Wind power plant torque and variable pitch cooperative control method for reducing fatigue load

Abstract

The invention relates to the technical field of fatigue load reduction, in particular to a wind power plant torque and variable pitch cooperative control method with fatigue load reduction as the target. The method comprises the following steps: acquiring real-time and historical operation data of each fan in a wind power plant, and preprocessing the operation data; based on the preprocessed operation data, load characteristic indexes reflecting structural fatigue are extracted, and a load state vector reflecting the current fatigue accumulation trend of the draught fan is constructed; and based on the load state vector, establishing a coupling model of the influence of the generator torque and the variable pitch angle on the rotating speed response and the fatigue load, wherein the coupling model is used for reflecting the influence mechanism of torque adjustment on the low-frequency load and the inhibition effect of variable pitch adjustment on the high-frequency load. According to the method, the load state vector fusing the low-frequency mechanical load and the high-frequency pneumatic excitation is constructed, and the rain flow counting and equivalent damage theory is combined, so that the load component which has the greatest influence on the fatigue life of key parts such as a main shaft, a transmission chain, a tower and a blade root is accurately identified.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of reducing fatigue load, in particular to a torque and variable pitch cooperative control method for wind farm aiming at reducing fatigue load. BACKGROUND

[0002] In the operation process of wind power generation system, the wind turbine is exposed to complex and changeable wind environment for a long time, and continuously bears dynamic aerodynamic load caused by turbulence, wind shear, tower shadow effect and yaw error, etc., which leads to repeated experience of alternating stress of key structural components such as main shaft, transmission chain, tower and blade, and thus significant fatigue damage. The traditional wind power control strategy usually takes maximizing energy capture or maintaining power / speed stability as the main target, and pays insufficient attention to structural fatigue load, and often uses independent torque control (used for adjusting speed and power) and variable pitch control (used for limiting power or stabilizing speed), and lacks systematic design of the cooperative effect of the two in load suppression. Especially in the medium and low wind speed area, although the rated power has not been reached, high-frequency aerodynamic excitation and low-frequency mechanical resonance can still cause serious fatigue accumulation; and in high wind speed or turbulent working conditions, frequent variable pitch action may exacerbate transmission chain impact and tower vibration if not coordinated with torque regulation. In addition, the existing control method relies on a simplified linear model, which is difficult to accurately depict the coupling influence mechanism of torque and variable pitch on fatigue load in different frequency bands, and also lacks real-time perception and prediction ability of the fatigue state of key components, resulting in disconnection between control decision and actual structural health state. Therefore, a torque and variable pitch cooperative control method for wind farm aiming at reducing fatigue load is provided. SUMMARY

[0003] The present application aims to provide a torque and variable pitch cooperative control method for wind farm aiming at reducing fatigue load, to solve the problem of ignoring the fatigue load suppression effect of torque and variable pitch cooperation in the traditional wind power control strategy and being difficult to balance power generation efficiency and structural life span.

[0004] To achieve the above-mentioned purpose, the present application aims to provide a torque and variable pitch cooperative control method for wind farm aiming at reducing fatigue load, comprising the following steps:

[0005] S1, obtaining real-time and historical operation data of each wind turbine in the wind farm, and preprocessing the operation data;

[0006] S2, based on the preprocessed operation data, extracting load characteristic indexes reflecting structural fatigue, and constructing a load state vector reflecting the current fatigue accumulation trend of the wind turbine;

[0007] S3, a coupling model of generator torque and variable pitch angle on the response of the speed and the influence of the fatigue load is established based on the load state vector, which is used to reflect the influence mechanism of torque adjustment on low-frequency load and the inhibition effect of variable pitch adjustment on high-frequency load;

[0008] S4, based on the coupling model, the control objective function of torque and variable pitch cooperative optimization is formed by taking the minimum equivalent fatigue load or load fluctuation index of the key components as the control objective and combining with the operation constraints;

[0009] S5, based on the load state vector, the particle swarm optimization method is used to solve the control objective function, and the optimal generator torque reference value and variable pitch angle reference value are generated;

[0010] S6, the optimal generator torque reference value and variable pitch angle reference value are converted into control instructions and sent to the generator controller and the variable pitch actuator respectively for cooperative control.

[0011] As a further improvement of the technical solution, in S1, the operating data at least includes the speed, the generated power, the main shaft torque, the blade pitch angle and the wind speed.

[0012] As a further improvement of the technical solution, in S2, the load characteristic index reflecting the structure fatigue is extracted, and the load state vector reflecting the current fatigue accumulation trend of the wind turbine is constructed, including the following steps:

[0013] S2.1, based on the preprocessed operating data, the load signal for characterizing the stress state of the structure is selected;

[0014] S2.2, time-frequency analysis method is used to divide the load signal into frequency bands;

[0015] S2.3, based on the frequency band division result, the statistical indexes of the main shaft torque and the speed are calculated respectively;

[0016] S2.4, based on the load signal corresponding to the key structure part of the wind turbine, the rain flow counting and the equivalent damage theory are combined to calculate the equivalent fatigue load value of each key component in the statistical time window;

[0017] S2.5, the statistical indexes and the equivalent fatigue load values are statistically processed in the time window to form a load characteristic set in a unified scale;

[0018] S2.6, the load characteristic set is combined according to the predetermined order to construct the load state vector reflecting the current load level and the fatigue accumulation trend of the wind turbine.

[0019] As a further improvement of the technical solution, in S3, the coupling model of generator torque and variable pitch angle on the response of the speed and the influence of the fatigue load is established based on the load state vector, including the following steps:

[0020] S3.1, normalizing the load state vector;

[0021] S3.2, defining the control input vector , and defining constraints for the control input vector ;

[0022] S3.3, based on the continuous dynamics equation, establishing a mapping relationship of the control input vector to the speed response ;

[0023] S3.4, based on the historical load state vector and the control input sequence, using a long short-term memory network model to establish a prediction model of the load state vector changing with the generator torque and the variable pitch angle;

[0024] S3.5, combining the mapping relationship of S3.3 with the prediction model of step S3.4 to form a complete coupling model of the generator torque and the variable pitch angle on the speed response and the fatigue load.

[0025] As a further improvement of the technical solution, in S3.3, based on the continuous dynamics equation, a mapping relationship of the control input vector to the speed response is established, including the following steps:

[0026] S3.31, at each time step , based on the current wind condition and the blade pitch angle , calculating the effective incoming wind speed , and calculating the blade aerodynamic torque according to the aerodynamic model;

[0027] S3.32, combining the blade aerodynamic torque with the load state vector;

[0028] S3.33, discretizing the continuous dynamics equation by using a first-order explicit Euler method;

[0029] S3.34, combining steps S3.32 and S3.33 into an input-output relationship to form a mapping relationship of the control input vector to the speed response .

[0030] As a further improvement of the technical solution, in S3.4, based on the historical load state vector and the control input sequence, a long short-term memory network model is used to establish a prediction model of the load state vector changing with the generator torque and the variable pitch angle, including the following steps:

[0031] S3.41, according to the historical time window length , selecting past the load state vector at the current moment and the corresponding control input sequence as the input sequence of the long short-term memory network model, and the load state vector at the future

[0032] S3.42, dividing the constructed input sequence and output sequence into a training set and a validation set;

[0033] S3.43, training the long short-term memory network model using the training set, wherein the loss function adopts the mean square error between the predicted load vector and the actual load vector, and the constraint on the fatigue-sensitive load feature is fused;

[0034] S3.44, using the trained long short-term memory network model as a prediction model of the load state vector varying with the generator torque and the pitch angle, to predict the load state vector at the future

[0035] As a further improvement of the technical solution, in S3.43, the loss function adopts the mean square error between the predicted load vector and the actual load vector, and the constraint on the fatigue-sensitive load feature is fused, including the following steps:

[0036] Extracting fatigue-related indicators for each load state vector, including equivalent load and stress amplitude, assigning weights to different load intervals; comparing the predicted load vector with the actual load vector, calculating the weighted mean square error, and combining the weighted mean square error with the fatigue-sensitive constraint term to form the final loss function.

[0037] As a further improvement of the technical solution, in S4, the control objective function of torque and pitch cooperative optimization is formed by minimizing the equivalent fatigue load or load fluctuation index of the key components and combining with the operation constraints, including the following steps:

[0038] S4.1, based on the load state vector, analyzing the influence degree of various loads on the fatigue life of the wind turbine structure, and selecting the key component load as the optimization object;

[0039] S4.2, based on the coupling model, and combining the load state vector at the current moment, the generator torque and the pitch angle control input, obtaining the predicted sequence of the key component load in the prediction time domain

[0040] S4.3, taking the minimization of the fluctuation level of the key component load in the prediction time domain as the control target, constructing the target function of torque and pitch cooperative optimization;

[0041] S4.4, introducing constraint conditions in the optimization process of the target function.

[0042] ​​​As a further improvement of the technical solution, in S5, the control objective function is solved based on the load state vector using a particle swarm optimization method to generate optimal generator torque reference values and variable pitch angle reference values, including the following steps:

[0043] S5.1, based on the load state vector at the current time, determine the control decision variable set in the prediction time domain The generator torque reference value sequence and the variable pitch angle reference value sequence are used as optimization variables, and are jointly encoded as a particle position vector;

[0044] S5.2, randomly initialize the particle position vector and velocity vector of the particle swarm;

[0045] S5.3, in each iteration process, for each particle in the particle swarm, input the torque and variable pitch control sequence corresponding to the particle into the load prediction model and the coupling model, calculate the key component load response in the prediction time domain, and obtain the fitness value of the particle based on the control objective function; compare the current fitness value with the historical optimal fitness of the particle;

[0046] S5.4, according to the velocity and position update rules of the particle swarm optimization algorithm, comprehensively update the velocity vector and position vector of each particle based on the individual optimal position and group optimal position information;

[0047] S5.5, perform constraint consistency check on the iteratively updated particle position vector;

[0048] S5.6, when the particle swarm reaches the preset maximum iteration number h, it is determined that the particle swarm optimization process converges, and the corresponding optimal generator torque reference value and variable pitch angle reference value are output.

[0049] As a further improvement of the technical solution, in S6, the optimal generator torque reference value and variable pitch angle reference value are converted into control instructions, including the following steps:

[0050] S6.1, perform reference value smoothing and dynamic feasibility check on the optimal generator torque reference value and variable pitch angle reference value;

[0051] S6.2, input the generator torque reference value after smoothing and dynamic feasibility check into the generator torque controller as the set value of the torque inner loop control, calculate the corresponding electromagnetic torque control instruction according to the current generator operating state, and further convert it into the current control instruction on the converter side;

[0052] S6.3. The pitch angle reference value after smoothing and dynamic feasibility verification is used as the target value of the pitch position controller. Combined with the current blade pitch angle feedback information, the pitch angle drive command corresponding to each blade is calculated. After limiting the pitch angle drive command and pitch rate, it is sent to the pitch actuator.

[0053] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0054] 1. This invention relates to a wind farm torque and pitch coordinated control method aimed at reducing fatigue loads. By constructing a load state vector that integrates low-frequency mechanical loads and high-frequency aerodynamic excitations, and combining rainflow counting and equivalent damage theory, it accurately identifies the load components that have the greatest impact on the fatigue life of key components such as the main shaft, drive train, tower, and blade roots. Based on this, a coupled influence model of torque and pitch on fatigue loads is established, and the optimization objective is to minimize the equivalent fatigue load or load fluctuation, thereby achieving coordinated control of generator torque and pitch angle. Compared to traditional independent control strategies, this method can effectively suppress the superimposed damage to the structure caused by loads in different frequency bands, significantly slow down the fatigue accumulation rate, thereby extending the service life of key wind turbine components and reducing operation and maintenance costs.

[0055] 2. This invention relates to a wind farm torque and pitch coordinated control method aimed at reducing fatigue loads. It integrates data-driven (LSTM prediction model) and mechanism modeling (dynamic equations) to optimize future control sequences in real time within the prediction time domain. A particle swarm optimization algorithm is used to solve for the optimal coordinated command that satisfies multiple operational constraints (such as torque / pitch amplitude and rate of change limits, safe speed range, and power grid connection requirements). Simultaneously, a smoothing process and dynamic feasibility verification mechanism are introduced before command issuance to ensure that the control output conforms to the physical characteristics of the actuator while avoiding mechanical shocks caused by high-frequency jitter or sudden changes. Therefore, while effectively reducing fatigue loads, it ensures the stable, safe, and efficient operation of the wind turbine under complex wind conditions. Attached Figure Description

[0056] Figure 1 This is a flowchart of the overall method of the present invention. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0058] Example: Please refer to Figure 1As shown, this embodiment provides a wind farm torque and pitch coordinated control method aimed at reducing fatigue load, including the following steps:

[0059] S1. Obtain real-time and historical operating data of each wind turbine in the wind farm, and preprocess the operating data;

[0060] In this embodiment, the operating data includes at least rotational speed, power generation, main shaft torque, blade pitch angle, wind speed, tower acceleration, and wind direction.

[0061] S2. Based on the preprocessed operating data, extract load characteristic indicators that reflect structural fatigue, and construct a load state vector that reflects the current fatigue accumulation trend of the wind turbine.

[0062] In this embodiment, load characteristic indicators reflecting structural fatigue are extracted, and a load state vector reflecting the current fatigue accumulation trend of the wind turbine is constructed, including the following steps:

[0063] S2.1 Based on the preprocessed operating data, load signals for characterizing the structural stress state are selected. These load signals include at least the main shaft torque, rotational speed, tower load, and blade root load. Specifically: First, the physical quantities most sensitive to fatigue accumulation in key structural components (such as the main shaft, transmission chain, tower, and blade root) are identified as candidate signals, including main shaft torque, rotational speed, tower acceleration, and blade root load. Second, mechanical low-frequency loads and aerodynamic high-frequency excitation loads are distinguished according to different load sources and frequency characteristics to ensure that the selected signals can comprehensively reflect the stress state of the wind turbine structure under different operating conditions. Finally, combined with statistical analysis and historical operating experience, redundant or highly correlated signals are eliminated to achieve effective characterization of key fatigue behaviors and form a high-information signal set that can be used to construct load state vectors.

[0064] S2.2 The load signal is divided into frequency bands using time-frequency analysis, decomposing it into low-frequency mechanical load components and mid-to-high-frequency aerodynamic excitation load components to distinguish fatigue mechanisms from different sources. Specifically: First, the pre-processed load signal (such as spindle torque, speed, tower acceleration, and blade root load) is discretized, and appropriate time-frequency analysis tools, such as Short-Time Fourier Transform (STFT), wavelet transform, or Hilbert-Huang Transform (HHT), are selected. The frequency resolution and window parameters are set according to the wind turbine structural characteristics and fatigue-sensitive frequency range. Then, the load signal is decomposed into low-frequency mechanical load components and mid-to-high-frequency aerodynamic excitation load components on the time-frequency plane. The boundaries of each frequency band are determined through amplitude spectrum or energy distribution analysis, and statistical features (such as root mean square value and fluctuation amplitude) are extracted from each decomposed frequency band signal, thereby achieving the distinction and quantification of fatigue mechanisms from different sources.

[0065] S2.3 Calculate the statistical indices of spindle torque and speed based on the frequency band division results, including fluctuation amplitude and root mean square (RMS) values, to characterize the fatigue load intensity related to the transmission chain. Specifically: First, the low-frequency mechanical load component and the mid-to-high-frequency aerodynamic excitation component are respectively truncated and their time series processed to ensure that the samples of each frequency band cover the entire statistical time window. Then, for each frequency band signal, the fluctuation amplitude is calculated, usually the difference between the maximum and minimum values ​​of the signal or the peak-to-peak value, to characterize the instantaneous fluctuation range of the load. At the same time, the root mean square (RMS) value is calculated, which is obtained by averaging the squares of the discrete sample data of the signal and then taking the square root, to quantify the load energy intensity or overall vibration level. For multi-channel or multi-component signals, calculations can be performed separately for each component or channel, and normalization or weighting can be performed when necessary to form a statistical feature set of a uniform scale, which serves as the input for constructing the load state vector.

[0066] S2.4. Based on the load signals corresponding to key structural parts of the wind turbine (tower and blade roots, etc.), and combined with rainflow counting and equivalent damage theory, the equivalent fatigue load values ​​of each key component within the statistical time window are calculated to characterize the degree of fatigue accumulation in different structural parts. Specifically: First, the preprocessed load signal is segmented according to the statistical time window to ensure that each time period contains sufficient cyclic features to represent fatigue behavior. Then, the rainflow counting algorithm is applied to each segment of the load signal to extract the cyclic load amplitude and average value, identify the positive and negative cycles, and form a complete cyclic load sequence. Next, based on the selected material S-N curve or fatigue damage parameters, the equivalent damage contribution corresponding to each cyclic load is calculated. Usually, the Miner linear accumulation method is used to proportionally superimpose the cyclic damage to obtain the total equivalent damage within the statistical window. Finally, the total equivalent damage is converted into an equivalent fatigue load value, which is represented by an equivalent single amplitude or RMS load to reflect the degree of fatigue accumulation of the key component within the time window, providing fatigue-sensitive features for the load state vector.

[0067] S2.5 Perform time window statistical processing on statistical indicators and equivalent fatigue load values, and eliminate the dimensional differences and scale effects between different features through normalization, standardization or weighted fusion to form a load feature set under a unified scale.

[0068] S2.6. Combine the load feature set in a predetermined order (structural part order) to construct a load state vector that reflects the current load level and fatigue accumulation trend of the wind turbine.

[0069] S3. Based on the load state vector, establish a coupled model of the influence of generator torque and pitch angle on speed response and fatigue load, which is used to reflect the influence mechanism of torque regulation on low-frequency load and the suppression effect of pitch regulation on high-frequency load.

[0070] In this embodiment, a coupled model of the influence of generator torque and pitch angle on speed response and fatigue load is established based on the load state vector, including the following steps:

[0071] S3.1 Standardize (i.e., subtract the mean and divide by the standard deviation to make it follow a zero-mean unit variance distribution) or normalize (i.e., linearly map to a fixed interval, such as [0,1], according to the maximum and minimum values) the load state vector to eliminate dimensional differences and improve the stability of model training or recognition.

[0072] S3.2, Define the control input vector ,in, This is the generator torque command. This is the blade pitch angle command. For transpose operation, For time, and for the control input vector Define constraints; specifically: for generator torque commands Set maximum and minimum value constraints, and define an upper limit for the torque change rate to prevent excessively rapid adjustment from causing mechanical shock or control instability; for blade pitch angle... First, upper and lower limits for the pitch angle and pitch rate are set to ensure smooth adjustment of the pitch angle within the allowable range of the blade structure. Then, these amplitude and rate-of-change constraints are explicitly defined mathematically, such as... ,and ( This is the minimum torque constraint for the generator, expressed in N·m, representing the minimum allowable torque for the control input. This is a constraint on the maximum generator torque, expressed in N·m, representing the maximum allowable torque for the control input. This refers to the change in generator torque, measured in N·m, and typically indicates the torque change within adjacent control cycles or time steps. (for time step); finally, all constraint information is integrated into the control input vector constraint definition for subsequent coupled model calculation and optimization solution;

[0073] S3.3. Based on the continuous dynamic equation, establish the control input vector. Response to rotational speed The mapping relationship;

[0074] Among them, the control input vector is established based on the continuous dynamic equation. Response to rotational speed The mapping relationship includes the following steps:

[0075] S3.31, For each time step Based on the current wind conditions and blade pitch angle Calculate the effective inflow velocity The aerodynamic torque of the blade is calculated based on the aerodynamic model (blade element momentum theory, BEM). :

[0076] ;

[0077] S3.32, the aerodynamic torque of the blades Combined with the load state vector;

[0078] S3.33. Discretize the continuous dynamic equations using the first-order explicit Euler method:

[0079] ;

[0080] S3.34. Combine steps S3.32 and S3.33 into an input-output relationship to form a control input vector. Response to rotational speed Mapping relationship:

[0081] ;

[0082] In the formula, This is the blade aerodynamic model function, describing the nonlinear mapping relationship between wind speed and pitch angle on aerodynamic torque, defined by the blade element momentum theory (BEM). Rotor speed (unit: rad / s) at time step The value, Rotor speed (unit: rad / s) at time step The value, The time step (in seconds) is used to discretize continuous-time dynamic equations. The equivalent moment of inertia of the rotor and drive train (unit: kg·m²) describes the rotor's ability to resist acceleration or deceleration. The generator torque (unit: N·m) is used as the control input to adjust the output speed. This is the load state vector. This is a dynamic mapping function that maps control inputs and load states to rotational speed outputs, obtained through discretized dynamic equations or data-driven methods.

[0083] S3.4. Based on the historical load state vector and control input sequence, a prediction model for the load state vector as a function of generator torque and pitch angle is established using a long short-term memory network model.

[0084] The process involves establishing a predictive model for the load state vector as a function of generator torque and pitch angle, based on historical load state vectors and control input sequences, using a long short-term memory network model. This includes the following steps:

[0085] S3.41, Based on the length of the historical time window Select the past The load state vector at each time step and the corresponding control input sequence are used as the input sequence of the long short-term memory network model, and the future... The load state vector at each time step is used as the output sequence of the long short-term memory network model. The input and output sequences are normalized or standardized to accelerate training convergence and reduce the impact of differences in the dimensions of different features.

[0086] The Long Short-Term Memory (LSTM) network model consists of an input layer, several LSTM hidden layer units, and an output layer. The input layer receives the length of the historical time window. The load state vector within the LSTM and the corresponding control input sequence form a multi-dimensional time series input. The hidden layers of the LSTM use gating mechanisms (input gate, forget gate, output gate) to remember and update the state of the input sequence, thereby capturing long-term dependencies. A single-layer or multi-layer stacked structure can be configured according to task requirements. Each hidden layer contains several memory units used to extract temporal dynamic features and nonlinear mapping relationships. The output layer maps the final state of the LSTM hidden layers to the future prediction time domain. The load state vector sequence within the model is typically multidimensionally output through fully connected layers or linear transformations. The entire model can combine standardized or normalized input preprocessing, time step expansion, and backpropagation training strategies. By minimizing the weighted mean square error between the predicted load vector and the actual load vector, as well as fatigue sensitivity constraints, it learns the dynamic mapping relationship between the control input and the load response, thereby achieving accurate prediction of future load states.

[0087] S3.42. Divide the constructed input and output sequences into a training set and a validation set;

[0088] S3.43. The long short-term memory network model is trained using the training set. Its loss function adopts the mean square error between the predicted load vector and the actual load vector, and incorporates the constraint on the characteristics of fatigue-sensitive loads, so that the model can reduce the impact of structural fatigue while fitting the load sequence.

[0089] Furthermore, the loss function addresses the real-world problem of wind turbines operating under complex wind conditions. Traditional prediction models often focus only on the fitting accuracy of the overall load sequence, neglecting the cumulative fatigue impact of key structural components (such as the main shaft, tower, and blade roots) in high-load or fatigue-sensitive regions. This makes it difficult to effectively reduce structural fatigue during control optimization. By integrating fatigue-sensitive load feature constraints into the mean square error, this method can assign higher weights to high-fatigue-risk regions when predicting load sequences. This allows the model to not only fit the overall load trend but also focus on load changes that significantly affect structural lifespan. Its advantage over existing technologies lies in its ability to quantify and optimize structural fatigue accumulation, providing more accurate fatigue-guided predictions for torque and pitch coordinated control, thereby extending the lifespan of key components and reducing maintenance costs.

[0090] Its loss function uses the mean square error between the predicted load vector and the actual load vector, and incorporates constraints on fatigue-sensitive load characteristics, including the following steps:

[0091] For each load state vector, fatigue-related indices are extracted, including equivalent load and stress amplitude. Weights are assigned to different load ranges (such as high load or fatigue-sensitive ranges) and to critical load ranges (such as high load or fatigue-sensitive ranges). For the ordinary load range, it is set to A weighted vector is formed to highlight loads that have a significant impact on structural fatigue. The critical load range refers to the range of load amplitudes or cycles that contribute significantly to the cumulative fatigue of the wind turbine during operation and may shorten its lifespan. For example, for main shaft torque or tower bending moment signals, rainflow counting analysis can statistically analyze the stress amplitude and cycle number of each cycle, and calculate the contribution of that cycle to the total fatigue damage based on the material's S-N curve. When the fatigue damage contribution of a certain cycle exceeds a preset threshold (e.g., accounting for more than 5% of the total damage in the statistical window), the load amplitude range to which that cycle belongs is defined as the critical load range. The ordinary load range refers to the load range with a smaller contribution to fatigue damage and no significant impact on structural lifespan, i.e., the part with low stress amplitude or few cycles, whose cumulative equivalent damage proportion is below the threshold. The predicted load vector is compared with the actual load vector to calculate the weighted mean square error, and the weighted mean square error is combined with the fatigue-sensitive constraint term to form the final loss function.

[0092] Among them, the mean square error (MSE) term for:

[0093] ;

[0094] Fatigue sensitivity constraint for:

[0095] ;

[0096] Final loss function for:

[0097] ;

[0098] In the formula, This refers to the total number of training samples or the total number of time steps. For sample or time step indexing, For the first The weights of each sample, The model predicts the first The load state vector at each time step. For the first actual observation The load state vector at each time step. For fatigue constraint weights, To map load vectors to functions of characteristic indices related to structural fatigue, including but not limited to equivalent loads, stress amplitudes, or load cycle counts, for quantifying the impact of predicted load sequences on structural fatigue accumulation;

[0099] S3.44. A trained long short-term memory network model is used as a prediction model for the load state vector as a function of generator torque and pitch angle to predict future changes. The load state vector at each moment;

[0100] S3.5. Combine the mapping relationship of S3.3 with the prediction model of step S3.4 to form a complete coupled model of the influence of generator torque and pitch angle on speed response and fatigue load.

[0101] S4. Based on the coupled model, the control objective is to minimize the equivalent fatigue load or load fluctuation index of key components and combine it with the operating constraints to form a control objective function for torque and pitch co-optimization.

[0102] In this embodiment, the control objective is to minimize the equivalent fatigue load or load fluctuation index of key components, combined with operational constraints, to form a control objective function for torque and pitch co-optimization, including the following steps:

[0103] S4.1 Based on the load state vector, analyze the impact of various loads on the fatigue life of the wind turbine structure, and select the loads of key components that have a significant impact on life as optimization objects, including at least: torque fatigue load of the main shaft and drive train, forward / backward / lateral bending moment load of the tower, and equivalent fatigue load at the blade root; specifically: first, extract the load signals of each key structural component (such as the main shaft, drive train, tower, and blade root) from the load state vector, and calculate the amplitude and number of cycles of each cyclic load using the rainflow counting method within the statistical time window; then, combine the material S-N curve or equivalent fatigue damage theory to calculate the contribution of each cycle to the cumulative fatigue damage, and normalize or standardize the equivalent fatigue load of each component to quantify its sensitivity to structural life; finally, based on the set threshold, determine the component loads that have a significant impact on life as key component loads, and use them as optimization targets to ensure that the fatigue accumulation and load fluctuation of these components are reduced first during the torque and pitch coordinated control process;

[0104] S4.2. Based on the coupled model, and combined with the load state vector, generator torque, and pitch angle control input at the current moment, in the prediction time domain... Predicted sequence of loads for key components obtained internally. Specifically, it means: within the prediction time domain, according to time steps Iterate step by step, with each step first calculating the rotational speed response through dynamic mapping. Then, the LSTM prediction model is used to obtain the load state vector at future time points. Repeat the above iterations until the entire prediction time domain is covered. The load signals corresponding to key components are extracted from the predicted load state vector to form a prediction sequence. ;

[0105] S4.3. With minimizing the load fluctuation level of key components in the predicted time domain as the control objective, construct an objective function for torque and pitch co-optimization. In the formula, Number of key load types To reflect the fatigue sensitivity of different components, This is the fatigue equivalent mapping function or fluctuation metric function corresponding to the load, used to uniformly map the predicted load to an optimizable fatigue evaluation index space;

[0106] S4.4. Constraints are introduced during the objective function optimization process, including generator torque constraints and torque change rate constraints, pitch angle and pitch rate constraints, safe operating speed range constraints, and power deviation or grid-connected operation constraints. Specifically: First, upper and lower limits for the amplitude and upper limit for the rate of change of the generator torque are set to prevent mechanical shock and excessively rapid adjustment. Second, upper and lower limits for the blade pitch angle and the maximum pitch rate are set to ensure smooth blade adjustment within the structural allowable range. Then, a safe operating speed range constraint is applied to the rotor speed to avoid overspeed or underspeed. Finally, deviation constraints are set for the output power or grid-connected operation to ensure that the power output meets the grid connection requirements. These constraints are integrated into the objective function optimization through mathematical expression or penalty functions, so that particle swarm optimization or other optimization algorithms focus on minimizing fatigue load while meeting the safety requirements of wind turbine structure and operation when searching for the optimal torque and pitch control sequence, forming a safe and feasible collaborative optimization logic.

[0107] S5. Based on the load state vector, the control objective function is solved using the particle swarm optimization method to generate the optimal generator torque reference value and pitch angle reference value.

[0108] In this embodiment, based on the load state vector, the control objective function is solved using the particle swarm optimization method to generate the optimal generator torque reference value and pitch angle reference value, including the following steps:

[0109] S5.1 Determine the prediction time domain based on the load state vector at the current moment. The set of control decision variables within the system uses the generator torque reference value sequence and the pitch angle reference value sequence as optimization variables, and jointly encodes them into a particle position vector. ;

[0110] Each particle corresponds to a set of candidate torque and pitch control sequences;

[0111] S5.2 Under the conditions of satisfying the constraints of generator torque amplitude and rate of change, pitch angle and pitch rate, the particle position vector and velocity vector of the particle swarm are randomly initialized; at the same time, the speed safety range, power deviation or grid-connected operation constraints are mapped to the particle search space by penalty function or feasibility correction method to ensure that the initial particles meet the basic operating requirements.

[0112] S5.3 In each iteration, for each particle in the particle swarm, its corresponding torque and pitch control sequence are input into the load prediction model and the coupled model to calculate the load response of key components in the prediction time domain, and the fitness value of the particle is obtained based on the control objective function (the fatigue load or load fluctuation index of each particle is calculated using the objective function and used as the fitness value of the particle); the current fitness value is compared with the particle's historical best fitness, and if the current fitness is better, its individual optimal position is updated; at the same time, the position of the particle with the best fitness among all particles is selected as the group optimal position of the current iteration; the fitness value of each particle in the current iteration is compared with its historical best fitness.

[0113] S5.4 Based on the velocity and position update rules of the particle swarm optimization algorithm, the velocity vector and position vector of each particle are iteratively updated by combining the optimal position information of the individual and the optimal position information of the group, so that the particles gradually move towards the cooperative control area with lower fatigue load or smaller load fluctuation in the search space.

[0114] S5.5 Perform a constraint consistency check on the iteratively updated particle position vector. When the particle position exceeds the constraint range of torque, pitch angle or its rate of change, the particle is corrected by boundary truncation, bounce correction or penalty function enhancement to ensure that the control sequence meets the constraints of wind turbine safe operation.

[0115] S5.6 When the particle swarm reaches the preset maximum number of iterations h, the particle swarm optimization process is determined to have converged, and the corresponding optimal generator torque reference value and pitch angle reference value are output.

[0116] S6. Convert the optimal generator torque reference value and pitch angle reference value into control commands and send them to the generator controller and pitch actuator respectively for coordinated control.

[0117] In this embodiment, converting the optimal generator torque reference value and pitch angle reference value into control commands includes the following steps:

[0118] S6.1. Smoothing and dynamic feasibility verification of the optimal generator torque reference value and pitch angle reference value; Specifically, when smoothing and verifying the dynamic feasibility of the optimal generator torque reference value and pitch angle reference value, the discrete reference value sequence obtained by optimization is first smoothed by low-pass filtering or moving weighted average to reduce high-frequency fluctuations introduced by prediction errors or instability of optimization solution; then, combined with the current operating conditions of the wind turbine and the physical constraints of the actuator, the smoothed reference value is dynamically verified. Specifically, rated torque and instantaneous overload capacity constraints are applied to the generator torque reference value, pitch angle upper and lower limits and maximum pitch rate constraints are applied to the pitch angle reference value, and a rate of change limit is introduced between adjacent control moments; when the reference value exceeds the constraint range, it is corrected online by saturation mapping or constraint projection, so as to obtain feasible reference values ​​for generator torque and pitch angle that satisfy the dynamic characteristics of the actuator in terms of amplitude and rate of change, providing stable and executable input for subsequent control command generation;

[0119] S6.2. The smoothed and dynamically feasible generator torque reference value is used as the setpoint for the inner loop torque control and input to the generator torque controller. The corresponding electromagnetic torque control command is calculated based on the current generator operating state and further converted into current or voltage control commands on the converter side to drive the generator to output the target torque. Specifically, the smoothed and feasible generator torque reference value, along with real-time operating parameters such as the current generator speed, rotor position, and excitation state, are input to the generator torque controller. The required electromagnetic torque output is calculated using a closed-loop control algorithm (such as PI or feedforward-feedback composite control). Then, based on the generator's mechanical model, the electromagnetic torque control quantity is mapped to the corresponding stator current or voltage reference value. Considering the converter's rated current, voltage limiting, and switching frequency constraints, the generated current / voltage commands are limited and dynamically adjusted, and finally output to the converter to achieve precise control of the generator's electromagnetic torque, ensuring that the actual output torque tracks the reference torque curve while guaranteeing that the converter operates within a safe and controllable range.

[0120] S6.3. The smoothed and dynamically feasible reference value of the pitch angle is used as the target value of the pitch position controller. Combined with the current blade pitch angle feedback information, the corresponding pitch angle drive command for each blade is calculated. After limiting the amplitude and pitch rate of the pitch angle drive command, it is sent to the pitch actuator to realize the actual adjustment of the blade pitch angle. Specifically, the smoothed and feasible reference value of the pitch angle and the current actual pitch angle feedback information of each blade are input into the pitch position controller. The deviation of each blade is calculated by a closed-loop control algorithm (such as PID or feedforward-feedback composite control) and the corresponding drive torque or motor speed reference value is generated. Then, combined with the maximum drive torque, speed limit and mechanical limit of the blade actuator, the calculated drive command is limited and rate constrained to ensure that the output command is within the safe and executable range. Finally, the processed drive command is sent to the pitch motor or hydraulic actuator of each blade to realize the precise adjustment of the blade pitch angle, so that the actual pitch angle can smoothly track the reference value, while meeting the system dynamic response and safety constraint requirements.

[0121] After issuing torque and pitch control commands, feedback information such as generator speed, output power, and blade pitch angle is collected to update the wind turbine operating status, providing initial values ​​for load prediction and optimization in the next control cycle, thus forming a closed-loop control process that coordinates torque control and pitch control.

[0122] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A wind farm torque and pitch coordinated control method aimed at reducing fatigue load, characterized in that, Includes the following steps: S1. Obtain real-time and historical operating data of each wind turbine in the wind farm, and preprocess the operating data; S2. Based on the preprocessed operating data, extract load characteristic indicators that reflect structural fatigue, and construct a load state vector that reflects the current fatigue accumulation trend of the wind turbine. S3. Based on the load state vector, establish a coupled model of the influence of generator torque and pitch angle on speed response and fatigue load, which is used to reflect the influence mechanism of torque regulation on low-frequency load and the suppression effect of pitch regulation on high-frequency load. S4. Based on the coupled model, the control objective is to minimize the equivalent fatigue load or load fluctuation index of key components and combine it with the operating constraints to form a control objective function for torque and pitch co-optimization. S5. Based on the load state vector, the control objective function is solved using the particle swarm optimization method to generate the optimal generator torque reference value and pitch angle reference value. S6. Convert the optimal generator torque reference value and pitch angle reference value into control commands and send them to the generator controller and pitch actuator respectively for coordinated control.

2. The wind farm torque and pitch coordinated control method according to claim 1, aimed at reducing fatigue load, is characterized in that: In S1, the operating data includes at least rotational speed, power generation, main shaft torque, blade pitch angle, and wind speed.

3. The wind farm torque and pitch coordinated control method according to claim 2, aimed at reducing fatigue load, is characterized in that: In step S2, load characteristic indicators reflecting structural fatigue are extracted, and a load state vector reflecting the current fatigue accumulation trend of the wind turbine is constructed, including the following steps: S2.

1. Based on the preprocessed running data, select load signals to characterize the stress state of the structure; S2.

2. Use time-frequency analysis to divide the load signal into frequency bands; S2.3 Calculate the statistical indices of spindle torque and speed based on the frequency band division results; S2.4 Based on the load signals corresponding to the key structural parts of the wind turbine, and combined with rainflow counting and equivalent damage theory, calculate the equivalent fatigue load value of each key component within the statistical time window; S2.

5. Perform time window statistical processing on statistical indicators and equivalent fatigue load values ​​to form a load characteristic set under a unified scale; S2.

6. Combine the load feature set in a predetermined order to construct a load state vector that reflects the current load level and fatigue accumulation trend of the wind turbine.

4. The wind farm torque and pitch coordinated control method according to claim 1, aimed at reducing fatigue load, is characterized in that: In step S3, a coupled model of the influence of generator torque and pitch angle on speed response and fatigue load is established based on the load state vector, including the following steps: S3.1 Standardize the load state vector; S3.2, Define the control input vector and control input vector Define constraints; S3.

3. Based on the continuous dynamic equation, establish the control input vector. Response to rotational speed The mapping relationship; S3.

4. Based on the historical load state vector and control input sequence, a prediction model for the load state vector as a function of generator torque and pitch angle is established using a long short-term memory network model. S3.

5. Combine the mapping relationship of S3.3 with the prediction model of step S3.4 to form a complete coupled model of the influence of generator torque and pitch angle on speed response and fatigue load.

5. The wind farm torque and pitch coordinated control method according to claim 4, aimed at reducing fatigue load, is characterized in that: In step S3.3, a control input vector is established based on the continuous dynamics equation. Response to rotational speed The mapping relationship includes the following steps: S3.31, For each time step Based on the current wind conditions and blade pitch angle Calculate the effective inflow velocity Calculate the aerodynamic torque of the blades based on the aerodynamic model. ; S3.32, the aerodynamic torque of the blades Combined with the load state vector; S3.

33. Discretize the continuous dynamic equations using the first-order explicit Euler method; S3.

34. Combine steps S3.32 and S3.33 into an input-output relationship to form a control input vector. Response to rotational speed The mapping relationship.

6. The wind farm torque and pitch coordinated control method according to claim 4, aimed at reducing fatigue load, is characterized in that: In step S3.4, based on the historical load state vector and control input sequence, a prediction model for the load state vector as a function of generator torque and pitch angle is established using a long short-term memory network model, including the following steps: S3.41, Based on the length of the historical time window Select the past The load state vector at each time step and the corresponding control input sequence are used as the input sequence of the long short-term memory network model, and the future... The load state vector at each time step is used as the output sequence of the long short-term memory network model; S3.

42. Divide the constructed input and output sequences into a training set and a validation set; S3.

43. Train the Long Short-Term Memory Network model using the training set. Its loss function adopts the mean square error between the predicted load vector and the actual load vector, and incorporates constraints on fatigue-sensitive load characteristics. S3.

44. A trained long short-term memory network model is used as a prediction model for the load state vector as a function of generator torque and pitch angle to predict future changes. The load state vector at each time step.

7. The wind farm torque and pitch coordinated control method according to claim 6, aimed at reducing fatigue load, is characterized in that: In S3.43, the loss function uses the mean square error between the predicted load vector and the actual load vector, and incorporates constraints on fatigue-sensitive load characteristics, including the following steps: For each load state vector, fatigue-related indices are extracted, including equivalent load and stress amplitude, and weights are assigned to different load intervals. The predicted load vector is compared with the actual load vector, the weighted mean square error is calculated, and the weighted mean square error is combined with the fatigue-sensitive constraint term to form the final loss function.

8. The wind farm torque and pitch coordinated control method according to claim 1, aimed at reducing fatigue load, is characterized in that: In step S4, the control objective function for torque and pitch coordination optimization is formed by minimizing the equivalent fatigue load or load fluctuation index of key components and combining it with operational constraints. This includes the following steps: S4.1 Based on the load state vector, analyze the influence of various loads on the fatigue life of the wind turbine structure, and select the loads of key components as the optimization objects; S4.

2. Based on the coupled model, and combined with the load state vector, generator torque, and pitch angle control input at the current moment, in the prediction time domain... The predicted sequence of loads on key components was obtained internally; S4.

3. With minimizing the load fluctuation level of key components in the predicted time domain as the control objective, construct an objective function for torque and pitch co-optimization. S4.4 Introduce constraints during the objective function optimization process.

9. The wind farm torque and pitch coordinated control method according to claim 1, aimed at reducing fatigue load, is characterized in that: In step S5, based on the load state vector, the control objective function is solved using the particle swarm optimization method to generate the optimal generator torque reference value and pitch angle reference value, including the following steps: S5.1 Determine the prediction time domain based on the load state vector at the current moment. The set of control decision variables within the system uses the generator torque reference value sequence and the pitch angle reference value sequence as optimization variables, and jointly encodes them into a particle position vector. S5.

2. Randomly initialize the particle position vector and velocity vector of the particle swarm; S5.3 In each iteration, for each particle in the particle swarm, its corresponding torque and pitch control sequence are input into the load prediction model and the coupled model to calculate the load response of key components in the prediction time domain, and the fitness value of the particle is obtained based on the control objective function; the current fitness value is compared with the particle's historical best fitness. S5.4 Based on the velocity and position update rules of the particle swarm optimization algorithm, and combining the optimal position information of the individual and the optimal position information of the group, the velocity vector and position vector of each particle are iteratively updated. S5.5 Perform a constraint consistency check on the iteratively updated particle position vector; S5.6 When the particle swarm reaches the preset maximum number of iterations h, the particle swarm optimization process is determined to have converged, and the corresponding optimal generator torque reference value and pitch angle reference value are output.

10. The wind farm torque and pitch coordinated control method according to claim 1, aimed at reducing fatigue load, is characterized in that: In step S6, converting the optimal generator torque reference value and pitch angle reference value into control commands includes the following steps: S6.1 Perform reference value smoothing and dynamic feasibility verification on the optimal generator torque reference value and pitch angle reference value; S6.

2. The generator torque reference value after smoothing and dynamic feasibility verification is used as the set value of the torque inner loop control and input to the generator torque controller. The corresponding electromagnetic torque control command is calculated according to the current generator operating status and further converted into the current control command on the converter side. S6.

3. The pitch angle reference value after smoothing and dynamic feasibility verification is used as the target value of the pitch position controller. Combined with the current blade pitch angle feedback information, the pitch angle drive command corresponding to each blade is calculated. After limiting the pitch angle drive command and pitch rate, it is sent to the pitch actuator.

Images