Wind turbine generator intelligent control system and method based on fusion of rotary laser radar and artificial intelligence

Abstract

The invention discloses a wind turbine generator intelligent control system and method based on rotation laser radar and artificial intelligence fusion. The wind turbine generator intelligent control system comprises a dynamic scanning monitoring module which can carry out three-dimensional point cloud data collection on the azimuth angle range of 0-300 m and + / -90 degrees in front of a generator; the space-time feature processing module comprises a voxel grid denoising unit, a coordinate conversion unit and a space-time convolutional neural network ST-CNN; and a hybrid optimization control module. A variable-speed rotating multi-line laser radar is adopted, a dynamic scanning strategy is combined, and accurate monitoring of a large-range three-dimensional wind field in front of a unit is achieved. A space-time convolutional neural network ST-CNN is constructed, spatial features and time sequence information are creatively fused, and future wind conditions are accurately predicted. A hybrid control algorithm combining model predictive control MPC and deep reinforcement learning DRL is proposed, and control parameters are dynamically generated by taking power generation efficiency and unit load as optimization targets. From hardware deployment to software configuration, a complete and highly integrated intelligent control system is formed.

Description

Technical Field

[0001] The present invention relates to the field of wind turbine control technology, and in particular to an intelligent control system and method for wind turbines based on the fusion of rotating laser radar and artificial intelligence. Background Art

[0002] Traditional anemometers can only measure the wind speed at a single point at the height of the nacelle and are unable to obtain the vertical wind shear and horizontal turbulence intensity distribution. Measured data show that when there is a 3m / s vertical wind shear, the bending moment fluctuation amplitude at the root of the blade can reach 25% of the rated load, while traditional sensors have no warning for such risks. In addition, the existing PID control algorithm relies on a fixed parameter table, and the response time from wind speed change to pitch angle adjustment is usually more than 8 seconds. In a rapidly changing turbulent environment, this lag can cause the generator speed to fluctuate by more than 10%, seriously affecting the power quality and exacerbating gearbox wear. Therefore, in response to the shortcomings of existing technologies, this solution proposes an intelligent control system and method for wind turbines based on the fusion of rotating lidar and artificial intelligence. Summary of the Invention

[0003] The present invention aims to address at least one of the technical problems existing in the prior art by providing an intelligent control system and method for wind turbines based on the integration of rotating laser radar and artificial intelligence. This system utilizes a variable-speed rotating multi-line laser radar, combined with a dynamic scanning strategy, to accurately monitor the large, three-dimensional wind field in front of the turbine. This system minimizes blind spots and significantly improves the accuracy and lead time of wind shear and turbulence monitoring, providing sufficient time for turbine control. An integrated dual-axis inertial measurement unit (IMU) compensates for cabin vibration errors in real time, while an anti-icing coating and thermal airflow circulation system ensure stable operation in extreme environments and reliable data acquisition.

[0004] The team constructed a spatiotemporal convolutional neural network (ST-CNN), which innovatively integrates spatial features with time series information to accurately predict future wind conditions. Through transfer learning and incremental learning, the system rapidly adapts to the characteristics of diverse wind fields, achieving an accuracy rate of over 92% for complex wind forecasts, providing a strong basis for proactively adjusting turbine control strategies.

[0005] A hybrid control algorithm combining model predictive control (MPC) and deep reinforcement learning (DRL) is proposed. This algorithm dynamically generates control parameters based on power generation efficiency and unit load as optimization objectives. MPC ensures basic safety constraints, while DRL enables real-time fine-tuning. Together, these two algorithms reduce unit load fluctuations and increase power generation. For example, under different typical operating conditions, control parameters can be specifically optimized to significantly improve unit performance.

[0006] From hardware deployment to software configuration, a complete and highly integrated intelligent control system is formed. The hardware enables collaborative operation of the LiDAR and turbine sensors, while the software implements closed-loop optimization of data processing, prediction, control, and feedback. Equipped with comprehensive fault diagnosis and fault tolerance mechanisms, this system ensures safe and stable operation of the turbine in the event of LiDAR, control algorithm, or sensor failures, minimizing downtime and improving system reliability.

[0007] The present invention also provides a wind turbine intelligent control system and method based on the above-mentioned integration of rotating laser radar and artificial intelligence, including: dynamic scanning monitoring module: composed of a multi-line laser radar with variable speed rotation, capable of collecting three-dimensional point cloud data in the range of 0-300 meters in front of the unit and ±90° azimuth, and having an environmental adaptive scanning mode, combining high-density core area scanning with low-density far-field scanning; high-density scanning is performed on the core area 100 meters in front of the unit, the main area affecting the blade force, with a point cloud density of 50 points / m 3 , the far field area is reduced to 10 points / m 3 , while ensuring accuracy, reducing data volume by 30%, and synchronously collecting echo intensity signals for inversion of atmospheric particulate matter concentration, assisting in judging severe weather such as sandstorms, haze, etc.

[0008] Spatiotemporal feature processing module: This module includes a voxel grid denoising unit, a coordinate conversion unit, and a spatiotemporal convolutional neural network (ST-CNN). It can convert point cloud data into regular grid data, extract the spatial structural characteristics and temporal evolution laws of the wind field, and realize multi-parameter prediction of complex wind conditions such as wind shear and turbulence within the next 10 seconds. The voxel grid-based denoising algorithm, combined with outlier detection (ROPS), increases the noise point rejection rate to 98%, establishes a local coordinate system with the blade rotation center as the origin, and realizes direct mapping of wind field data and blade aerodynamic model. The coordinate conversion error is less than 0.05 meters.

[0009] Spatial feature extraction: 3D convolution layer (kernel size 3×3×3) extracts the spatial structure features of the wind field;

[0010] Time series modeling: A bidirectional LSTM layer (256 hidden nodes) captures the temporal evolution of wind speed and direction;

[0011] Output layer: Multi-task learning simultaneously predicts the wind shear index in the next 10 seconds with an error of <0.08, turbulence intensity 19 with an error of <5%, and main wind direction deviation angle with an error of <2°

[0012] Training strategy: Using transfer learning, the pre-trained model is based on 200TB of historical data from 100 wind farms around the world, and incremental learning is used to adapt to the specific wind farm characteristics during on-site deployment.

[0013] Hybrid optimization control module: Integrates model predictive control (MPC) and deep reinforcement learning (DRL) algorithms, with the goal of maximizing power generation efficiency and minimizing unit load. It dynamically generates control parameters such as pitch angle, yaw angle, and generator torque, and has real-time fine-tuning and safety constraint mechanisms for control parameters.

[0014] Model predictive control layer: A multi-physics prediction model is established, which includes the blade aerodynamic model, the transmission chain dynamic model, and the generator electromagnetic model. The prediction time domain is 10 seconds and the control period is 0.2 seconds. The optimization objective function is:

[0015] in, is the power deviation, Δβ 2 is the pitch angle change rate, M flap The blade swings the bending moment.

[0016] State space: Contains 12-dimensional wind farm parameters and 8-dimensional unit states.

[0017] Action space: pitch angle ±15°, accuracy 0.1°, yaw rate ±10° / s, generator torque ±20%

[0018] Reward function: R = 0.8P gain -0.15M rms -0.05|yaw_error|

[0019] Among them, P gain is the excess power generation, M rms is the root mean square of the bending moment at the blade root, and yaw_error is the yaw error. Network Structure: Adopts an Actor-Critic architecture, with the Actor network outputting continuous control variables and the Critic network evaluating state values. The experience replay buffer has a capacity of 100,000 entries.

[0020] According to the present invention, an intelligent control system for wind turbines based on the fusion of rotating lidar and artificial intelligence is provided. The rotating lidar integrates a dual-axis inertial measurement unit and an anti-icing coating, which can compensate for measurement errors caused by cabin vibration in real time and operate stably in an environment of -40°C to +80°C.

[0021] According to the present invention, an intelligent control system for wind turbines based on the fusion of rotating lidar and artificial intelligence is provided. The spatiotemporal convolutional neural network ST-CNN adopts a cascade structure of 3D convolutional layers and bidirectional LSTM layers, inputs point cloud data containing three-dimensional coordinates and radial velocity, and outputs the wind shear index, turbulence intensity and main wind direction deviation angle for the next 10 seconds, with prediction errors not exceeding 0.08, 5% and 2°, respectively.

[0022] According to the present invention, an intelligent control system for wind turbines based on the fusion of rotating lidar and artificial intelligence is provided. The objective function of the hybrid optimization control module includes power deviation, pitch angle change rate and blade flapping bending moment. The control instructions are quickly generated through model parallel computing technology, and the calculation time does not exceed 150ms.

[0023] According to the present invention, a wind turbine intelligent control method based on the integration of rotating laser radar and artificial intelligence includes the following steps:

[0024] Dynamic data acquisition: 3D point cloud data and unit operating status data from the rotating lidar are collected in a 0.1 second cycle. A variable density scanning strategy is used to reduce data volume. Data preprocessing is also achieved through voxel grid denoising and coordinate transformation.

[0025] Wind forecast modeling: The preprocessed grid data is fed into the spatiotemporal convolutional neural network (ST-CNN). Using a training strategy that combines transfer learning with incremental learning, the model outputs a 10-second wind parameter forecast sequence. The forecast lead time is positively correlated with the scanning distance.

[0026] Control parameter optimization: Model predictive control (MPC) is used to generate basic control instructions that meet safety constraints. Deep reinforcement learning (DRL) is used to dynamically fine-tune control parameters. The state space includes 12-dimensional wind farm parameters and 8-dimensional turbine state, and the action space includes continuous pitch angle, yaw rate, and generator torque.

[0027] Closed-loop feedback correction: Collect the actual operating data after the actuator moves, compare it with the control target, correct the prediction model error, and update the experience buffer of reinforcement learning to form an online optimization of the control strategy.

[0028] According to the present invention, a method for intelligent control of wind turbines based on the fusion of rotating laser radar and artificial intelligence is provided. In the dynamic data acquisition step, the atmospheric particulate matter concentration is inverted by the echo intensity signal to trigger the anti-pollution working mode and adjust the scanning frequency and power output of the laser radar.

[0029] According to the present invention, a method for intelligent control of wind turbines based on the fusion of rotating lidar and artificial intelligence is provided. In the control parameter optimization step, a blade segmentation control strategy is adopted for vertical wind shear conditions, and the control bandwidth and yaw prediction advance amount are adjusted for turbulent conditions to achieve coordinated optimization of load balancing and power generation efficiency.

[0030] According to the present invention, a method for intelligent control of wind turbines based on the fusion of rotating lidar and artificial intelligence is provided. In the closed-loop feedback correction step, when a lidar failure is detected, the system automatically switches to a wind field reconstruction algorithm based on historical data to maintain the system running in fault-tolerant mode for no less than 48 hours.

[0031] Compared to existing technologies, the present invention's intelligent wind turbine control system and method, based on the integration of rotating laser radar and artificial intelligence, utilizes a variable-speed rotating multi-line laser radar, combined with a dynamic scanning strategy, to achieve precise, three-dimensional monitoring of the wind field in front of the turbine over a large area. This minimizes blind spots and significantly improves the accuracy and lead time of wind shear and turbulence monitoring, providing ample time for turbine control. An integrated dual-axis inertial measurement unit provides real-time compensation for nacelle vibration errors, while an anti-icing coating and thermal airflow circulation system ensure stable operation in extreme environments and reliable data acquisition.

[0032] Compared to existing technologies, the present invention's intelligent wind turbine control system and method, based on the integration of rotating lidar and artificial intelligence, constructs a spatiotemporal convolutional neural network (ST-CNN), innovatively integrating spatial features with time series information to accurately predict future wind conditions. Through transfer learning and incremental learning, it rapidly adapts to diverse wind field characteristics, achieving an accuracy rate exceeding 92% for complex wind condition predictions, providing a powerful basis for proactively adjusting turbine control strategies.

[0033] Compared to existing technologies, the present invention's intelligent wind turbine control system and method, based on the integration of rotating LiDAR and artificial intelligence, forms a complete and highly integrated intelligent control system, from hardware deployment to software configuration. The hardware implements collaborative operation between the LiDAR and turbine sensors, while the software implements closed-loop optimization of data processing, prediction, control, and feedback. Equipped with comprehensive fault diagnosis and fault tolerance mechanisms, this system ensures safe and stable operation of the turbine in the event of LiDAR, control algorithm, or sensor failures, reducing downtime and improving system reliability.

[0034] Compared to existing technologies, the present invention proposes an intelligent control system and method for wind turbines based on the integration of rotating lidar and artificial intelligence. This hybrid control algorithm combines model predictive control (MPC) with deep reinforcement learning (DRL). This algorithm dynamically generates control parameters based on power generation efficiency and turbine load as optimization objectives. MPC ensures basic safety constraints, while DRL enables real-time fine-tuning. The two work together to reduce turbine load fluctuations and increase power generation. For example, under different typical operating conditions, control parameters can be specifically optimized to significantly improve turbine performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The present invention will be further described below with reference to the accompanying drawings and embodiments;

[0036] Figure 1 This is an architecture diagram of an intelligent control system for wind turbines based on the integration of rotating laser radar and artificial intelligence in the present invention;

[0037] Figure 2This is a flow chart of an intelligent control method for wind turbines based on the integration of rotating laser radar and artificial intelligence in the present invention;

[0038] Figure 3 This is a flow chart of a hybrid control algorithm for an intelligent control system of a wind turbine generator system based on the fusion of rotating laser radar and artificial intelligence in the present invention. DETAILED DESCRIPTION

[0039] This section will describe in detail the specific embodiments of the present invention. The preferred embodiments of the present invention are shown in the accompanying drawings. The purpose of the accompanying drawings is to supplement the description of the text part of the specification with graphics, so that people can intuitively and vividly understand each technical feature and the overall technical solution of the present invention, but it should not be understood as a limitation on the scope of protection of the present invention.

[0040] Reference Figure 1-3 The present invention discloses an intelligent wind turbine control system and method based on the integration of a rotating laser radar and artificial intelligence. The system includes system installation and commissioning. The hardware deployment involves installing a rotating laser radar, which is secured to the center of the nacelle's roof using a three-point shock-absorbing bracket. The bracket's natural frequency is designed to be 3.5 Hz, lower than the turbine's main vibration frequency of 5-8 Hz, minimizing vibration interference. The horizontal distance between the radar center and the blade's rotation center is precisely adjusted to 2.0 ± 0.05 meters, and spatial positioning is performed using a total station with an accuracy of ± 1 mm.

[0041] Cables and Interfaces: Signal transmission utilizes dual-core single-mode fiber slip rings, supporting 360° unrestricted rotation and a data transmission rate of 10Gbps. The power supply utilizes double-shielded cable with a shielding efficiency of 65dB and a surge protector installed. Its response time is less than 1ns, making it suitable for the strong electromagnetic environment at sea. Environmentally Adaptable Design: The radar housing is coated with a 50μm thick polytetrafluoroethylene anti-icing coating, with a temperature resistance range of -40°C to +80°C. A built-in heating element with a power of 50W operates through the cabin's thermal air circulation system, achieving a temperature rise rate of 2°C / min and enabling automatic de-icing in -20°C environments. Sensor Joint Calibration: Coordinate System Calibration: RTK-GPS is used to obtain the cabin's absolute position. Combined with the INS inertial navigation system, the least-squares method is used to calculate the conversion matrix between the radar coordinate system and the crew coordinate system, with an error controlled within 0.01°. Wind speed calibration: Install an ultrasonic anemometer on the side of the cabin, select 10 typical working conditions with wind speed of 4-25m / s and wind direction of 0-360°, conduct comparative tests, and establish a radar wind speed inversion model: v corrected =1.02·v lidar +0.15·sin(θ uind )

[0042] Among them, v lidar is the wind speed measured by radar, θ uind is the main wind direction angle, and the correlation coefficient R after calibration2 =0.985.

[0043] The software configuration builds an edge computing platform, with the hardware configuration using industrial-grade edge servers, supporting real-time data processing and model inference, with a task response time of less than 10μs. The operating system uses Ubuntu 22.04LTS real-time version, integrated with the OPC UA server, and a data throughput of 50MB / s.

[0044] Model training and updating, including eliminating invalid points with echo intensity less than 100mV in data preprocessing, generating regular grid data with a resolution of 5 meters through K-nearest neighbor interpolation method, and constructing a training set containing 100,000 groups of samples.

[0045] Network training: The ST-CNN model uses the AdamW optimizer and a combined mean squared error (MSE) and cross-entropy loss function. Training is performed on an 8-GPU cluster for 72 hours, achieving a validation set prediction accuracy of >92%.

[0046] Reinforcement learning simulation: A digital twin model was built based on Bladed wind power simulation software. 100,000 simulation training cycles were conducted under 100 extreme wind conditions. The average reward value of the DRL algorithm reached 0.78, with a full score of 1.0.

[0047] Online incremental learning: Automatically synchronize the latest hour's operating data every 2 hours, adopt a dynamic weight update strategy, and the incremental training time is less than 3 minutes, without affecting real-time control.

[0048] Typical operating condition control strategy, where the wind speed gradient in vertical wind shear conditions is >0.2m / s / m.

[0049] Detection and warning: When the ST-CNN model predicts that there will be upward wind shear at an altitude of 100 meters within the next 6 seconds (wind speed of 8m / s in the lower layer (0-50 meters), wind speed of 12m / s in the upper layer (50-100 meters), and gradient of 0.08m / s / m), the load balancing control mode is triggered.

[0050] Control parameter adjustment: Blade segment control: Based on the flow field simulation results (calculated using CFD software Fluent), differentiated pitch angle adjustment is implemented for the three blades: the pitch angle of blade #1 (windward side) is increased by 2.0°, and the angle of attack is reduced to reduce the aerodynamic load; the pitch angle of blade #2 (middle position) remains unchanged; the pitch angle of blade #3 (leeward side) is reduced by 1.5° to compensate for airflow separation losses.

[0051] Generator torque compensation: Based on the transmission chain dynamics model, the generator torque is temporarily increased by 5% of the rated torque (from 50kNm to 52.5kNm) to offset the speed fluctuation caused by blade aerodynamic imbalance (control target: speed fluctuation <±0.5rpm).

[0052] Actuator Operation: Pitch angle adjustment is driven by a servo motor (model: Stober DRS60, torque accuracy ±0.5 Nm), with an adjustment rate limited to 3° / s (safety threshold 5° / s), ensuring completion within 4 seconds. The yaw system maintains the current heading to avoid additional loads caused by wind direction fluctuations.

[0053] The turbulence intensity in strong turbulence conditions is greater than 15%.

[0054] Feature recognition: The DRL algorithm determines whether a vehicle has entered a strong turbulent environment by calculating the wind speed standard deviation (σ>3m / s) and wind direction change rate (dθ / dt>10° / s) in real time.

[0055] The control strategy was switched, in which the pitch angle control was optimized, and the bandwidth of the PID controller was reduced from 5 Hz to 2 Hz. By adjusting the proportional coefficient Kp = 0.8 → 0.4 and the integral time Ti = 1.2s → 3.0s, high-frequency adjustment actions were reduced. In actual measurements, the blade flapping moment fluctuation can be reduced by 18%.

[0056] Predictive Yaw Control: Based on the three-second wind direction sequence predicted by the ST-CNN, a feedforward control algorithm is used to pre-adjust the nacelle angle. The yaw rate is limited to 8° / s, with a positioning accuracy of ±1.2°. Generator power smoothing: A torque filter with a cutoff frequency of 1Hz is enabled to control power fluctuations within ±5% of the rated power.

[0057] Power generation optimization maximum power tracking under normal wind conditions: Based on the DRL algorithm, the optimal tip speed ratio λ = 8.5 ± 0.3 is calculated in real time, and the pitch angle β = 0°-15° and the generator speed n = 12-22 rpm are dynamically adjusted to make the unit operate on the Cp_max curve (power coefficient > 0.48).

[0058] Yaw alignment control: The main wind direction data monitored by the lidar has a resolution of 0.5°, and yaw correction is performed every 10 seconds to ensure that the deviation between the cabin and the main wind direction is less than 2°, reducing aerodynamic losses by more than 3%.

[0059] Fault diagnosis and fault tolerance mechanism, including the lidar fault handling fault detection rule, if the point cloud density drops by more than 40% for three consecutive scanning cycles, or the vibration acceleration output by the inertial measurement unit (IMU) is more than 1.5g, it is judged as a radar abnormality.

[0060] Fault-tolerance strategy, in which the first level of fault tolerance starts the heating and de-icing program. If the ambient temperature is less than 0°C, the scanning speed is increased from 10 rpm to 15 rpm, the data collection density is increased, and monitoring is continued for 1 minute.

[0061] Level 2 fault tolerance: If the fault is not restored, the system switches to the backup wind farm model and reconstructs the wind farm through the Kalman filter algorithm based on the historical data of the past hour and the operating status of the unit to maintain system operation for 48 hours.

[0062] Fault reporting: sending early warning information to the operation and maintenance platform through the remote transmission module, including the fault code, occurrence time and current operating status.

[0063] The control algorithm has fault-tolerant abnormal command detection, which monitors the control parameters output by the DRL in real time. If the pitch angle command is greater than 15° or less than -2°, or the generator torque is greater than 120% of the rated torque, it is determined to be an abnormal command.

[0064] The safety mode is switched by immediately deactivating the DRL module and switching to the MPC minimum control strategy. The control target is adjusted to "load minimization" with the following constraints: pitch angle change rate < 5° / s, blade root bending moment < 1.5 times the rated load, generator speed fluctuation < ±10% of the rated speed. At the same time, the algorithm self-diagnosis program is triggered to clean abnormal data in the DRL experience buffer, and an attempt is made to reactivate the DRL module after 3 minutes.

[0065] Sensor fault handling utilizes triple-redundancy for key sensors, with a majority voting algorithm determining valid data. A single sensor failure does not impact system operation. If the sensor data on a channel deviates by more than 10% from that of a redundant channel for two consecutive sampling cycles, the channel is automatically flagged as faulty and interpolation is used to generate virtual data until manual repair and replacement are performed.

[0066] Data collection and storage, including real-time data collection with a 0.1 second cycle to collect the following data:

[0067] LiDAR: 3D point cloud coordinates (X, Y, Z), radial velocity Vr, echo intensity I

[0068] Unit status: blade pitch angle (accuracy 0.1°), generator speed (resolution 0.1rpm), gearbox oil temperature (accuracy ±0.5°C), tower vibration acceleration (range ±10g, accuracy 0.1%FS)

[0069] Data format: Binary protocol encoding is used, the single-cycle data volume is approximately 150KB, and data is temporarily stored in a ring buffer to avoid packet loss.

[0070] Historical data storage, where important data is encrypted and stored on the local hard disk at a frequency of 50Hz, and uploaded to the cloud server via FTP for later algorithm optimization and fault analysis.

[0071] System testing and verification, including laboratory testing in a wind tunnel laboratory, simulated 10 typical wind conditions, including vertical wind shear and turbulence, to test the system's wind prediction accuracy and control parameter response time. Results showed a wind shear prediction lead time of 8.2±1.5 seconds, a control command generation time of 120±30ms, and a pitch angle adjustment accuracy of ±0.15°.

[0072] The embodiments of the present invention are described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Various changes can be made within the scope of knowledge possessed by ordinary technicians in the technical field without departing from the scope of the present invention.

Claims

1. An intelligent control system for wind turbines based on the integration of rotating laser radar and artificial intelligence, characterized in that: include: Dynamic scanning monitoring module: Consists of a multi-line laser radar with variable speed rotation, capable of collecting 3D point cloud data within a range of 0-300 meters and ±90° azimuth angle in front of the unit. It also has an environmentally adaptive scanning mode that combines high-density core area scanning with low-density far-field scanning. Spatiotemporal feature processing module: This module includes a voxel grid denoising unit, a coordinate conversion unit, and a spatiotemporal convolutional neural network (ST-CNN). It can convert point cloud data into regular grid data and extract the spatial structural characteristics and temporal evolution of the wind field, enabling multi-parameter prediction of complex wind conditions such as wind shear and turbulence within the next 10 seconds. Hybrid optimization control module: Integrates model predictive control (MPC) and deep reinforcement learning (DRL) algorithms, with the goal of maximizing power generation efficiency and minimizing unit load. It dynamically generates control parameters such as pitch angle, yaw angle, and generator torque, and has real-time fine-tuning and safety constraint mechanisms for control parameters.

2. The wind turbine intelligent control system based on the fusion of rotating laser radar and artificial intelligence according to claim 1 is characterized in that: The rotating lidar integrates a dual-axis inertial measurement unit and an anti-icing coating, which can compensate for measurement errors caused by cabin vibration in real time and operate stably in an environment of -40°C to +80°C.

3. The wind turbine intelligent control system based on the fusion of rotating laser radar and artificial intelligence according to claim 1 is characterized in that: The spatiotemporal convolutional neural network (ST-CNN) adopts a cascade structure of 3D convolutional layers and bidirectional LSTM layers. It inputs point cloud data containing three-dimensional coordinates and radial velocity, and outputs the wind shear index, turbulence intensity, and main wind direction deviation angle for the next 10 seconds, with prediction errors not exceeding 0.08, 5%, and 2°, respectively.

4. The wind turbine intelligent control system and method based on the integration of rotating laser radar and artificial intelligence according to claim 1 is characterized in that: The objective function of the hybrid optimization control module includes power deviation, pitch angle change rate and blade flapping bending moment. The control instructions are quickly generated through model parallel computing technology, and the calculation time does not exceed 150ms.

5. A method for intelligent control of a wind turbine generator system based on the fusion of a rotating laser radar and artificial intelligence, using the intelligent control system for a wind turbine generator system based on the fusion of a rotating laser radar and artificial intelligence as described in any one of claims 1 to 4, characterized in that: The following steps are involved: Dynamic data acquisition: 3D point cloud data and unit operating status data from the rotating lidar are collected in a 0.1 second cycle. A variable density scanning strategy is used to reduce data volume. Data preprocessing is also achieved through voxel grid denoising and coordinate transformation. Wind forecast modeling: The preprocessed grid data is fed into the spatiotemporal convolutional neural network (ST-CNN). Using a training strategy that combines transfer learning with incremental learning, the model outputs a 10-second wind parameter forecast sequence. The forecast lead time is positively correlated with the scanning distance. Control parameter optimization: Model predictive control (MPC) is used to generate basic control instructions that meet safety constraints. Deep reinforcement learning (DRL) is used to dynamically fine-tune control parameters. The state space includes 12-dimensional wind farm parameters and 8-dimensional turbine state, and the action space includes continuous pitch angle, yaw rate, and generator torque. Closed-loop feedback correction: Collect the actual operating data after the actuator moves, compare it with the control target, correct the prediction model error, and update the experience buffer of reinforcement learning to form an online optimization of the control strategy.

6. The intelligent control method for wind turbines based on the integration of rotating laser radar and artificial intelligence according to claim 5 is characterized in that: In the dynamic data acquisition step, the atmospheric particulate matter concentration is inverted through the echo intensity signal, which is used to trigger the anti-pollution working mode and adjust the scanning frequency and power output of the lidar.

7. The intelligent control method for wind turbines based on the integration of rotating laser radar and artificial intelligence according to claim 5 is characterized in that: In the control parameter optimization step, a blade segmentation control strategy is adopted for vertical wind shear conditions, and the control bandwidth and yaw prediction advance amount are adjusted for turbulent conditions to achieve coordinated optimization of load balancing and power generation efficiency.

8. A method based on the integration of rotating laser radar and artificial intelligence according to claim 5 The wind turbine intelligent control method is characterized in that: In the closed-loop feedback correction step, When a lidar failure is detected, it automatically switches to a wind field reconstruction algorithm based on historical data. Maintain system operation in fault-tolerant mode for at least 48 hours.

Images