Unmanned aerial vehicle flight control method based on self-adaptive overturning of photovoltaic cell panel

Abstract

The invention provides an unmanned aerial vehicle flight control method based on self-adaptive overturning of a photovoltaic cell panel, and relates to the technical field of unmanned aerial vehicles. According to the method, a flight control system is constructed, and flight control and photovoltaic panel overturning are optimized by means of multi-modal meteorological and topographic data and fusion of a reinforcement learning advanced technology. The sensing layer obtains terrain, photovoltaic panel state, weather and federal learning model parameter data through a sensor and transmits the data to the central control system. And the decision-making layer adjusts the flight attitude by using a PID algorithm, calculates the optimal overturning angle of the photovoltaic panel and dynamically distributes energy. The reward function is designed based on the wind resistance stability, the photovoltaic panel safety coefficient, the task completion rate and the energy consumption penalty so as to optimize the control strategy. The core algorithm comprises a dynamic grid attention mechanism, a digital twin pre-training-online fine tuning strategy and an anti-interference hierarchical control algorithm. After the system is initialized, perception, prediction, decision making, execution and verification are completed in real-time control circulation, and it is ensured that the unmanned aerial vehicle efficiently and stably completes tasks in a complex environment.

Description

Technical Field

[0001] The invention relates to the technical field of unmanned aerial vehicles, and in particular to a flight control method for unmanned aerial vehicles based on adaptive flipping of photovoltaic panels. Background Art

[0002] According to a UAV remote flight control system and method disclosed in China Publication No. "CN118642514B", it includes an environmental perception module, a data processing module, a flight control module, an anti-interference module, a task management module and a communication module; wherein: the environmental perception module: real-time monitoring of environmental parameters around the UAV; the data processing module: used to generate comprehensive environmental information; the flight control module: real-time adjustment of the UAV flight parameters according to the comprehensive environmental information provided by the data processing module; the anti-interference module: used to identify and suppress external electromagnetic interference signals; the task management module: used to dynamically plan and coordinate the multi-task execution of the UAV. The present invention significantly improves the adaptability, flight safety and efficiency and coordination of multi-task execution of the UAV in complex environments by integrating advanced environmental data acquisition, data fusion processing, dynamic flight parameter adjustment and real-time task management and communication.

[0003] The above patent documents and prior art have the following technical problems when used: Problem 1: Existing technologies have significant defects in the multimodal fusion of terrain and meteorological data. Traditional methods use static grid division or a single sensor data source, resulting in insufficient accuracy in wind field modeling and the inability to adapt to sudden meteorological changes in real time. At the same time, there is a lack of a dynamic parameter correction mechanism supported by federated learning, which makes it impossible to effectively integrate regional meteorological characteristics (such as Gobi sand attenuation and coastal salt spray corrosion) into the control strategy, resulting in a high equipment damage rate and a response delay of more than 0.5 seconds, making it difficult to meet the control needs in a highly dynamic environment; Problem two: The generalization ability and anti-interference performance of existing control strategies are insufficient. Traditional methods rely on a single digital twin model or a PID controller with fixed parameters. Control instability is prone to occur in scenarios where extreme weather (such as instantaneous gusts exceeding 12m / s) and complex terrain (slope > 30°) are coupled. Especially under the impact of gusts of wind, the fixed switching gain of the sliding mode control often causes the actuator to vibrate, resulting in fatigue damage to the photovoltaic panel connectors. Summary of the invention

[0004] Technical issues solved In view of the shortcomings of the prior art, the present invention provides a UAV flight control method based on adaptive flipping of photovoltaic panels, which solves the following problems: 1. The traditional method uses static grid division or a single sensor data source, which leads to insufficient wind field modeling accuracy and inability to adapt to sudden meteorological changes in real time; 2. Traditional methods rely on a single digital twin model or a PID controller with fixed parameters, which is prone to control instability in scenarios where extreme weather and complex terrain are coupled.

[0005] Technical Solution

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: A UAV flight control method based on adaptive flipping of photovoltaic panels, comprising the following steps: Sp1: Dynamic Grid Attention Modeling: Based on the digital surface model DSM grid elevation data and real-time wind speed gradient, combined with the real-time distance between the drone and the grid center, the grid importance score of terrain-wind field coupling is calculated; Sp2: Digital twin strategy iteration: Rehearse meteorological scenarios in a virtual environment, generate an expert strategy library, and train the initial control strategy through reinforcement learning; compare the actual control effect with the twin prediction result in real time, and when the deviation exceeds the set threshold, trigger the federated learning parameter update and strategy online fine-tuning; the federated learning parameter update obtains the latest regional meteorological feature library from the cloud, and the strategy fine-tuning superimposes the real-time error term and the federated gradient term on the basic strategy network according to the dynamic learning rate; Sp3: Hierarchical anti-disturbance control execution: Generate flight paths and photovoltaic panel flip instructions based on global weather forecasts, combine regional safety stay time calculation and sliding mode control algorithm to drive the drone motor and shape memory alloy joints to perform dynamic adjustments; Sp4: Multi-source data assimilation verification: Integrate lidar point cloud, three-dimensional ultrasonic anemometer and satellite meteorological data, update dynamic grid model parameters and verify the effectiveness of the control strategy; data update requirements are that the lidar point cloud data update cycle does not exceed 5 seconds, the three-dimensional ultrasonic anemometer sampling frequency is not less than 10Hz, and the spatial resolution of satellite meteorological data is less than 1 km; data assimilation verification feeds back the residual of the actual flip angle and the twin prediction value to the federated learning model for parameter calibration.

[0007] Preferably, the calculation method of the grid importance score in the Sp1 is: multiply the absolute value of the wind speed change intensity in the grid by the terrain slope value, and then divide it by the real-time distance between the UAV and the center of the grid to obtain the initial score of each grid; the initial score is normalized, and the normalization process uses a Softmax function to convert the score into a weight distribution in the range of 0-1, and the weight distribution is multiplied and corrected by the regional turbulence characteristic coefficient output by the federated learning model.

[0008] Preferably, the method for solving the optimal flip angle of the photovoltaic panel in the Sp1 is: weighted summing the wind speed vector of each grid according to the attention weight, and calculating the weighted average wind speed direction; taking the angle between the normal direction of the current photovoltaic panel of the drone and the weighted average wind speed direction as the target adjustment angle, and calculating the optimal flip angle by the inverse tangent function; at the same time, the method also includes the design and optimization of the reward function, and the reward function is designed based on the following factors to optimize the overall energy efficiency of the system: Wind resistance stability: Evaluate flight attitude stability based on wind speed, turbulence intensity, and the roll and pitch angles of the drone in real-time meteorological data; Photovoltaic panel safety factor: The damage risk is assessed by the deviation between the actual and predicted flip angles of the photovoltaic panel, and the flip strategy is optimized using a reward and punishment mechanism; Task completion rate: Real-time monitoring of photovoltaic panel flipping and drone flight conditions, calculation of task completion and positive rewards for meeting the target behavior; Energy consumption penalty: According to the real-time battery status and the power consumption of flight, flipping, communication, etc., excessive energy consumption is punished to optimize energy allocation; the control strategy is dynamically adjusted through this reward function to balance photovoltaic panel protection, task execution efficiency and energy utilization efficiency.

[0009] Preferably, the triggering condition for online fine-tuning of the strategy in Sp2 is: the absolute deviation between the actual flip angle of the photovoltaic panel monitored in real time and the digital twin prediction value exceeds 2 degrees; the federated learning parameter update obtains the latest regional meteorological feature library from the cloud, and the strategy fine-tuning superimposes the real-time error term and the federated gradient term on the basic strategy network according to the dynamic learning rate.

[0010] Preferably, the calculation method of the safe stay time of the area in Sp3 is: dividing the maximum dynamic stress value allowed by the photovoltaic panel material by the weighted sum of the wind impact force in the current area, wherein the wind impact force is calculated by weighted summing the normal components of the wind speed of each grid according to the attention weight, and considering the proportional relationship between air density and the square of wind speed.

[0011] Preferably, the sliding mode control algorithm described in Sp3 is implemented as follows: constructing a sliding surface formed by a linear combination of trajectory tracking errors and their rate of change, and determining a corresponding sliding surface parameter matrix based on the sliding surface; the control output is composed of a proportional differential term and an adaptive switching gain term, wherein the switching gain value increases dynamically with the real-time turbulence intensity, and the sliding surface parameter matrix avoids the regional wind vibration main frequency band identified by the federated learning model.

[0012] Preferably, the regional characteristic parameters of the federated learning model are: wind and sand attenuation factor in the Gobi region, turbulence intensity coefficient in mountainous terrain, and salt spray corrosion impact factor in coastal areas, and the model parameters are synchronously updated every 24 hours through distributed edge nodes.

[0013] Preferably, the digital twin pre-trained scenario library covers the following extreme working conditions: wind speed range is 8-15 m / s, wind direction changes continuously distributed from 0 to 360 degrees, terrain slope is 5-45 degrees, and photovoltaic panel layout density is 0.5-2.5 groups per square meter.

[0014] Beneficial Effects

[0015] The present invention provides a UAV flight control method based on the adaptive flipping of photovoltaic panels. It has the following beneficial effects: 1. This solution significantly improves the fusion accuracy and real-time performance of terrain-meteorological data through the synergy of the dynamic grid attention mechanism (DGAM) and the update of federated learning parameters. Specifically, it calculates the coupling effect of terrain slope and wind field changes in real time based on the product operation of the digital surface model grid elevation data and the real-time wind speed gradient. Combined with the dynamic distance weight distribution between the drone and the grid center, the wind field modeling error is reduced by more than 67%; by superimposing the regionalized turbulence correction coefficient output by the federated learning model, the wind field characteristics in different geographical environments are adaptively calibrated, effectively suppressing the control deviation caused by data fusion lag, and accurately adapting to the control needs in complex terrain and highly dynamic meteorological scenarios.

[0016] 2. This solution relies on the digital twin pre-training-online fine-tuning (DP-OT) and hierarchical disturbance rejection control (DHC) architecture to greatly enhance the generalization ability and anti-interference performance of the control strategy. Specifically, by building a pre-training scenario library covering extreme working conditions with wind speeds of 8-15m / s and slopes of 5-45 degrees, the working condition coverage rate of the expert strategy library is increased to 98%, solving the control instability problem caused by the single scenario of the traditional method; a hierarchical control architecture is used to separate global path planning and local execution instructions, and the adaptive switching gain of the sliding mode control is dynamically adjusted. The gain value increases linearly with the real-time turbulence intensity, eliminating the tremor of the actuator under the impact of sudden wind, and increasing the task completion rate of the mountain photovoltaic power station from 80% to 94%. At the same time, the algorithm memory usage is controlled within 256MB and the single decision time is less than 50ms, which is fully adapted to the resource constraints of edge computing devices and achieves a dynamic balance between photovoltaic panel protection and task efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a flow chart of the method of the present invention; Figure 2 It is a system framework diagram of the present invention; Figure 3 It is a schematic diagram of the adaptive flipping mechanism of the present invention; Figure 4 It is a control schematic diagram of the flight method of the present invention. DETAILED DESCRIPTION

[0018] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. 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 creative work are within the scope of protection of the present invention. Specific embodiment one:

[0020] like Figures 1 to 4 As shown in the figure, a UAV flight control method based on adaptive flipping of photovoltaic panels: This technical solution describes a UAV flight control system that uses multimodal meteorological and terrain data, combined with advanced technologies such as reinforcement learning, digital twins, and federated learning to optimize the flight control and photovoltaic panel flipping of UAVs. Through the dynamic grid attention mechanism, digital twin pre-training strategy, and hierarchical anti-disturbance control algorithm, it ensures that UAVs can complete tasks efficiently and stably in complex environments.

[0021] Sp1: Core algorithm module interacts with data, and the perception layer is described in detail: At the perception layer, terrain data is obtained through a high-precision digital elevation model (DSM). This data includes terrain changes (such as slope and undulation) in the power station area, which is usually obtained through lidar. The elevation data of the target area is obtained through laser scanning to generate a digital elevation model (DSM) to accurately depict the slope and undulation of the terrain. The high-precision GPS sensor carried by the drone provides real-time geographic location data for correcting DSM data and real-time monitoring of the drone position; the data is stored in the form of grid coordinates, and the slope and undulation values ​​of each grid are expressed in digital form. The sampling period can be adjusted according to the task requirements, usually once per second or every time the task is updated. DSM grid data can be updated to the control system in real time through a geographic information system (GIS) or sensor. The data is transmitted to the central control system through edge computing nodes or communication protocols MQTT or HTTP.

[0022] Photovoltaic panel layout and status acquisition, through the installation of tilt sensors to monitor the flip angle of photovoltaic panels in real time, to ensure that the photovoltaic panels operate within the predetermined angle range. Provide real-time feedback to the flip control system to ensure that the flipping of photovoltaic panels will not cause damage.

[0023] Acquisition of real-time meteorological field data (wind speed / wind direction / turbulence intensity): Meteorological data is collected in real time through sensors (such as anemometers, ultrasonic anemometers, and lidar). Each data source sets a different sampling frequency according to the region, and wind speed and turbulence intensity are usually updated every second. Sensor data is transmitted to the central control system through wireless communication protocols ZigBee, Wi-Fi, and 5G. The system denoises, interpolates, and standardizes the data, and after processing, it is input into the decision-making layer.

[0024] Federated learning model parameters (regional turbulence feature library): This part of the data is obtained from multiple distributed nodes through federated learning, including the turbulence feature library in the region (such as the wind and sand attenuation factor in the Gobi region and the turbulence intensity coefficient in the mountainous terrain). The federated learning model parameters are synchronized from the cloud to the drone system through the secure communication protocol HTTPS and MQT. The data will be updated in real time according to the regional meteorological conditions.

[0025] Sp2: Detailed description of the decision-making layer: Flight attitude control (roll / pitch angle): The flight controller uses a PID (proportional-integral-differential) control algorithm or a control algorithm based on reinforcement learning to adjust the flight attitude (roll and pitch angle) of the drone. Meteorological data (such as wind speed, wind direction, and turbulence intensity) will affect the calculation of attitude control instructions. The meteorological data transmitted from the perception layer is calculated by the dynamic grid attention mechanism (DGAM) to generate control instructions, which are input into the flight control algorithm to adjust the flight attitude through the PID or LQR algorithm; Photovoltaic panel flip angle and stiffness adjustment: During the flipping process of the photovoltaic panel, the control system uses the inverse model algorithm to calculate the optimal flip angle. The control system will dynamically adjust the flip angle based on the real-time data of wind speed, turbulence intensity and panel attitude. The real-time collected wind speed and meteorological data are calculated by the control algorithm to generate the optimal flip angle instruction, and transmitted to the actuator such as the motor or SMA joint through wireless signals for adjustment. Energy allocation strategy (flight / flip / communication power consumption): The system dynamically adjusts the power consumption according to the mission requirements (flight, battery, flip, communication) to ensure the maximum energy efficiency of the system. Sensors (such as battery power monitoring and flight status sensors) collect data in real time and transmit it back to the central control system, which optimizes energy allocation based on this data.

[0026] Sp3: Design and optimization of the reward function of the system: The reward function is designed based on the following four factors to dynamically adjust the control strategy: Wind resistance stability: Dynamically evaluate the wind resistance of the current attitude based on real-time meteorological data such as wind speed, turbulence intensity, and the roll and pitch angles of the drone attitude. Photovoltaic panel safety factor: Calculate whether the panel surface may be damaged based on the flip angle of the photovoltaic panel and the wind strength. Adjust the reward based on the deviation between the actual flip angle and the predicted flip angle. Task completion rate: Calculate the task completion rate by real-time monitoring of the flipping of the photovoltaic panel and the flight status of the drone, and give rewards for task completion. Energy consumption penalty: Optimize the energy efficiency of the system based on the real-time battery status and flight power consumption to avoid excessive energy consumption.

[0027] Sp4: Detailed description of the core points of the system: Dynamic Grid Attention Mechanism (DGAM): The system first calculates the wind speed gradient and DSM slope of the grid, generates the importance score of the grid through these data, and weights it based on the distance between the drone and the grid. Subsequently, the score is normalized by the Softmax function to generate the attention weight of the grid; Data processing and transmission: The calculated grid importance weight is corrected with the regional turbulence correction coefficient of the federated learning model, and finally used to control the system to calculate the optimal flip angle. Specifically, when calculating the optimal flip angle of the photovoltaic panel, the wind speed vector of each grid is first weighted and summed according to the weight based on the above grid importance weight. Assume that there is grid, The wind speed vector of the grid is , the corresponding weight is , then the weighted average wind speed vector The calculation formula is: , thus obtaining the weighted average wind speed direction. Then, compare the normal direction of the current photovoltaic panel of the drone with the weighted average wind speed direction, and set the normal direction vector of the photovoltaic panel to be , according to the vector angle formula Get the angle between the two ,Then the angle is further processed by the inverse tangent function to convert the angle into an ,actually operable flip angle command, and finally the optimal flip angle of the ,photovoltaic panel is obtained, which is used to control the control system to ,adjust the flip of the photovoltaic panel.

[0028] in: : It represents the total number of grids. The entire area to be analyzed will be divided into multiple grids. That's the number of these grids. : As an index, it is used to represent a specific grid. The value range of is from 1 to n, so that each grid can be analyzed in turn. : indicates the The wind speed vector of the grid. This vector contains the magnitude and direction of the wind speed, which is used to describe the wind speed of the grid. The wind conditions at each grid point. :It is The importance weight of each grid is calculated by the Dynamic Grid Attention Mechanism (DGAM), which takes into account factors such as the wind speed gradient of the grid, the DSM slope, and the distance between the drone and the grid. : Refers to the weighted average wind speed vector. It is obtained by weighted summing the wind speed vectors of all grids according to their respective weights, and can comprehensively reflect the average wind conditions in the entire area. : This is the summation symbol, representing arrive For accumulation. , is to move from the first grid to the The products of the weights of each grid and the wind speed vector are all added together; The weights of all grids are added together. : Indicates the normal direction vector of the current photovoltaic panel of the drone. This vector is perpendicular to the surface of the photovoltaic panel and can be used to describe the orientation of the photovoltaic panel. : is the normal direction vector of the photovoltaic panel and the weighted average wind speed direction vector By calculating this angle, we can know the relative direction between the photovoltaic panel and the wind. and Represents vectors and The modulus of , which is the size of the vector.

[0029] Digital twin pre-training-online fine-tuning (DP-OT) strategy: During the digital twin pre-training phase, simulated data of meteorology and terrain are generated through the simulation platform to form an expert strategy library. In actual operation, the difference between the flip angle of the photovoltaic panel and the predicted value of the digital twin is monitored in real time, and fine-tuning is performed when deviation occurs. When fine-tuning is triggered, the model parameters are updated through federated learning, and these updates are fed back to the strategy through the control network for real-time adjustment.

[0030] Disturbance Hierarchical Control (DHC) algorithm: Upper control: Generate flight path and photovoltaic panel flip instructions based on global weather forecast results, and calculate regional safety stay time. Regional safety stay time is calculated based on wind impact force and corrected in combination with attention weight and air density. Lower control: Compensate for prediction errors through sliding mode control algorithm and adjust control gain so that the system can cope with sudden weather changes. The control algorithm is dynamically adjusted during actual execution to avoid the risk of wind vibration resonance.

[0031] Detailed description of system integration and workflow: Initialization process: Load the DSM data and federated learning model plug-in of the target power station, create a digital twin environment and generate an expert strategy library; Real-time control loop: Each control cycle (usually 100Hz) completes the process of perception, prediction, decision-making, execution and verification. The perception stage obtains environmental data through lidar and weather stations, the prediction stage calculates the risk distribution of the grid through DGAM, and the decision layer generates flip angle instructions and flight paths based on the MMTRL algorithm. Execution and verification: The execution layer drives the motor and shape memory alloy joints to adjust, and compares the actual flip result with the twin prediction. If the error exceeds the standard, the parameter update of federated learning is triggered. Specific embodiment 2: like Figures 1 to 4 As shown, according to the content of the above specific embodiments, the following contents are further disclosed: This system is an innovative UAV flight control system, which is specially designed for the adaptive flipping of photovoltaic panels and the optimized management of UAV flight missions. Its workflow includes multiple links, each of which involves complex data processing, decision making and execution adjustment. The following is a detailed description of the algorithm workflow.

[0033] Sp1: Initialization phase: loading and data preparation Loading target power station data: When the system is started, the terrain data of the target power station is first loaded from the digital surface model (DSM). The DSM data provides elevation change information of the power station area, helping the drone understand the slope and undulating geographical features of the power station terrain. These data are usually derived from high-precision lidar scanning or remote sensing technology. Loading federated learning model parameters: At the same time, the federated learning model parameters related to the power station area are obtained from the cloud, which contain the turbulence feature library of the area. This feature library is obtained through distributed learning and contains information on wind, turbulence, and wind and sand attenuation factors under different terrain and meteorological conditions. Digital twin model construction and training: Through digital twin technology, the meteorological and terrain information of the target power station is input into the simulation platform to generate a flight and flipping behavior model in a virtual environment. At this stage, preliminary flight control strategies and photovoltaic panel flipping strategies are trained through reinforcement learning.

[0034] Sp2: Real-time control loop: data acquisition and processing Meteorological data acquisition and processing: Real-time meteorological data is acquired through sensors such as lidar, three-dimensional ultrasonic anemometers, and satellite meteorological data. This data includes information on wind speed, wind direction, and turbulence intensity. The system transmits the data to the central control unit in real time through data transmission protocols such as ZigBee, Wi-Fi, and 5G. The data undergoes preprocessing such as denoising, standardization, and interpolation to ensure its quality meets the requirements of the control system. The meteorological data will be fed into the dynamic grid attention mechanism (DGAM) to calculate the wind speed gradient of each grid, and combined with the real-time distance between the drone and the grid and the terrain slope characteristics to generate grid importance weights. After calculating the grid importance weights, it is further used to solve the optimal flipping angle of the photovoltaic panel. Specifically, the wind speed vectors of each grid are weighted and summed according to these weights to obtain the weighted average wind speed direction, and then the normal direction of the current photovoltaic panel of the drone is compared with the weighted average wind speed direction, and the optimal flipping angle is calculated through the arctangent function. In areas with large terrain undulations and frequent wind speed changes, this method can more accurately determine the flipping angle of the photovoltaic panel, reduce the adverse effects of wind on the photovoltaic panel, and ensure a certain power generation efficiency. Photovoltaic panel layout and status acquisition: The real-time layout information of the photovoltaic panel, such as position, orientation, and angle, is updated in real time through sensors or manual input. The system combines this information with the meteorological data to analyze the actual status of the current photovoltaic panel and the possible wind force impacts. Interaction between terrain data and photovoltaic panel layout: According to the interaction between terrain data, such as slope and undulation, and the photovoltaic panel layout, the system further adjusts the decision-making of the flight path and flipping angle. In areas with large slopes or strong wind speeds, the system may need to specially adjust the flipping angle of the photovoltaic panel to ensure safety.

[0035] Sp3: Decision-making layer: Control instruction generation Flight attitude control: MMTRL-Control adjusts the flight attitude of the drone through flight controllers such as PID controllers or reinforcement learning controllers. Specifically, according to the real-time wind speed and turbulence intensity meteorological data, the roll angle and pitch angle of the drone are adjusted to maintain the best flight stability. Data flow process: After being processed by the dynamic grid attention mechanism (DGAM), control instructions related to wind speed and turbulence intensity are generated. The control instructions are adjusted by the PID or LQR (linear quadratic regulator) algorithm to control the flight attitude, ensuring that the drone maintains stable flight under various meteorological conditions.

[0036] Photovoltaic panel flip angle and stiffness adjustment: According to the real-time changes in wind speed, turbulence intensity and panel posture, the flip angle of the photovoltaic panel is optimized through the inverse model algorithm MPC. MPC (model predictive control) will calculate the optimal flip angle in real time, taking into account the structural strength and wind impact of the photovoltaic panel to avoid damage to the photovoltaic panel. Data flow process: The control system adjusts the flip angle of the photovoltaic panel based on real-time wind speed data and meteorological data. Sensor feedback such as motor status and photovoltaic panel angle are transmitted to actuators such as motors or SMA joints through wireless signals to achieve precise adjustment.

[0037] Energy allocation strategy: The system also monitors energy consumption in real time and dynamically adjusts energy allocation according to mission requirements such as flight, flipping, and communication. By optimizing battery usage, the system balances power consumption between modules to ensure the continuation of missions. Data flow process: Real-time data provided by battery monitoring and flight status sensors will be fed back to the energy management module. The energy management module dynamically adjusts power consumption according to real-time requirements such as flight and flipping to maximize the energy efficiency of the system.

[0038] Sp4: Reward Function Design and Optimization: Reinforcement Learning Tuning Wind resistance stability: This part calculates the wind resistance of the current attitude based on real-time meteorological data such as wind speed, turbulence intensity, and the roll and pitch angles of the flight attitude. The higher the wind speed and turbulence intensity, the higher the stability requirements of the control system. Photovoltaic panel safety factor: Based on the error between the actual flip angle and the expected flip angle, evaluate whether the photovoltaic panel may be damaged due to excessive wind. The system will adjust the reward based on this error to prompt the algorithm to optimize the safe flip strategy of the photovoltaic panel. Mission completion rate: The mission completion rate is calculated based on the actual execution of flight and flipping. When the error between the flip angle of the photovoltaic panel and the flight attitude of the drone and the predetermined target is less than the set threshold, the system will give a positive reward. Energy consumption penalty: The system will penalize excessive energy consumption based on the real-time battery power and flight power consumption factors to optimize the overall energy efficiency of the system.

[0039] Sp5: Implementation of the algorithm Dynamic Grid Attention Mechanism (DGAM): Through the combined calculation of wind speed gradient and terrain slope, the grid importance score is dynamically adjusted, and then the corresponding control weight is calculated according to the distance between the drone and the grid to optimize the flip angle of the photovoltaic panel. In the actual calculation, the optimal flip angle of the photovoltaic panel is determined in detail according to the weighted sum and the calculated angle, providing accurate control instructions for the actuator. Digital Twin Strategy Pre-training and Online Fine-tuning (DP-OT): Preview meteorological and terrain scenarios in a virtual environment, generate an expert strategy library, and correct the control strategy in real time through online fine-tuning. Real-time feedback is combined with the update of the federated learning model to improve the adaptability and control accuracy of the system. Disturbance-Resistant Hierarchical Control (DHC) Algorithm: The upper control system generates flight paths and photovoltaic panel flip instructions based on global meteorological forecasts, and the lower control adjusts the gain in real time through the sliding mode control algorithm to compensate for errors and ensure the system's anti-disturbance capability.

[0040] Sp6: Execution and Verification: Real-time Adjustment and Feedback At the execution layer, all control commands are transmitted to the drone's motor and photovoltaic panel actuators via wireless signals. During the execution process, the system monitors the flip angle of the photovoltaic panel and the attitude status of the drone in real time, and compares the actual results with the digital twin prediction values. Data verification and fine-tuning: If the error between the real-time flip angle and the twin prediction value exceeds the set threshold, the strategy fine-tuning is triggered. At this time, the model parameters are updated through federated learning and fed back to the control network for adjustment. Execution and safety check: The dynamically adjusted flight and flip commands will drive the drone to adjust its attitude, while ensuring that the photovoltaic panel flips within a safe range to avoid damage caused by excessive wind pressure. Specific embodiment three: like Figures 1 to 4 As shown, according to the content in the above specific embodiments, the following contents are further disclosed: The key algorithms mentioned in the above embodiments are analyzed in detail below, including their core mathematical formulas and explanations: Dynamic Grid Attention Mechanism (DGAM): Goal: Assign different importance weights to different spatial regions, guide the drone to avoid high turbulence areas, and adjust the flip angle of the photovoltaic panel.

[0042] Core formula: The wind speed gradient tensor is expressed as:

[0043] Grid importance weight calculation formula:

[0044] in: Wind speed gradient tensor, expressed at a point in space The spatial variation rate of wind speed. : Direction wind speed Position derivative of direction. : Direction wind speed Position derivative of direction. : Direction wind speed Position derivative of direction. : No. The wind speed gradient norm of each grid indicates the instability and turbulence of the wind in the area. :UAVs and Real-time Euclidean distance between grid centers. : No. The terrain slope (inclination angle) corresponding to the center of each grid. : Slope adjustment function, used to adjust the amplifying effect of terrain undulation on wind instability. : Terrain slope magnification factor, empirically taken in . : Weight sensitivity parameter, which adjusts the response intensity of attention weight to wind speed gradient. The empirical value is . : No. The importance attention weight of each grid.

[0045] Function: This formula calculates the "risk" value of each spatial area by combining wind speed gradient, terrain slope and flight distance, guiding the drone to avoid areas of strong turbulence and regulating the flip direction of the photovoltaic panel.

[0046] Flight attitude control model (roll / pitch): Control goal: Stably control the attitude (roll / pitch) of the drone to cope with sudden airflow.

[0047] Attitude control formula (simplified model of PID controller):

[0048] in: : The current flight attitude control value (such as roll angle or pitch angle). : Attitude error, that is, the deviation between the target angle and the current angle. : Desired attitude angle. : Current real-time attitude angle. : Proportional gain coefficient, direct influence of regulation error. : Integral gain coefficient, used to eliminate steady-state error. : Differential gain coefficient, suppresses oscillation caused by too fast system response.

[0049] Function: Dynamically adjust the flight attitude according to the attitude error fed back by the real-time sensor to improve the stability in the wind.

[0050] Photovoltaic panel flipping optimization model (Model Predictive Control - MPC): Optimization objective: Based on the wind speed change trend and the panel attitude, optimize the future flipping angle path to ensure the safety of the panel.

[0051] State - control model:

[0052] Where: : System state vector, which may include the current angle and angular velocity states of the photovoltaic panel. : Control input, such as the PWM signal or voltage of the flipping motor. : Observable output of the system, i.e., the current angle sensor reading. 、 System state - space dynamic matrix. State - transition matrix. Control - input matrix. Output matrix.

[0053] MPC optimization objective function:

[0054] Where: Overall optimization objective function. : Prediction horizon length (predict the future N time steps). : From the current time Start predicting the th step output. Desired angle trajectory. Error penalty weight matrix. : Control energy consumption penalty weight matrix.

[0055] Function: Predict the wind force fluctuation trend in advance, intelligently plan the panel flipping process, and improve wind resistance and safety.

[0056] Disturbance - resistant hierarchical control (DHC): Control objective: Resist strong wind disturbances and enhance the control robustness through a two - layer structure. Upper - layer command generation: Using the task trajectory and wind prediction as inputs, plan the ideal state

[0057] Lower - layer sliding - mode controller (SMC):

[0058] Where: Desired system state trajectory. Current state. : First - order derivative of the state and the desired state. : Sliding - mode surface function. sign : Output control variable of sliding mode controller. Sliding surface convergence rate control parameter. Switch the gain to control the speed at which the sliding surface is reached. Symbolic function, used to implement switching control.

[0059] Function: To achieve rapid suppression of high-frequency wind disturbances and enhance system responsiveness and robustness.

[0060] Reward Function Design and Reinforcement Learning (RL) Overall reward function structure:

[0061] Definition of each sub-function: - Wind resistance index:

[0062] Rollover Safety:

[0063] Task completion rate:

[0064] Energy Consumption Penalty:

[0065] in: :Total reward value. : Wind resistance bonus reflecting posture stability. : Safety bonus for flip angle accuracy. :Task completion status. : Penalty term for energy / consumption. , , , : Weight factor of each item in the reward function. Usually set to: 0.3, 0.4, 0.2, 0.1 (adjustable). : Calculate wind resistance performance based on attitude changes and wind speed instability. : The smaller the deviation between the posture and the target, the higher the reward. (if the deviation is less than the threshold ε), otherwise it is 0. :The proportion of battery energy consumed by flight and motor power consumption.

[0066] Function: The reward function constructs a multi-objective optimization strategy to guide RLagent to optimize wind resistance, safety, and energy efficiency simultaneously.

[0067] Federated Learning and Digital Twin Updates FedAvg parameter update formula:

[0068] in: No. The model parameters of each client. :No. The number of local samples of each client. : Total number of samples from all clients. Global model parameters after aggregation.

[0069] When the system detects that the rollover error exceeds the threshold , trigger online fine-tuning:

[0070] in: The actual flip angle. The twin model predicts the flip angle. : Model update step size. Learning rate. : Loss function, usually the mean square error (MSE).

[0071] Function: Maintain the adaptability and robustness of the system by continuously adjusting the twin model and control strategy parameters online. Specific embodiment four: like Figures 1 to 4 As shown, according to the content in the above specific embodiments, the following content is further disclosed: The present invention relates to the field of intelligent control systems and renewable energy technology, and is particularly suitable for intelligent collaborative scenarios in which drone platforms perform fine flipping and real-time attitude stabilization control of photovoltaic panels in complex wind field environments. Through the collaborative work of multiple modules, efficient, safe and intelligent collaboration of photovoltaic panel flipping attitude control and drone flight attitude control in complex disturbance environments is achieved, with good real-time, robust and scalability.

[0073] To achieve the above purpose, the present invention is implemented by the following technical means: Dynamic Grid Attention Mechanism (DGAM) Module: Build a three-dimensional wind field dynamic perception network, divide the working space into multiple grid units, assign attention weights to each grid according to wind speed, turbulence, and terrain factors, build a dynamic environmental risk map, and provide accurate perception input for subsequent control modules. Attitude Control Module (PID): Adopt an adaptive PID controller to adjust the roll angle, pitch angle, and yaw angle of the drone in real time during the photovoltaic panel flip mission to ensure the stability of the flight platform and provide balance support for the flip action. Photovoltaic Panel Flip Prediction Control Module (MPC): Build a control optimization model based on the flight status of the drone and the output of DGAM to predict and optimize the flip action trajectory of the photovoltaic panel in the future, and achieve flip control with the lowest energy consumption, the smoothest path, and the best dynamic response. Anti-disturbance Sliding Mode Control Module (DHC): In the execution of MPC, if there is a strong disturbance or abnormal prediction deviation, it will automatically switch to the sliding mode control mode, use the robust sliding surface for error compensation control, achieve rapid recovery of the system state, and enhance disturbance robustness. Multi-objective reinforcement learning controller (MMTRL-RL): Construct a reward function with wind resistance, flip accuracy, energy consumption control and response speed as joint objectives, adaptively optimize the control strategy through deep reinforcement learning, and output joint control instructions (path planning + attitude adjustment + flip control). Federated learning and digital twin module (FL+DT): Multiple drone platforms share model parameters through federated learning to avoid data privacy leakage; the digital twin system builds a virtual mapping simulation system to reproduce the physical system and algorithm behavior with high fidelity and perform online error correction to improve the generalization ability and adaptability of the control strategy. Inter-module coordination mechanism: The system coordinates the operating timing and information flow of each module through the main control scheduling logic to ensure the stable closed-loop operation of the perception-decision-making-control chain, and realize the dual-objective joint intelligent control of the drone attitude and photovoltaic flip angle.

[0074] The present invention uses the DGAM mechanism to perform high-precision dynamic modeling of wind fields and terrain, improve the system's ability to predict sudden environmental disturbances, effectively avoid dangerous areas, and improve mission safety. Combining MPC with sliding mode control, it is possible to switch control strategies in a timely manner when predictive control fails or encounters strong disturbances, thereby ensuring system stability and continuous control capabilities. The MMTRL reinforcement learning controller can adaptively balance energy consumption, execution efficiency, and attitude stability, output the optimal joint control strategy, and improve the overall performance and execution efficiency of the system. Multi-machine model sharing is achieved through federated learning, and combined with real-time error correction of digital twins, the system's long-term self-learning ability is achieved, and it has good task migration capabilities and long-term stability. Each module has a clear structure and distinct logic, which is convenient for flexible nesting and expansion in other intelligent control platforms, and has good platform independence and versatility. Specific embodiment five: like Figures 1 to 4As shown, according to the content in the above specific embodiments, the following content is further disclosed: In the present UAV flight control method based on the adaptive flipping of photovoltaic panels, each module not only relies on precise control algorithms, but also needs to be equipped with corresponding hardware systems for data collection, calculation, transmission and execution control. The overall system consists of a perception layer, a computing control layer and an execution layer, covering a variety of sensors, embedded processing platforms, communication equipment and actuators, ensuring that the system can operate stably in a complex environment with high real-time and robustness.

[0076] The perception layer of the system mainly includes: high-precision inertial measurement unit (IMU), GNSS positioning module, wind speed and direction sensor, laser radar (LiDAR), visual camera, ultrasonic ranging sensor and environmental temperature and humidity sensor. IMU is used to detect the attitude angular velocity and acceleration information of the drone in real time, and provide flight status input for the PID attitude controller; GNSS module is used for high-precision positioning and trajectory tracking; wind speed and direction sensor is installed in front of or on the side of the drone to measure wind field parameters in real time, and provide DGAM module for grid weight calculation; LiDAR and camera are used together for environmental three-dimensional reconstruction, terrain modeling and obstacle identification, forming multi-modal spatial perception input; ultrasonic ranging is used for close-range precision landing and altitude maintenance control; temperature and humidity sensors assist in evaluating the impact of climate on the flip efficiency of photovoltaic panels, and participate as environmental influencing factors in multi-objective optimization.

[0077] The computing control layer is centered on the edge computing unit, and generally uses high-performance embedded processing platforms, such as NVIDIA Jetson Xavier, Raspberry Pi 4B, STM32 H7 series MCU and FPGA acceleration cards. Jetson Xavier mainly carries deep learning reasoning tasks and DGAM, MMTRL-RL module calculations, and has high concurrent tensor processing capabilities; the Raspberry Pi module is responsible for coordinating sensor data acquisition, communication relay and medium and low-speed data processing tasks; the STM32 H7 microcontroller is integrated in the flight control motherboard to perform attitude solution and PID real-time control tasks, and has high interrupt response and high-precision PWM output capabilities; the FPGA module is used to accelerate the operation of the MPC and sliding mode control modules, especially in tasks with short control cycles or intensive data processing. It has significant advantages.

[0078] At the execution control layer, the system is equipped with a high-response ratio electronic speed controller (ESC) and a brushless motor for UAV flight propulsion control, and an integrated photovoltaic panel flipping mechanism driven by a servo motor or a stepper motor. The flipping mechanism is connected to the platform or independent bracket below the UAV through an articulated structure. The servo motor is directly driven by the PWM or CAN signal output by the controller to ensure the smoothness and accuracy of the flipping action. The system also includes a power management module equipped with a high-density lithium battery pack and an energy monitoring chip (such as INA226) for real-time monitoring of the power consumption of the entire machine and providing data support for the energy consumption optimization algorithm.

[0079] In terms of communication modules, the system integrates LoRa, Wi-Fi6, 5G modules and RTK-GNSS differential base station communication interfaces. The 5G module and the digital twin platform perform high-bandwidth and low-latency communication for real-time state synchronization and remote command issuance; the LoRa module is used for long-distance low-power federated learning model parameter exchange to improve energy efficiency and security when multiple machines are working together; the Wi-Fi module is mainly used for close-range debugging and firmware upgrades. The RTK-GNSS base station combines satellite positioning data to provide sub-meter positioning accuracy, supporting the MPC module for high-precision path prediction.

[0080] The entire system is highly modular in hardware selection and architecture, with good integration, maintainability and scalability. It can flexibly adjust the perception or execution hardware combination according to different application scenarios, ensuring that the UAV flight control method based on adaptive flipping of photovoltaic panels can be stably and efficiently applied in UAV and photovoltaic flipping missions. Specific embodiment six: like Figures 1 to 4 As shown, according to the content of the above specific embodiment, the following contents are further disclosed: The following are use cases of the UAV flight control method based on adaptive flipping of photovoltaic panels in different situations: Mountain photovoltaic power station scenario: The mountain terrain is complex, the slope varies greatly, the wind field is significantly affected by the terrain, the wind speed and wind direction in different areas vary greatly, and there may be unstable meteorological conditions due to sudden winds. Workflow: Initialization phase: Load the DSM data of the mountain power station to clearly understand the slope and undulating terrain characteristics of the mountain. For example, the slope in some areas can reach 30-45 degrees. Obtain the federated learning model parameters of the mountain area from the cloud, including the turbulence intensity coefficient information of the mountain terrain. Use digital twin technology to build a virtual environment, and train flight control strategies and photovoltaic panel flipping strategies that adapt to complex mountain terrain; Real-time control: Meteorological data processing: Use lidar and three-dimensional ultrasonic anemometer sensors to obtain mountain meteorological data in real time, such as wind speeds may vary greatly in different valleys and ridges. The importance weight of each grid is calculated through the dynamic grid attention mechanism (DGAM), and higher weights are given to grids with large slopes and drastic wind speed changes; Photovoltaic panel layout and status acquisition: Real-time update of the layout and status information of photovoltaic panels at different locations in the mountains. Considering the mountainous terrain, the orientation and angle of photovoltaic panels may need to be specially adjusted according to the terrain; Terrain and photovoltaic panel layout interaction: According to the slope and undulation of the mountain, combined with the photovoltaic panel layout, the flight path and flip angle are adjusted. In areas with large slopes, the flip angle of the photovoltaic panel is appropriately increased to reduce wind impact.

[0082] Decision-making control: Flight attitude control: According to real-time meteorological data, such as when a gust of wind hits, the flight controller (such as PID controller) is used to timely adjust the roll angle and pitch angle of the drone to ensure stable flight of the drone in complex mountain wind fields; Photovoltaic panel flip angle and stiffness adjustment: The reverse model algorithm (MPC) is used to optimize the flip angle of the photovoltaic panel in real time according to wind speed, turbulence intensity and panel attitude to avoid damage to the photovoltaic panel due to excessive wind; Energy allocation strategy: Since mountain flight may require more energy to overcome the influence of terrain and wind, the system will dynamically adjust energy allocation to give priority to the energy requirements of flight and photovoltaic panel flipping; Reward function optimization: Wind resistance stability: Real-time evaluation of the UAV's wind resistance in mountain wind fields, adjusting the control strategy according to wind speed and turbulence intensity to ensure stable flight attitude; Photovoltaic panel safety factor: Pay close attention to the error between the actual flip angle of the photovoltaic panel and the expected angle to avoid damage to the photovoltaic panel due to excessive wind; Mission completion rate: Real-time monitoring of the flipping of the photovoltaic panel and the flight status of the UAV to ensure the completion of the mission in complex mountain environments; Energy consumption penalty: Reasonably control energy consumption to avoid mission interruption due to excessive energy consumption; Execution and verification: Execute control instructions to drive the drone motor and photovoltaic panel actuator to make adjustments. Monitor the flip angle of the photovoltaic panel and the drone attitude in real time and compare them with the digital twin prediction value.

[0083] If the error exceeds the set threshold, the model parameters are updated through federated learning and the control strategy is adjusted to ensure the safe and stable operation of photovoltaic panels in mountainous environments.

[0084] Coastal photovoltaic power station scenario: Coastal areas are affected by the marine climate, with high and stable wind speeds. There is also the problem of salt spray corrosion, which has a certain impact on drones and photovoltaic panel equipment.

[0085] Usage process: Initialization phase: Load the DSM data of coastal power stations to understand the characteristics of coastal terrain, such as the possible existence of certain slopes and sandy land. Obtain the federated learning model parameters of coastal areas from the cloud, including information on the factors affecting salt spray corrosion in coastal areas. Build a digital twin environment and train control strategies that adapt to coastal meteorological and terrain conditions.

[0086] Real-time control loop: Meteorological data processing: Obtain coastal wind speed, wind direction and turbulence intensity meteorological data in real time through sensors. Due to the high wind speed along the coast, the dynamic grid attention mechanism (DGAM) will focus on areas with large wind speed changes and calculate the grid importance weights. Photovoltaic panel layout and status acquisition: Real-time update of the layout and status information of photovoltaic panels in coastal environments, considering the impact of salt spray corrosion on photovoltaic panels, and timely adjustment of monitoring frequency. Interaction between terrain and photovoltaic panel layout: According to the coastal terrain and photovoltaic panel layout, adjust the flight path and flip angle to ensure stable operation of photovoltaic panels under strong winds.

[0087] Decision-making layer: Flight attitude control: According to the strong wind weather conditions along the coast, the flight controller timely adjusts the flight attitude of the drone to maintain stable flight. Photovoltaic panel flip angle and stiffness adjustment: The reverse model algorithm (MPC) is used to optimize the flip angle of the photovoltaic panel in real time according to the wind speed and turbulence intensity along the coast, while considering the impact of salt spray corrosion on the structural strength of the photovoltaic panel. Energy allocation strategy: Since coastal flight needs to overcome strong wind resistance, the system will reasonably allocate energy to ensure efficient flight and photovoltaic panel flipping.

[0088] Reward function optimization: Wind resistance stability: Real-time evaluation of the UAV's wind resistance in coastal strong wind environments to ensure stable flight attitude. Photovoltaic panel safety factor: Consider the impact of salt spray corrosion on photovoltaic panels, pay close attention to the error between the actual flip angle of photovoltaic panels and the expected angle, and ensure the safety of photovoltaic panels. Mission completion rate: Real-time monitoring of the flipping of photovoltaic panels and the flight of drones to ensure the completion of missions in coastal environments. Energy consumption penalty: Reasonably control energy consumption to avoid excessive energy consumption due to strong winds.

[0089] Execution and verification: Execute control instructions to drive the drone motor and photovoltaic panel actuator to make adjustments. Monitor the flip angle of the photovoltaic panel and the drone attitude in real time and compare them with the digital twin prediction value.

[0090] If the error exceeds the set threshold, the model parameters are updated through federated learning, the impact of salt spray corrosion is taken into account, and the control strategy is adjusted to ensure the normal operation of photovoltaic panels in coastal environments.

[0091] Gobi photovoltaic power station scenario: The Gobi region has strong sandstorms and frequent changes in wind speed. The terrain is relatively flat but there is wind and sand erosion, which causes certain damage to drones and photovoltaic panels.

[0092] Usage process: Load the DSM data of the Gobi power station to understand the Gobi terrain characteristics, such as sand and dunes. Obtain the federated learning model parameters of the Gobi region from the cloud, including the wind and sand attenuation factor information of the Gobi region. Build a digital twin environment and train control strategies that adapt to the Gobi meteorological and terrain conditions.

[0093] Real-time control loop: Meteorological data processing: Get the Gobi's wind speed, wind direction and turbulence intensity meteorological data in real time through sensors. Due to the influence of wind and sand, the dynamic grid attention mechanism (DGAM) will focus on areas with larger wind and sand and calculate the grid importance weight. Photovoltaic panel layout and status acquisition: Real-time update of the layout and status information of photovoltaic panels in the Gobi environment, considering the impact of wind and sand erosion on photovoltaic panels, and timely adjust the monitoring frequency. Interaction between terrain and photovoltaic panel layout: According to the Gobi terrain and photovoltaic panel layout, adjust the flight path and flip angle to ensure the stable operation of photovoltaic panels in wind and sand environments.

[0094] Decision-making layer: Flight attitude control: According to the wind and sand weather conditions in the Gobi Desert, the flight controller adjusts the flight attitude of the drone in time to avoid interference from wind and sand on flight. Photovoltaic panel flip angle and stiffness adjustment: The reverse model algorithm (MPC) is used to optimize the flip angle of the photovoltaic panel in real time according to the wind speed and turbulence intensity in the Gobi Desert, while considering the impact of wind and sand erosion on the structural strength of the photovoltaic panel. Energy allocation strategy: Since Gobi flight may be affected by wind and sand resistance, the system will reasonably allocate energy to ensure efficient flight and photovoltaic panel flipping.

[0095] Reward function optimization: Wind resistance stability: Real-time evaluation of the UAV's wind resistance in the Gobi desert sand environment to ensure stable flight posture. Photovoltaic panel safety factor: Consider the impact of wind and sand erosion on photovoltaic panels, pay close attention to the error between the actual flip angle of the photovoltaic panel and the expected angle, and ensure the safety of the photovoltaic panel.

[0096] Mission completion rate: Real-time monitoring of the flipping of photovoltaic panels and the flight status of drones to ensure the completion of missions in the Gobi environment. Energy consumption penalty: Reasonably control energy consumption to avoid excessive energy consumption caused by wind and sand.

[0097] Execution and verification: Execute control instructions to drive the drone motor and photovoltaic panel actuator to make adjustments. Monitor the flip angle of the photovoltaic panel and the drone attitude in real time and compare them with the digital twin prediction value.

[0098] If the error exceeds the set threshold, the model parameters are updated through federated learning, the impact of wind and sand attenuation is taken into account, and the control strategy is adjusted to ensure the normal operation of photovoltaic panels in the Gobi environment.

[0099] The following is a table of specific experimental data based on the above different scenarios: Equipment scenario indicator experimental data table:

[0100] The above data are simulation experimental data based on theory and common practical situations. The actual data may vary due to specific geographical environment and meteorological conditions.

[0101] It should be noted that, in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the statement "comprising a reference structure" do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0102] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A UAV flight control method based on adaptive flipping of photovoltaic panels, characterized in that: Includes the following: Sp1: Dynamic Grid Attention Modeling: Based on the digital surface model DSM grid elevation data and real-time wind speed gradient, combined with the real-time distance between the drone and the grid center, the grid importance score of terrain-wind field coupling is calculated; Sp2: Digital twin strategy iteration: Rehearse meteorological scenarios in a virtual environment, generate an expert strategy library, and train the initial control strategy through reinforcement learning; Compare the actual control effect with the twin prediction result in real time. When the deviation exceeds the set threshold, the federated learning parameter update and strategy online fine-tuning are triggered; The parameter update of federated learning obtains the latest regional meteorological feature library from the cloud, and the strategy fine-tuning superimposes the real-time error term and the federated gradient term on the basic strategy network according to the dynamic learning rate; Sp3: Hierarchical anti-disturbance control execution: Generate flight paths and photovoltaic panel flip instructions based on global weather forecasts, combine regional safety stay time calculation and sliding mode control algorithm to drive the drone motor and shape memory alloy joints to perform dynamic adjustments; Sp4: Multi-source data assimilation verification: Integrate lidar point cloud, three-dimensional ultrasonic anemometer and satellite meteorological data, update dynamic grid model parameters and verify the effectiveness of the control strategy; data update requirements are that the lidar point cloud data update cycle does not exceed 5 seconds, the three-dimensional ultrasonic anemometer sampling frequency is not less than 10Hz, and the spatial resolution of satellite meteorological data is less than 1 km; data assimilation verification feeds back the residual of the actual flip angle and the twin prediction value to the federated learning model for parameter calibration.

2. The method for controlling the flight of a UAV based on the adaptive flipping of photovoltaic panels according to claim 1, characterized in that: The calculation method of the grid importance score in the Sp1 is: multiply the absolute value of the wind speed change intensity in the grid by the terrain slope value, and then divide it by the real-time distance between the drone and the center of the grid to obtain the initial score of each grid; The initial score is normalized, and the normalization process uses a Softmax function to convert the score into a weight distribution in the range of 0-1, and the weight distribution is multiplied and corrected by the regional turbulence characteristic coefficient output by the federated learning model.

3. The method for controlling the flight of a UAV based on the adaptive flipping of photovoltaic panels according to claim 1, characterized in that: The method for solving the optimal flip angle of the photovoltaic panel in the Sp1 is: weighted summing the wind speed vector of each grid according to the attention weight, and calculating the weighted average wind speed direction; taking the angle between the normal direction of the current photovoltaic panel of the drone and the weighted average wind speed direction as the target adjustment angle, and calculating the optimal flip angle through the inverse tangent function; at the same time, the method also includes the design and optimization of the reward function, and the reward function is designed based on the following factors to optimize the overall energy efficiency of the system: Wind resistance stability: Evaluate flight attitude stability based on wind speed, turbulence intensity, and the roll and pitch angles of the drone in real-time meteorological data; Photovoltaic panel safety factor: The damage risk is assessed by the deviation between the actual and predicted flip angles of the photovoltaic panel, and the flip strategy is optimized using a reward and punishment mechanism; Task completion rate: Real-time monitoring of photovoltaic panel flipping and drone flight conditions, calculation of task completion and positive rewards for meeting the target behavior; Energy consumption penalty: According to the real-time battery status and the power consumption of flight, flipping, communication, etc., excessive energy consumption is punished to optimize energy allocation; the control strategy is dynamically adjusted through this reward function to balance photovoltaic panel protection, task execution efficiency and energy utilization efficiency.

4. The method for controlling the flight of a UAV based on the adaptive flipping of photovoltaic panels according to claim 1, characterized in that: The triggering conditions for online fine-tuning of the strategy in Sp2 are: The absolute deviation between the actual flip angle of the photovoltaic panel monitored in real time and the predicted value by the digital twin exceeds 2 degrees; the federated learning parameter update obtains the latest regional meteorological feature library from the cloud, and the strategy fine-tuning superimposes the real-time error term and the federated gradient term on the basic strategy network at a dynamic learning rate.

5. The method for controlling the flight of a UAV based on the adaptive flipping of photovoltaic panels according to claim 1, characterized in that: The calculation method of the safe stay time in the area described in Sp3 is: The maximum allowable dynamic stress value of the photovoltaic panel material is divided by the weighted sum of the wind impact force in the current area. The calculation of the wind impact force is weighted summation of the normal components of the wind speed of each grid according to the attention weight, and the proportional relationship between air density and wind speed squared is considered.

6. The method for controlling the flight of a UAV based on the adaptive flipping of photovoltaic panels according to claim 1, characterized in that: The implementation of the sliding mode control algorithm in Sp3 is: A sliding surface formed by a linear combination of trajectory tracking errors and their rate of change is constructed, and the corresponding sliding surface parameter matrix is ​​determined based on the sliding surface; the control output consists of a proportional differential term and an adaptive switching gain term, wherein the switching gain value increases dynamically with the real-time turbulence intensity, and the sliding surface parameter matrix avoids the regional wind vibration main frequency band identified by the federated learning model.

7. The method for controlling the flight of a UAV based on the adaptive flipping of photovoltaic panels according to claim 1, characterized in that: The regional characteristic parameters of the federated learning model are: wind and sand attenuation factor in the Gobi region, turbulence intensity coefficient in mountainous terrain, and salt spray corrosion impact factor in coastal areas, and the model parameters are synchronously updated every 24 hours through distributed edge nodes.

8. The method for controlling the flight of a UAV based on the adaptive flipping of photovoltaic panels according to claim 1, characterized in that: The digital twin pre-training scenario library covers the following extreme working conditions: wind speed range is 8-15 m / s, wind direction changes continuously distributed from 0 to 360 degrees, terrain slope is 5-45 degrees, and photovoltaic panel layout density is 0.5-2.5 groups per square meter.

Images