Wind power variable pitch intelligent control method and controller based on DRL-PID fusion nonlinear filtering

Abstract

The invention discloses a wind power variable pitch intelligent control method and controller based on DRL-PID fusion nonlinear filtering, and relates to the field of wind power control, and the method comprises the steps: collecting the real-time wind speed of a wind turbine generator and a generator rotating speed signal through a wind speed sensor and a rotating speed sensor, obtaining a blade load estimation value according to a load estimator, and constructing a control input state set; performing joint estimation processing on the signals in the input state set by applying an extended Kalman filter, suppressing high-frequency noise and nonlinear disturbance, and outputting standardized state sensing data; the multi-dimensional operation data is collected and optimized, the noise and nonlinear disturbance influence is effectively reduced, the state sensing accuracy is improved, the control parameters are dynamically adjusted by means of autonomous learning, the control instruction is made to accurately adapt to the real-time working condition, and the flexibility and adaptability of variable pitch control are enhanced.

Description

Technical Field

[0001] This invention relates to the field of wind power control technology, specifically to a wind power pitch intelligent control method and controller based on DRL-PID fusion nonlinear filtering. Background Technology

[0002] Intelligent pitch control for wind turbines is a core technology of wind power systems. It uses sensors to collect real-time data such as wind speed and turbine rotation speed, and combines this data with PID, fuzzy control, or machine learning algorithms to dynamically adjust the blade angle. Its core objective is to adapt to changes in wind speed, maximizing wind energy capture efficiency at low wind speeds and limiting power output for stable loads at high wind speeds, while simultaneously reducing turbine fatigue damage and ensuring safe equipment operation. This technology features fast real-time response and high control precision, significantly improving the stability and power generation efficiency of wind power systems, and is a key support for intelligent operation and maintenance of wind power systems.

[0003] Patent application number 202110782430.3 discloses a method for controlling the pitch of a motor used in wind turbine pitch control. This application aims to address the problem that "currently, wind turbine pitch control uses a separate motor and drive system. This system involves extensive wiring, complex installation, and requires significant time for high-altitude work. To reduce installation space and high-altitude work time, Fischer has developed an integrated wind turbine pitch control system. This new type of integrated drive and control motor features high power density, small size, light weight, simple transportation and wiring, and short installation and commissioning time, significantly reducing customer costs. The integrated system needs to incorporate algorithms adapted to wind turbine pitch control functions to solve the control and safety logic issues of wind turbine pitch control."

[0004] However, traditional PID control is widely used in wind turbine pitch control systems, but its parameters are fixed, making it difficult to cope with sudden changes in wind speed, model uncertainties, and nonlinear disturbances. Although there are some adaptive PID or model predictive control (MPC) schemes, they have the following problems:

[0005] Control parameters rely on expert experience or offline modeling;

[0006] The response to nonlinear wind speed is lagging, which can easily cause speed fluctuations;

[0007] Lacking embedded AI deployment capabilities, it is difficult to implement in industrial controllers;

[0008] There is a lack of active suppression mechanisms for blade load fluctuations.

[0009] Therefore, there is an urgent need for a wind power pitch intelligent control method and controller based on DRL-PID fusion nonlinear filtering. Summary of the Invention

[0010] In view of the above-mentioned shortcomings of the existing technology, the present invention provides a wind power pitch intelligent control method and controller based on DRL-PID fusion nonlinear filtering, which can effectively solve the problems of the existing technology.

[0011] To achieve the above objectives, the present invention is implemented through the following technical solutions;

[0012] This invention discloses a wind power pitch intelligent control method based on DRL-PID fusion nonlinear filtering, comprising:

[0013] Real-time wind speed and generator speed signals of the wind turbine are collected by wind speed and speed sensors. Blade load estimates are obtained using a load estimator to construct a control input state set. An extended Kalman filter is applied to jointly estimate the signals in the input state set, suppressing high-frequency noise and nonlinear disturbances, and outputting standardized state-aware data. A lightweight deep reinforcement learning network is constructed, and the standardized state-aware data is input into the network. The network autonomously learns the control law and outputs PID parameter adjustment increments. Based on the PID parameter adjustment increments, the proportional coefficient, integral coefficient, and derivative coefficient of the PID controller are dynamically updated to generate a pitch angle control command adapted to the current operating conditions. The pitch angle control command is transmitted to the pitch actuator through the CANopen / EtherCAT communication interface of the embedded intelligent controller to drive the blades to perform angle adjustment actions. The control effect feedback value is calculated according to a preset reward function. The state data, parameter adjustment amounts, and feedback values ​​are stored as experience samples and used for online training of the DRL network.

[0014] Preferably, the load estimator senses the real-time operating status of the wind turbine, integrates the airflow speed signal collected by the wind speed sensor and the rotor rotation signal collected by the speed sensor, and combines the aerodynamic characteristic parameters related to the blade airfoil lift coefficient, drag coefficient and aspect ratio. By coupling the aeroelasticity principle with the real-time operating parameters, a multi-dimensional load evaluation model is constructed, and based on the evaluation model, a comprehensive blade load estimate including the load peak value, mean value and rate of change is output.

[0015] Preferably, the joint estimation process of the extended Kalman filter is adapted to the nonlinear dynamic characteristics of the wind turbine, and the state equation is constructed using real-time wind speed, generator speed, and blade load estimation as the core state vector.

[0016] The observation equation is constructed using the raw signals collected by each sensor as the observation vector;

[0017] During the iteration process, the initial value of the error covariance matrix is ​​preset according to the rated operating parameters of the wind turbine. After each iteration, the weight of the observation vector is dynamically adjusted through the Kalman gain matrix to output standardized state perception data that has been noise suppressed and bias corrected.

[0018] Preferably, the lightweight deep reinforcement learning network adopts a working condition-adaptive network architecture;

[0019] The input layer dimension is precisely matched with the feature dimension of the standardized state-aware data, and the input data is processed by feature normalization to improve training efficiency.

[0020] The hidden layer adopts a structure that alternates between convolutional layers and fully connected layers. The convolutional layer extracts local features related to the working conditions through the local receptive field mechanism, while the fully connected layer is responsible for fusing global features. At the same time, an adaptive ReLU function is configured as a non-linear activation function to avoid gradient vanishing.

[0021] The output layer outputs the adjustment increment corresponding to the PID parameters through linear mapping, and the built-in output constraint module controls the adjustment increment to always stay within the preset range.

[0022] Preferably, when adjusting the incremental dynamic update of the PID controller parameters based on the PID parameters, the following conditions are met:

[0023] The updated value of the proportional coefficient is the weighted sum of the current proportional coefficient and the corresponding adjustment increment. The weight coefficient is dynamically adjusted according to the magnitude of the current control error. When the error is large, the weight of the adjustment increment is increased, and when the error is small, the weight is decreased.

[0024] The updated value of the integral coefficient is adjusted and corrected according to the trend of the cumulative control error.

[0025] The updated value of the differential coefficient is dynamically adjusted based on the magnitude of the error rate of change and the smoothness index, with the weight of the increment increasing when the rate of change fluctuates drastically and decreasing when the fluctuation is gentle.

[0026] Among them, the proportional coefficient, integral coefficient, and derivative coefficient all satisfy the preset upper and lower limit constraints of parameters and the proportional relationship constraints between parameters after the update.

[0027] Preferably, the transmission process of the CANopen / EtherCAT communication interface follows a preset strategy that prioritizes control commands;

[0028] The pitch angle control command is encapsulated into a standardized data frame, which includes a uniquely encoded command identifier field, a target pitch angle field encoded with 16-bit precision, a timestamp field recording the command generation time, and a check field generated by a cyclic redundancy check algorithm.

[0029] The instruction identifier field is used to distinguish control instructions for different blades, and the target pitch angle field with 16-bit precision encoding ensures the accuracy of angle control.

[0030] The transmission cycle of control commands is dynamically adjusted based on the wind speed change rate and load fluctuation amplitude.

[0031] Preferably, the operating logic of the pitch actuator is as follows:

[0032] After receiving the pitch angle control command, the pitch actuator first obtains the current actual blade angle through the built-in absolute encoder and Hall sensor dual detection structure. After calculating the deviation between the actual angle and the target angle, it eliminates measurement noise interference through moving average filtering.

[0033] The drive mode is adaptively adjusted according to the magnitude and trend of the deviation value. When the deviation value is large, the high-speed preset drive mode is used, and when the deviation value is small, the preset low-speed adjustment mode is switched. At the same time, the output torque is adjusted according to the current force state of the blade.

[0034] During execution, the actual angle data is fed back to the control terminal in real time through the CANopen / EtherCAT communication interface, and the control terminal dynamically corrects subsequent control commands based on the feedback data.

[0035] Preferably, the online training process of the DRL network follows the following:

[0036] The stored experience samples are managed through a priority experience replay mechanism. The sample priority is determined by the absolute value of the control effect feedback value and the sample diversity coefficient. The sample diversity coefficient is obtained by calculating the Euclidean distance of the sample feature space. The experience samples are stored through a circular buffer structure. When the buffer reaches the storage limit, the low-priority samples are replaced according to the sample priority.

[0037] The network parameter update adopts the mini-batch stochastic gradient descent algorithm. The number of samples in each batch is dynamically adjusted according to the computing resources of the embedded device. After accumulating a preset number of empirical samples, a parameter update is triggered. During the update process, an L2 regularization term is introduced to constrain the magnitude of the network parameters. At the same time, the parameter update stops when the control error index on the validation set no longer decreases.

[0038] Preferably, the calculation formula for the preset reward function is as follows:

[0039] ;

[0040] In the formula: The control effect feedback value is a quantitative indicator that comprehensively evaluates the pitch control accuracy and actuator energy consumption. Weighting coefficients for controlling accuracy; This refers to the control accuracy item, specifically the quantified value of the precision of the pitch angle control. Energy consumption weighting coefficient; This is the energy consumption penalty term, which is the quantified value of energy consumption during the pitch actuator adjustment process;

[0041] ;

[0042] in, Indicates the actual angle of the blade. Indicates the target pitch angle. Indicates the maximum adjustable angle of the propeller blade;

[0043] Based on the preset control priority of the wind turbine unit. Pre-set according to the energy-saving requirements of wind turbine units The value is positively correlated with the square of the actuator drive current and the settling time.

[0044] On the other hand, a controller includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the computer program is executed by the processor, it implements the execution steps of a wind power pitch intelligent control method based on DRL-PID fused nonlinear filtering.

[0045] Compared with the known prior art, the technical solution provided by this invention has the following beneficial effects:

[0046] This invention provides a wind turbine pitch control method based on DRL-PID fusion nonlinear filtering. During execution, this method effectively reduces the impact of noise and nonlinear disturbances by collecting and optimizing multi-dimensional operational data, improving the accuracy of state perception. It dynamically adjusts control parameters through autonomous learning, enabling control commands to precisely adapt to real-time operating conditions and enhancing the flexibility and adaptability of pitch control. A high-efficiency communication transmission strategy ensures rapid and stable transmission of control commands, while an actuator adaptive drive mode and real-time feedback correction mechanism further improve angle adjustment accuracy and response speed. Through scientific sample management and online training, the control effect is continuously optimized. While ensuring stable operation of the wind turbine, this method balances control accuracy and execution energy consumption, reduces operating losses, significantly improves the overall operating efficiency and reliability of the unit, and adapts to the complex and ever-changing wind power operating environment. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 This is a flowchart illustrating a wind turbine pitch control method based on DRL-PID fusion nonlinear filtering. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0050] The present invention will be further described below with reference to embodiments.

[0051] Example:

[0052] This embodiment presents a wind power pitch control method based on DRL-PID fusion nonlinear filtering, such as... Figure 1 As shown, it includes:

[0053] Real-time wind speed and generator speed signals of the wind turbine are collected by wind speed sensor and speed sensor, and blade load estimation value is obtained by load estimator to construct control input state set;

[0054] The load estimator senses the real-time operating status of the wind turbine, integrates the airflow speed signal collected by the wind speed sensor and the rotor rotation signal collected by the speed sensor, and combines the aerodynamic characteristic parameters related to the blade airfoil lift coefficient, drag coefficient and aspect ratio. By coupling the aeroelasticity principle with the real-time operating parameters, a multi-dimensional load assessment model is constructed. Based on the assessment model, a comprehensive blade load estimate value including the load peak, mean and rate of change is output.

[0055] An extended Kalman filter is applied to perform joint estimation processing on the signals in the input state set, suppressing high-frequency noise and nonlinear disturbances, and outputting standardized state-aware data.

[0056] The joint estimation process of the extended Kalman filter is adapted to the nonlinear dynamic characteristics of the wind turbine. The state equation is constructed using real-time wind speed, generator speed, and blade load estimates as the core state vectors. The nonlinear aerodynamic damping term and structural stiffness coefficient are introduced into the equation to adapt the dynamic response of the unit.

[0057] The observation equation is constructed using the raw signals collected by each sensor as the observation vector, and the deviation of the raw signal is corrected by adding the sensor error compensation coefficient.

[0058] During the iteration process, the initial value of the error covariance matrix is ​​preset according to the rated operating parameters of the wind turbine. After each iteration, the weight of the observation vector is dynamically adjusted through the Kalman gain matrix to output standardized state perception data that has been noise suppressed and bias corrected.

[0059] The lightweight deep reinforcement learning network adopts a working condition-adaptive network architecture;

[0060] The input layer dimension is precisely matched with the feature dimension of the standardized state-aware data, and the input data is processed by feature normalization to improve training efficiency.

[0061] The hidden layer adopts a structure that alternates between convolutional layers and fully connected layers. The convolutional layer extracts local features related to the working conditions through the local receptive field mechanism, while the fully connected layer is responsible for fusing global features. At the same time, an adaptive ReLU function is configured as a non-linear activation function to avoid gradient vanishing.

[0062] The output layer outputs the adjustment increment corresponding to the PID parameters through linear mapping, and the built-in output constraint module controls the adjustment increment to always be within the preset range;

[0063] The network as a whole is optimized through targeted pruning, selectively removing redundant connections and low-weight parameters while retaining the core feature extraction path. This reduces computational complexity to adapt to embedded deployment requirements while maintaining high-precision parameter adjustment decision-making capabilities.

[0064] A lightweight deep reinforcement learning network is constructed. Standardized state-aware data is input into the network, which then learns control laws autonomously and outputs PID parameters to adjust the increment.

[0065] Based on the incremental adjustment of PID parameters, the proportional coefficient, integral coefficient and derivative coefficient of the PID controller are dynamically updated to generate a pitch angle control command that is adapted to the current operating conditions.

[0066] When dynamically updating PID controller parameters incrementally based on PID parameters, the following conditions must be met:

[0067] The updated value of the proportional coefficient is the weighted sum of the current proportional coefficient and the corresponding adjustment increment. The weight coefficient is dynamically adjusted according to the magnitude of the current control error. When the error is large, the weight of the adjustment increment is increased, and when the error is small, the weight is decreased.

[0068] The updated value of the integral coefficient is adapted and corrected to the adjustment increment based on the changing trend of the cumulative control error. At the same time, an integral saturation suppression factor is introduced to avoid system overshoot caused by excessive cumulative error.

[0069] The updated value of the differential coefficient is dynamically adjusted based on the magnitude of the error rate of change and the smoothness index, with the weight of the increment increasing when the rate of change fluctuates drastically and decreasing when the fluctuation is gentle.

[0070] Among them, the proportional coefficient, integral coefficient, and derivative coefficient all satisfy the preset upper and lower limit constraints of parameters and the proportional relationship constraints between parameters after the update.

[0071] The pitch angle control command is transmitted to the pitch actuator through the CANopen / EtherCAT communication interface of the embedded intelligent controller, driving the blades to perform the angle adjustment action;

[0072] The transmission process of the CANopen / EtherCAT communication interface follows a preset strategy that prioritizes control commands.

[0073] The pitch angle control command is encapsulated into a standardized data frame. The data frame includes a uniquely coded command identifier field, a target pitch angle field with 16-bit precision encoding, a timestamp field recording the command generation time, and a check field generated by a cyclic redundancy check algorithm.

[0074] The instruction identifier field is used to distinguish control instructions for different blades, and the target pitch angle field with 16-bit precision encoding ensures the accuracy of angle control.

[0075] The transmission cycle of control commands is dynamically adjusted according to the wind speed change rate and load fluctuation amplitude. Non-control data, namely status monitoring data and fault diagnosis data, are transmitted in real time and with a fast response by reducing the transmission frequency and priority, so as to avoid communication delays affecting the pitch control effect.

[0076] The control effect feedback value is calculated based on the preset reward function, and the state data, parameter adjustment amount and feedback value are stored as experience samples and used for online training of DRL network.

[0077] The operating logic of the pitch actuator is as follows:

[0078] After receiving the pitch angle control command, the pitch actuator first obtains the current actual blade angle through the built-in absolute encoder and Hall sensor dual detection structure. After calculating the deviation between the actual angle and the target angle, it eliminates measurement noise interference through moving average filtering.

[0079] The drive mode is adaptively adjusted according to the magnitude and trend of the deviation value. When the deviation value is large, the high-speed preset drive mode is used, and when the deviation value is small, the preset low-speed adjustment mode is switched. At the same time, the output torque is adjusted according to the current force state of the blade.

[0080] During execution, the actual angle data is fed back to the control terminal in real time through the CANopen / EtherCAT communication interface, and the control terminal dynamically corrects subsequent control commands based on the feedback data.

[0081] The online training process of DRL networks follows the following rules:

[0082] The stored experience samples are managed through a priority experience replay mechanism. The sample priority is determined by the absolute value of the control effect feedback value and the sample diversity coefficient. The sample diversity coefficient is obtained by calculating the Euclidean distance of the sample feature space. The experience samples are stored through a circular buffer structure. When the buffer reaches the storage limit, the low-priority samples are replaced according to the sample priority.

[0083] The network parameter update adopts the mini-batch stochastic gradient descent algorithm. The number of samples in each batch is dynamically adjusted according to the computing resources of the embedded device. After accumulating a preset number of empirical samples, a parameter update is triggered. During the update process, an L2 regularization term is introduced to constrain the magnitude of the network parameters. At the same time, the parameter update stops when the control error index on the validation set no longer decreases.

[0084] The formula for calculating the preset reward function is:

[0085] ;

[0086] In the formula: The control effect feedback value is a quantitative indicator that comprehensively evaluates the pitch control accuracy and actuator energy consumption. Weighting coefficients for controlling accuracy; This refers to the control accuracy item, specifically the quantified value of the precision of the pitch angle control. Energy consumption weighting coefficient; This is the energy consumption penalty term, which is the quantified value of energy consumption during the pitch actuator adjustment process;

[0087] ;

[0088] in, Indicates the actual angle of the blade. Indicates the target pitch angle. Indicates the maximum adjustable angle of the propeller blade;

[0089] Based on the preset control priority of the wind turbine unit. Pre-set according to the energy-saving requirements of wind turbine units The value is positively correlated with the square of the actuator drive current and the settling time;

[0090] By comprehensively considering the core requirements of wind turbine pitch control through a reward function, and by introducing control accuracy weighting coefficients and energy consumption weighting coefficients, the degree of fit between the actual blade angle and the target angle, as well as the energy consumption during the pitch actuator adjustment process, are quantitatively evaluated. The control accuracy item is accurately calculated based on the maximum adjustable blade angle. At the same time, the square of the actuator drive current and the adjustment time are taken as key influencing factors of the energy consumption penalty item. This not only meets the different control priorities and energy-saving requirements of wind turbines, but also achieves a dynamic balance between control accuracy and actuator energy consumption, so that the reward feedback can comprehensively and accurately reflect the control effect.

[0091] The method described in the above embodiments can accurately capture operating condition information and suppress interference in wind power operation scenarios, dynamically adapt and adjust the blade angle, improve control accuracy and response speed, reduce load fluctuations, reduce execution energy consumption, and ensure stable operation of the unit. At the same time, it can further improve operating efficiency and reliability through continuous optimization, and adapt to the operating requirements under complex wind conditions.

[0092] It should be noted that:

[0093] Example of #DRL-PID parameter update logic (pseudocode):

[0094] state = get_state()

[0095] action=actor_network.predict(state)

[0096] kp+=action[0];ki+=action[1];kd+=action[2]

[0097] pitch_cmd =pid_controller(err,kp,ki,kd)

[0098] send_to_actuator(pitch_cmd)

[0099] reward = compute_reward()

[0100] store_transition(state,action,reward)

[0101] train_network().

[0102] Example 2:

[0103] A controller includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When executed by the processor, the computer program implements the execution steps of a wind power pitch intelligent control method based on DRL-PID fused nonlinear filtering.

[0104] In summary, the methods described in the above embodiments effectively reduce the impact of noise and nonlinear disturbances and improve the accuracy of state perception by collecting and optimizing multi-dimensional operational data during execution. They also dynamically adjust control parameters through autonomous learning, enabling control commands to accurately adapt to real-time operating conditions and enhancing the flexibility and adaptability of pitch control. Furthermore, efficient communication transmission strategies ensure rapid and stable transmission of control commands, while the actuator adaptive drive mode and real-time feedback correction mechanism further improve angle adjustment accuracy and response speed. Scientific sample management and online training continuously optimize the control effect, balancing control accuracy and execution energy consumption while ensuring stable wind turbine operation, reducing operational losses, significantly improving the overall operating efficiency and reliability of the unit, and adapting to the complex and ever-changing wind power operating environment.

[0105] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A wind turbine variable pitch intelligent control method based on DRL-PID fusion nonlinear filtering, characterized in that, The method comprises the following steps: Collecting real-time wind speed and generator speed signals of the wind turbine through wind speed sensors and speed sensors, obtaining blade load estimation values according to a load estimator, and constructing a control input state set; Applying an extended Kalman filter to jointly estimate and process signals in the input state set, suppress high-frequency noise and nonlinear disturbances, and output standardized state perception data; Constructing a lightweight deep reinforcement learning network, inputting the standardized state perception data into the network, autonomously learning control rules by the network, and outputting PID parameter adjustment increments; Based on the PID parameter adjustment increments, dynamically updating the proportional coefficient, integral coefficient and differential coefficient of the PID controller to generate a pitch angle control instruction that adapts to the current working condition; Transmitting the pitch angle control instruction to the pitch actuator through the CANopen / EtherCAT communication interface of the embedded intelligent controller to drive the blade to perform angle adjustment actions; According to a preset reward function, calculating a control effect feedback value, and storing the state data, parameter adjustment amount and feedback value as experience samples for online training of the DRL network.

2. The wind power variable pitch intelligent control method based on DRL-PID fusion nonlinear filtering according to claim 1, characterized in that, The load estimator perceives the real-time running state of the wind turbine, fuses the airflow speed signal collected by the wind speed sensor and the rotor rotation signal collected by the speed sensor, combines the blade airfoil lift coefficient, drag coefficient and aspect ratio related aerodynamic characteristic parameters, constructs a multi-dimensional load evaluation model by coupling the aerodynamic elasticity mechanics principle and the real-time working condition parameters, and outputs comprehensive blade load estimation values including load peak value, average value and change rate based on the evaluation model.

3. The wind power variable pitch intelligent control method based on DRL-PID fusion nonlinear filtering according to claim 1, characterized in that, The joint estimation and processing process of the extended Kalman filter is adapted to the nonlinear dynamic characteristics of the wind turbine, and the real-time wind speed, generator speed and blade load estimation value are used as the core state vector to construct the state equation; The original signals collected by each sensor are used as the observation vector to construct the observation equation; In the iteration process, the initial value of the error covariance matrix is preset according to the rated working condition parameters of the wind turbine, and the weight of the observation vector is dynamically adjusted through the Kalman gain matrix after each iteration to output standardized state perception data after noise suppression and bias correction.

4. The wind power variable pitch intelligent control method based on DRL-PID fusion nonlinear filtering according to claim 1, characterized in that, The lightweight deep reinforcement learning network adopts a working condition adaptive network architecture; The input layer dimension is accurately matched with the feature dimension of the standardized state perception data, and feature normalization is performed on the input data to improve the training efficiency; The hidden layer adopts an alternating connection structure of convolutional layer and fully connected layer, the convolutional layer extracts working condition related local features through the local receptive field mechanism, the fully connected layer is responsible for fusing global features, and an adaptive ReLU function is configured as a nonlinear activation function to avoid gradient disappearance; The output layer outputs the adjustment increment corresponding to the PID parameter through linear mapping, and an output constraint module is built-in to control the adjustment increment to always be within a preset range.

5. The wind power variable pitch intelligent control method based on DRL-PID fusion nonlinear filtering according to claim 1, characterized in that, When the PID controller parameters are dynamically updated based on the PID parameter adjustment increments, the following conditions are met: The update value of the proportional coefficient is the weighted sum of the current proportional coefficient and the corresponding adjustment increment, and the weight coefficient is dynamically adjusted according to the amplitude of the current control error, that is, the adjustment increment weight is increased when the error is large, and the weight is reduced when the error is small. The updated value of the integral coefficient is adapted and corrected according to the trend of the cumulative control error; The updated value of the differential coefficient dynamically adjusts the weight of the adjustment increment according to the amplitude and smoothness of the error change rate, and the weight is increased when the change rate fluctuates sharply and is reduced when the change rate fluctuates gently; The proportional coefficient, the integral coefficient and the differential coefficient after updating all satisfy the preset parameter upper and lower limit constraints and the parameter interval proportionality constraints.

6. The wind power variable pitch intelligent control method based on DRL-PID fusion nonlinear filtering according to claim 1, characterized in that, The transmission process of the CANopen / EtherCAT communication interface complies with the preset strategy of giving priority to control instructions; The pitch angle control instruction is packaged into a standardized data frame, which includes a uniquely encoded instruction identification field, a 16-bit precision encoded target pitch angle field, a timestamp field recording the generation time of the instruction, and a check field generated by a cyclic redundancy check algorithm; The instruction identification field is used to distinguish control instructions for different blades, and the 16-bit precision encoded target pitch angle field ensures the accuracy of angle control. The transmission period of the control instruction is dynamically adjusted according to the wind speed change rate and the load fluctuation amplitude.

7. The wind power variable pitch intelligent control method based on DRL-PID fusion nonlinear filtering according to claim 1, characterized in that, The operation logic of the variable pitch actuator is as follows: After receiving the pitch angle control instruction, the variable pitch actuator first acquires the current actual angle of the blade through the built-in absolute value encoder and the Hall sensor double detection structure, calculates the deviation value between the actual angle and the target angle, and then eliminates measurement noise interference through sliding average filtering processing; According to the size and trend of the deviation value, the driving mode is adjusted adaptively, the high-speed preset driving mode is adopted when the deviation value is large, and the preset low-speed adjustment mode is switched when the deviation value is small, and the output torque is adjusted according to the current stress state of the blade; During the execution process, the actual angle data is fed back to the control end in real time through the CANopen / EtherCAT communication interface, and the control end dynamically corrects the subsequent control instruction according to the feedback data.

8. The wind power variable pitch intelligent control method based on DRL-PID fusion nonlinear filtering according to claim 1, characterized in that, The online training process of the DRL network complies with the following: The stored experience samples are managed through a priority experience replay mechanism, the sample priority is determined by the absolute value of the control effect feedback value and the sample diversity coefficient, the sample diversity coefficient is obtained by calculating the Euclidean distance of the sample feature space, the experience samples are stored through a ring buffer structure, and when the buffer reaches the storage upper limit, low-priority samples are replaced according to the sample priority; The network parameter update adopts a small-batch stochastic gradient descent algorithm, the number of samples in each batch is dynamically adjusted according to the computing resources of the embedded device, the parameter update is triggered once every time a preset number of experience samples are accumulated, and an L2 regularization term is introduced to constrain the network parameter amplitude during the update process. When the control error index on the validation set no longer decreases, the parameter update is stopped.

9. The wind power variable pitch intelligent control method based on DRL-PID fusion nonlinear filtering according to claim 1, characterized in that, The calculation formula of the preset reward function is as follows: ； In the formula: is a control effect feedback value, that is, a quantitative index for comprehensively evaluating the control accuracy of the variable pitch and the energy consumption of the actuator; is a control accuracy weight coefficient; is a control accuracy term, that is, a quantitative value of the accuracy degree of the pitch angle control; is an energy consumption weight coefficient; is an energy consumption penalty term, that is, a quantitative value of the energy consumption in the adjustment process of the variable pitch actuator; ； wherein, represents the actual angle of the blade, represents the target pitch angle, represents the maximum adjustable angle of the blade; According to the wind turbine control priority preset, According to the wind turbine energy saving demand preset, The value is positively related to the square of the actuator driving current and the adjustment time.

10. A controller characterized by comprising: The controller includes a memory, a processor, and a computer program stored on the memory and executable on the processor, which, when executed by the processor, implements the execution steps of the wind power variable pitch intelligent control method based on DRL-PID fusion nonlinear filtering according to any one of claims 1-9.

Images