the following steps: taking an equivalent disturbance vector as a known input, explicitly introducing the equivalent disturbance vector into a discrete state prediction equation of MPC, and constructing a disturbance prediction sequence; at each sampling time, obtaining an optimal control sequence based on a current state estimation, a reference trajectory and the disturbance prediction sequence; and based on the equivalent disturbance vector, explicitly calculating a feedforward compensation amount according to an unmanned aerial vehicle dynamics model. The gust disturbance estimated by a nonlinear disturbance observer is explicitly introduced into a prediction model of model predictive control, the feedforward compensation amount and the nominal control amount are superposed based on the dynamics model to output, so that the controller can predict the disturbance influence in a prediction time domain and plan the control in advance; by constructing a zero-order hold disturbance prediction sequence and introducing the control amount amplitude, the control increment and the state quantity constraint into an optimization problem, the control instruction can always satisfy the motor saturation and the safe flight envelope limit.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) flight control technology, and more specifically, to a method and system for suppressing UAV gust disturbances based on model predictive control. Background Technology

[0002] Civilian drones, especially multi-rotor and VTOL fixed-wing drones, are widely used in aerial photography, inspection, plant protection, and logistics delivery due to their advantages such as hovering, VTOL, and low-speed cruising. However, these drones are typically small, lightweight, and have low inertia, making them extremely sensitive to wind disturbances. Statistics show that wind disturbance is one of the main causes of flight accidents involving small drones, accounting for more than 30%.

[0003] In complex low-altitude environments (such as urban canyons, areas surrounding building complexes, and offshore platforms), wind fields exhibit distinct gust characteristics: wind speeds change drastically within a short period (typically 0.5 to 5 seconds) (peak wind speeds can reach over 15 m / s), with large speed gradients (≥10 m / s²) and random directions. These short-duration strong gusts cannot be accurately predicted using conventional wind field models, and their frequency components often fall within the attitude control bandwidth of UAVs (approximately 2–10 Hz), easily triggering attitude resonance or control divergence, leading to severe attitude fluctuations, position drift, or even loss of control and crashes. Therefore, how to achieve rapid and effective suppression of short-duration sudden gusts is a key technical problem urgently needing to be solved in this field.

[0004] To address the above problems, numerous technical solutions have been proposed, mainly including the following categories: (1) Wind resistance method based on classical PID control: improve the system speed by tuning PID parameters or adopt cascade PID structure. Its advantages are simple structure and small amount of calculation, but its disadvantages are that the response is lagging when facing sudden strong winds, and it is easy to overshoot or even oscillate after integral saturation, making it difficult to balance dynamic response and steady-state accuracy.

[0005] (2) Wind resistance method based on sliding mode control (SMC): The sliding mode is fully robust to matched disturbances and has a fast response speed. However, the inherent chattering phenomenon of sliding mode control is difficult to eliminate. When the gusts change frequently, it is easy to excite high-frequency unmodeled dynamics of propellers and motors. Moreover, the control input is discontinuous and it is not suitable for actual motor speed regulation.

[0006] (3) Wind resistance method based on adaptive control: By estimating wind disturbance parameters online and adjusting controller parameters, it can adapt to slowly changing wind fields. However, the convergence speed of the adaptive law is limited. For short-term (1~2 seconds) sudden gusts, the gusts often disappear before convergence, and the coupling of parameter estimation and control may lead to transient performance deterioration.

[0007] (4) Wind-resistant methods based on disturbance observers: Designing nonlinear disturbance observers (NDOB) and extended state observers (ESO) to estimate wind disturbances online and perform feedforward compensation can estimate disturbances in real time and achieve good compensation results. However, most of these methods only use a simple superposition of "observer + feedback controller" without deep integration with predictive control; when there is control input saturation, the estimation accuracy of the observer decreases; and there is a lack of utilization of future disturbances in the prediction time domain, making it impossible to predict the impact of disturbances and plan control in advance.

[0008] (5) Wind-resistant methods based on wind field perception and feedforward: MPC can explicitly handle physical constraints such as actuator saturation and attitude angle limitation, and obtain optimal control through rolling optimization. However, most existing MPC methods do not introduce the real-time estimation of the disturbance observer as a known input into the prediction model, or only treat the disturbance as output noise and passively correct it through feedback correction, resulting in a large dynamic error under gusts. In addition, the prediction time domain and control time domain of existing MPC methods are mostly fixed values ​​preset offline, which cannot be dynamically adjusted according to the real-time changes in gust intensity, resulting in computational redundancy in light winds and insufficient prediction in strong winds.

[0009] In summary, existing technologies generally suffer from the following common problems: insufficient response speed: unable to respond to short-term strong gusts, by which time the gusts have already caused significant deviations; buffeting / high-frequency oscillations: sliding mode control and other methods easily excite unmodeled dynamics; lack of constraint handling: not considering practical constraints such as motor saturation and angular velocity limiting; lack of predictive capability: unable to predict the impact of disturbances and plan control in advance within the prediction time domain; insufficient composite coordination: disturbance observers and MPCs are mostly designed independently, without deep integration; fixed MPC time domain parameters: unable to dynamically adjust the prediction and control time domains according to gust intensity; insufficient position and attitude coordination: most methods only handle the attitude loop or position loop separately. Therefore, there is an urgent need to provide a UAV wind-resistant control method and system that has fast response speed, strong disturbance resistance, robustness to model uncertainty, ability to handle practical physical constraints, and is particularly suitable for suppressing short-term strong gusts. Summary of the Invention

[0010] To address the aforementioned technical problems in related technologies, this invention proposes a method and system for suppressing gust disturbances from unmanned aerial vehicles (UAVs) based on model predictive control, which can overcome the above-mentioned shortcomings of existing technologies.

[0011] To achieve the above-mentioned technical objectives, the technical solution of the present invention is implemented as follows: A method for suppressing gust disturbances from unmanned aerial vehicles (UAVs) based on model predictive control includes the following steps: S1 establishes a dynamic model of the UAV, which includes the equivalent disturbance force and equivalent disturbance torque generated by wind disturbance on the position subsystem and attitude subsystem; S2 designs a nonlinear disturbance observer that uses the UAV state estimate and control input to estimate the equivalent disturbance vector at the current moment online. S3 builds a model predictive controller (MPC), including: S301 takes the equivalent disturbance vector estimated in step S2 as a known input and explicitly introduces it into the discrete state prediction equation of MPC to construct a disturbance prediction sequence, so that MPC can predict the impact of the disturbance on the future state of the system in the prediction time domain. S302 at each sampling time Based on the current state estimate, the reference trajectory, and the disturbance prediction sequence, a finite-time quadratic programming optimization problem with physical constraints is solved to obtain the optimal control sequence. S303 calculates the feedforward compensation amount explicitly based on the equivalent interference vector and the UAV dynamics model. S4 superimposes the first control quantity in the optimal control sequence obtained by the MPC solution with the feedforward compensation quantity to generate a total control quantity and applies it to the UAV to suppress gust disturbances.

[0012] Furthermore, the nonlinear disturbance observer in step S2 is constructed using the auxiliary variable method to avoid using acceleration signals. Defined as ,in This is the disturbance estimate. The observer gain matrix is... , These are the attitude angle vector and the angular velocity vector, respectively.

[0013] Furthermore, the method for constructing the perturbation prediction sequence in step S301 is as follows: The zero-order preservation assumption is adopted, assuming that the perturbation value estimated at the current sampling time remains unchanged throughout the entire prediction time domain, i.e. , For prediction in the time domain.

[0014] Further, in step S303, the feedforward compensation amount is explicitly calculated. The methods include: Feedforward compensation torque of attitude channel ,in Here is the rotational inertia matrix. The estimated equivalent disturbance torque; Feedforward compensation force of position channel ,in The estimated equivalent disturbance force is mapped to the body coordinate system through a rotation matrix to obtain the total thrust feedforward component.

[0015] Furthermore, the physical constraints in step S302 include at least one of the following: control magnitude constraints, control increment constraints, and state constraints; the state constraints include attitude angle constraints and angular velocity constraints; the objective function of the optimization problem includes weights on the output error, control quantity, and control increment.

[0016] Furthermore, it also includes: S5 calculates the instantaneous energy, cumulative energy, and rate of change of the gust disturbance in real time based on the current disturbance estimate and its historical sequence, and classifies the gust intensity; it dynamically adjusts the prediction time domain length and control time domain length of the model prediction controller according to the gust intensity classification, wherein the stronger the gust, the longer the time domain.

[0017] Furthermore, the Model Predictive Controller (MPC) adopts a full state-space model, and the state vector includes at least three-axis position, three-axis velocity, three-axis attitude angle, and three-axis angular velocity to achieve coordinated disturbance rejection control of the position loop and attitude loop.

[0018] A model predictive control-based UAV gust disturbance suppression system, used to execute the method, characterized in that it includes: Sensor data acquisition module, used to collect UAV flight data; The state estimation module is used to fuse data from multiple sensors and calculate the current state estimate of the UAV. The disturbance observer module is used to estimate the magnitude of the current gust disturbance online based on the UAV dynamics model and state estimates. The Model Predictive Control (MPC) module receives reference trajectory, state estimate, and disturbance estimate, constructs a predictive model containing a disturbance prediction sequence, and outputs the nominal control quantity by solving an optimization problem with physical constraints. The feedforward compensation calculation module is used to explicitly calculate the feedforward compensation amount based on the disturbance estimate. The control allocation module is used to distribute the total control quantity to each actuator.

[0019] Furthermore, the MPC module also includes a time-domain adaptive adjustment unit, which identifies the gust energy level online based on the disturbance estimate output by the disturbance observer module, and dynamically adjusts the prediction time domain and control time domain of the MPC according to a preset scheduling strategy.

[0020] The beneficial effects of this invention are as follows: By explicitly introducing the gust disturbance estimated by the nonlinear disturbance observer into the prediction model of the model predictive control, and calculating the feedforward compensation and nominal control output based on the dynamic model, the controller can anticipate the impact of the disturbance and plan the control in advance within the prediction time domain. Compared with traditional PID and sliding mode control, the response speed is improved by 3-5 times and there is no chattering phenomenon. By constructing a zero-order hold disturbance prediction sequence and simultaneously introducing control amplitude, control increment and state constraints into the optimization problem, the control command always meets the motor saturation and safe flight envelope limits, avoiding integral saturation and command over-limit. By identifying the gust energy characteristics online and dynamically adjusting the prediction time domain and control time domain, the computational redundancy is reduced in light winds and the prediction capability is enhanced in strong winds. By uniformly optimizing the position and attitude coupling relationship through a twelve-dimensional full state vector, the hovering accuracy under gusts is improved to 0.1-0.2 meters and the attitude fluctuation is reduced by about 60%, achieving efficient suppression of short-term strong gusts and coordinated protection of flight safety. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is an overall architecture block diagram of the UAV gust disturbance suppression system based on model predictive control according to an embodiment of the present invention; Figure 2 This is a composite control block diagram based on MPC and a disturbance observer for the UAV gust disturbance suppression method based on model predictive control according to an embodiment of the present invention. Figure 3 This is a control flowchart of the UAV gust disturbance suppression method based on model predictive control according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the UAV dynamic model coordinate system for the UAV gust disturbance suppression method based on model predictive control according to an embodiment of the present invention; Figure 5 This is a diagram of the dual-loop control structure of the UAV position subsystem and attitude subsystem in the UAV gust disturbance suppression method based on model predictive control according to an embodiment of the present invention. Figure 6 This is a time-domain schematic diagram of MPC rolling optimization for the UAV gust disturbance suppression method based on model predictive control according to an embodiment of the present invention. Figure 7This is a comparison chart of simulation experiment / flight verification results (attitude angle tracking error, position tracking error, etc.) of the UAV gust disturbance suppression method based on model predictive control according to the embodiments of the present invention. Figure 8 This is a schematic diagram of the hardware module connection of the UAV gust disturbance suppression method based on model predictive control according to an embodiment of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention are within the scope of protection of the present invention.

[0024] like Figure 1-8 As shown in the embodiment of the present invention, a method for suppressing gust disturbances of unmanned aerial vehicles based on model predictive control includes the following steps: S1 establishes a dynamic model of the UAV, which includes the equivalent disturbance force and equivalent disturbance torque generated by wind disturbance on the position subsystem and attitude subsystem; S2 designs a nonlinear disturbance observer that uses the UAV state estimate and control input to estimate the equivalent disturbance vector at the current moment online. S3 builds a model predictive controller (MPC), including: S301 takes the equivalent disturbance vector estimated in step S2 as a known input and explicitly introduces it into the discrete state prediction equation of MPC to construct a disturbance prediction sequence, so that MPC can predict the impact of the disturbance on the future state of the system in the prediction time domain. S302 at each sampling time Based on the current state estimate, the reference trajectory, and the disturbance prediction sequence, a finite-time quadratic programming optimization problem with physical constraints is solved to obtain the optimal control sequence. S303 calculates the feedforward compensation amount explicitly based on the equivalent interference vector and the UAV dynamics model. S4 superimposes the first control quantity in the optimal control sequence obtained by the MPC solution with the feedforward compensation quantity to generate a total control quantity and applies it to the UAV to suppress gust disturbances.

[0025] Preferably, the nonlinear disturbance observer in step S2 is constructed using the auxiliary variable method to avoid using acceleration signals. Defined as ,in This is the disturbance estimate. The observer gain matrix is... , These are the attitude angle vector and the angular velocity vector, respectively.

[0026] Preferably, the method for constructing the perturbation prediction sequence in step S301 is as follows: The zero-order preservation assumption is adopted, which assumes that the perturbation value estimated at the current sampling time remains unchanged throughout the entire prediction time domain. .

[0027] Preferably, in step S303, the feedforward compensation amount is explicitly calculated. The methods include: Feedforward compensation torque of attitude channel ,in Here is the rotational inertia matrix. The estimated equivalent disturbance torque; Feedforward compensation force of position channel ,in The estimated equivalent disturbance force is mapped to the body coordinate system through a rotation matrix to obtain the total thrust feedforward component.

[0028] Preferably, the physical constraints in step S302 include at least one of the following: control magnitude constraints, control increment constraints, and state constraints; the state constraints include attitude angle constraints and angular velocity constraints; the objective function of the optimization problem includes weights on the output error, control quantity, and control increment.

[0029] Preferably, it further includes: S5 calculates the instantaneous energy, cumulative energy, and rate of change of the gust disturbance in real time based on the current disturbance estimate and its historical sequence, and classifies the gust intensity; dynamically adjusts the prediction time domain length and control time domain length of the model prediction controller according to the gust intensity classification, wherein the stronger the gust, the longer the time domain.

[0030] Preferably, the prediction model adopts a full state space model, and the state vector includes at least three-axis position, three-axis velocity, three-axis attitude angle and three-axis angular velocity, so as to realize the cooperative disturbance rejection control of the position loop and the attitude loop.

[0031] A model predictive control-based UAV gust disturbance suppression system, used to execute the method, characterized in that it includes: Sensor data acquisition module, used to collect UAV flight data; The state estimation module is used to fuse data from multiple sensors and calculate the current state estimate of the UAV. The disturbance observer module is used to estimate the magnitude of the current gust disturbance online based on the UAV dynamics model and state estimates. The Model Predictive Control (MPC) module receives reference trajectory, state estimate, and disturbance estimate, constructs a predictive model containing a disturbance prediction sequence, and outputs the nominal control quantity by solving an optimization problem with physical constraints. The feedforward compensation calculation module is used to explicitly calculate the feedforward compensation amount based on the disturbance estimate. The control allocation module is used to distribute the total control quantity to each actuator.

[0032] Preferably, the MPC module further includes a time-domain adaptive adjustment unit, which identifies the gust energy level online based on the disturbance estimate output by the disturbance observer module, and dynamically adjusts the prediction time domain and control time domain of the MPC according to a preset scheduling strategy.

[0033] To facilitate understanding of the above technical solutions of the present invention, the following detailed description of the above technical solutions of the present invention will be provided through specific usage methods.

[0034] In practical application, the UAV gust disturbance suppression system based on model predictive control according to the present invention includes the following functional modules:

[0035] It also includes a feedforward compensation calculation module, which is used to explicitly calculate the feedforward compensation amount based on the disturbance estimate. Figure 1 In the diagram, solid arrows represent signal flow, while dashed arrows represent feedback / correction signal flow.

[0036] Step 1: Establishing the UAV Dynamics Model 1.1 Coordinate System Definition like Figure 4 As shown, the following coordinate system is defined:

[0037] The rotation relationship from the body coordinate system to the ground coordinate system is described by Euler angles: roll angle φ, pitch angle θ, and yaw angle ψ. The rotation matrix R(φ,θ,ψ) is defined as follows:

[0038] in, .

[0039] 1.2 Dynamic Model of Position Subsystem In the ground coordinate system, the position dynamics equation of the UAV can be expressed as: (1) in: This is the position vector of the UAV in the ground coordinate system; The total mass of the drone (unit: kg); Total thrust generated by the four rotors (unit: N). ; Let gravitational acceleration be (unit: m / s²), and take... ; The equivalent disturbance force (in N) generated by wind on the position subsystem.

[0040] 1.3 Dynamic Model of Attitude Subsystem The dynamic equations of the attitude subsystem (in body coordinates) are: (2) in: This is the body's angular velocity vector (unit: rad / s). , , These are the roll, pitch, and yaw angular velocities, respectively. Here is the rotational inertia matrix of the UAV (unit: kg·m²). The control torque for the three channels (unit: N·m); The equivalent disturbance torque (unit: N·m) generated by wind disturbance on the attitude subsystem. This represents the cross product operation of vectors.

[0041] The kinematic relationship between Euler angles and body angular velocity is as follows: (3) For hovering or small-angle flight conditions, it can be approximated as , , .

[0042] 1.4 Gust Disturbance Modeling To describe the characteristics of short-duration, sudden gusts, this invention uses the following "1-cos" type gust model as the standard input for simulation verification: (4) in: Gust wind speed (unit: m / s); This refers to the peak wind speed of the gust. The starting time of the gust (unit: seconds); The duration of the gust (in seconds) is usually taken as... The gust of wind acts on the drone's body, generating an equivalent interference force. and disturbance torque : (5) in: The density of air (unit: kg / m³) is taken under standard atmospheric conditions. ; The overall aerodynamic drag coefficient of the UAV; The reference windward area of ​​the drone (unit: m²); This is a unit vector representing wind direction. The position vector of the wind pressure center relative to the center of gravity of the UAV (unit: m).

[0043] It should be noted that, in the actual control process, the present invention does not require explicit knowledge of the physical parameters in equation (5), but rather obtains them through the disturbance observer described below. and Perform real-time estimation.

[0044] Step 2: Gust Disturbance Observation and Estimation 2.1 Disturbance Observer Design This invention employs a nonlinear disturbance observer (NDOB) to measure the equivalent disturbance torque generated by wind disturbance. and interference Perform online estimation.

[0045] Taking the attitude subsystem as an example, equation (2) can be rewritten in the following state-space form: (6) in: This is the attitude angle vector; It is the angular velocity vector; For control torque input; This refers to the nonlinear term of the system. The input matrix; This is the perturbation term equivalent to the angular acceleration layer.

[0046] The nonlinear disturbance observer is designed as follows: (7) To avoid using acceleration signals (Usually contains a lot of noise), introduce auxiliary state variables. : (8) in This represents the observer gain matrix. Unlike conventional fixed-gain designs, this invention employs a nonlinear variable-gain strategy, dynamically adjusting the observer gain based on the gust gradient to balance rapid disturbance tracking and measurement noise suppression.

[0047] Variable gain design: The observer gain matrix is ​​taken as Each element Calculate in real time using the following formula: (8a) in: These are the lower and upper limits of the gain (preset constants), respectively. It is a saturation function of the Sigmoid type, defined as ; This is the scaling factor for the rate of change of the disturbance; is the steepness coefficient of the Sigmoid function; This is the bias parameter.

[0048] The principle of variable gain mechanism: When the wind is relatively strong (rapidly rising / falling in gusts), the gain is higher. Approaching The observer achieves fast disturbance tracking; when the disturbance tends to stabilize, the gain... Approaching The observer achieves low-noise, high-precision estimation.

[0049] (9) Finally, the disturbance estimate is given by the following formula: (10) Similarly, for the position subsystem, a disturbance observer with the same structure is designed to estimate the equivalent disturbance force. .

[0050] Replacing the nonlinear disturbance observer with a linear extended state observer (LESO) or a nonlinear extended state observer (NESO) can also achieve the estimation of wind disturbances and connect to the feedforward channel of MPC.

[0051] 2.2 Observer Convergence Analysis Define the perturbation estimation error In disturbance Under the assumption that the rate of change is bounded (i.e. , (where is a bounded constant), and from equation (7), the error dynamic equation can be obtained: (11) Solving equation (11) yields: (12) From equation (12), it can be seen that when It converges to a bounded region, and this is achieved by increasing the observer gain. This allows for arbitrarily small steady-state errors. For short-duration gusts (duration... (s), the observer can make an effective estimate of the disturbance before the gust reaches its peak.

[0052] Step 3: Design of a Gust Disturbance Suppression Controller Based on MPC 3.0 Construction of the Spatiotemporal Distribution and Evolution Model of Wind Field (Prerequisite Module for MPC Prediction Model) Traditional MPC methods (including the aforementioned schemes in this application) rely solely on a nonlinear disturbance observer (NDOB) to estimate wind disturbance at the current point and extrapolate the time-domain disturbance using the zero-order preserve assumption. This method is effective in spatially uniform or slowly changing wind fields, but when the UAV traverses spatially non-uniform wind fields such as gust gradient regions, terrain-induced turbulence regions, and building wake regions, it cannot predict the wind field conditions in areas not yet reached along the flight path and can only respond passively.

[0053] To address the aforementioned issues, this invention introduces a wind field spatiotemporal distribution and evolution model at the front end of the MPC prediction model, embedding the spatial distribution characteristics of the three-dimensional wind field as prior knowledge into the prediction time domain of MPC to achieve advanced feedforward compensation.

[0054] 3.0.1 Offline Construction of 3D Wind Field Spatiotemporal Distribution Model Before executing the mission, the following offline preprocessing was performed on the UAV's flight airspace (altitude ≤ 300m, horizontal range covering the area surrounding the mission route): (1) Spatial discretization and CFD simulation database construction The spatial domain is discretized into a three-dimensional structured mesh. The grid resolution is set according to the accuracy requirements of the task (typical values: 5~20m horizontally, 2~5m vertically). A high-fidelity CFD simulation database is constructed using the following data sources: Topographic / building geometry data: digital surface model (DSM) acquired from GIS systems, satellite imagery, or LiDAR point clouds; Meteorological boundary conditions: Collect historical meteorological observation data (wind speed, wind direction, turbulence intensity) of the task area to form a set of boundary conditions; CFD simulation: Large eddy simulation (LES) or Reynolds-averaged Navier-Stokes (RANS) methods are used to perform offline simulations under different boundary conditions, and the reference wind field vector at each grid node is calculated. ), forming a CFD simulation database .

[0055] (2) Wind field order reduction model based on intrinsic orthogonal decomposition (POD) To facilitate real-time airborne computation, a reduced-order wind field model (ROM) is constructed using the Proper Orthogonal Decomposition (POD) technique, projecting the high-dimensional CFD data into a low-dimensional subspace. (Equation 3.0.1-1) in: For a moment Location Wind field vector at the location; The wind field is the time-averaged field. For the first The order POD basis functions (spatial modes) characterize the spatial distribution features of the wind field; For the first The time coefficient characterizes the time-varying evolution characteristics of the wind field; To truncate the mode numbers, take The role of M is to control the balance between the accuracy and computational complexity of the wind field reduction model. The larger the value, the more accurate the reconstructed wind field, but the greater the computational load; this invention takes... This is to meet the requirements of airborne real-time computing while ensuring sufficient accuracy.

[0056] The POD basis functions are obtained by solving the following eigenvalue problem: (Equation 3.0.1-2) in The spatial covariance function of the wind field These are the eigenvalues.

[0057] The aforementioned offline construction work is completed in one go before the task execution, and the calculated POD basis functions are... It is loaded into the onboard computer for online use.

[0058] 3.0.2 Online Prediction of Wind Field Time Coefficient Based on LSTM During drone flight, meteorological boundary conditions may change in real time. To dynamically correct the wind field evolution model, this invention deploys a lightweight Long Short-Term Memory (LSTM) network on the airborne terminal to predict the POD time coefficient in real time. Future evolution: (Equation 3.0.1-3) in: From the current moment For the future Predicted values ​​of the POD time coefficient after the step; This is the POD time coefficient vector at the current moment; The time window length of the LSTM (take) (corresponding to historical data of 0.2~0.5 seconds). These are the weight parameters of the LSTM network.

[0059] The LSTM network was trained offline using CFD simulation data, with the training objective being to minimize the prediction error. (Equation 3.0.1-4) During online flight, the LSTM network takes the currently observed POD time coefficient sequence as input and outputs the future time coefficient sequence in a rolling manner. Predicted value of time coefficient for each step .

[0060] 3.0.3 Reconstruction of 3D Wind Field Spatiotemporal Distribution and Generation of Disturbance Prediction Sequences Based on the time coefficients obtained from the POD basis function and LSTM prediction, the spatiotemporal distribution of the three-dimensional wind field in the future prediction time domain can be reconstructed. (Equation 3.0.1-5) Furthermore, based on the reference trajectory of the UAV in the prediction time domain... (Given by the task planning), extract the predicted wind field disturbance values ​​at each prediction node to form the wind field disturbance prediction sequence required by the MPC prediction model: (Equation 3.0.1-6) in The predicted future wind disturbance vector (equivalent disturbance force and disturbance moment) is used as the known input to the MPC prediction model.

[0061] 3.0.4 Synergy with Gust Energy Adaptive Time Domain The wind field spatiotemporal distribution model constructed in this section is coordinated with the online identification of gust energy characteristics in step five: when the gust energy level is level 0 or 1 (no wind / light wind), the update frequency of the wind field spatiotemporal distribution model can be reduced to once every 10 control cycles; when the gust energy level is level 2 or 3 (moderate / strong gust), the wind field spatiotemporal distribution model is updated in every control cycle, and the LSTM time window T_w is automatically shortened to improve the response capability to rapidly changing wind fields.

[0062] 3.1 Basic Principles and Framework of MPC like Figure 6 As shown, Model Predictive Control (MPC) at each sampling time... Based on the current state estimate and disturbance estimates In the prediction time domain The system predicts the future output of the internal system and obtains the control time domain by solving a constrained finite-time optimization problem. ( Optimal control sequence within ) Then only the first control variable The control is applied to the controlled object. At the next sampling time, the above "prediction-optimization-control" process is repeated to achieve rolling optimization.

[0063]

[0064] 3.2 Prediction Model Construction This invention uses a discretized linear time-varying model as the prediction model for MPC. Equations (1) and (2) are linearized near the hovering operating point, and the estimated values ​​from the disturbance observer are considered. Given the input, the following discrete state-space equations are obtained: (13) (14) in: This is the wind disturbance vector predicted for the (k+1)th future step based on the wind field spatiotemporal distribution model in Section 1.4.0; The wind disturbance input matrix maps the predicted wind field values ​​to the UAV state space. This is the estimated disturbance value for the current point of NDOB (as a residual compensation term).

[0065] State vector ; Perturbation input vector Output vector (Position + Yaw Angle); , 、 For the system matrix of the corresponding dimension, the subscripts are... This indicates that the matrix can be updated as the operating point changes (time-varying); The output matrix takes values ​​of .

[0066] In the prediction time domain Within, the predicted output sequence can be obtained by iterating according to equation (13): (15) in: To predict the output sequence; Let this be the state vector of the drone at the current moment. The control sequence to be optimized; For perturbation prediction sequences; , For the system matrix , , The resulting prediction coefficient matrix.

[0067] Apply equation (13) to the prediction time domain The inner iteration yields a predicted output sequence that considers the spatiotemporal distribution of the wind field: (15-1) in This is a wind field spatiotemporal distribution prediction sequence constructed from [the original text]. This is the corresponding prediction coefficient matrix.

[0068] Unlike existing technologies, this invention explicitly embeds wind field spatiotemporal distribution prediction sequences into the MPC prediction model. This enables the controller to know the wind field state of the UAV at each prediction node in the prediction time domain in advance during the optimization solution, thus realizing a fundamental leap from "posterior compensation" to "advanced feedforward".

[0069] It is particularly important to note that the future changes of short-term gust disturbances within the prediction time domain cannot be accurately predicted. Using complex time-series prediction models (such as Kalman filtering or neural networks) would increase the computational burden and have limited accuracy in predicting gusts that change abruptly on the order of seconds. To address this contradiction, this invention proposes the zeroth-order preservation assumption: (16) This means that the disturbance value estimated at the current sampling time is assumed to remain unchanged in the subsequent short-term prediction time domain (typically 0.1~0.4 seconds). This assumption is based on the fact that the frequency of gust changes (approximately 0.2~2Hz) is much lower than the sampling / update frequency of MPC (100Hz), therefore the disturbance can be approximated as constant within a single prediction time domain. This zero-order hold-at-the-moment disturbance prediction sequence construction method is a unique design of this invention specifically for short-term gust characteristics, fundamentally different from existing technologies. Combined with the rolling optimization mechanism of MPC, the disturbance estimate is refreshed at each sampling time, thereby achieving effective tracking of time-varying gusts.

[0070] 3.3 Design of the rolling optimization objective function The MPC optimization problem of this invention can be formulated as: at each sampling time... Solve the following constrained quadratic programming problem: (17) in: The desired output sequence (reference trajectory, including the desired location) and desired yaw angle ); The output error weight matrix is ​​a positive definite diagonal matrix used to adjust the tracking accuracy of position and yaw angle. This is the control weight matrix, a positive definite diagonal matrix used to adjust control energy consumption; The control increment weight matrix is ​​a positive definite diagonal matrix used to suppress drastic changes in the control quantity (with a low-pass filtering effect). To control the increment; This is the weighted Euclidean norm. Here, R represents the set of real numbers, which is the standard notation in mathematics. Represents an n-dimensional real vector space. express Real matrix space. Specifically, in this paragraph: This indicates that the reference trajectory vector is a 4-dimensional real vector (containing the desired position). and desired yaw angle ; This indicates that the output error weight matrix is ​​a A real matrix.

[0071] Expanding equation (17) and substituting it into equation (15), we can transform it into the following standard quadratic form: (18) in It is the Hessian matrix (positive definite). These are linear term coefficient vectors, both of which are calculated from the prediction coefficient matrix and the weight matrix.

[0072] 3.4 Constraint Handling This invention fully considers the physical limitations of UAV actuators and explicitly introduces the following constraints into the optimization problem: (1) Control amplitude constraint (actuator saturation): (19) in: ; .

[0073] Taking a quadcopter drone as an example, typical values ​​are: (The motor cannot be reversed.) (single motor) (maximum thrust) ,in This refers to the rotor wheelbase; ,in , These are the thrust coefficient and the torque coefficient, respectively.

[0074] (2) Controlling incremental constraints (preventing sudden changes): (20) (3) State constraints (safe flight envelope): (twenty one) Of particular interest is attitude angle constraint. , (usually taken) (within); angular velocity constraints: , .

[0075] The above constraints constitute a convex quadratic programming (QP) problem, which can be solved in real time on an embedded platform using the effective set method, interior point method, or alternating direction multiplier method (ADMM).

[0076] For embedded platforms with limited computing power, multi-parameter quadratic programming (mp-QP) can be solved offline. The optimal control law of MPC is represented as a piecewise affine function of the state, which can be queried online without the need for real-time QP solution.

[0077] 3.5 Feedback Correction Mechanism Due to model mismatch and unmodeled perturbations, there is an error between the predicted output and the actual output. This invention employs the following feedback correction strategy: (twenty two) in The current time is the output prediction error, and it is assumed that this error remains constant in the prediction time domain.

[0078] Step 4: Disturbance Feedforward Compensation and MPC Fusion Control Strategy 4.1 Composite Control Architecture like Figure 2 As shown, this invention adopts a three-layer composite control architecture of "disturbance observer + MPC + feedforward compensation":

[0079] Figure 2 middle, and The difference is adjusted through the optimization objective function of MPC, while the output of the perturbation observer is... Simultaneously used for: Feedforward compensation: directly added to the control quantity; Prediction model correction: In equation (13) of MPC, it is used as a known input to participate in the prediction of future states.

[0080] 4.2 Calculation of Feedforward Compensation for Disturbance Estimation Based on the disturbance observer (Disturbing torque) and (Disturbance force) The calculation principle of the feedforward compensation is to directly offset the influence of the disturbance at the control level.

[0081] For the attitude channel, the feedforward compensation torque is: (twenty three) in This is the equivalent angular acceleration disturbance estimated in equation (7).

[0082] For the position channel, the feedforward compensation force is: (twenty four).

[0083] Due to the control quantity of the position loop The total thrust is coupled with the attitude angle (see equation (1)), and it is necessary to... Mapped to the body coordinate system: (25) in , Let be the rotation matrix in equation (1).

[0084] The explicit calculation of the feedforward compensation amount (Equations (23)-(25)) is entirely based on the derivation of the UAV dynamics model, which differs from the empirical proportional compensation or the method that relies solely on MPC feedback correction in the prior art. In this invention, the disturbance estimate... Used simultaneously for: 1. MPC prediction model correction (in equation (13) (term), enabling the controller to anticipate the future impact of disturbances; feedforward compensation (in equation (26) This directly cancels out the current disturbance. This dual-utilization mechanism enables the present invention to achieve a faster response speed than the "observer + traditional feedback controller" approach.

[0085] In addition to the feedforward compensation based on the NDOB current point disturbance estimation mentioned above, this invention also calculates the advance feedforward compensation amount based on the wind field spatiotemporal distribution prediction value in Section 1.4.0: (25 new) in Wind disturbance input matrix The pseudo-inverse maps the predicted future wind disturbances to the feedforward control variable.

[0086] The final control quantity is the sum of three items: (26 new) in: The nominal control quantity for MPC optimization (already obtained from equation (13)) The term considers the impact of the spatiotemporal distribution of the wind field on future states. For feedforward compensation based on the current point estimate of NDOB (see Equation (25)); This is a feedforward compensation based on wind field spatiotemporal distribution prediction.

[0087] The triple compensation mechanism forms a complete anti-disturbance chain of "advanced prediction → advanced planning → advanced compensation → current compensation", which fundamentally solves the shortcomings of existing methods that can only respond passively.

[0088] 4.3 Generation of final control commands like Figure 3 As shown, the complete control flow of the gust disturbance suppression method of the present invention is as follows:

[0089] The final control quantity applied to the drone is: (26) in This is the first component of the solution to the optimization problem of equation (18).

[0090] Step 5: Online Identification of Gust Energy Characteristics and Adaptive Time-Domain Adjustment of MPC 5.1 Technical Issues In gust disturbance suppression, the prediction time domain of MPC and control time domain The selection of [aspect name] directly affects control performance and computational burden. If... If the period is too short, the controller cannot predict the impact of long-period disturbances; if If the time period is too long, it increases the computational burden and creates redundancy in the prediction of short-term gusts. Conventional MPC uses fixed time-domain parameters, which cannot adapt to real-time changes in gust intensity.

[0091] 5.2 Online Identification of Gust Energy Characteristics This invention designs a lightweight gust energy characteristic identification module based on the estimated value output by a disturbance observer. And its historical sequence, and calculate the following features in real time: (1) Instantaneous energy of gust disturbance: (27) in This is the estimated disturbance value at the current moment.

[0092] (2) Accumulated energy of gust disturbances (within a short-term window): (28) in Let the length of the sliding window be [value]. (Corresponding to 0.2~0.5 seconds).

[0093] (3) Rate of change of gust disturbance (characterizing gust gradient): (29) 5.3 Gust Intensity Classification Based on Energy Characteristics Based on instantaneous energy and accumulated energy The current gust conditions are classified into the following levels:

[0094] in , , , , The preset threshold can be calibrated according to the drone model and mission scenario.

[0095] 5.4 Adaptive Scheduling Strategy for MPC Time Domain Parameters Based on the gust level, the following adaptive scheduling strategy is adopted to adjust the prediction time domain of MPC. and control time domain :

[0096] The above scheduling strategy can use a hysteresis comparison method to avoid frequent switching: (30) in For threshold hysteresis band, For continuous counting threshold (e.g.) ).

[0097] 1.6.5 Prediction Model Update in Dynamic Time Domain when or When changes occur, the prediction coefficient matrix in equation (15) needs to be updated synchronously. , , Its update method is as follows:

[0098] (31)

[0099] To ensure real-time performance, the above recalculation is performed quickly using pre-stored analytical formulas.

[0100] By incorporating online identification of UAV mass and moment of inertia into the predictive model parameters, the adaptability to operating conditions such as load changes and icing can be further improved.

[0101] 6. System Hardware Composition and Module Function Description like Figure 8 As shown, the system hardware components of the present invention include:

[0102] Figure 8 In the process, each sensor is connected to the flight controller via buses such as I²C, SPI, and UART. After processing, the flight controller outputs a PWM signal to the ESC.

[0103] In summary, by employing the technical solutions described above, the gust disturbance estimated by the nonlinear disturbance observer is explicitly introduced into the prediction model of the model predictive control. The feedforward compensation and nominal control quantities are then superimposed and output based on the dynamic model. This allows the controller to anticipate the disturbance's impact and plan control in advance within the prediction time domain. Compared to traditional PID and sliding mode control, the response speed is improved by 3-5 times without chattering. By constructing a zero-order hold disturbance prediction sequence and simultaneously introducing control magnitude, control increment, and state constraints into the optimization problem, the control commands always satisfy motor saturation and safe flight envelope limits, avoiding integral saturation and command overruns. By identifying gust energy characteristics online and dynamically adjusting the prediction and control time domains, computational redundancy is reduced in light winds, and predictive capability is enhanced in strong winds. By uniformly optimizing the position and attitude coupling relationship using a twelve-dimensional full-state vector, the hovering accuracy under gust conditions is improved to 0.1-0.2 meters, and attitude fluctuations are reduced by approximately 60%, achieving efficient suppression of short-term strong gusts and coordinated protection of flight safety.

[0104] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for suppressing gust disturbances from unmanned aerial vehicles (UAVs) based on model predictive control, characterized in that, Includes the following steps: S1 establishes a dynamic model of the UAV, which includes the equivalent disturbance force and equivalent disturbance torque generated by wind disturbance on the position subsystem and attitude subsystem; S2 designs a nonlinear disturbance observer that uses UAV state estimates and control inputs to estimate the equivalent disturbance vector at the current moment online. S3 builds a model predictive controller (MPC), including: S301 takes the equivalent disturbance vector estimated in step S2 as a known input and explicitly introduces it into the discrete state prediction equation of MPC to construct a disturbance prediction sequence, so that MPC can predict the impact of the disturbance on the future state of the system in the prediction time domain. S302 at each sampling time Based on the current state estimate, the reference trajectory, and the disturbance prediction sequence, a finite-time quadratic programming optimization problem with physical constraints is solved to obtain the optimal control sequence. S303 calculates the feedforward compensation amount explicitly based on the equivalent interference vector and the UAV dynamics model. S4 superimposes the first control quantity in the optimal control sequence obtained by the MPC solution with the feedforward compensation quantity to generate a total control quantity and applies it to the UAV to suppress gust disturbances.

2. The method for suppressing UAV gust disturbances based on model predictive control according to claim 1, characterized in that, The nonlinear disturbance observer in step S2 is constructed using the auxiliary variable method to avoid using acceleration signals. The auxiliary variables... Defined as ,in This is the disturbance estimate. The observer gain matrix is... , These are the attitude angle vector and the angular velocity vector, respectively.

3. The method for suppressing UAV gust disturbances based on model predictive control according to claim 1, characterized in that, The method for constructing the perturbation prediction sequence in step S301 is as follows: The zero-order preservation assumption is adopted, which assumes that the perturbation value estimated at the current sampling time remains unchanged throughout the entire prediction time domain. , For prediction in the time domain.

4. The method for suppressing UAV gust disturbances based on model predictive control according to claim 1, characterized in that, In step S303, the feedforward compensation amount is explicitly calculated. The methods include: Feedforward compensation torque of attitude channel ,in Here is the rotational inertia matrix. The estimated equivalent disturbance torque; Feedforward compensation force of position channel ,in The estimated equivalent disturbance force is mapped to the body coordinate system through a rotation matrix to obtain the total thrust feedforward component.

5. The method for suppressing UAV gust disturbances based on model predictive control according to claim 1, characterized in that, The physical constraints in step S302 include at least one of the following: control magnitude constraints, control increment constraints, and state constraints; the state constraints include attitude angle constraints and angular velocity constraints; the objective function of the optimization problem includes weights on the output error, control quantity, and control increment.

6. The method for suppressing UAV gust disturbances based on model predictive control according to claim 1, characterized in that, Also includes: S5 calculates the instantaneous energy, cumulative energy, and rate of change of the gust disturbance in real time based on the current disturbance estimate and its historical sequence, and classifies the gust intensity; it dynamically adjusts the prediction time domain length and control time domain length of the model prediction controller according to the gust intensity classification, wherein the stronger the gust, the longer the time domain.

7. The method for suppressing UAV gust disturbances based on model predictive control according to claim 1, characterized in that, The Model Predictive Controller (MPC) adopts a full state-space model, and the state vector includes at least three-axis position, three-axis velocity, three-axis attitude angle, and three-axis angular velocity to achieve coordinated disturbance rejection control between the position loop and the attitude loop.

8. A model predictive control-based unmanned aerial vehicle (UAV) gust disturbance suppression system, used to execute the method according to any one of claims 1 to 7, characterized in that, include: Sensor data acquisition module, used to collect UAV flight data; The state estimation module is used to fuse data from multiple sensors and calculate the current state estimate of the UAV. The disturbance observer module is used to estimate the magnitude of the current gust disturbance online based on the UAV dynamics model and state estimates. The Model Predictive Control (MPC) module receives reference trajectory, state estimate, and disturbance estimate, constructs a predictive model containing a disturbance prediction sequence, and outputs the nominal control quantity by solving an optimization problem with physical constraints. The feedforward compensation calculation module is used to explicitly calculate the feedforward compensation amount based on the disturbance estimate. The control allocation module is used to distribute the total control quantity to each actuator.

9. The UAV gust disturbance suppression system based on model predictive control according to claim 8, characterized in that, The MPC module also includes a time-domain adaptive adjustment unit, which identifies the gust energy level online based on the disturbance estimate output by the disturbance observer module, and dynamically adjusts the prediction time domain and control time domain of the MPC according to a preset scheduling strategy.

Images