An intelligent variable pitch control system and method for a wind turbine

Abstract

The present disclosure provides a wind turbine intelligent variable pitch control system and method, which first utilizes the time series of feedforward wind field map obtained by laser radar, performs ECG event recognition through space-time sequence analysis, performs wind condition diagnosis, and accurately predicts whether the incoming flow wind field is a regular fluctuation or a high-risk extreme gust. Based on the diagnosis result, the control system can make intelligent decisions: when normal working conditions are identified, the normal mode is automatically run to optimize power generation and fatigue load; when a high-confidence ECG event is identified, the emergency mode is switched to as the primary target of rapid unloading and ensuring limit safety. This dynamic mode switching mechanism based on event driving enables the controller to call the optimal dedicated control law for different scenarios, thereby breaking the inherent limitations of static control and achieving the collaborative optimization of safety and efficiency in the full working condition range.

Description

Technical Field

[0001] This invention relates to the field of wind turbine pitch control, and particularly to an intelligent pitch control system and method for wind turbines. Background Technology

[0002] Wind energy, as a clean and renewable energy source, plays a crucial role in the global energy structure transformation. Wind turbines are the core equipment for converting wind energy into electricity, and their economic efficiency and safety directly impact the healthy development of the wind power industry. Among the many control subsystems of wind turbines, the pitch control system is the primary actuator for regulating aerodynamic torque, controlling generator power and speed, and suppressing loads on critical components. Therefore, constructing an efficient and reliable intelligent pitch control scheme for wind turbines is crucial for improving power generation performance, extending service life, and reducing the cost per kilowatt-hour. The pitch control system needs to strike a balance between two mutually constraining objectives: first, under normal wind conditions, finely adjusting the pitch angle to maximize wind energy capture efficiency and ensure stable output power; second, in extreme wind conditions (such as sudden gusts or strong wind shear), rapidly adjusting the pitch angle to quickly unload the load and ensure the safety of critical structural components such as blades, drive trains, and towers.

[0003] Traditional wind turbine pitch control methods primarily rely on PID control strategies based on generator speed or power feedback. These methods are essentially passive responses; the controller only begins adjustment after wind conditions have already affected the rotor and caused the speed or power to deviate from the setpoint. This inherent time lag makes them unresponsive to severe wind speed fluctuations, especially destructive extreme gusts, easily leading to instantaneous overspeeding and load shocks. To overcome the delay inherent in feedback control, the industry has introduced predictive control technology based on feedforward lidar. By detecting the incoming wind field ahead of the turbine, the controller can anticipate and adjust pitch accordingly. However, existing feedforward control schemes generally suffer from a deep-seated technical flaw: they often employ a static control logic to handle all predicted wind changes. This singular control strategy cannot effectively address the distinct needs of normal operation and emergency situations. Specifically, control laws designed to optimize power generation and reduce conventional fatigue loads typically have a relatively gentle response. When encountering extreme gusts, they become overly conservative, failing to provide sufficiently rapid and large pitch angle changes to effectively suppress ultimate loads. Conversely, if the control law is designed to be extremely aggressive to cope with extreme gusts, it will cause frequent and excessive pitch angle movements under normal wind conditions, affecting not only the stability of power generation but also exacerbating wear on actuators such as pitch bearings. This static nature of the control strategy means that when facing complex and variable inflow wind fields, the controller constantly struggles to balance power generation efficiency and ultimate safety, making it difficult to achieve global optimization.

[0004] Therefore, an optimized pitch control scheme for wind turbines is desired. Summary of the Invention

[0005] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide an intelligent pitch control system and method for wind turbine generators.

[0006] According to one aspect of this application, a smart pitch control method for wind turbine generators is provided, comprising: The acquired raw lidar data and raw unit status data are preprocessed to obtain the feedforward wind field map and the unit status vector. ECG event identification based on spatiotemporal series analysis was performed on the time series of feedforward wind field maps to obtain event identifiers, event confidence levels, and gust parameters; Control mode decisions are made based on event identifiers, event confidence levels, and gust parameters to obtain control modes; In response to the control mode being normal, the final pitch angle command is generated based on the feedforward wind field diagram; In response to the emergency control mode, the final pitch angle command is generated based on the gust parameters and the unit state vector.

[0007] According to another aspect of this application, a smart pitch control system for a wind turbine is provided, comprising: The raw data preprocessing module is used to preprocess the acquired raw lidar data and raw unit status data to obtain the feedforward wind field map and unit status vector. The wind field map time series analysis and identification module is used to perform ECG event identification based on spatiotemporal series analysis on the time series of feedforward wind field maps to obtain event identifiers, event confidence and gust parameters; The control mode decision module is used to make control mode decisions based on event identifiers, event confidence levels, and gust parameters to obtain control modes. The normal mode response module is used to generate the final pitch angle command based on the feedforward wind field diagram when the control mode is normal. The emergency mode response module is used to respond to the control mode being emergency, and to generate the final pitch angle command based on gust parameters and unit state vector.

[0008] Compared with existing technologies, the intelligent pitch control system and method for wind turbines provided in this application first utilizes the time series of feedforward wind field maps acquired by lidar, and performs ECG event identification through spatiotemporal sequence analysis to diagnose wind conditions and accurately predict whether the incoming wind field is a normal fluctuation or a high-risk extreme gust. Based on this diagnostic result, the control system can make intelligent decisions: when identified as a normal operating condition, it automatically operates in the normal mode aimed at optimizing power generation and fatigue load; when a high-confidence ECG event is identified, it decisively switches to the emergency mode with the primary goals of rapid unloading and ensuring ultimate safety. This event-driven dynamic mode switching mechanism enables the controller to call the optimal dedicated control law for different scenarios, thereby breaking the inherent limitations of static control and achieving synergistic optimization of safety and efficiency across the entire operating range. Attached Figure Description

[0009] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0010] Figure 1 This is a flowchart of a wind turbine intelligent pitch control method according to an embodiment of this application; Figure 2 This is a schematic diagram of the data flow of the intelligent pitch control method for wind turbines according to an embodiment of this application; Figure 3 The flowchart illustrates the process of performing ECG event identification based on spatiotemporal sequence analysis on the time series of the feedforward wind field map to obtain event identifiers, event confidence levels, and gust parameters in the intelligent pitch control method for wind turbines according to embodiments of this application. Figure 4 This is a flowchart illustrating the process of performing time-series encoding on the time series of the feedforward wind field spatial feature encoding vector to obtain the feedforward wind field spatiotemporal feature encoding vector according to the intelligent pitch control method for wind turbines in this application. Figure 5 This is a block diagram of a wind turbine intelligent pitch control system according to an embodiment of this application. Detailed Implementation

[0011] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0012] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0013] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0014] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0015] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0016] To address the inherent contradiction in existing wind turbine pitch control systems where a single control strategy struggles to balance optimized power generation under normal operating conditions with safety assurance under extreme conditions, this application proposes an intelligent pitch control method for wind turbines. The implementation process begins by using a feedforward lidar and the turbine's own sensors to construct a high-resolution feedforward wind field map time series and turbine state vector in real time. Furthermore, by introducing an ECG event recognition model based on spatiotemporal sequence analysis, this model can deeply learn and understand the dynamic evolution of the wind field map. This allows for accurate event identification, confidence assessment, and quantification of key parameters (such as gust amplitude and arrival time) before high-risk events like extreme gusts reach the rotor. Based on this forward-looking diagnostic result, the control system can make intelligent control mode decisions: when the wind condition is identified as normal, the system will automatically execute a multi-objective predictive control strategy based on the complete wind field map to coordinate the optimization of power generation and fatigue load; once an extreme gust is predicted with high confidence, the system will immediately switch to emergency control mode, calling a dedicated feedforward compensation algorithm based on specific gust parameters and the real-time status of the unit to generate a faster and more accurate pitch angle command to actively unload aerodynamics. Through this dynamic and adaptive control paradigm of diagnosis first and decision later, this solution effectively solves the limitations of static control and achieves optimal control performance under different wind conditions.

[0017] Figure 1 This is a flowchart of a wind turbine intelligent pitch control method according to an embodiment of this application. Figure 2 This is a schematic diagram of the data flow in the intelligent pitch control method for wind turbines according to an embodiment of this application. Figure 1 and Figure 2 As shown, the intelligent pitch control method for wind turbines according to an embodiment of this application includes the following steps: S100, preprocessing the acquired raw lidar data and raw turbine state data to obtain a feedforward wind field map and a turbine state vector; S200, performing ECG event identification based on spatiotemporal sequence analysis on the time series of the feedforward wind field map to obtain event identifiers, event confidence levels, and gust parameters; S300, performing control mode decision-making based on the event identifiers, event confidence levels, and gust parameters to obtain a control mode; S400, generating a final pitch angle command based on the feedforward wind field map in response to a normal control mode; S500, generating a final pitch angle command based on the gust parameters and the turbine state vector in response to an emergency control mode.

[0018] Specifically, in step S100, the acquired raw lidar data and raw unit status data are preprocessed to obtain a feedforward wind field map and a unit status vector. It should be understood that, due to significant differences in physical dimensions, data structure, time synchronization, and numerical scale between the raw lidar data and raw unit status data, they are essentially heterogeneous and unstructured raw signals that cannot be directly utilized by subsequent spatiotemporal sequence analysis models. Therefore, in the technical solution of this application, the acquired raw lidar data and raw unit status data are preprocessed to obtain a feedforward wind field map and a unit status vector, thereby transforming the multi-source heterogeneous raw data into a unified, regular, and information-rich characteristic input. This provides a high-quality and highly reliable data foundation for subsequent ECG event recognition and control mode decision-making, thereby ensuring the analytical accuracy and decision robustness of the entire intelligent control system.

[0019] More specifically, in this embodiment, the acquired raw lidar data and raw turbine state data are preprocessed to obtain a feedforward wind field map and a turbine state vector. This includes: performing coordinate transformation on the radial wind speed in the raw lidar data to obtain a three-dimensional wind speed field; using spatial interpolation to mesh the three-dimensional wind speed field to generate a feedforward wind field map; and performing timestamp alignment and normalization on the generator speed, pitch angle, yaw angle, and generator power in the raw turbine state data to obtain a turbine state vector. Specifically, in a particular example of this application, the preprocessing process is divided into two parallel processing paths. For the lidar data path, a coordinate transformation is first performed. The radial wind speed data directly measured by the lidar is transformed from the lidar's own spherical coordinate system to the Cartesian coordinate system of the wind turbine hub center, based on the azimuth and pitch angle information of each measurement point during the scanning process, thereby resolving the three-dimensional wind speed field of a series of discrete points in space. Subsequently, to obtain a continuous and structured wind field representation, spatial interpolation methods, such as Kriging interpolation, are used to process the aforementioned discrete three-dimensional wind speed field data points. The wind speed value of each grid point is estimated on a pre-defined two-dimensional grid plane (i.e., the feedforward wind field), ultimately generating a feedforward wind field map that intuitively reflects the wind speed distribution in front of the wind turbine. For the unit state data flow path, firstly, the collected key state variables, such as generator speed, pitch angle, yaw angle, and generator power, are timestamped. Through methods such as linear interpolation, it is ensured that all state variables accurately correspond on the time segment of each control cycle, eliminating asynchronous problems introduced by different sensor sampling frequencies or data transmission delays. Next, the time-aligned multidimensional state data is normalized, for example, using the max-min normalization method to uniformly map variables with different physical units and numerical ranges to the interval [0,1] or [-1,1], forming the final unit state vector to eliminate the influence of dimensional differences on subsequent model analysis.

[0020] Specifically, in step S200, ECG event identification based on spatiotemporal sequence analysis is performed on the time series of the feedforward wind field map to obtain event identifiers, event confidence levels, and gust parameters. It should be understood that since the feedforward wind field map time series obtained after preprocessing is essentially a high-dimensional, continuous visual data stream, it does not directly reveal the event attributes of future wind conditions, such as the existence of a structured, potentially destructive extreme gust. Therefore, in the technical solution of this application, ECG event identification based on spatiotemporal sequence analysis is further performed on the time series of the feedforward wind field map to obtain event identifiers, event confidence levels, and gust parameters. This allows for the automatic and forward-looking identification, classification, and quantification of key wind events from the complex dynamic evolution of the wind field. In this way, raw, non-semantic wind field data can be transformed into structured event information with clear physical meaning, providing a direct and reliable basis for subsequent control mode decisions, thereby achieving an intelligent leap from perceiving wind to understanding wind.

[0021] Figure 3 This is a flowchart illustrating the process of performing ECG event identification based on spatiotemporal sequence analysis on the time series of a feedforward wind field map to obtain event identifiers, event confidence levels, and gust parameters, according to the intelligent pitch control method for wind turbines in this application. Figure 3 As shown, step S200 includes: S210, extracting spatial features from each feedforward wind field map in the time series of the feedforward wind field map to obtain the time series of the feedforward wind field spatial feature encoding vector; S220, performing time series encoding on the time series of the feedforward wind field spatial feature encoding vector to obtain the feedforward wind field spatiotemporal feature encoding vector; S230, inputting the feedforward wind field spatiotemporal feature encoding vector into a classification head to obtain event identifiers and event confidence; S240, inputting the feedforward wind field spatiotemporal feature encoding vector into a regression head to obtain the gust parameters.

[0022] Accordingly, in step S210, spatial features are extracted from each feedforward wind field map in the time series of the feedforward wind field map to obtain a time series of feedforward wind field spatial feature encoding vectors. It should be understood that since the original feedforward wind field map is a high-dimensional, image-like raster data, directly using it as input to a time series model would lead to the curse of dimensionality and low computational efficiency, and the key spatial structural features it contains (such as wind speed gradients and vortex morphology) are implicit. Therefore, in the technical solution of this application, spatial features are further extracted from each feedforward wind field map in the time series of the feedforward wind field map to obtain a time series of feedforward wind field spatial feature encoding vectors, thereby automatically learning and extracting deep, low-dimensional feature representations that can characterize the core spatial distribution characteristics from each frame of the wind field map. In this way, while effectively compressing the data dimensionality, the multi-scale spatial information crucial for ECG event recognition is preserved to the greatest extent, thus transforming the original image sequence into a feature vector sequence with higher information density and more suitable for time series modeling.

[0023] Specifically, in a concrete example of this application, the time series of spatial feature extraction for each feedforward wind field map in the time series of feedforward wind field maps to obtain feedforward wind field spatial feature encoding vectors includes: passing the feedforward wind field map through a spatial feature extractor based on a dilated convolutional neural network model to obtain the feedforward wind field spatial feature encoding vector. That is, this spatial feature extraction process is implemented through a pre-trained dilated convolutional neural network model. For each frame of the feedforward wind field map in the time series, the model uses it as input and performs a series of forward propagation calculations. This process first feeds a single frame of the feedforward wind field map into the input layer of the model. Subsequently, the data passes sequentially through multiple stacked dilated convolutional layers. These convolutional layers, by setting different dilation rates, enable the convolutional kernels to obtain an exponentially increased receptive field without increasing the number of parameters or reducing the resolution. Specifically, shallow networks use smaller dilation rates to accurately capture local wind speed gradient details, while deep networks use larger dilation rates to integrate large-scale wind field structure information, thereby achieving effective multi-scale capture of wind field features. After each convolutional layer, a non-linear activation function (e.g., ReLU) is applied to enhance the model's non-linear expressive power. After processing through all convolutional layers, the model outputs a high-dimensional multi-channel feature map. Finally, a global average pooling layer integrates this feature map spatially, transforming it into a fixed-length one-dimensional vector. This vector is the feedforward wind field spatial feature encoding vector corresponding to the feedforward wind field map of that frame.

[0024] Accordingly, in step S220, the time series of the feedforward wind field spatial feature encoding vector is time-series encoded to obtain the feedforward wind field spatiotemporal feature encoding vector. It should be understood that when the time series of the feedforward wind field spatial feature encoding vector is encoded by a deep nonlinear temporal model (such as a single LSTM), the learned high-level temporal features may deviate from their original semantics representing the physical structure of the wind field, i.e., conceptual drift occurs. This drift can lead to the model's final output features performing well on the training task, but potentially deviating significantly from their original physical meaning, resulting in poor generalization ability and weak interpretability for unseen extreme wind conditions. Therefore, in the technical solution of this application, the time series of the feedforward wind field spatial feature encoding vector is further time-series encoded to obtain the feedforward wind field spatiotemporal feature encoding vector, thereby introducing an explicit, closed-loop self-calibration mechanism to actively monitor and dynamically correct the semantic fidelity during the representation learning process. In this way, the temporal coding process can be transformed from an open, unidirectional feature extraction pipeline into a sophisticated system with internal feedback and control, which manages the quality of representation in a closed loop. This ensures the dynamic fidelity and robustness of the final spatiotemporal feature coding vector, and guarantees that its description of ECG events is accurate and physically reliable.

[0025] Figure 4 This is a flowchart illustrating the process of time-series encoding the time series of the feedforward wind field spatial feature encoding vector to obtain the feedforward wind field spatiotemporal feature encoding vector, according to the intelligent pitch control method for wind turbines in this application. Figure 4 As shown, step S220 includes: S221, inputting the time series of the feedforward wind field spatial feature encoding vector into the primary LSTM model to obtain the temporal distribution of the feedforward wind field spatial semantic latent encoding vector; S222, calculating the offset metric coefficient between the temporal distribution of the feedforward wind field spatial semantic latent encoding vector and the corresponding feedforward wind field spatial feature encoding vector in the time series of each group to obtain a set of feedforward wind field spatial offset metric coefficients; S223, generating a set of feedforward wind field spatial inverse offset calibration coefficients based on the set of feedforward wind field spatial offset metric coefficients; S224, performing inverse offset calibration on the temporal distribution of the feedforward wind field spatial semantic latent encoding vector based on the set of feedforward wind field spatial inverse offset calibration coefficients to obtain a sequence of feedforward wind field spatial enhanced semantic latent encoding vectors; S225, inputting the sequence of feedforward wind field spatial enhanced semantic latent encoding vectors into the secondary LSTM model to obtain the feedforward wind field spatiotemporal feature encoding vector.

[0026] In step S221, the time series of the feedforward wind field spatial feature encoding vector is input into the primary LSTM model to obtain the temporal distribution of the feedforward wind field spatial semantic latent encoding vector. This is expressed by the following formula:

[0027] in, The time series of the spatial feature encoding vectors of the feedforward wind field. The first time series of the spatial feature encoding vector of the feedforward wind field Each feedforward wind field spatial feature encoding vector The number of vectors in the time series that encode the spatial features of the feedforward wind field. For the first-level LSTM model, The temporal distribution of the feedforward wind field spatial semantic latent encoding vector. The th in the temporal distribution of the feedforward wind field spatial semantic latent encoding vector A feedforward wind field spatial semantic latent encoding vector.

[0028] It is understandable that, since the time series of the feedforward wind field spatial feature encoding vector is merely a series of independent feature snapshots representing instantaneous spatial structures, it lacks an explicit representation of the dynamic evolution of wind field events over time. Therefore, in the technical solution of this application, the time series of the feedforward wind field spatial feature encoding vector is input into the primary LSTM model to obtain the temporal distribution of the feedforward wind field spatial semantic latent encoding vector. This allows for preliminary semantic encoding of the sequence distribution, effectively learning and memorizing the long-range dependencies in the evolution of wind field spatial features using the inherent gating mechanism of the LSTM network. In this way, the original feature vectors representing instantaneous spatial structures can be mapped to a higher-dimensional, more abstract semantic space. The product is the temporal distribution of the feedforward wind field spatial semantic latent encoding vector. This distribution not only encodes the wind field spatial features at each moment but also incorporates their dynamic evolution information within the entire time series context, providing a temporally logical input foundation for subsequent, more refined concept anti-drift calibration.

[0029] In step S222, the offset metric coefficients between the feedforward wind field spatial semantic latent coding vector and the feedforward wind field spatial feature coding vector in the time series of each corresponding group are calculated to obtain the set of feedforward wind field spatial offset metric coefficients. This is expressed by the following formula:

[0030] in, For the projection matrix, Let covariance matrix be the variance matrix. and The learnable adjustment coefficient, The trace of the matrix, ∈[0,1], is The corresponding feedforward wind field spatial offset metric coefficient.

[0031] It is understandable that, when the primary LSTM model performs deep nonlinear transformations to capture temporal dependencies, there is an inherent risk of semantic deviation between the generated feedforward wind field spatial semantic latent encoding vector and the original feedforward wind field spatial feature encoding vector. This semantic deviation is abstract and unquantified. Therefore, in the technical solution of this application, the temporal distribution of the feedforward wind field spatial semantic latent encoding vector and the offset metric coefficients between each corresponding group of feedforward wind field spatial semantic latent encoding vectors and feedforward wind field spatial feature encoding vectors in the time series are further calculated to obtain a set of feedforward wind field spatial offset metric coefficients. This establishes an explicit concept drift quantification step, transforming the semantic difference of features between the transform domain and the original domain into a measurable and specific index. In this way, a set of quantification factors that accurately describes the degree of semantic information loss or distortion at each step of temporal feature extraction can be obtained, providing a precise and operable quantitative basis for subsequent active and adaptive anti-drift correction.

[0032] In step S223, a set of feedforward wind field spatial inverse offset calibration coefficients is generated based on the set of feedforward wind field spatial offset metric coefficients. This is expressed by the following formula:

[0033] in, For temperature coefficient, It is a smoothing constant. Let e ​​be the value of the logarithmic function with the natural constant e as the base. for The corresponding spatial inverse offset calibration coefficient for the feedforward wind field.

[0034] It should be understood that, since the set of feedforward wind field spatial offset metric coefficients only quantifies the degree of semantic drift, it is not itself an operational signal that can be directly used to perform correction. Therefore, in the technical solution of this application, a set of feedforward wind field spatial inverse offset calibration coefficients is further generated based on the set of feedforward wind field spatial offset metric coefficients. This is used to construct an adaptive adjustment strategy, transforming the aforementioned quantified drift degree into a specific and operable correction signal. In this way, a precise set of correction factors can be provided for the subsequent anti-drift constraint stage, ensuring that the subsequent correction intensity can accurately match the quantified semantic drift degree at each time step, thereby avoiding the problems of under-correction or over-correction.

[0035] In step S224, based on the set of feedforward wind field spatial inverse offset calibration coefficients, the temporal distribution of the feedforward wind field spatial semantic latent coding vector is subjected to inverse offset calibration to obtain a sequence of feedforward wind field spatial enhanced semantic latent coding vectors. This is expressed by the following formula:

[0036] in, for The corresponding feedforward wind field spatially enhanced semantic latent encoding vector.

[0037] It should be understood that although the temporal distribution of the feedforward wind field spatial semantic latent coding vector captures the deep temporal dependencies, it may still contain semantic noise and biases caused by nonlinear transformations. The set of feedforward wind field spatial inverse offset calibration coefficients provides a precise strategy for correcting these biases. Therefore, in the technical solution of this application, the temporal distribution of the feedforward wind field spatial semantic latent coding vector is further inverse offset calibrated based on the set of feedforward wind field spatial inverse offset calibration coefficients to obtain a sequence of feedforward wind field spatial enhanced semantic latent coding vectors. This is used to perform the core inverse offset calibration step, by using calibration coefficients to impose stronger constraints on latent coding vectors that have experienced severe semantic drift, or to pull them back to a representation space that is closer to the original wind field physical semantics. In this way, a sequence of feedforward wind field spatially enhanced semantic latent encoding vectors can be generated. The significant advantage of this new sequence is that it retains the deep temporal dependencies captured by the first-level LSTM network, while filtering out semantic noise and biases generated during the transformation process through the inverse offset calibration mechanism. This achieves a balance between representational ability and semantic fidelity, providing a purified and enhanced input for subsequent higher-quality information aggregation and encoding.

[0038] In step S225, the sequence of feedforward wind field spatially enhanced semantic latent encoding vectors is input into the secondary LSTM model to obtain the feedforward wind field spatiotemporal feature encoding vector. This is expressed by the following formula:

[0039] in, For secondary LSTM models, This is the spatiotemporal feature encoding vector for the feedforward wind field.

[0040] It is understandable that although the sequence of feedforward wind field spatially enhanced semantic latent encoding vectors has undergone anti-drift calibration to ensure the semantic fidelity of features at each time step, it is still a distributed, temporally arranged set of features and has not yet been integrated into a global, singular representation that can represent the entire wind field event evolution process. Therefore, in the technical solution of this application, the sequence of feedforward wind field spatially enhanced semantic latent encoding vectors is further input into a secondary LSTM model to obtain a feedforward wind field spatiotemporal feature encoding vector. This allows for final information aggregation and encoding based on a higher quality and more semantically stable representation, thereby further refining and integrating the dynamic information of the entire sequence. In this way, a single, highly condensed feedforward wind field spatiotemporal feature encoding vector can be output. This vector is not only a global representation of the entire feedforward wind field evolution process, but also a robust and reliable representation that has undergone rigorous self-examination and calibration. It can be directly used for downstream ECG event classification and gust parameter regression tasks, significantly improving their prediction accuracy and stability.

[0041] Accordingly, in step S230, the feedforward wind field spatiotemporal feature encoding vector is input into the classification head to obtain the event identifier and event confidence level. It should be understood that although the feedforward wind field spatiotemporal feature encoding vector highly condenses the spatiotemporal dynamic information of wind field evolution, it is essentially still an abstract numerical representation located in a latent feature space, and does not directly provide an explicit, discrete classification judgment regarding the attributes of wind events. Therefore, in the technical solution of this application, the feedforward wind field spatiotemporal feature encoding vector is further input into the classification head to obtain the event identifier and event confidence level, thereby mapping this high-dimensional, continuous feature vector to a predefined, discrete wind event category space, realizing the final decoding and judgment of the nature of wind field events. In this way, implicit spatiotemporal features can be transformed into explicit event identifiers and quantified event confidence levels, providing a direct, clear, and reliability-assessed decision basis for the control system to make subsequent mode switching decisions.

[0042] More specifically, in a concrete example of this application, the classification process is performed through a classification head composed of a multi-layer fully connected neural network. First, the feedforward wind field spatiotemporal feature encoding vector, obtained through temporal encoding, is fed as input into the first layer of the fully connected network of the classification head. Subsequently, the vector passes through one or more hidden layers. Each hidden layer linearly transforms the input features using its weight matrix, superimposes a bias vector, and then processes it through a non-linear activation function (e.g., ReLU) to learn and extract higher-order, more discriminative feature combinations used to distinguish different wind event categories. Finally, the data is passed to the output layer, where the number of neurons corresponds to the preset total number of event categories (e.g., normal operating conditions, ECG events, etc.). The output layer uses the Softmax activation function, which transforms the logical values ​​output by the previous layer into a probability distribution, where each output value represents the probability that the input vector belongs to the corresponding event category, and the sum of all output values ​​is 1. Accordingly, the category corresponding to the neuron with the highest probability value is determined as the final event identifier, and this highest probability value itself serves as the event confidence level for this identification.

[0043] Accordingly, in step S240, the feedforward wind field spatiotemporal feature encoding vector is input into the regression head to obtain the gust parameters. It should be understood that although the feedforward wind field spatiotemporal feature encoding vector contains a deep representation of future wind field events, it is itself a high-dimensional, continuous latent feature and does not directly provide continuous numerical values ​​describing the key physical properties of the event (especially extreme gust events). Therefore, in the technical solution of this application, the feedforward wind field spatiotemporal feature encoding vector is further input into the regression head to obtain the gust parameters, thereby establishing a nonlinear mapping from the condensed spatiotemporal feature space to the specific physical parameter space, realizing quantitative prediction of key continuous variables of wind field events. In this way, while identifying extreme operating conditions, its key physical properties, such as gust amplitude, arrival time, and duration, can be accurately quantified, providing the necessary quantitative input parameters for generating rapid and accurate feedforward compensation commands for subsequent emergency control modes.

[0044] More specifically, in a concrete example of this application, the regression process is implemented through a regression head consisting of a multi-layer fully connected neural network. First, the same feedforward wind field spatiotemporal feature encoding vector (the same vector as the input to the classification head) is fed into the first layer of this regression head. This vector then passes through a series of hidden layers, each undergoing an affine transformation of the weight matrix and bias vector, and processed by a nonlinear activation function (e.g., ReLU) to learn the complex nonlinear mapping relationship between abstract spatiotemporal features and specific physical parameters. Finally, the features are passed to the output layer. The number of neurons in this layer strictly corresponds to the number of gust parameters to be predicted (e.g., three neurons are used if predicting gust amplitude, arrival time, and duration). Unlike the classification head, the output layer uses a linear activation function, enabling it to directly output unbounded continuous values. Therefore, the output values ​​of each neuron in this layer are directly parsed as the specific values ​​of the predicted gust parameters.

[0045] Specifically, in step S300, control mode decision-making is performed based on event identifiers, event confidence levels, and gust parameters to obtain the control mode. It should be understood that since the event identifiers, event confidence levels, and gust parameters output by the ECG event identification module are merely structured descriptions of future wind conditions, they do not directly form the final instructions for the wind turbine control strategy. Therefore, in the technical solution of this application, control mode decision-making is further performed based on event identifiers, event confidence levels, and gust parameters to obtain the control mode, thereby establishing a clear, rule-based decision-making center that transforms forward-looking wind condition perception information into specific, executable control system operation state instructions. This enables the entire control system to have intelligent switching capabilities, dynamically selecting between a conventional mode prioritizing power generation efficiency and an emergency mode prioritizing load safety based on the predicted future wind conditions, thus achieving optimal response to different wind conditions.

[0046] More specifically, in a concrete example of this application, the control mode decision-making process is implemented as a preset logical decision-making module. This module first parses the input event identifier. If the event identifier is determined to be a normal wind condition or a non-critical wind event, the module directly outputs a normal operating condition control mode command, instructing the system to adopt or maintain the conventional pitch angle control strategy. Conversely, if the event identifier is determined to be an ECG event, such as an extreme gust, the module further reads the event confidence associated with the event and compares it with a preset confidence threshold. If the event confidence is lower than the threshold, the decision-making module determines it as a low-confidence warning, and the system will maintain or switch to normal operating conditions to avoid misoperation due to prediction uncertainty. If the event confidence is higher than or equal to the threshold, the decision-making module determines it as a high-confidence, impending extreme wind condition event, and the module will make the final decision and output an emergency operating condition control mode command. Simultaneously, the gust parameters associated with the event are marked and prepared to be passed to the pitch angle command generation process in the emergency mode as the basis for calculating the feedforward compensation.

[0047] Specifically, in step S400, in response to the control mode being normal, a final pitch angle command is generated based on the feedforward wind field map. It should be understood that when the control mode is determined to be normal operating condition, the core objective of the system is to maximize power generation efficiency and suppress the accumulation of fatigue loads while ensuring basic safety. Traditional control strategies based on a single feedback variable are difficult to use forward-looking wind condition information for fine-tuning. Therefore, in the technical solution of this application, in response to the control mode being normal, a final pitch angle command is generated based on the feedforward wind field map. This directly integrates the predictive wind field information with high spatiotemporal resolution provided by the lidar into the closed loop of pitch angle control command generation, achieving a feedforward-based, multi-objective optimization-based fine-tuning control. This enables the wind turbine to not only passively respond to speed deviations during normal operation but also actively and predictively adjust the blade attitude to adapt to upcoming wind speed fluctuations, thereby achieving a better balance between power generation and turbine lifespan.

[0048] More specifically, in this embodiment of the application, in response to the control mode being normal, generating the final pitch angle command based on the feedforward wind field map includes: calculating the predicted power generation, predicted fatigue load of key components, and pitch rate of the i-th candidate pitch angle control sequence based on the feedforward wind field map; performing a cost evaluation on the i-th candidate pitch angle control sequence based on the power generation, predicted fatigue load of key components, and pitch rate to obtain the i-th cost coefficient; selecting the candidate pitch angle control sequence corresponding to the minimum cost coefficient as the optimal pitch angle control sequence, and taking the first pitch angle in the optimal pitch angle control sequence as the final pitch angle. Specifically, the cost evaluation is performed using the following formula to obtain the i-th cost coefficient, wherein the formula is:

[0049] in, To predict the time domain, For the predicted power generation, For reference power, For predicting the fatigue load of key components, For pitch rate, , , These are the weighting coefficients. Let be the i-th cost coefficient.

[0050] Specifically, the command generation process under the normal mode is implemented through an optimizer based on model predictive control. First, the optimizer receives the current control mode command as normal and obtains the latest feedforward wind field map. Then, this feedforward wind field map, as an external disturbance input, is fed into a built-in high-fidelity predictive model that describes the aerodynamic and structural dynamic characteristics of the wind turbine. This model is used in a finite future time domain. Within this framework, key performance indicators of the unit, such as generator power output, are predicted under different candidate pitch angle control sequences. Fatigue loads on critical components (such as blade roots) Next, an optimization solver is initiated, aiming to find a pitch angle control sequence that minimizes a predefined cost function while satisfying unit operating constraints (such as pitch angle rate of change limits and maximum pitch angle limits). This cost function is designed as a weighted sum, with its components representing the values ​​of the reference power. The tracking error (aimed at maximizing power generation) and the penalty for fatigue load (aimed at reducing losses) and the pitch rate The suppression of [something]. Finally, the first element in the optimal control sequence output by the optimizer is adopted as the final pitch angle command for the current moment and sent to the pitch actuator.

[0051] Specifically, in a concrete example of this application, the process of predicting key performance indicators of the turbine unit under different candidate pitch angle control sequences based on the feedforward wind field map is implemented through the following steps: In general, the core of this process lies in using a high-fidelity internal prediction model of the wind turbine unit, taking the current unit state as the initial condition and the future wind conditions provided by the feedforward wind field map as the external excitation, to deduce a series of candidate pitch angle control sequences, thereby obtaining the performance characteristics that these control sequences will lead to in the future. Specifically, the first step is that at the beginning of each control cycle, the MPC controller obtains the current state vector of the unit (including the current speed, pitch angle, generator torque, etc.) as the initial condition for prediction. Simultaneously, it extracts the wind speed distribution data corresponding to each time step k (from t to t+H) within the future prediction time domain H from the feedforward wind field map. The second step is that the optimizer generates a set or an initial candidate pitch angle control sequence { This sequence defines the planned pitch rate for each step from the current time t to the future time t+H. The third step involves the controller initiating its internal predictive model. In the first step of the simulation (k=t), the model uses the current unit state, wind data at time t provided by the wind farm diagram, and the pitch rate at time t from the candidate sequence. As input, the unit's state at time t+1 is calculated by solving the unit's aerodynamic and dynamic equations, and the predicted power generation at time t is also calculated simultaneously. and predict fatigue load Fourth, in the second step of the simulation (k=t+1), the model uses the predicted state at time t+1 calculated in the previous step as the new initial condition, combined with the wind condition data at time t+1 provided by the wind field map and the pitch rate at time t+1 in the candidate sequence. The solution is then performed again to obtain the unit status at time t+2 and the performance indicators at time t+1. and This process is repeated until the entire prediction time domain H is covered, thus obtaining the complete prediction performance index trajectory corresponding to the candidate control sequence. }and{ Through the above steps, the MPC controller can accurately predict the performance of any given candidate pitch angle control sequence under future wind conditions. The optimizer will repeatedly execute this process, continuously adjusting the candidate control sequences until it finds the optimal sequence that minimizes the cost function, thus completing the generation of the final pitch angle command.

[0052] Specifically, in step S500, in response to the control mode being set to emergency, a final pitch angle command is generated based on gust parameters and the turbine state vector. It should be understood that when the control mode is decided to be in emergency mode, the wind turbine faces a foreseen, high-intensity transient impact. Traditional feedback control based on speed error has a delayed response and cannot take measures before the impact arrives, easily leading to the turbine bearing enormous aerodynamic loads and dangerous overspeed events. Therefore, in the technical solution of this application, in response to the control mode being set to emergency, a final pitch angle command is generated based on gust parameters and the turbine state vector to execute a forward-looking information-based feedforward control strategy. The core of this strategy is to proactively and in advance calculate and execute an optimal unloading action using known gust characteristics and the current state of the turbine. This allows the wind turbine to pre-adjust the blade angle before the extreme gusts arrive, thereby effectively reducing aerodynamic loads at the moment of impact, minimizing speed spikes, and ensuring the structural safety of the turbine.

[0053] More specifically, in this embodiment of the application, in response to the control mode being emergency, generating a final pitch angle command based on gust parameters and the unit state vector includes: calculating the final pitch angle based on the gust parameters and the unit state vector using the following formula, wherein the formula is:

[0054]

[0055] in, For feedforward compensation of pitch angle, This represents the gust amplitude. This refers to the arrival time of the gusts. The current rotational speed, This is the gust compensation gain coefficient. Rated speed, To predict the time, The time decay constant, Based on the pitch angle, This is the final pitch angle in emergency mode.

[0056] Specifically, the command generation process in this emergency mode is implemented as a dual-channel control command synthesis module. First, upon receiving the emergency mode command, the controller immediately acquires the associated gust parameters (including gust amplitude, arrival time, and duration) and the real-time unit state vector (including current generator speed, pitch angle, etc.). Then, it calculates two pitch angle components in parallel. The first component is the base pitch angle, generated by a standard feedback controller (e.g., a PI controller), which is based on the current speed in the unit state vector. With rated speed The deviation is calculated to maintain the system's basic stability and response to unforeseen disturbances. The second component is the feedforward compensated pitch angle. Its calculations are entirely based on predictive information. This module will output the gust amplitude from the ECG event identification module. and the arrival time of the gusts Together with the current speed in the unit state vector The module inputs a pre-defined analytical model. The compensation calculated by this model is proportional to the gust amplitude and is adjusted based on the difference between the current rotational speed and the rated rotational speed. Simultaneously, the timing of its application is precisely shaped using an exponential decay function related to the gust arrival time. Finally, the module algebraically sums the calculated base pitch angle and the feedforward compensated pitch angle to obtain the final pitch angle command. And send it to the pitch actuator.

[0057] In summary, the intelligent pitch control method for wind turbines according to the embodiments of this application is explained. It first utilizes the time series of the feedforward wind field map acquired by lidar, and then performs ECG event identification through spatiotemporal sequence analysis to diagnose wind conditions and accurately predict whether the incoming wind field is a normal fluctuation or a high-risk extreme gust. Based on this diagnostic result, the control system can make intelligent decisions: when identified as a normal operating condition, it automatically operates in a normal mode aimed at optimizing power generation and fatigue load; when a high-confidence ECG event is identified, it decisively switches to an emergency mode with rapid unloading and ensuring ultimate safety as its primary objectives. This event-driven dynamic mode switching mechanism enables the controller to call the optimal dedicated control law for different scenarios, thereby breaking the inherent limitations of static control and achieving coordinated optimization of safety and efficiency across the entire operating range.

[0058] Furthermore, an intelligent pitch control system for wind turbines is also provided.

[0059] Figure 5 This is a block diagram of a smart pitch control system for a wind turbine according to an embodiment of this application. Figure 5As shown, the intelligent pitch control system 100 for wind turbines according to an embodiment of this application includes: a raw data preprocessing module 110, used to preprocess the acquired raw lidar data and raw turbine state data to obtain a feedforward wind field map and a turbine state vector; a wind field map time series analysis and identification module 120, used to perform ECG event identification based on spatiotemporal sequence analysis on the time series of the feedforward wind field map to obtain event identifiers, event confidence levels, and gust parameters; a control mode decision module 130, used to perform control mode decision based on event identifiers, event confidence levels, and gust parameters to obtain a control mode; a normal mode response module 140, used to generate a final pitch angle command based on the feedforward wind field map in response to a normal control mode; and an emergency mode response module 150, used to generate a final pitch angle command based on gust parameters and a turbine state vector in response to an emergency control mode.

[0060] As described above, the wind turbine intelligent pitch control system 100 according to the embodiments of this application can be implemented in various wireless terminals, such as servers with wind turbine intelligent pitch control algorithms. In one possible implementation, the wind turbine intelligent pitch control system 100 according to the embodiments of this application can be integrated into the wireless terminal as a software module and / or hardware module. For example, the wind turbine intelligent pitch control system 100 can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the wind turbine intelligent pitch control system 100 can also be one of many hardware modules of the wireless terminal.

[0061] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for intelligent pitch control of wind turbine generators, characterized in that, include: The acquired raw lidar data and raw unit status data are preprocessed to obtain the feedforward wind field map and the unit status vector. ECG event identification based on spatiotemporal series analysis was performed on the time series of feedforward wind field maps to obtain event identifiers, event confidence levels, and gust parameters; Control mode decisions are made based on event identifiers, event confidence levels, and gust parameters to obtain control modes; In response to the control mode being normal, the final pitch angle command is generated based on the feedforward wind field diagram; In response to the emergency control mode, the final pitch angle command is generated based on the gust parameters and the unit state vector.

2. The intelligent pitch control method for wind turbines according to claim 1, characterized in that, The acquired raw lidar data and raw unit status data are preprocessed to obtain the feedforward wind field map and unit status vector, including: The radial wind speed in the original lidar data is transformed by coordinates to obtain a three-dimensional wind speed field. Spatial interpolation is used to grid the three-dimensional wind speed field to generate a feedforward wind field map; The generator speed, pitch angle, yaw angle, and generator power in the original unit state data are timestamped and normalized to obtain the unit state vector.

3. The intelligent pitch control method for wind turbines according to claim 1, characterized in that, ECG event identification based on spatiotemporal series analysis was performed on the time series of feedforward wind field maps to obtain event identifiers, event confidence scores, and gust parameters, including: Spatial features are extracted from each feedforward wind field map in the time series to obtain the time series of feedforward wind field spatial feature encoding vectors. Time series encoding is performed on the time series of the feedforward wind field spatial feature encoding vector to obtain the feedforward wind field spatiotemporal feature encoding vector; The feedforward wind field spatiotemporal feature encoding vector is input into the classification head to obtain the event identifier and event confidence. The feedforward wind field spatiotemporal feature encoding vector is input into the regression head to obtain the gust parameters.

4. The intelligent pitch control method for wind turbines according to claim 3, characterized in that, The time series of spatial feature extraction for each feedforward wind field map in the time series of feedforward wind field maps to obtain the feedforward wind field spatial feature encoding vector includes: passing the feedforward wind field map through a spatial feature extractor based on a dilated convolutional neural network model to obtain the feedforward wind field spatial feature encoding vector.

5. The intelligent pitch control method for wind turbines according to claim 3, characterized in that, The time series of the feedforward wind field spatial feature encoding vector is time-series encoded to obtain the feedforward wind field spatiotemporal feature encoding vector, including: The time series of the feedforward wind field spatial feature encoding vector is input into the primary LSTM model to obtain the temporal distribution of the feedforward wind field spatial semantic latent encoding vector. The set of feedforward wind field spatial offset metric coefficients is obtained by calculating the temporal distribution of the feedforward wind field spatial semantic latent coding vector and the offset metric coefficient between each group of feedforward wind field spatial feature coding vector in the time series of the feedforward wind field spatial semantic latent coding vector and the feedforward wind field spatial feature coding vector. Based on the set of feedforward wind field spatial offset metric coefficients, a set of feedforward wind field spatial inverse offset calibration coefficients is generated; Based on the set of inverse offset calibration coefficients of the feedforward wind field space, the temporal distribution of the feedforward wind field space semantic latent coding vector is subjected to inverse offset calibration to obtain the sequence of feedforward wind field space enhanced semantic latent coding vectors. The sequence of feedforward wind field spatially enhanced semantic latent encoding vectors is input into the secondary LSTM model to obtain the feedforward wind field spatiotemporal feature encoding vector.

6. The intelligent pitch control method for wind turbines according to claim 1, characterized in that, In response to the control mode being normal, the final pitch angle command is generated based on the feedforward wind field diagram, including: The predicted power generation, predicted fatigue load of key components, and pitch rate are calculated based on the feedforward wind field map for the i-th candidate pitch angle control sequence. Based on power generation, predicted fatigue load of key components and pitch rate, a cost evaluation is performed on the i-th candidate pitch angle control sequence to obtain the i-th cost coefficient. The candidate pitch angle control sequence corresponding to the minimum cost coefficient is selected as the optimal pitch angle control sequence, and the first pitch angle in the optimal pitch angle control sequence is taken as the final pitch angle.

7. The intelligent pitch control method for wind turbines according to claim 6, characterized in that, A cost evaluation is performed on the i-th candidate pitch angle control sequence based on power generation, predicted fatigue load of key components, and pitch rate to obtain the i-th cost coefficient, including: The cost evaluation is performed using the following formula to obtain the i-th cost coefficient, where the formula is: in, To predict the time domain, For the predicted power generation, For reference power, For predicting the fatigue load of key components, For pitch rate, , , These are the weighting coefficients. Let be the i-th cost coefficient.

8. The intelligent pitch control method for wind turbines according to claim 1, characterized in that, In response to the control mode being emergency, a final pitch angle command is generated based on gust parameters and the unit state vector, including: calculating the final pitch angle based on the gust parameters and the unit state vector using the following formula, wherein the formula is: in, For feedforward compensation of pitch angle, This represents the gust amplitude. This refers to the arrival time of the gusts. The current rotational speed, This is the gust compensation gain coefficient. Rated speed, To predict the time, The time decay constant, Based on the pitch angle, This is the final pitch angle in emergency mode.

9. A smart pitch control system for wind turbine generators, characterized in that, include: The raw data preprocessing module is used to preprocess the acquired raw lidar data and raw unit status data to obtain the feedforward wind field map and unit status vector. The wind field map time series analysis and identification module is used to perform ECG event identification based on spatiotemporal series analysis on the time series of feedforward wind field maps to obtain event identifiers, event confidence and gust parameters; The control mode decision module is used to make control mode decisions based on event identifiers, event confidence levels, and gust parameters to obtain control modes. The normal mode response module is used to generate the final pitch angle command based on the feedforward wind field diagram when the control mode is normal. The emergency mode response module is used to respond to the control mode being emergency, and to generate the final pitch angle command based on gust parameters and unit state vector.

Images