Control method and system for variable pitch system of wind generating set

Abstract

The invention relates to the technical field of wind driven generators, in particular to a control method and system for a variable pitch system of a wind generating set, and the control method is based on a generative adversarial network architecture comprising a generator and a multi-target membership discriminator and comprises the following steps: collecting operation state data of the wind generating set, and obtaining a multi-dimensional control target; the generator generates a candidate variable pitch control instruction sequence according to the operation state data, and predicts an expected effect vector generated for a multi-dimensional control target after execution of the candidate variable pitch control instruction sequence; and the multi-target membership discriminator receives the expected effect vector, calculates a comprehensive membership score based on a multi-target membership function set corresponding to the multi-dimensional control target, selects the highest comprehensive membership score as a final variable pitch control instruction, and drives a variable pitch system of the wind generating set to execute the variable pitch control instruction. According to the method, the fatigue load of key components of the unit can be remarkably reduced, and global optimal control is realized.

Description

Technical Field

[0001] This invention relates to the field of wind turbine technology, and more specifically, to a control method and system for a pitch control system of a wind turbine generator set. Background Technology

[0002] As the global energy structure transitions towards cleaner and lower-carbon energy, wind power, as a renewable energy source, is experiencing rapid development in both industrial scale and technological level. Larger scale, lighter weight, and greater intelligence have become irreversible design trends for modern wind turbine generators. In this process, the generator's control system, especially the pitch control system as one of the core actuators, plays a crucial role. It is responsible not only for limiting captured energy above rated wind speeds by adjusting the blade pitch angle to ensure stable generator output power, but also for coordinating various tasks such as generator start-up and shutdown, power regulation, speed control, and handling extreme loads under various complex wind conditions. A wind turbine generator itself is a complex nonlinear system encompassing aerodynamics, power electronics, mechanical transmission, and structural mechanics. Therefore, its control objectives naturally exhibit multidimensionality and conflict; for example, the goal of improving power generation efficiency often conflicts with the goals of suppressing structural loads and reducing component fatigue damage. How to design a pitch control strategy that can accurately and intelligently coordinate these inherent contradictions and achieve global performance optimization while ensuring safe operation has become a key issue of continuous focus and in-depth research in the field of wind power technology. Existing pitch control technology still has significant limitations in addressing multi-objective optimization problems. On the one hand, many traditional control methods, such as the classic PID (Proportional-Integral-Derivative) control, are primarily designed around the precise tracking of a single core variable (usually generator speed). For other key performance indicators such as power generation, tower vibration, and main bearing load, they often only passively constrain these indicators by setting hard protection thresholds or introducing auxiliary feedforward and compensation mechanisms. This makes it difficult to actively and flexibly seek optimization and balance among multiple objectives, resulting in suboptimal overall system performance under varying operating conditions. On the other hand, even when some advanced control strategies attempt to incorporate the concept of multi-objective optimization, their evaluation and decision-making mechanisms for comprehensive performance are relatively rudimentary. For example, most of them use a simple linear weighted summation method to integrate various performance indicators. This method not only fails to accurately reflect the nonlinear contribution of each objective to the overall system performance but also ignores the balance of satisfaction distribution among different objectives. It is even more unable to address the strong coupling and conflict between specific targets, such as power generation and gearbox load, which can easily cause the control system to "hesitate" between multiple conflicting targets. This not only sacrifices potential performance improvement space but may also cause unnecessary mechanical vibrations and exacerbate the wear of key components. Summary of the Invention

[0003] The purpose of this invention is to provide a control method and system for a pitch control system of a wind turbine generator set. By constructing a comprehensive membership evaluation model that integrates nonlinear aggregation and dual structured penalties, the pitch control system can perform deeper intelligent trade-offs and online optimization for multi-dimensional conflicting objectives. This significantly reduces the fatigue load on key components of the generator set while increasing power generation, thereby achieving global optimal control.

[0004] This invention is achieved through the following technical solution: A control method for a wind turbine pitch system, based on a generative adversarial network architecture including a generator and a multi-objective membership discriminator, includes the following steps: Collect operating status data of wind turbine generator sets and obtain multi-dimensional control objectives including power generation, component load and grid adaptability; The generator generates candidate pitch control command sequences based on the wind turbine generator's operating status data. For each candidate pitch control command sequence, the generator predicts the expected effect vector on the multi-dimensional control objectives after execution using a rapid prediction sub-model of the generator's status. The multi-objective membership discriminator receives the expected effect vector and calculates the comprehensive membership score based on the multi-objective membership function set corresponding to the multi-dimensional control objective. The comprehensive membership score is used to evaluate the quality of the candidate pitch control command sequence. The command with the highest overall membership score is selected as the final pitch control command. The target pitch angle value is generated and sent to the pitch drive system for execution, thereby adjusting the blade pitch angle of the wind turbine generator and completing the precise control of the wind turbine generator pitch system.

[0005] Optionally, obtaining the multidimensional control target specifically includes: Based on the real-time wind speed and unit operation mode in the wind turbine generator operation status data, the reference power value at the current moment is determined from the preset wind speed-power optimal curve and is represented as the power generation. Real-time monitoring values ​​of tower bottom bending moment, blade root flapping moment and gearbox torque are obtained from load sensors installed on key components. Short-term limit thresholds are set based on the unit design limit load database. A set of constraints consisting of the limit thresholds and the fluctuation variance of the loads of the monitored key components is defined as the load of the components. The real-time values ​​of grid frequency and grid connection point voltage are obtained from the grid monitoring unit. When a frequency deviation or voltage deviation is detected, the required active power adjustment and reactive current injection are calculated according to the preset droop coefficient and grid connection guidelines to form a grid support demand vector, which is characterized as the grid adaptability.

[0006] Optionally, the generator is constructed as follows: Construct an encoder-decoder structure based on a long short-term memory network; The encoder consists of at least one LSTM layer, which receives a time series of multivariate running state data within a past time window as input and compresses the time series of multivariate running state data into a fixed-length context vector. The decoder consists of at least one LSTM layer, whose input at each time step includes the context vector, a random noise vector sampled from a standard normal distribution, and the pitch angle generated in the previous time step. It generates a pitch control command sequence in the future control time domain through autoregression, which is characterized as a candidate pitch control command sequence.

[0007] Optionally, the generation of candidate pitch control command sequences specifically includes: The time series of multivariate running state data at a set sampling time is constructed into an input tensor; The input tensor is fed into the encoder part of the constructed generator, and after forward propagation, the context vector is output at the last time step. To generate K distinct candidate sequences, execute K loops. In the k-th loop: Generate a random noise vector; The context vector and random noise vector are used as the initial input to the decoder to initiate the autoregressive generation process; In the future control time domain, the decoder generates the pitch angle at step t and uses it as the input at step t+1 until a complete candidate instruction sequence is generated, and finally obtains the candidate pitch control instruction sequence.

[0008] Optionally, the specific construction process of the unit status rapid prediction sub-model is as follows: A feedforward neural network is constructed as the prediction sub-model. Its input layer receives the concatenated vector of the current unit state vector and the candidate pitch control command sequence, and the output layer outputs the expected effect vector that matches the multi-dimensional control target dimension. Offline, using the wind turbine generator simulation module, a training dataset consisting of initial state, control sequence, and final state effect is formed by executing multiple candidate pitch control command sequences of different forms. Wherein, the component of the initial state is the current operating state data of the wind turbine generator set, the control sequence is the candidate pitch control command sequence, and the final state effect is the expected effect vector, whose component is the specific value output by the wind turbine generator set simulation module corresponding to the multi-dimensional control target after the control sequence is executed in the initial state. A supervised learning approach is adopted, using the initial state and control sequence in the training dataset as input and the final state effect as label, and a feedforward neural network is trained through the backpropagation algorithm to form the unit state fast prediction sub-model.

[0009] Optionally, the multi-target membership discriminator is constructed as follows: A feedforward neural network consisting of multiple fully connected layers is constructed as a discriminator. Define an input layer, wherein the number of neurons in the input layer is the same as the dimension of the expected effect vector; Define a hidden layer, which includes a hidden layer with the LeakyReLU activation function, for non-linear feature extraction of the input expected effect vector; Define an output layer consisting of a single neuron and employing a sigmoid activation function to ensure that the overall membership score of its output is a scalar value between 0 and 1.

[0010] Optionally, the multi-target membership discriminator specifically aims to maximize the discriminator's ability to distinguish between real and fake expected effect vectors, and its calculation formula is as follows:

[0011] in, Let be the objective function. For mathematical expectation, This represents the true expected effect vector. The data distribution represents the true expected effect vector. The total number of multidimensional control targets. To control the index of the target, For the first The dynamic weights corresponding to each control objective. For scalar hyperparameters, For the discriminator's internal processing of the expected effect vector, the first... The intermediate evaluation score given by each target component. As a scalar penalty factor, This is the vector of the fake expected effect generated by the candidate pitch control command sequence generated by the generator. The data distribution for the fake expected effect vector is determined by the generator.

[0012] Optionally, the calculation of the comprehensive membership score is specifically as follows: For each effect component in the given expected effect vector, the corresponding individual membership function is called to calculate and output the individual satisfaction score. The weighted sum of the individual satisfaction scores of the group is then obtained using dynamic weights. The individual satisfaction scores, dynamic weights, and nonlinear aggregation adjustment factors of this group are input into a preset nonlinear aggregation function. Through a combination of exponential transformation, weighted summation, and logarithmic transformation, the baseline aggregated satisfaction is calculated and output. The individual satisfaction scores, weighted arithmetic mean, dynamic weights, and directional interaction weights of this group are input into the structured penalty module in parallel to calculate the global dispersion penalty value that quantifies the overall goal dispersion and the directional interaction conflict penalty value that quantifies the inconsistency between specific goals. The total penalty value is obtained by multiplying the global dispersion penalty value and the targeted interaction conflict penalty value by their respective penalty coefficients and summing them. Then, the total penalty value is subtracted from the value 1 and compared with 0 to generate the comprehensive penalty factor. The benchmark aggregated satisfaction is multiplied by the comprehensive penalty factor, and the product is output as the final result, namely the comprehensive membership score.

[0013] Optionally, the formula for calculating the comprehensive membership score is as follows:

[0014] in, This represents the final comprehensive membership score corresponding to the expected effect vector. For the first The satisfaction score is calculated using the individual membership function of each objective. This is the weighted arithmetic mean of the satisfaction scores for all individuals. It is a nonlinear aggregation adjustment factor. This is the global dispersion penalty coefficient. The penalty coefficient for targeted interaction conflicts. For targeted interaction weights, For the first The satisfaction score is calculated using the individual membership function of each objective.

[0015] A control system for a wind turbine pitch system is configured to execute a control method for a wind turbine pitch system.

[0016] The technical solution of the present invention has at least the following advantages and beneficial effects: This invention establishes a unified multidimensional performance evaluation framework. Instead of viewing each control objective in isolation, it introduces the concept of a comprehensive membership score, mapping multiple sub-objectives, including power generation, unit load, and component lifespan, to a single scalar that comprehensively reflects the current operational health of the unit. This allows the control system to make decisions from a more macroscopic and fundamental perspective. Furthermore, the comprehensive membership score function constructed in this invention employs a nonlinear aggregation function to calculate baseline satisfaction, intelligently amplifying the contribution of high-scoring items and incentivizing the controller to optimize towards excellence rather than maintaining average performance. On the other hand, it innovatively introduces a dual structured penalty mechanism. This mechanism not only penalizes the overall imbalance in the achievement of all objectives, ensuring the system does not become unbalanced, but also targets and penalizes conflicts between pre-identified key objectives, such as power generation and tower load. This achieves precise reconciliation of core contradictions, enabling the final control decision to pursue higher overall benefits while ensuring stable and harmonious unit operation, significantly improving power generation efficiency and extending unit lifespan. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the control method for a wind turbine pitch system provided by the present invention. Figure 2 This is a schematic diagram of the control system for a wind turbine pitch system provided by the present invention. Detailed Implementation

[0018] 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 embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0019] like Figure 1 As shown, embodiments of the present invention provide a control method and corresponding control system for a wind turbine generator pitch system. The core idea of ​​this method is to predict and comprehensively evaluate the future unit state that a series of candidate pitch actions could trigger within each control cycle, thereby finding a dynamically optimal balance point among numerous conflicting performance objectives (such as power generation, structural load, and operational stability), and executing corresponding control actions. This method can significantly improve the overall operational efficiency, stability, and safety of wind turbine generators under complex and variable wind conditions.

[0020] like Figure 2As shown, one embodiment of the present invention relates to a pitch control system for a wind turbine generator. This system is integrated into the main controller of the wind turbine generator or a dedicated high-performance edge computing unit. Physically, the system includes, but is not limited to, the following components: Data acquisition module: Composed of a series of sensors deployed at various key locations within the wind turbine generator set. These sensors are responsible for real-time monitoring of the unit's operating status. For example, the anemometer and wind vane on the top of the nacelle measure the wind speed and direction of the incoming wind; the rotary encoder or tachometer motor at the generator end measures the generator's rotational speed; the power transmitter measures the generator's output power; strain gauges and accelerometers at the blade roots, hub, main shaft, and tower measure the structural loads and vibrations of key components; and encoders acquire the current pitch angle of each of the three blades.

[0021] Central Processing Unit: This can be a high-performance programmable logic controller (PLC), an industrial computer (IPC), or an embedded system with a powerful computing core (such as a CPU, GPU, or dedicated AI chip). This unit is the computing core that enables the control method of this invention to operate, and is responsible for performing all data processing, model prediction, optimization decision-making, and instruction generation tasks, which will be detailed later.

[0022] Control command execution module: This mainly refers to the pitch drive system of the wind turbine generator. This system receives pitch angle setting commands from the central processing unit and drives the electric motor or hydraulic system at the root of each blade to precisely rotate the blade to the target angle.

[0023] Data storage unit: used to store various model files, parameter configurations, historical running data, and the algorithm program itself required by the method of this invention.

[0024] At the software level, a sophisticated control algorithm runs in the central processing unit. This program is executed cyclically at a fixed high frequency (e.g., 10 to 50 times per second, i.e., a control frequency of 10Hz-50Hz), ensuring rapid response to changes in wind conditions and continuous optimized control of the unit's status.

[0025] Detailed implementation process of the control method: Step 1: Comprehensive perception of the unit's real-time operating status.

[0026] At the start of each control cycle, for example, at time t, the control system's data acquisition module synchronously locks onto and reads a series of key sensor data, thereby constructing a multi-dimensional state vector x(t) that comprehensively describes the current operating state of the wind turbine generator. This vector forms the basis for all subsequent predictions and decisions. A typical state vector x(t) contains at least the following data components: Wind condition information: such as average wind speed, wind direction, and turbulence intensity at hub height. Wind condition information can be measured directly by anemometers on the nacelle, or more predictive wind field information can be obtained by combining it with feedforward measurement equipment such as lidar.

[0027] Aerodynamic and mechanical conditions: including current generator speed, generator output power, average pitch angle of the three blades, yaw azimuth angle, etc.

[0028] Structural load information: used to reflect the health status of the unit, including but not limited to the acceleration at the top of the tower or the tower top displacement calculated therefrom, the bending moment of the tower base, the torque and bending moment of the main shaft, and the bending moment in the flapping and oscillation direction at the blade root, etc.

[0029] All these raw sensor signals undergo necessary preprocessing before being synthesized into a state vector. High-frequency noise is filtered out using a low-pass filter, and state estimation is performed using a Kalman filter to obtain smoother and more accurate state values. Finally, the processed data is organized into a vector with fixed dimensions and consistent order, such as x(t) = [wind speed, rotational speed, power, blade pitch angle, tower top acceleration, ...].

[0030] Step 2: Generation of candidate pitch control action sets.

[0031] After obtaining the precise state x(t) of the current unit, the control system does not directly calculate a single control command. Instead, it first generates a set of candidate pitch control actions containing multiple possible control options. The purpose is to proactively explore the potential future impacts of different control decisions, providing a basis for subsequent optimal decision-making.

[0032] The candidate action set U is typically a series of small adjustments around the current pitch angle. Understandably, if the current pitch angle is 12.0 degrees, the control system can generate a candidate adjustment set as follows: U = {-0.5°, -0.3°, -0.1°, 0.0°, +0.1°, +0.3°, +0.5°}. These adjustments represent the various possibilities for increasing the pitch angle to reduce energy capture, maintaining it unchanged, or decreasing the pitch angle to increase energy capture in the next control cycle.

[0033] In this embodiment, the range and fineness of the candidate set can be dynamically adaptive. As an example, when the wind speed is stable and the unit is operating well, a finer adjustment range (such as ±0.1°) can be used to achieve fine adjustment; while when a sharp change in wind speed or large unit vibration is detected, the range of the adjustment range can be appropriately expanded (such as ±1.0°) to allow the system to make a more decisive and larger-amplitude response to ensure safety.

[0034] Step 3: Predict future effects based on the predictive model.

[0035] The control system will sequentially extrapolate each candidate action u in the candidate action set U generated in the second step. Specifically, the control system will use a state prediction model f to calculate what the key performance indicators of the unit will be like at the end of the next control cycle or in the next few seconds if a specific candidate pitch action u is executed.

[0036] The input to the state prediction model f is the current state x(t) and the candidate action u, and its output is the expected effect vector y.

[0037] The expected effect vector y contains the predicted values ​​of all key performance indicators used to evaluate the performance of the generator unit. A typical expected effect vector y contains the following components: expected generator output power, expected generator speed, a measure of expected tower fatigue load or ultimate load (e.g., the magnitude of tower base bending moment), the magnitude of expected blade root bending moment, expected drive train torque or gearbox load, and the intensity of the blade motion corresponding to the action (e.g., pitch rate).

[0038] Regarding the specific implementation of the state prediction model f, embodiments of the present invention may employ one or a combination of the following two technical approaches: Path A: A simplified physical model based on mechanisms.

[0039] This approach employs a simplified mathematical model based on fundamental principles of aerodynamics and structural dynamics. Understandably, a blade element momentum theory (BEM) can be used to construct an aerodynamic model that can quickly calculate the aerodynamic torque acting on the blades based on the input wind speed and candidate blade pitch angle u. Then, combined with a simplified transmission chain model—for example, representing the entire rotating part as a single-mast or double-mast block model with one or two moments of inertia—the generator speed variation can be predicted based on the aerodynamic torque and generator electromagnetic torque. Similarly, simplified tower and blade dynamic models can be established to predict tower vibration and blade load variations. The advantage of this type of model is its strong interpretability and low dependence on historical data; however, the model accuracy depends on the accuracy of the modeling.

[0040] Path B: Data-driven model based on historical data.

[0041] This approach utilizes advanced machine learning techniques, particularly recurrent neural networks (RNNs) or their variants in deep learning, such as long short-term memory networks (LSTM) or gated recurrent units (GRUs). In practice, a large amount of input-output data pairs are first collected from the historical operational database of the wind turbine generator (SCADA data and high-frequency vibration data). The input consists of a sequence of state vectors and control actions over a past time period, while the output is the expected effect vector for a future time step. This massive amount of data is used to train a deep neural network. After training, this network becomes a high-precision nonlinear prediction model f. In actual control, simply inputting the current real-time state x(t) and a candidate action u into this trained network allows for the instantaneous output of the corresponding expected effect vector y within milliseconds. The advantages of this type of model are high accuracy and the ability to capture complex dynamic characteristics that are difficult for physical models to describe, but it requires a large amount of historical data for training.

[0042] By repeating this step for each u in the candidate action set U, the control system will eventually obtain a series of expected effect vectors that correspond one-to-one with the candidate actions.

[0043] Step 4: Calculation of multi-objective comprehensive membership score.

[0044] After obtaining a series of expected effect vectors representing different future possibilities, the control system needs a mechanism to evaluate which future is best. Since the expected effect vector y is multi-dimensional, encompassing multiple conflicting objectives such as power generation, load, and stability, it cannot be simply compared.

[0045] This invention innovatively proposes a comprehensive membership evaluation system. Its goal is to map a multi-dimensional expected effect vector y to a single comprehensive membership score ranging from 0 to 1 through a nonlinear function. The comprehensive membership score intuitively represents the overall satisfaction of the control system with the future state of the unit described by this expected effect vector y; a higher score indicates a more ideal future state.

[0046] The calculation process of the overall membership score can be broken down into the following closely related sub-steps: The system needs to transform the predicted value of each physical dimension in the expected effect vector y into a dimensionless individual satisfaction score ranging from 0 to 1. This transformation is accomplished through an individual membership function pre-defined for each objective. The design of the individual membership function reflects the engineer's expectations of good and bad.

[0047] In this embodiment, the following design examples of individual membership functions are provided: For power generation: This is an ascending-order objective where more is better. Its membership function can be designed as a sigmoid function or a piecewise linear function. Understandably, when the expected power is far below the theoretical maximum power achievable at the current wind speed, the satisfaction level is 0; when the expected power is close to or exceeds the rated power, the satisfaction level is 1; in the intermediate region, the satisfaction level smoothly transitions from 0 to 1. This encourages the system to capture as much energy as possible whenever possible.

[0048] For tower loads: This is a descending objective where less is better. Its membership function can be designed as an inverse S-shaped function. Understandably, when the expected load is very low, the satisfaction level is 1; when the expected load exceeds a certain warning value, the satisfaction level drops rapidly to 0; in the intermediate region, the satisfaction level smoothly transitions from 1 to 0. This makes the system averse to control actions that would induce severe structural vibrations.

[0049] For generator speed: This is an optimal range-bound target that remains stable around a certain value. Its membership function can be designed as a Gaussian function or a trapezoidal function. Understandably, when the expected speed is exactly equal to its rated speed, the satisfaction level is 1; when the expected speed deviates from the rated speed, whether higher or lower, the satisfaction level decreases, ensuring the generator can operate stably at its optimal operating point.

[0050] By applying the corresponding membership function to each component in the expected effect vector y, the system transforms the physical prediction vector into an individual satisfaction score vector.

[0051] After obtaining the individual satisfaction score vector, the system needs to assign a dynamic weight to each objective. The dynamic weight represents the importance of each objective under the current specific operating conditions. Unlike traditional methods that use fixed weights, the weights in this invention are dynamically adjusted based on the real-time operating status of the unit.

[0052] As an example: If the current wind speed is significantly higher than the rated wind speed and the turbulence intensity is very high, the system determines that the primary task is to ensure safety and suppress loads. Therefore, it will significantly increase the weight of targets related to various loads while decreasing the weight of power generation targets.

[0053] If the current wind speed is stable near the rated wind speed and the unit vibration is very small, the system determines that the primary task at present is to accurately track the rated power. Therefore, it will increase the weight of power generation and speed targets, and decrease the weight of load targets.

[0054] Based on the above, the control system can automatically focus its attention on the most critical task at any given time under different operating scenarios.

[0055] Finally, the system will use a composite function to synthesize all individual satisfaction scores and dynamic weights into a final single scalar score, which reflects the encouragement of excellent performance and the punishment of internal imbalance. The calculation formula is as follows:

[0056] in, This represents the final comprehensive membership score corresponding to the expected effect vector. For the first The satisfaction score is calculated using the individual membership function of each objective. This is the weighted arithmetic mean of the satisfaction scores for all individuals. It is a nonlinear aggregation adjustment factor. This is the global dispersion penalty coefficient. The penalty coefficient for targeted interaction conflicts. For targeted interaction weights, For the first The satisfaction score is calculated using the individual membership function of each objective. In this embodiment, the calculation formula can be understood as the product of the benchmark aggregated satisfaction and the comprehensive penalty factor: Calculation of benchmark aggregated satisfaction: The upper part of the calculation formula calculates the weighted average satisfaction after nonlinear amplification. It is a positive non-linear aggregation adjustment factor, which can be set to 1.5 or 2.0. When When the value is greater than 0, this function has an important characteristic: it disproportionately amplifies the contributions of targets with higher satisfaction scores. In other words, it doesn't just calculate a simple weighted average, but rather favors the aggregation of high-scoring items. This incentivizes the controller to seek solutions that make certain performance metrics perform exceptionally well, rather than simply making all metrics passable.

[0057] Calculation of the comprehensive penalty factor: The lower part of the calculation formula is the penalty multiplier, ranging from 0 to 1. The closer the value is to 1, the smaller the penalty; the closer it is to 0, the larger the penalty. It consists of two structured parts: Global dispersion penalty: by This is represented by a standard deviation, which quantifies the dispersion of satisfaction scores for all objectives. Understandably, if an alternative results in high satisfaction for some objectives while low satisfaction for others, the standard deviation will be large, leading to a greater penalty. It is the global dispersion penalty coefficient, which can be set to 0.5 to adjust the penalty for uneven performance, ensuring the comprehensiveness of control decisions and avoiding short-sighted behavior of sacrificing all other objectives for a single goal.

[0058] Targeted interaction conflict penalty: by To embody this, here It is a preset targeted interaction weight, which defines the first... The first goal and the first Are there any known inherent conflicts between the objectives? It's understandable that increasing power generation (i=1) and reducing blade load (j=2) are a typical conflicting pair of objectives. Therefore, significant interaction weights can be assigned. For two unrelated objectives, such as yaw error and gearbox oil temperature, the interaction weight is 0. The directional interaction conflict penalty is calculated by weighting the absolute values ​​of the differences in satisfaction scores among all predefined conflict pairs. If an alternative solution leads to a significant difference in satisfaction scores between a pair of conflicting objectives—for example, power generation satisfaction is 0.9 while load satisfaction is only 0.2—the difference |0.9-0.2|=0.7 will be substantial, resulting in a large penalty. δ is the directional interaction conflict penalty coefficient, which can be set to 1.0 to adjust the penalty intensity for internal conflicts, enabling the controller to intelligently coordinate and reconcile the most critical performance conflicts, finding a harmonious operating point.

[0059] Finally, the calculated baseline aggregated satisfaction is multiplied by the comprehensive penalty factor to obtain the final comprehensive membership score of the expected effect vector y.

[0060] Step 5: Selecting the optimal control action.

[0061] After the fourth step of calculation, the control system calculates a comprehensive membership score for each candidate pitch action u and its corresponding expected effect vector y, thus obtaining a score list.

[0062] The control system selects the candidate action that has obtained the highest comprehensive membership score as the optimal pitch control action for the current control cycle. The selection process is the theoretically best choice that can be found after comprehensively considering future power generation benefits, structural safety, operational stability, target balance, and the degree of reconciliation of core contradictions.

[0063] Step 6: Execution of control commands.

[0064] Finally, the central processing unit applies the optimal pitch control action to the current pitch angle, forming a new target pitch angle value. This target value is then sent to the control command execution module, i.e., the pitch drive system. The pitch drive system precisely controls the drive motors of the three blades, causing the blades to rotate quickly and smoothly to the new target angle.

[0065] This completes the closed-loop control cycle, from state perception to prediction, evaluation, decision-making, and execution. At the next control moment, the entire process restarts with a new unit state, repeating the cycle continuously to achieve continuous and precise online optimization control of the wind turbine generator.

[0066] Through the implementation of the above steps, the wind turbine pitch control method provided by this invention constructs an intelligent decision-making framework that far surpasses traditional PID control. It is no longer a passive response based on error, but rather an active optimization based on future prediction. Its core advantage lies in transforming a high-dimensional, nonlinear, and conflict-ridden control problem into a clear and quantifiable single-objective optimization problem through a cleverly designed comprehensive membership function. This enables the controller to make intelligent decisions that are both forward-looking and cost-effective amidst a multitude of complex objectives, ultimately maximizing the operational efficiency and reliability of the wind turbine throughout its entire lifecycle.

[0067] The above are merely preferred embodiments of the present invention and are not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A control method for a pitch control system of a wind turbine generator set, characterized in that, It is based on a generative adversarial network architecture that includes a generator and a multi-target membership discriminator, and includes the following steps: Collect operating status data of wind turbine generator sets and obtain multi-dimensional control objectives including power generation, component load and grid adaptability; The generator generates candidate pitch control command sequences based on the wind turbine generator's operating status data. For each candidate pitch control command sequence, the generator predicts the expected effect vector on the multi-dimensional control objectives after execution using a rapid prediction sub-model of the generator's status. The multi-objective membership discriminator receives the expected effect vector and calculates a comprehensive membership score based on the multi-objective membership function set corresponding to the multi-dimensional control objective. The comprehensive membership score is used to evaluate the quality of candidate pitch control command sequences. The command with the highest overall membership score is selected as the final pitch control command and driven to execute by the wind turbine pitch system.

2. The control method for a pitch control system of a wind turbine generator set according to claim 1, characterized in that, The acquisition of the multidimensional control target specifically refers to: Based on the real-time wind speed and unit operation mode in the wind turbine generator operation status data, the reference power value at the current moment is determined from the preset wind speed-power optimal curve and is represented as the power generation. Real-time monitoring values ​​of tower bottom bending moment, blade root flapping moment and gearbox torque are obtained from load sensors installed on key components. Short-term limit thresholds are set based on the unit design limit load database. A set of constraints consisting of the limit thresholds and the fluctuation variance of the loads of the monitored key components is defined as the load of the components. The real-time values ​​of grid frequency and grid connection point voltage are obtained from the grid monitoring unit. When a frequency deviation or voltage deviation is detected, the required active power adjustment and reactive current injection are calculated according to the preset droop coefficient and grid connection guidelines to form a grid support demand vector, which is characterized as the grid adaptability.

3. The control method for a pitch control system of a wind turbine generator set according to claim 2, characterized in that, The generator is constructed as follows: Construct an encoder-decoder structure based on a long short-term memory network; The encoder consists of at least one LSTM layer, which receives a time series of multivariate running state data within a past time window as input and compresses the time series of multivariate running state data into a fixed-length context vector. The decoder consists of at least one LSTM layer, whose input at each time step includes the context vector, a random noise vector sampled from a standard normal distribution, and the pitch angle generated in the previous time step. It generates a pitch control command sequence in the future control time domain through autoregression, which is characterized as a candidate pitch control command sequence.

4. The control method for a pitch control system of a wind turbine generator set according to claim 3, characterized in that, The specific steps for generating candidate pitch control command sequences are as follows: The time series of multivariate running state data at a set sampling time is constructed into an input tensor; The input tensor is fed into the encoder part of the constructed generator, and after forward propagation, the context vector is output at the last time step. To generate K distinct candidate sequences, execute K loops. In the k-th loop: Generate a random noise vector; The context vector and random noise vector are used as the initial input to the decoder to initiate the autoregressive generation process; In the future control time domain, the decoder generates the pitch angle at step t and uses it as the input at step t+1 until a complete candidate instruction sequence is generated, and finally obtains the candidate pitch control instruction sequence.

5. The control method for a pitch control system of a wind turbine generator set according to claim 4, characterized in that, The specific construction process of the rapid prediction sub-model of unit status is as follows: A feedforward neural network is constructed as the prediction sub-model. Its input layer receives the concatenated vector of the current unit state vector and the candidate pitch control command sequence, and the output layer outputs the expected effect vector that matches the multi-dimensional control target dimension. Offline, using the wind turbine generator simulation module, a training dataset consisting of initial state, control sequence, and final state effect is formed by executing multiple candidate pitch control command sequences of different forms. Wherein, the component of the initial state is the current operating state data of the wind turbine generator set, the control sequence is the candidate pitch control command sequence, and the final state effect is the expected effect vector, whose component is the specific value output by the wind turbine generator set simulation module corresponding to the multi-dimensional control target after executing the control sequence in the initial state. A supervised learning approach is adopted, using the initial state and control sequence in the training dataset as input and the final state effect as label, and a feedforward neural network is trained through the backpropagation algorithm to form the unit state fast prediction sub-model.

6. The control method for a pitch control system of a wind turbine generator set according to claim 5, characterized in that, The construction process of the multi-target membership discriminator is as follows: A feedforward neural network consisting of multiple fully connected layers is constructed as a discriminator. Define an input layer, wherein the number of neurons in the input layer is the same as the dimension of the expected effect vector; Define a hidden layer, which includes a hidden layer with the LeakyReLU activation function, for non-linear feature extraction of the input expected effect vector; Define an output layer consisting of a single neuron and employing a sigmoid activation function to ensure that the overall membership score of its output is a scalar value between 0 and 1.

7. The control method for a pitch control system of a wind turbine generator set according to claim 6, characterized in that, The multi-objective membership discriminator specifically aims to maximize the discriminator's ability to distinguish between real and fake expected effect vectors. Its calculation formula is as follows: in, Let be the objective function. For mathematical expectation, This represents the true expected effect vector. The data distribution represents the true expected effect vector. The total number of multidimensional control targets. To control the index of the target, For the first The dynamic weights corresponding to each control objective. For scalar hyperparameters, For the discriminator's internal processing of the expected effect vector, the first... The intermediate evaluation score given by each target component. As a scalar penalty factor, This is the vector of the fake expected effect generated by the candidate pitch control command sequence generated by the generator. The data distribution for the fake expected effect vector is determined by the generator.

8. The control method for a pitch control system of a wind turbine generator set according to claim 7, characterized in that, The calculation of the comprehensive membership score is as follows: For each effect component in the given expected effect vector, the corresponding individual membership function is called to calculate and output the individual satisfaction score. The weighted sum of the individual satisfaction scores of the group is then obtained using dynamic weights. The individual satisfaction scores, dynamic weights, and nonlinear aggregation adjustment factors of this group are input into a preset nonlinear aggregation function. Through a combination of exponential transformation, weighted summation, and logarithmic transformation, the baseline aggregated satisfaction is calculated and output. The individual satisfaction scores, weighted arithmetic mean, dynamic weights, and directional interaction weights of this group are input into the structured penalty module in parallel to calculate the global dispersion penalty value that quantifies the overall goal dispersion and the directional interaction conflict penalty value that quantifies the inconsistency between specific goals. The global dispersion penalty value and the targeted interaction conflict penalty value are multiplied by their respective penalty coefficients and summed to obtain the total penalty value. The total penalty value is then subtracted from the value 1 and compared with 0 to generate the comprehensive penalty factor. The benchmark aggregated satisfaction is multiplied by the comprehensive penalty factor, and the product is output as the final result, namely the comprehensive membership score.

9. The control method for a pitch control system of a wind turbine generator set according to claim 8, characterized in that, The formula for calculating the comprehensive membership score is as follows: in, This represents the final comprehensive membership score corresponding to the expected effect vector. For the first The satisfaction score is calculated using the individual membership function of each objective. This is the weighted arithmetic mean of the satisfaction scores for all individuals. It is a nonlinear aggregation adjustment factor. This is the global dispersion penalty coefficient. The penalty coefficient for targeted interaction conflicts. For targeted interaction weights, For the first The satisfaction score is calculated using the individual membership function of each objective.

10. A control system for a pitch system of a wind turbine generator set, characterized in that, It is configured to perform the control method for a wind turbine pitch system as described in any one of claims 1 to 9.

Images