Variable speed and variable pitch control method and system for wind generating set

Abstract

The invention relates to the field of variable-speed variable-pitch control of wind generating sets, and provides a variable-speed variable-pitch control method and system of a wind generating set, which are used for acquiring and fusing wind speed, rotor speed, blade root load, pitch and orientation to form a system state, generating collective pitch for constant speed and constant power, and predicting feed-forward compensation of asymmetric load by short-term load. In cooperation with closed-loop correction driven by a real-time bending moment error, three paths of information and a collective instruction are superposed to synthesize an independent variable pitch instruction, and the instruction simultaneously comprises basic stability control information, information for actively inhibiting a known periodic load, correcting a prediction model error and information for inhibiting unknown random disturbance; fatigue and 1P fluctuation caused by asymmetric loads are solved, dynamic response of a tower top and a transmission chain is remarkably reduced, stable power and rotating speed are considered, compensation accuracy and response speed are kept under the condition that model deviation and execution bandwidth are limited, collective and independent ring coupling is weakened, the availability of a wind generating set is improved, and the service life of the wind generating set is prolonged.

Description

Technical Field

[0001] This invention relates to the field of variable speed and pitch control technology for wind turbine generator sets, and particularly to a variable speed and pitch control method and system for wind turbine generator sets. Background Technology

[0002] As wind turbines evolve towards larger and more gigantic rotors, variable pitch control must not only achieve maximum power capture in low-wind-speed regions and maintain constant power and stable speed in rated areas, but also suppress the structural load deterioration caused by the increased size. Currently, the engineering challenge lies in suppressing asymmetric loads: during rotor rotation, wind shear causes significant differences in wind speed between the upper and lower swept zones, tower shadows cause instantaneous load reduction on the blades as they pass over the tower, and the superimposed uneven turbulence space results in blade flapping moment offsets, dominated by periodic fluctuations of 1P and random disturbances, leading to hub overturning, drivetrain torsional vibration, and fatigue accumulation. Relying solely on collective pitch control to maintain overall steady state is insufficient to achieve load balancing at the individual blade level, becoming a key contradiction limiting lifespan and availability.

[0003] Most existing variable speed pitch control methods employ a PI loop driven by generator / rotor speed errors to generate collective pitch commands, achieving constant power and overspeed suppression in the rated range. For structural loads, common independent pitch paths in the industry mainly fall into two categories: one is feedback PI / PID control based on blade root moment measurements, which performs closed-loop correction by subtracting the average moment from the individual blade moments; the other is 1P / 2P harmonic feedforward or multi-blade coordinate methods based on rotor azimuth angle, achieving periodic compensation by transforming blade and stator coordinates with fixed phase and gain. The former relies on measurement error-driven control, is limited by the bandwidth of the pitch actuator and system time delay, and is prone to phase lag, requiring only post-processing correction, with limited active suppression capability for known periodic loads; the latter is sensitive to operating conditions, with harmonic amplitude and phase drifting with wind shear strength, yaw, and speed changes, making it difficult for fixed parameters to adapt in a timely manner, and the uncertainties of modeled tower shadows and turbulence often lead to compensation mismatch. Furthermore, while predictive solutions can improve feedforward performance to some extent, their large-scale deployment is limited by cost, maintenance, and availability. More importantly, existing solutions often separate the design of collective pitch stability control from that of independent pitch load suppression. They lack a collaborative mechanism based on multi-source time-series data that simultaneously addresses the active suppression of known periodic loads and the robust suppression of unknown random disturbances. This results in coupling interference between independent pitch compensation and collective stability control, making it difficult to simultaneously address voltage drop caused by residual asymmetric loads and power / speed fluctuations. Summary of the Invention

[0004] The present invention aims to solve at least one of the problems existing in the prior art, and to provide a method and system for variable speed and pitch control of wind turbine generator sets.

[0005] One aspect of the present invention provides a variable speed and pitch control method for a wind turbine generator set, the method comprising: Acquire raw timing signals from anemometers, generator encoders, blade root load sensors, pitch drive encoders, and nacelle azimuth sensors; Based on the original time series signal, the time series of the system state vector is calculated. The system state vector includes the current wind speed, rotor speed, flapping moment of each blade, current blade pitch angle and rotor azimuth angle. Generate collective pitch command based on rotor speed; The time series of the system state vector is used to predict the bending moment offset of each blade through the trained load prediction model to obtain the predicted asymmetric load vector. The predicted asymmetric load offset vector is transformed by a proportional controller to calculate the pitch angle increment of each blade to obtain the feedforward independent pitch compensation vector. The real-time load error is calculated based on the flapping moment of each blade, and the real-time load error is used by a PI controller to calculate the pitch angle correction of each blade to obtain a feedback independent pitch correction vector. The collective pitch command, the feedforward independent pitch compensation vector, and the feedback independent pitch correction vector are superimposed and synthesized to obtain the independent pitch command.

[0006] Optionally, the original timing signals include anemometer signals, generator encoder signals, blade root strain signals, pitch drive encoder signals, and nacelle azimuth signals.

[0007] Optionally, a collective pitch command is generated based on the rotor speed, including: Calculate the difference between the rotor speed and the rated speed to obtain the speed error; The speed error is used to generate a collective pitch command via a PI controller.

[0008] Optionally, the time series of the system state vector is passed through the trained load prediction model to predict the bending moment offset of each blade to obtain the predicted asymmetric load vector, including: The system state vectors in the time series of the system state vectors are passed through the encoding module of the trained load prediction model to obtain the system state correlation time series encoded vector. The system state-related time-series encoded vector is passed through the decoding module of the trained load prediction model to obtain the predicted asymmetric load vector composed of the bending moment offsets of each blade.

[0009] Optionally, the encoding module includes an MLP unit and an LSTM unit; The system state vectors in the time series of the system state vectors are passed through the encoding module of the trained load prediction model to obtain the system state correlation time series encoded vector, including: Each system state vector in the time series of the system state vector is input into the MLP unit to obtain the time series of the system state-related feature vectors. The time series of system state-related feature vectors are input into LSTM units to obtain system state-related time-series encoded vectors.

[0010] Optionally, the real-time load error is calculated based on the flapping moment of each blade, and the real-time load error is used by a PI controller to calculate the pitch angle correction of each blade to obtain a feedback independent pitch correction vector, including: Calculate the average value of the flapping moments of multiple blades to obtain the average bending moment; The real-time load error of each blade is obtained by subtracting the flapping moment of each blade from the average bending moment. The real-time load error of multiple blades is passed through a PI controller to obtain a feedback independent pitch correction vector.

[0011] Optionally, the collective pitch command, the feedforward independent pitch compensation vector, and the feedback independent pitch correction vector are superimposed and synthesized to obtain the independent pitch command, including: The position-weighted sum of the collective pitch command, the elements in the feedforward independent pitch compensation vector, and the elements in the feedback independent pitch correction vector is calculated to obtain the independent pitch command.

[0012] In another aspect, the present invention provides a variable speed and pitch control system for a wind turbine generator set, the variable speed and pitch control system comprising: The raw timing signal acquisition module is used to acquire raw timing signals collected from the anemometer, generator encoder, blade root load sensor, pitch drive encoder and nacelle azimuth sensor. The system state vector time series module is used to calculate the time series of the system state vector based on the original time series signal. The system state vector includes the current wind speed, rotor speed, flapping moment of each blade, current pitch angle of each blade and rotor azimuth angle. Collective pitch command generation module, used to generate collective pitch commands based on rotor speed; The module for predicting asymmetric load vectors is used to obtain the predicted asymmetric load vectors by passing the time series of the system state vectors through the trained load prediction model to predict the bending moment offset of each blade. The feedforward independent pitch compensation vector acquisition module is used to convert the predicted asymmetric load offset vector through a proportional controller to calculate the pitch angle increment of each blade to obtain the feedforward independent pitch compensation vector. The feedback independent pitch correction vector acquisition module is used to calculate the real-time load error based on the flapping moment of each blade, and to calculate the pitch angle correction of each blade through the PI controller to obtain the feedback independent pitch correction vector. The independent pitch command acquisition module is used to superimpose and synthesize the collective pitch command, the feedforward independent pitch compensation vector, and the feedback independent pitch correction vector to obtain the independent pitch command.

[0013] Optionally, the module for predicting and obtaining the asymmetric load vector includes: The system state correlation time series coding vector acquisition unit is used to obtain the system state correlation time series coding vector by passing each system state vector in the time series of the system state vector through the coding module of the trained load prediction model. The predictive asymmetric load vector acquisition unit is used to obtain the predicted asymmetric load vector composed of the bending moment offsets of each blade by passing the system state correlation time-series encoded vector through the decoding module of the trained load prediction model.

[0014] Optionally, the feedback independent pitch correction vector acquisition module includes: The average bending moment calculation unit is used to calculate the average value of the flapping bending moments of multiple blades to obtain the average bending moment; The real-time load error calculation unit is used to subtract the flapping moment of each blade from the average bending moment to obtain the real-time load error of each blade. The feedback independent pitch correction vector acquisition unit is used to obtain the feedback independent pitch correction vector by passing the real-time load error of multiple blades through a PI controller.

[0015] Compared to existing technologies, this invention collects and integrates wind speed, rotor speed, blade root load, pitch, and azimuth to form a system state. It generates a collective pitch in the stable layer for constant speed and power, and provides feedforward compensation for asymmetric loads based on short-term load prediction. Combined with closed-loop correction driven by real-time bending moment error, the three information sources and collective commands are weighted by position to synthesize independent pitch commands. The independent pitch commands simultaneously include basic stability control information, information on actively suppressing known periodic loads, and information on correcting prediction model errors and suppressing unknown random disturbances. This solves the fatigue and 1P fluctuations caused by asymmetric loads, significantly reduces the dynamic response of the tower top and drivetrain, balances power and speed stability, and maintains compensation accuracy and response speed under conditions of model bias and execution bandwidth limitations. It also weakens the coupling between the collective and independent loops, improving the availability and lifespan of the wind turbine generator. Attached Figure Description

[0016] One or more embodiments are illustrated by way of example with reference numerals in the accompanying drawings. These illustrations do not constitute a limitation on the embodiments. Elements with the same reference numerals in the drawings are denoted as similar elements. Unless otherwise stated, the figures in the drawings are not to be limited by scale.

[0017] Figure 1 A flowchart of a variable speed and pitch control method for a wind turbine generator set provided in an embodiment of the present invention; Figure 2 This is a data flow diagram of a variable speed and pitch control method for a wind turbine generator set provided in another embodiment of the present invention; Figure 3 A flowchart of a wind turbine generator variable speed and pitch control method provided in another embodiment of the present invention is used to predict the bending moment offset of each blade by passing the time series of the system state vector through a trained load prediction model to obtain the predicted asymmetric load vector. Figure 4 A flowchart of a wind turbine generator variable speed and pitch control method provided in another embodiment of the present invention is shown below, which describes a method for obtaining a system state correlation time-series encoded vector by passing each system state vector in the time series of the system state vector through the encoding module of a trained load prediction model. Figure 5 This is a block diagram of a wind turbine generator variable speed and pitch control system provided in another embodiment of the present invention. Detailed Implementation

[0018] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. 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 described embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0019] Unless otherwise specifically stated, the technical or scientific terms used in the embodiments of this invention should be understood in their ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains. The terms "comprising" or "including," as used in the embodiments of this invention, do not limit the shapes, numbers, steps, actions, operations, components, elements, and / or groups thereof mentioned, nor do they exclude the appearance or addition of one or more other different shapes, numbers, steps, actions, operations, components, elements, and / or groups thereof, or the inclusion of these.

[0020] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention. It should also be understood that, for ease of description, the dimensions of the various parts shown in the drawings are not drawn to actual scale, and techniques, methods, and apparatus known to those skilled in the art may not be discussed in detail; however, where appropriate, the illustrated techniques, methods, and apparatus should be considered part of the specification. In all the examples shown and discussed herein, any other specific example may have different values. It should be noted that similar symbols and letters in the following figures denote similar items; therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.

[0021] In the description of the embodiments of the present invention, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In the embodiments of the present invention, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in a suitable manner in any one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in the embodiments of the present invention, as well as the features of different embodiments or examples.

[0022] Hereinafter, exemplary embodiments according to the present invention will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments of the present invention. It should be understood that the present invention is not limited to the exemplary embodiments described herein.

[0023] Existing variable speed and pitch control systems mostly generate collective pitch based on speed errors, which is insufficient to effectively suppress asymmetric blade loads caused by wind shear and tower shadow superposition turbulence. This leads to coexistence of IP fluctuations and random disturbances, increased fatigue, and power / speed fluctuations. Feedback-based IPC is limited by execution bandwidth and phase lag, and harmonic feedforward is sensitive to operating conditions and coupled with collective stability control, making it difficult to balance steady-state control and load suppression. Therefore, the technical solution of this invention proposes a variable speed and pitch control method for wind turbine generators. Specifically, it first integrates the original time-series signals from the anemometer, generator encoder, blade root load, pitch encoder, and nacelle orientation to form a time series of the system state vector. In the constant power region, the collective pitch is generated by the speed error PI to ensure foundation stability. At the same time, a trained load prediction model is introduced to decode and predict the short-term bending moment offset of each blade, capture the periodic load fluctuations caused by wind shear and tower shadow, and obtain feedforward independent pitch compensation by proportional transformation. Then, the real-time deviation of the relative average value of the bending moment of each blade is used as input, and the independent pitch correction is calculated by PI to offset model errors and unknown disturbances. Finally, the collective pitch, feedforward compensation, and feedback correction are weighted by position to synthesize an independent pitch command. This independent pitch command simultaneously contains information on foundation stability control, active suppression of known periodic loads, and robust suppression of random disturbances, thereby reducing asymmetric loads and 1P vibration, reducing the dynamic response of the tower top and drive train, improving power and speed stability, and extending service life.

[0024] Figure 1 This is a flowchart of a wind turbine generator variable speed and pitch control method according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the data flow in a wind turbine generator variable speed and pitch control method according to an embodiment of the present invention. (In conjunction with...) Figure 1 and Figure 2According to an embodiment of the present invention, a wind turbine generator variable speed and pitch control method includes the following steps: S100, acquiring raw time-series signals collected from an anemometer, generator encoder, blade root load sensor, pitch drive encoder, and nacelle azimuth sensor; S200, calculating the time series of the system state vector based on the raw time-series signals, wherein the system state vector includes the current wind speed, rotor speed, flapping moment of each blade, current pitch angle of each blade, and rotor azimuth angle; S300, generating a collective pitch command based on the rotor speed; S400, applying the time series of the system state vector to a trained load sensor. The prediction model predicts the moment offset of each blade to obtain the predicted asymmetric load vector; S500, the predicted asymmetric load offset vector is converted by a proportional controller to calculate the pitch angle increment of each blade to obtain the feedforward independent pitch compensation vector; S600, the real-time load error is calculated based on the flapping moment of each blade, and the real-time load error is used by a PI controller to calculate the pitch angle correction of each blade to obtain the feedback independent pitch correction vector; S700, the collective pitch command, the feedforward independent pitch compensation vector, and the feedback independent pitch correction vector are superimposed and synthesized to obtain the independent pitch command.

[0025] Specifically, in step S100, raw time-series signals are acquired from the anemometer, generator encoder, blade root load sensor, pitch drive encoder, and nacelle azimuth sensor. It should be understood that a single operating parameter cannot fully characterize the complex dynamic characteristics of a large wind turbine under the combined effects of wind shear, tower shadow effect, and turbulence, especially the asymmetric load state borne by each blade. Therefore, in the technical solution of this invention, raw time-series signals are acquired from the anemometer, generator encoder, blade root load sensor, pitch drive encoder, and nacelle azimuth sensor to construct a multi-dimensional data foundation that can reflect the wind field input, wind turbine response, and component status in real time and synchronously. This provides comprehensive and necessary data input for the subsequent accurate calculation of the system state vector, effective prediction of asymmetric loads, and coordinated generation of control commands, ensuring the control strategy's environmental awareness and self-state recognition capabilities.

[0026] More specifically, in a particular example of the invention, the acquisition of the original timing signals is achieved through a synchronous data acquisition process. Specifically, the original timing signals include anemometer signals, generator encoder signals, blade root strain signals, pitch drive encoder signals, and nacelle azimuth signals. The electrical signal output by the anemometer located on top of the nacelle is sampled and converted into an incoming wind speed value in meters per second. An encoder mounted on the high-speed shaft of the generator generates a pulse sequence; by counting the pulses per unit time and combining this with the transmission chain gear ratio, the rotor speed in revolutions per minute (rpm) is calculated. Strain gauges attached to the roots of the three blades form a measuring bridge; their weak voltage signals are amplified and filtered, and then converted into bending moment values ​​characterizing the flapping load of each blade, in Newton-meters (Nm), based on calibration coefficients. The encoders within the pitch drive mechanism of each blade provide real-time feedback of its angular position, which is interpreted as the current pitch angle of each blade in degrees. The azimuth sensor on the rotor main shaft provides the instantaneous rotation angle of the rotor in degrees, used to identify the specific position of the blade in the plane of rotation. All signals are acquired under a unified clock reference, forming a timestamp-aligned data frame containing wind speed, rotational speed, three-blade bending moment, three-blade pitch angle, and rotor azimuth angle, which is then output as the raw timing signal.

[0027] Specifically, in step S200, based on the original time-series signals, a time series of the system state vector is calculated. The system state vector includes the current wind speed, rotor speed, flapping moments of each blade, current pitch angle of each blade, and rotor azimuth angle. It should be understood that the original time-series signals collected from various sensors are discrete and independent data streams. Directly using them for complex control algorithms makes it difficult to effectively reveal the intrinsic relationships and dynamic evolution patterns between various physical quantities. Therefore, in the technical solution of this invention, a time series of the system state vector is further calculated based on the original time-series signals. The system state vector includes the current wind speed, rotor speed, flapping moments of each blade, current pitch angle of each blade, and rotor azimuth angle. This integrates the multi-source heterogeneous sensor data into a unified, structured mathematical expression that fully describes the operating state of the wind turbine generator at any given time. This provides a standardized, information-intensive input for subsequent load prediction models, facilitating the model's capture of system dynamic characteristics and temporal dependencies, thereby enabling accurate predictions.

[0028] More specifically, in a particular example of the present invention, the calculation of the system state vector and the construction of the time series are performed within each control cycle. At a time point k, the synchronization data frame acquired in the previous embodiment is used to populate a predefined vector structure. Specifically, the average wind speed, rotor speed, flapping moment of the three blades, current pitch angle of the three blades, and rotor azimuth angle at that moment are sequentially arranged and combined to form a system state vector of a defined dimension. The system state vector is stored in a first-in, first-out (FIFO) data buffer, which maintains the system state vectors for the past N control cycles. This forms a time series of system state vectors of length N, which fully records the dynamic evolution trajectory of the wind turbine generator from the past to the present moment, and serves as the input for the next control step.

[0029] Specifically, in step S300, a collective pitch command is generated based on the rotor speed. It should be understood that in high-wind-speed operating conditions above the rated wind speed, the wind energy entering the rotor swept surface exceeds the rated absorption and conversion capacity of the wind turbine generator set. If not controlled, the excess aerodynamic torque will cause the rotor speed to continuously increase, triggering overspeed protection shutdown or damaging the drivetrain and generator. Therefore, in the technical solution of this invention, a collective pitch command is further generated based on the rotor speed to directly manage the overall aerodynamic power captured by the rotor from the wind by uniformly adjusting the pitch angle of the three blades. This establishes a negative feedback control loop aimed at maintaining the rated speed, providing a fundamental guarantee for the constant power and stable operation of the wind turbine generator set in high-wind-speed areas, and forming a superposition benchmark for subsequent independent pitch load suppression commands.

[0030] More specifically, in a specific example of the present invention, generating a collective pitch command based on the rotor speed includes: calculating the difference between the rotor speed and the rated speed to obtain the speed error; and generating a collective pitch command using a PI controller based on the speed error. That is, more specifically, firstly, the rotor speed value at the current moment is extracted from the system state vector and subtracted from the rated speed target value determined during the design of the wind turbine generator set, for example, 10.0 rpm, to calculate the real-time speed error value. If this speed error value is positive, it indicates that the rotor speed at the current moment exceeds the rated speed target, requiring a reduction in aerodynamic torque; if it is negative, it indicates that the rotor speed at the current moment is lower than the rated speed target, requiring an increase. Subsequently, the speed error value is sent to a proportional-integral controller, i.e., a PI controller. The proportional part of the PI controller multiplies the speed error value by a proportional gain to generate an instantaneous pitch angle adjustment proportional to the magnitude of the error, used to quickly respond to speed changes caused by wind speed fluctuations. Simultaneously, the integral part of the PI controller accumulates the speed error over time to eliminate steady-state speed residuals caused by model inaccuracies or persistent wind speed deviations. The outputs of the proportional and integral parts of the PI controller are summed to form a single collective pitch command applicable to all blades. This collective pitch command is in degrees; increasing its value will cause the blades to deflect in the feathering direction, reducing wind energy capture efficiency, and vice versa.

[0031] In step S400, the time series of the system state vector is processed by a trained load prediction model to predict the bending moment offset of each blade, thereby obtaining a predicted asymmetric load vector. It should be understood that control methods relying solely on real-time error feedback have inherent phase lag due to the physical delay of the pitch actuator, making it impossible to proactively suppress asymmetric loads with deterministic periodic characteristics caused by wind shear and tower shadow effects. Therefore, in the technical solution of this invention, the time series of the system state vector, containing information such as wind speed, rotational speed, historical bending moment, pitch angle, and rotor azimuth angle, is further input into a pre-trained load prediction model to analyze the dynamic correlation in the time series data, predict the offset of each blade relative to the average bending moment caused by wind shear and tower shadow effects in the next one or more control cycles, and combine these into a predicted asymmetric load vector. This provides forward-looking information about future periodic load disturbances, making it possible to generate feedforward compensation commands. This transforms passive error correction into active disturbance suppression, overcoming the lag of traditional feedback control and laying the foundation for precise and timely asymmetric load reduction.

[0032] Figure 3 This is a flowchart illustrating the wind turbine generator variable speed and pitch control method according to an embodiment of the present invention, which uses a trained load prediction model to predict the bending moment offset of each blade to obtain a predicted asymmetric load vector. Figure 3 As shown, step S400 includes: S410, passing each system state vector in the time series of the system state vector through the encoding module of the trained load prediction model to obtain the system state correlation time series encoding vector; S420, passing the system state correlation time series encoding vector through the decoding module of the trained load prediction model to obtain the predicted asymmetric load vector composed of the bending moment offset of each blade.

[0033] In step S410, each system state vector in the time series of the system state vector is passed through the encoding module of the trained load prediction model to obtain a system state correlation time series encoded vector. It should be understood that the time series of the system state vector is a high-dimensional data stream, in which the nonlinear correlations and temporal dependencies between various physical quantities are complex, making it difficult to effectively extract key dynamic features when directly used for prediction. Therefore, in the technical solution of this invention, each system state vector in the time series of the system state vector is further passed through the encoding module of the trained load prediction model to extract features and compress information from the input time series data, mapping the high-dimensional original input sequence to a low-dimensional feature space that can characterize the dynamic evolution of the system. This generates a system state correlation time series encoded vector that contains historical state information and the inherent correlations between variables, providing a refined input with higher information density for the subsequent decoding module.

[0034] The encoding modules of the trained load prediction model include MLP units and LSTM units. Figure 4 This is a flowchart illustrating the process of obtaining a system state correlation time-series encoded vector by passing each system state vector in the time series of the system state vectors in the wind turbine generator variable speed and pitch control method according to an embodiment of the present invention through the encoding module of a trained load prediction model. For example... Figure 4 As shown, step S410 includes: S411, inputting each system state vector in the time series of the system state vector into the MLP unit to obtain the time series of the system state correlation feature vector; S412, inputting the time series of the system state correlation feature vector into the LSTM unit to obtain the system state correlation time-series encoding vector.

[0035] In step S411, each system state vector in the time series of the system state vector is input into the MLP unit to obtain a time series of system state-related feature vectors. It should be understood that the effects of various physical quantities in the system state vector, such as wind speed, rotor azimuth angle, and blade pitch angle, on blade bending moment at the same moment are mutually coupled and highly nonlinear. Directly inputting this original combination into the time series model may not fully reveal its instantaneous intrinsic correlation. Therefore, in the technical solution of this invention, each system state vector in the time series of the system state vector is further input into the MLP unit to perform nonlinear transformation and feature extraction on the system state at each time step, mapping the original physical quantities to a feature space that better characterizes their impact on the load. This generates a time series of system state-related feature vectors, where each vector contains a deep feature representation of the original state, thereby enhancing the subsequent time series model's understanding of the state at each time step while preserving temporal information.

[0036] More specifically, in a specific example of the present invention, at a time point k, the system state vector The input is fed as input to the input layer of a multilayer perceptron (MLP) unit. This MLP unit consists of at least one fully connected hidden layer and one output layer, where the neurons in the hidden layers employ non-linear activation functions. When the system state vector... After passing through the input layer, it propagates layer by layer during the forward propagation. In each hidden layer, the input from the previous layer is weighted and summed, a bias term is added, and then a non-linear mapping is performed through an activation function. After transformation through all hidden layers, a feature vector with dimensions that can be different from or the same as the input, associated with the system state, is finally generated at the output layer. This operation takes each system state vector in the input time series and starts from the first system state vector in the time series of the system state vector. arrive This process is repeated. Ultimately, a new time series of the same length as the input sequence, composed of feature vectors associated with the system state, is obtained. This sequence will be used for subsequent processing. This is the first system state-related feature vector in the time series of system state-related feature vectors.

[0037] In step S412, the time series of the system state-related feature vector is input into the LSTM unit to obtain the system state-related temporal encoding vector. It should be understood that although the time series of the system state-related feature vector processed by the MLP unit contains rich feature information at each time step, the inherent temporal dependencies of the sequence itself—that is, how historical states influence the dynamic evolution of future loads—have not been effectively extracted. Therefore, in the technical solution of this invention, the time series of the system state-related feature vector is further input into the LSTM unit to learn and capture long-distance temporal dependencies in the sequence data through the gated recurrent mechanism within the LSTM unit. In this way, the information of the entire time series can be compressed and encoded into a fixed-dimensional system state-related temporal encoding vector, which condenses the historical dynamic information of the system that is crucial for predicting future loads.

[0038] More specifically, in a specific example of the present invention, the encoding process sequentially processes the entire feature vector time series. First, the system state-associated feature vector sequence generated in the previous embodiment is used as input. This sequence is fed into a Long Short-Term Memory (LSTM) network unit step by step from beginning to end. In the first time step, the system state-associated feature vector... The input is processed by the LSTM unit, which updates its internal cell state and generates a hidden state through its input gate, forget gate, and output gate structure. In the next time step, the system state is associated with the feature vector. The hidden state, along with the one generated in the previous step, is fed into the LSTM unit to update the cell state and generate a new hidden state. This process is recursively performed along the time axis, with each step utilizing both the current input features and the hidden state from the previous time step, which contains all historical information. The last system state in the sequence is associated with the feature vector. After processing, the hidden state output by the LSTM unit is used as the system state correlation temporal coding vector. This vector is a comprehensive and compact mathematical representation of the dynamic evolution of the entire input sequence.

[0039] In step S420, the system state-related time-series encoded vector is passed through the decoding module of the trained load prediction model to obtain a predicted asymmetric load vector composed of the bending moment offsets of each blade. It should be understood that the system state-related time-series encoded vector generated by the encoding module is an abstract feature representation of compressed historical dynamic information, and it does not directly provide the specific values ​​of future loads. Therefore, in the technical solution of this invention, the system state-related time-series encoded vector is further passed through the decoding module of the trained load prediction model to decode the abstract time-series features and map them onto specific physical prediction targets, that is, to calculate the bending moment offsets of each blade relative to the average load at future times caused by factors such as wind shear and tower shadow effects. In this way, the model's understanding of system dynamics can be transformed into a clear, quantifiable, multi-dimensional prediction result, namely, a predicted asymmetric load vector composed of the bending moment offsets of each blade, providing direct input for subsequent feedforward compensation control.

[0040] More specifically, in a concrete example of the present invention, the decoding process is implemented through a dedicated decoder for calculating the bending moment offsets of different blades. The system state correlation time-series encoded vector generated in the previous embodiment is used as input and fed into this decoder. This decoder can be a multilayer perceptron structure, where the dimension of the input layer matches the dimension of the system state correlation time-series encoded vector, and the output layer precisely contains three neurons, corresponding to the predicted bending moment offset values ​​of the three blades, respectively. As the system state correlation time-series encoded vector propagates forward, the comprehensive information it contains about the historical operating state of the entire wind turbine is nonlinearly transformed and distributed to the three output neurons. Finally, the output layer produces a vector containing three elements, where the first element is the predicted bending moment offset of blade number one at the next moment, the second element is the predicted bending moment offset of blade number two, and the third element is the predicted bending moment offset of blade number three. The unit of all three elements is Newton-meter, and these three elements together form the predicted asymmetric load vector.

[0041] Specifically, in step S500, the predicted asymmetric load offset vector is converted using a proportional controller to calculate the pitch angle increment of each blade, thereby obtaining a feedforward independent pitch compensation vector. It should be understood that the predicted asymmetric load vector provides quantitative information on future load imbalance, with units of torque, while the control input of the pitch actuator is an angle command. There is a mismatch between the physical domain and units, making it unsuitable for direct control. Therefore, in the technical solution of this invention, the predicted asymmetric load offset vector is further converted using a proportional controller to calculate the pitch angle increment of each blade, thereby establishing a direct mapping from the predicted load domain to the control command domain. The predicted, impending bending moment offset is pre-converted into the active pitch adjustment required to offset this offset. This generates a feedforward independent pitch compensation vector, which provides a pure, prediction-based active suppression component for subsequent command synthesis, achieving precise and forward-looking control of periodic disturbances such as wind shear and tower shadows.

[0042] More specifically, in a concrete example of the present invention, the calculation of the feedforward independent pitch compensation vector is a direct proportional conversion process. First, the predicted asymmetric load vector output in the previous embodiment is obtained. Each element in this vector, i.e., the predicted bending moment offset corresponding to the blade, is multiplied by a pre-set proportional gain. This proportional gain is a conversion coefficient calibrated according to the blade aerodynamic characteristics, and its physical meaning is the inverse of the flapping bending moment change caused by a unit pitch angle change. After multiplying the three elements in the predicted asymmetric load vector by this proportional gain and performing calculations, a new three-dimensional vector is obtained, which is the feedforward independent pitch compensation vector.

[0043] Specifically, in step S600, the real-time load error is calculated based on the flapping moment of each blade, and the real-time load error is used by a PI controller to calculate the pitch angle correction of each blade to obtain a feedback independent pitch correction vector. It should be understood that the predicted asymmetric load vector provides quantitative information on future load imbalance, with its unit being torque, while the control input of the pitch actuator is an angle command; there is a mismatch between the physical domain and units. Therefore, in the technical solution of this invention, the predicted asymmetric load offset vector is further converted by a proportional controller to calculate the pitch angle increment of each blade, thereby establishing a direct mapping from the predicted load domain to the control command domain. The predicted, impending moment offset is pre-converted into the active pitch adjustment required to offset the offset. This generates a feedforward independent pitch compensation vector, which provides a pure, prediction-based active suppression component for subsequent command synthesis, achieving precise and forward-looking control of periodic disturbances such as wind shear and tower shadows.

[0044] More specifically, in a specific example of the present invention, the real-time load error is calculated based on the flapping moment of each blade, and the real-time load error is used by a PI controller to calculate the pitch angle correction of each blade to obtain a feedback independent pitch correction vector. This includes: calculating the average value of the flapping moments of multiple blades to obtain the average moment; subtracting the flapping moment of each blade from the average moment to obtain the real-time load error of each blade; and using the real-time load errors of multiple blades by a PI controller to obtain a feedback independent pitch correction vector.

[0045] Accordingly, the average value of multiple blade flapping moments is calculated to obtain the average bending moment; the difference between each blade flapping moment and the average bending moment is calculated to obtain the real-time load error of each blade. It should be understood that the prediction model itself has inherent accuracy limitations and cannot anticipate sudden, unmodeled random turbulent disturbances. Simply relying on feedforward control is insufficient to completely eliminate asymmetric loads, resulting in residual errors. Therefore, in the technical solution of this invention, the average value of multiple blade flapping moments is further calculated to obtain the average bending moment, and the difference between each blade flapping moment and the average bending moment is calculated to obtain the real-time load error of each blade. This separates the asymmetric component from the total measured blade load, which is a direct indicator of the degree of impeller load imbalance at the current moment. This provides an immediate, quantified error driving signal for subsequent feedback control, enabling rapid and effective compensation for deficiencies in the prediction model and unknown disturbances, thereby significantly improving the robustness of the entire control strategy and the final accuracy of load suppression.

[0046] Accordingly, the real-time load errors of multiple blades are processed by a PI controller to obtain a feedback independent pitch correction vector. It should be understood that the calculated real-time load error is merely a quantitative description of the state deviation and cannot be directly used as a control command for the pitch mechanism; a dynamic adjustment mechanism is needed to convert it into a specific execution action. Therefore, in the technical solution of this invention, the real-time load errors of multiple blades are further processed by a PI controller to generate a pitch angle adjustment that can effectively drive the error to converge to zero, based on the current error magnitude and historical accumulation. This yields a feedback independent pitch correction vector, which provides real-time, measurement-based closed-loop correction for the final control command. This compensates for residual errors in feedforward predictions and suppresses unknown disturbances such as unmodeled random turbulence, thereby enhancing the robustness and adaptability of the entire control strategy.

[0047] More specifically, in a concrete example of the present invention, the generation process of the feedback independent pitch correction vector is performed as follows. The vector consisting of three real-time load error values ​​obtained in the previous embodiment is input to a set of parallel proportional-integral (PI) controllers, with each error component corresponding to an independent PI control channel. Within each PI control channel, the current load error is first multiplied by a proportional gain to generate an instantaneous pitch angle correction proportional to the error amplitude, used to quickly respond to the current load imbalance. Simultaneously, the load error is also fed into an integrator for accumulation, and the integration result is multiplied by an integral gain to generate a correction amount used to eliminate long-term steady-state load deviations. Subsequently, the proportional and integral term outputs of each PI control channel are summed to obtain the final pitch angle correction value for the blade. Finally, the outputs of the three PI control channels are combined into a three-dimensional feedback independent pitch correction vector.

[0048] Specifically, in step S700, the collective pitch command, the feedforward independent pitch compensation vector, and the feedback independent pitch correction vector are superimposed and synthesized to obtain an independent pitch command. It should be understood that the collective pitch command, the feedforward independent pitch compensation vector, and the feedback independent pitch correction vector are independent control components generated in parallel to achieve different control objectives (global stability, active prediction and suppression, and real-time error correction, respectively). They must be effectively fused into a unified, executable final command. Therefore, in the technical solution of this invention, the collective pitch command, the feedforward independent pitch compensation vector, and the feedback independent pitch correction vector are further superimposed and synthesized to generate a unique composite command containing multiple control intentions for the pitch actuator of each blade. This ensures that the final output independent pitch command meets the overall power and speed stability requirements of the wind turbine generator set while simultaneously achieving refined, forward-looking, and robust suppression of asymmetric loads on each blade, solving the command fusion problem in multi-objective control.

[0049] More specifically, in a specific example of the present invention, the collective pitch command, the feedforward independent pitch compensation vector, and the feedback independent pitch correction vector are superimposed and synthesized to obtain the independent pitch command, including: calculating the position-weighted sum between the collective pitch command, each element in the feedforward independent pitch compensation vector, and each element in the feedback independent pitch correction vector to obtain the independent pitch command.

[0050] In summary, the variable speed and pitch control method for wind turbine generators according to embodiments of the present invention is explained. It collects and fuses wind speed, rotor speed, blade root load, pitch, and azimuth to form the system state, generates a collective pitch for constant speed and power in the stable layer, and simultaneously provides feedforward compensation for asymmetric loads using short-term load prediction. Combined with closed-loop correction driven by real-time bending moment error, the three information sources and the collective command are weighted by position to synthesize an independent pitch command. This independent pitch command simultaneously includes basic stability control information, information on actively suppressing known periodic loads, and information on correcting prediction model errors and suppressing unknown random disturbances. This solves the fatigue and 1P fluctuations caused by asymmetric loads, significantly reduces the dynamic response of the tower top and drivetrain, balances power and speed stability, maintains compensation accuracy and response speed under conditions of model bias and execution bandwidth limitations, weakens the coupling between the collective and independent loops, and improves the availability and lifespan of the wind turbine generator.

[0051] This invention also provides a variable speed and pitch control system for wind turbine generator sets.

[0052] Figure 5 This is a block diagram of a wind turbine generator variable speed and pitch control system according to an embodiment of the present invention. Figure 5 As shown, the wind turbine generator variable speed and pitch control system 100 according to an embodiment of the present invention includes: a raw timing signal acquisition module 110, used to acquire raw timing signals collected from an anemometer, generator encoder, blade root load sensor, pitch drive encoder, and nacelle azimuth sensor; a system state vector time sequence module 120, used to calculate the time sequence of the system state vector based on the raw timing signals, wherein the system state vector includes the current wind speed, rotor speed, flapping moment of each blade, current pitch angle of each blade, and rotor azimuth angle; a collective pitch command generation module 130, used to generate a collective pitch command based on the rotor speed; and a predicted asymmetric load vector acquisition module 140, used to train the time sequence of the system state vector. The post-training load prediction model predicts the bending moment offset of each blade to obtain the predicted asymmetric load vector; the feedforward independent pitch compensation vector acquisition module 150 is used to convert the predicted asymmetric load offset vector through a proportional controller to calculate the pitch angle increment of each blade to obtain the feedforward independent pitch compensation vector; the feedback independent pitch correction vector acquisition module 160 is used to calculate the real-time load error based on the flapping moment of each blade, and to calculate the pitch angle correction of each blade through a PI controller to obtain the feedback independent pitch correction vector; the independent pitch command acquisition module 170 is used to superimpose and synthesize the collective pitch command, the feedforward independent pitch compensation vector, and the feedback independent pitch correction vector to obtain the independent pitch command.

[0053] Specifically, the asymmetric load vector acquisition module 140 includes: a system state correlation time-series encoding vector acquisition unit, used to obtain a system state correlation time-series encoding vector by passing each system state vector in the time series of the system state vector through the encoding module of the trained load prediction model; and an asymmetric load vector acquisition unit, used to obtain a predicted asymmetric load vector composed of the bending moment offsets of each blade by passing the system state correlation time-series encoding vector through the decoding module of the trained load prediction model.

[0054] Specifically, the feedback independent pitch correction vector acquisition module 160 includes: an average bending moment calculation unit, used to calculate the average value of the flapping bending moments of multiple blades to obtain the average bending moment; a real-time load error calculation unit, used to subtract the average bending moment from the flapping bending moment of each blade to obtain the real-time load error of each blade; and a feedback independent pitch correction vector acquisition unit, used to obtain the feedback independent pitch correction vector by passing the real-time load error of multiple blades through a PI controller.

[0055] The wind turbine generator variable speed and pitch control system 100 according to embodiments of the present invention can be deployed in the edge computing unit at the wind turbine site, such as the main controller or dedicated industrial control computer deployed inside the wind turbine nacelle or tower base, and communicate in real time with the wind turbine's pitch drive system and SCADA monitoring system. In one possible implementation, the wind turbine generator variable speed and pitch control system 100 according to embodiments of the present invention can be integrated into the main control system of the wind turbine generator as a software module or hardware module. For example, core components in the wind turbine generator variable speed and pitch control system 100, such as the trained load prediction model and its network weights, as well as the PI controller parameters for collective pitch and feedback correction, can be trained offline, simulated, verified, and optimized on the back-end server of the wind farm control center. The optimized model and parameter package can then be sent to the front-end controller. Similarly, the complete algorithm for real-time control in the wind turbine generator variable speed and pitch control system 100, including the acquisition of multi-source time-series signals, the construction of system state vectors, forward inference of the load prediction model, calculation of real-time load errors, and the final superposition and synthesis of independent pitch commands, can also be embedded in dedicated industrial computing hardware, such as an embedded processor or AI acceleration module within the wind turbine main controller. This accelerates the processing of real-time data and the computation of the prediction model, ensuring low-latency output of the final independent pitch commands.

[0056] It is understood that the above embodiments are merely exemplary implementations used to illustrate the principles of the present invention, and the present invention is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A method for variable speed and pitch control of a wind turbine generator set, characterized in that, The wind turbine generator set variable speed and pitch control method includes: Acquire raw timing signals from anemometers, generator encoders, blade root load sensors, pitch drive encoders, and nacelle azimuth sensors; Based on the original time series signal, the time series of the system state vector is calculated. The system state vector includes the current wind speed, rotor speed, flapping moment of each blade, current blade pitch angle and rotor azimuth angle. Generate collective pitch command based on rotor speed; The time series of the system state vector is used to predict the bending moment offset of each blade through the trained load prediction model to obtain the predicted asymmetric load vector. The predicted asymmetric load offset vector is transformed by a proportional controller to calculate the pitch angle increment of each blade to obtain the feedforward independent pitch compensation vector. The real-time load error is calculated based on the flapping moment of each blade, and the real-time load error is used by a PI controller to calculate the pitch angle correction of each blade to obtain a feedback independent pitch correction vector. The collective pitch command, the feedforward independent pitch compensation vector, and the feedback independent pitch correction vector are superimposed and synthesized to obtain the independent pitch command.

2. The wind turbine generator set variable speed and pitch control method according to claim 1, characterized in that, The original timing signals include anemometer signals, generator encoder signals, blade root strain signals, pitch drive encoder signals, and nacelle azimuth signals.

3. The wind turbine generator set variable speed and pitch control method according to claim 1, characterized in that, Generate collective pitch commands based on rotor speed, including: Calculate the difference between the rotor speed and the rated speed to obtain the speed error; The speed error is used to generate a collective pitch command via a PI controller.

4. The wind turbine generator set variable speed and pitch control method according to claim 1, characterized in that, The time series of the system state vector is used to predict the bending moment offset of each blade through a trained load prediction model to obtain the predicted asymmetric load vector, including: The system state vectors in the time series of the system state vectors are passed through the encoding module of the trained load prediction model to obtain the system state correlation time series encoded vector. The system state-related time-series encoded vector is passed through the decoding module of the trained load prediction model to obtain the predicted asymmetric load vector composed of the bending moment offsets of each blade.

5. The wind turbine generator set variable speed and pitch control method according to claim 4, characterized in that, The encoding module includes MLP units and LSTM units; The system state vectors in the time series of the system state vectors are passed through the encoding module of the trained load prediction model to obtain the system state correlation time series encoded vector, including: Each system state vector in the time series of the system state vector is input into the MLP unit to obtain the time series of the system state-related feature vectors. The time series of system state-related feature vectors are input into LSTM units to obtain system state-related time-series encoded vectors.

6. The wind turbine generator set variable speed and pitch control method according to claim 1, characterized in that, The real-time load error is calculated based on the flapping moment of each blade, and the real-time load error is used by a PI controller to calculate the pitch angle correction of each blade to obtain a feedback independent pitch correction vector, including: Calculate the average value of the flapping moments of multiple blades to obtain the average bending moment; The real-time load error of each blade is obtained by subtracting the flapping moment of each blade from the average bending moment. The real-time load error of multiple blades is passed through a PI controller to obtain a feedback independent pitch correction vector.

7. The wind turbine generator set variable speed and pitch control method according to claim 1, characterized in that, The collective pitch command, the feedforward independent pitch compensation vector, and the feedback independent pitch correction vector are superimposed and synthesized to obtain the independent pitch command, including: The position-weighted sum of the collective pitch command, the elements in the feedforward independent pitch compensation vector, and the elements in the feedback independent pitch correction vector is calculated to obtain the independent pitch command.

8. A variable speed and pitch control system for a wind turbine generator set, characterized in that, The wind turbine generator variable speed and pitch control system includes: The raw timing signal acquisition module is used to acquire raw timing signals collected from the anemometer, generator encoder, blade root load sensor, pitch drive encoder and nacelle azimuth sensor. The system state vector time series module is used to calculate the time series of the system state vector based on the original time series signal. The system state vector includes the current wind speed, rotor speed, flapping moment of each blade, current pitch angle of each blade and rotor azimuth angle. Collective pitch command generation module, used to generate collective pitch commands based on rotor speed; The module for predicting asymmetric load vectors is used to obtain the predicted asymmetric load vectors by passing the time series of the system state vectors through the trained load prediction model to predict the bending moment offset of each blade. The feedforward independent pitch compensation vector acquisition module is used to convert the predicted asymmetric load offset vector through a proportional controller to calculate the pitch angle increment of each blade to obtain the feedforward independent pitch compensation vector. The feedback independent pitch correction vector acquisition module is used to calculate the real-time load error based on the flapping moment of each blade, and to calculate the pitch angle correction of each blade through the PI controller to obtain the feedback independent pitch correction vector. The independent pitch command acquisition module is used to superimpose and synthesize the collective pitch command, the feedforward independent pitch compensation vector, and the feedback independent pitch correction vector to obtain the independent pitch command.

9. The wind turbine generator set variable speed and pitch control system according to claim 8, characterized in that, The module for predicting and obtaining asymmetric load vectors includes: The system state correlation time series coding vector acquisition unit is used to obtain the system state correlation time series coding vector by passing each system state vector in the time series of the system state vector through the coding module of the trained load prediction model. The predictive asymmetric load vector acquisition unit is used to obtain the predicted asymmetric load vector composed of the bending moment offsets of each blade by passing the system state correlation time-series encoded vector through the decoding module of the trained load prediction model.

10. The wind turbine generator set variable speed and pitch control system according to claim 8, characterized in that, The independent pitch correction vector acquisition module includes: The average bending moment calculation unit is used to calculate the average value of the flapping bending moments of multiple blades to obtain the average bending moment; The real-time load error calculation unit is used to subtract the flapping moment of each blade from the average bending moment to obtain the real-time load error of each blade. The feedback independent pitch correction vector acquisition unit is used to obtain the feedback independent pitch correction vector by passing the real-time load error of multiple blades through a PI controller.

Images