Wind turbine control optimization method based on dynamic changes of wind speed and direction

Abstract

The application relates to the technical field of wind turbine control and discloses a wind turbine control optimization method based on dynamic changes of wind speed and wind direction. The method receives real-time wind speed time series data and three-dimensional wind direction vector field data of a target wind field, constructs a wind field dynamic analysis model by using a space-time convolutional neural network, generates a wind field energy density distribution matrix and a turbulence intensity probability graph. A multi-objective adaptive optimization model is constructed based on this, a set of wind turbine control parameter instructions is generated, a preset wind turbine load-power generation efficiency balance equation is used to optimize a cooperative regulation coefficient, the Bayesian optimization framework is iteratively updated, and an optimal control action sequence is output to a wind turbine main control system. The application can accurately perceive changes in a wind field, effectively balance wind turbine load and power generation efficiency, realize multi-wind turbine cooperative control, improve wind energy capture efficiency, and improve the stability and economic benefits of wind turbine operation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wind turbine control, in particular to a wind turbine control optimization method based on dynamic changes of wind speed and direction. BACKGROUND

[0002] In the field of wind power generation, the efficient and stable operation of wind turbines is crucial for improving power generation efficiency and reducing costs. However, the wind field environment is complex and variable, and the dynamic changes of wind speed and direction bring great challenges to the control of wind turbines.

[0003] Traditional wind turbine control methods are mostly based on fixed parameters or simple empirical models, which are difficult to adapt to complex wind conditions. For example, in the case of frequent wind speed fluctuations, the setting of fixed pitch angle and yaw angle cannot adjust the unit posture in time, resulting in low wind energy capture efficiency. Moreover, this "one-size-fits-all" control method does not fully consider the uniqueness of different wind fields, and cannot be optimized according to local meteorological conditions and geographical environment, so that wind turbines often operate in a non-optimal state in actual operation.

[0004] In addition, there is a complex trade-off between the load and power generation efficiency of wind turbines. In the pursuit of high power generation efficiency, the unit may bear excessive load, accelerate component wear, shorten equipment service life, and increase maintenance costs; conversely, excessive attention to load control will sacrifice power generation efficiency and reduce economic benefits. Existing technologies often fail to find the best balance between the two, resulting in the overall performance of wind turbines not being fully utilized.

[0005] At the same time, with the continuous expansion of wind farms, the demand for coordinated operation of multiple units is increasingly prominent. However, there is currently a lack of effective coordinated control mechanism between multiple units, and units interfere with each other, unable to form an efficient power generation cluster. For example, when the wind direction changes, the yaw action of adjacent units may affect each other, resulting in reduced wind energy capture efficiency, and even causing unit failure.

[0006] In addition, real-time monitoring and accurate prediction technology of wind field also needs to be improved. The wind field data obtained by existing monitoring equipment and methods is not comprehensive and accurate enough to reflect the subtle changes and complex characteristics of the wind field. This makes the wind turbine lag in responding to changes in wind conditions, unable to adjust the control strategy in time, further affecting the power generation efficiency and stability of the unit.

[0007] With the gradual rise of wind power in the energy field, the requirements for wind turbine control technology are also becoming higher and higher. The existing control methods have many shortcomings in dealing with dynamic changes of wind speed and direction, balancing unit load and power generation efficiency, realizing coordinated operation of multiple units, and accurately monitoring and predicting wind field, etc. Therefore, an intelligent and efficient control optimization method is urgently needed to improve the overall performance of wind turbines and meet the growing energy demand. SUMMARY

[0008] The present application aims to provide a wind turbine control optimization method based on dynamic changes of wind speed and direction to solve the problems raised in the background.

[0009] To achieve the above-mentioned purpose, the present application provides the following technical solution: a wind turbine control optimization method based on dynamic changes of wind speed and direction, the method comprising:

[0010] receiving real-time wind speed time series data and three-dimensional wind direction vector field data of a target wind field, the three-dimensional wind direction vector field data including laser radar wind measurement point cloud and atmospheric boundary layer turbulence spectrum;

[0011] constructing a wind field dynamic analysis model based on a spatio-temporal convolutional neural network, performing multi-scale feature extraction and fusion on the real-time data, and generating a wind field energy density distribution matrix;

[0012] constructing a multi-objective adaptive optimization model according to the energy density distribution matrix, generating a turbine control parameter instruction set, the instruction set including a yaw angle adjustment sequence and a blade pitch angle optimization strategy;

[0013] based on a preset turbine load-power generation efficiency balance equation, simulating the dynamic game relationship of turbine operating state in a future preset time window, and optimizing the cooperative adjustment coefficient in the control parameter instruction set;

[0014] iteratively updating the cooperative adjustment coefficient through a Bayesian optimization framework, and outputting an optimal control action sequence to a wind turbine main control system.

[0015] Preferably, the construction step of the wind field dynamic analysis model comprises:

[0016] collecting historical wind field data sets under multiple meteorological conditions, and constructing multi-dimensional training samples containing wind speed fluctuation characteristics, wind direction shear gradient, and turbine response delay time;

[0017] performing hidden variable inference on the multi-dimensional training samples through a pre-constructed deep probabilistic graph model, and extracting independent representations of wind field dominant modes and random disturbance factors;

[0018] combining a simplified form of Navier-Stokes equation to construct a partial differential constraint condition for the wind field evolution process;

[0019] embedding the partial differential constraint condition into the residual module of the spatio-temporal convolutional neural network to generate the wind field dynamic analysis model supporting online updating.

[0020] Preferably, the multi-objective adaptive optimization model comprises:

[0021] According to the spatio-temporal variation rate of the energy density distribution matrix, a wind energy capture priority area is dynamically divided;

[0022] Based on the mechanical fatigue accumulation index of the unit and the grid dispatching demand curve, the benefit-risk trade-off score of the control strategy is calculated;

[0023] The trade-off score and the priority area are nonlinearly weighted by a Sigmoid function to generate a unit-specific control threshold;

[0024] According to the threshold, the switching condition of the multi-stage regulation mode is triggered.

[0025] Preferably, the construction step of the unit load-power balance equation comprises:

[0026] Collecting unit operating state parameters and structural stress monitoring data, a load-power mapping relationship dataset is constructed;

[0027] Through a fuzzy logic controller, a membership function of blade aerodynamic load and a fuzzy rule base of power generation efficiency are generated;

[0028] Combined with a multi-objective genetic algorithm, a Pareto front search space is constructed to quantify the conflict intensity between control parameters;

[0029] The conflict intensity and real-time wind field disturbance factor are input into a dynamic game network to generate the unit load-power balance equation.

[0030] Preferably, the method further comprises:

[0031] According to the simulation results of the dynamic game network, a key parameter coupling node is identified;

[0032] In the control parameter instruction set, a game equilibrium strategy is configured for the node;

[0033] Based on the equilibrium strategy, a multi-unit cooperative control scheme is automatically generated, including yaw system phase synchronization instructions and pitch angle difference compensation strategies.

[0034] Preferably, the calculation of the benefit-risk trade-off score comprises:

[0035] Obtaining unit bearing vibration spectrum and gear box oil temperature time series data, a mechanical health state evaluation tensor is constructed;

[0036] Through a tensor decomposition algorithm, principal component weights of multi-dimensional features are extracted;

[0037] The principal component weights and the evaluation tensor are subjected to bilinear pooling operation to obtain a comprehensive trade-off score;

[0038] The calculation formula of the comprehensive trade-off score is:

[0039]

[0040] In the formula, Γ represents the comprehensive trade-off score value, α k represents the fatigue damage coefficient of the kth mechanical component, β k represents the maintenance cost weight of the kth component, γ represents the unit basic reliability constant, and m represents the total number of mechanical component classifications.

[0041] Preferably, the embedding of the partial differential constraint condition comprises:

[0042] Perform Lie group symmetry analysis on the hidden variable inference result to screen physically compatible wind field evolution paths;

[0043] Generate characteristic evolution trajectories conforming to fluid mechanics constraints through stochastic differential equation sampling;

[0044] Regularize and train the convolution kernel of the spatiotemporal convolutional neural network using trajectory data to ensure that the model output conforms to the laws of atmospheric motion.

[0045] Preferably, the execution of the Bayesian optimization framework comprises:

[0046] Define a Gaussian process prior distribution function containing a balance factor for control parameter exploration and utilization;

[0047] Select the optimal candidate parameter combination through expected improvement acquisition functions;

[0048] In each iteration, dynamically adjust the hyperparameters of the covariance matrix according to the parameter uncertainty;

[0049] Output the optimal control action sequence that satisfies the global convergence condition.

[0050] Preferably, the method further comprises:

[0051] After configuring the game equilibrium strategy, real-time monitor the sensitivity index of the key parameter coupling node;

[0052] If the sensitivity index exceeds the preset threshold, trigger the simulated annealing optimization mechanism to re-plan the spatiotemporal phase difference scheme for multi-unit cooperation.

[0053] Preferably, the quantification of the balance factor comprises:

[0054] Establish a dynamic shrinkage model of the control parameter confidence interval and define an overfitting penalty function;

[0055] According to the wind field prediction error distribution, calculate the information entropy gain value of each dimension parameter;

[0056] Use the information entropy gain value as the adaptive adjustment coefficient of the acquisition function;

[0057] The calculation formula of the dynamic shrinkage model is:

[0058]

[0059] In the formula, η represents the equilibrium factor value, λ(τ) represents the initial confidence of the τ-dimensional parameter, ν(τ) represents the shrinkage rate coefficient, δ(τ, t) represents the parameter activation state indication function at time t, and s represents the total number of parameter dimensions.

[0060] Compared with the prior art, the beneficial effects of the present application are:

[0061] From the perspective of wind field data processing and analysis, by receiving real-time wind speed time series data and three-dimensional wind direction vector field data of the target wind field, and combining a space-time convolutional neural network to construct a wind field dynamic analysis model, multi-scale feature extraction and fusion of complex wind field data can be performed. This process not only accurately captures the change characteristics of the wind field at different time and spatial scales, generates an accurate wind field energy density distribution matrix, and provides comprehensive and reliable data support for subsequent control decisions. Compared with traditional methods, it is no longer limited to simple wind speed and wind direction measurement values, but deeply excavates the dynamic characteristics of the wind field behind the data, greatly improving the perception ability of the wind field environment.

[0062] In terms of control model construction and optimization, the multi-objective adaptive optimization model constructed based on the energy density distribution matrix can dynamically divide the wind energy capture priority area according to the actual situation of the wind field. This means that the wind turbine can more intelligently capture wind energy in the high-energy area, avoiding wasting resources in low-energy areas. At the same time, by comprehensively considering the mechanical fatigue accumulation index of the unit and the grid dispatching demand curve to calculate the benefit-risk trade-off score, and combining the Sigmoid function to generate the unit-specific control threshold, the power generation efficiency and mechanical health risk in the unit operation process are effectively balanced. This fine-grained control strategy can reduce unnecessary wear and tear of the unit, prolong the service life of the equipment, reduce maintenance costs, and at the same time meet the stability and flexibility requirements of the power grid for power supply.

[0063] For the balance problem of unit load and power generation efficiency, the application builds a unit load-power generation efficiency balance equation, simulates the dynamic game relationship of unit operation state in the future preset time window, and optimizes the synergistic regulation coefficient. This innovative method can find the best balance point of unit load and power generation efficiency under different wind conditions, avoiding the situation of sacrificing one target for pursuing a single target. At high wind speed, it can ensure safe and stable operation of the unit, control the load within a reasonable range, and maximize power generation efficiency. At low wind speed, by optimizing the control parameters, the wind energy capture capacity is improved and the power generation capacity is increased. This precise balance control effectively improves the overall economic benefit and operation reliability of the wind turbine.

[0064] In terms of multi-unit cooperative control, key parameter coupling nodes are identified according to the simulation results of the dynamic game network, and game equilibrium strategies are configured, thereby automatically generating a multi-unit cooperative control scheme, including yaw system phase synchronization instructions and pitch angle difference compensation strategies. This enables multiple units in a wind farm to cooperate with each other, reducing wake interference and improving overall wind energy capture efficiency. Compared with traditional independent control methods, multi-unit cooperative control can fully utilize wind farm resources, improve the total power generation of the wind farm, and enhance the stability and reliability of power supply.

[0065] In terms of optimization algorithm application, the synergistic regulation coefficient is iteratively updated with the help of the Bayesian optimization framework to constantly find the optimal control action sequence. The Bayesian optimization framework can effectively balance the exploration and utilization of control parameters, ensuring global search capability while quickly converging to the optimal solution. By dynamically adjusting the hyperparameters of the covariance matrix, the control optimization accuracy and efficiency are further improved through adaptive optimization based on wind farm prediction error distribution. This efficient optimization algorithm ensures that the wind turbine always maintains the best operating state in complex and variable wind field environments, maximizing the utilization of wind energy. BRIEF DESCRIPTION OF DRAWINGS

[0066] Figure 1 A working principle diagram of the wind turbine control optimization method described in the application;

[0067] Figure 2 A flowchart for multi-objective adaptive optimization model construction and control threshold generation;

[0068] Figure 3 A flowchart for generating and optimizing a multi-unit cooperative control scheme based on dynamic game results;

[0069] Figure 4 A flowchart for the execution of the Bayesian optimization framework. DETAILED DESCRIPTION

[0070] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of the present application.

[0071] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of the present application. Figures 1-4 The present application provides a wind turbine control optimization method based on dynamic changes of wind speed and direction, and the specific implementation steps are as follows:

[0072] Real-time wind speed time series data and three-dimensional wind direction vector field data of the target wind field are obtained. The three-dimensional wind direction vector field data includes laser radar wind measurement point cloud and atmospheric boundary layer turbulence spectrum. The laser radar wind measurement point cloud can accurately measure wind speed and direction information at different heights and positions, and the atmospheric boundary layer turbulence spectrum reflects the turbulence characteristics of the wind field.

[0073] A wind field dynamic analysis model is constructed using a spatio-temporal convolutional neural network. Through this model, multi-scale feature extraction and fusion are performed on real-time data, and then a wind field energy density distribution matrix is generated. Multi-scale feature extraction can capture the changing characteristics of the wind field at different time and spatial scales, and the fusion of these features can more comprehensively describe the wind field state.

[0074] According to the generated energy density distribution matrix, a multi-objective adaptive optimization model is constructed. The model generates a set of turbine control parameter instructions, including a yaw angle adjustment sequence and a blade pitch angle optimization strategy. The yaw angle adjustment sequence is used to make the wind wheel of the wind turbine better align with the wind direction, improving the efficiency of wind energy capture; the blade pitch angle optimization strategy adjusts the blade angle according to the wind conditions to control wind energy capture and turbine load.

[0075] Based on the preset turbine load-power generation efficiency balance equation, the dynamic game relationship of the turbine operating state in the future preset time window is simulated, so as to optimize the cooperative adjustment coefficient in the control parameter instruction set. The turbine load-power generation efficiency balance equation considers various factors in the turbine operation process, and by simulating the dynamic game relationship, the optimal cooperative adjustment coefficient that balances the turbine load and power generation efficiency under different wind conditions can be found.

[0076] With the help of the Bayesian optimization framework, the cooperative adjustment coefficient is iteratively updated, and finally the optimal control action sequence is output to the wind turbine main control system. The Bayesian optimization framework can use the previous iteration results to continuously adjust the search direction, efficiently find the optimal solution, and ensure that the wind turbine can operate in the optimal state under various wind conditions.

[0077] In the wind turbine control optimization method involved in the present application, the synergistic regulation coefficient is an important parameter set for balancing and coordinating the relationship between different control parameters of the wind turbine. Its specific role is reflected in multiple key links, and plays a decisive role in realizing the efficient and stable operation of the wind turbine.

[0078] In simulating the dynamic game relationship of the unit operating state in the future preset time window based on the preset unit load-power generation efficiency balance equation, the synergistic regulation coefficient is used to quantify the influence weight of different control parameters (such as yaw angle adjustment sequence and blade pitch angle optimization strategy) on unit load and power generation efficiency. For example, when the energy density distribution of the wind farm changes, the yaw angle needs to be adjusted to better capture wind energy, and the blade pitch angle needs to be adjusted to control the unit load. At this time, the synergistic regulation coefficient determines the relative importance and synergistic change degree between the yaw angle adjustment and the blade pitch angle adjustment. If the synergistic regulation coefficient is set unreasonably, it may lead to excessive pursuit of power generation efficiency, resulting in excessive unit load, affecting the service life of the unit; or too conservative control of load, sacrificing too much power generation efficiency.

[0079] In the multi-objective adaptive optimization model, the synergistic regulation coefficient also participates in the generation process of the control strategy. By dynamically dividing the wind energy capture priority area and calculating the benefit-risk trade-off score of the control strategy based on the unit mechanical fatigue accumulation index and the grid scheduling demand curve, the synergistic regulation coefficient associates and adjusts the trade-off score with the priority area. For example, in the high wind speed area, in order to avoid excessive stress on the unit, the synergistic regulation coefficient will make the control strategy more inclined to limit the load and appropriately reduce the power generation efficiency; while in the case of low wind speed and large grid demand, the synergistic regulation coefficient will adjust the control strategy, and focus more on improving the power generation efficiency to meet the grid scheduling demand.

[0080] In actual application scenarios, assuming that the wind conditions of a certain wind farm are complex and changeable, the wind speed shows an unstable rising trend in a certain time period, and the wind direction also changes frequently. At this time, the wind turbine needs to respond quickly and adjust the control parameters. The wind field dynamic analysis model processes the laser radar wind point cloud and atmospheric boundary layer turbulence spectrum data in real time to generate a wind field energy density distribution matrix. The multi-objective adaptive optimization model determines the current wind energy capture priority area and the benefit-risk trade-off score of the control strategy based on this information. And the synergistic regulation coefficient, in this process, comprehensively considers the requirements of the unit load-power generation efficiency balance equation, dynamically adjusts the synergistic relationship between the yaw angle adjustment sequence and the blade pitch angle optimization strategy. By continuously adjusting the synergistic regulation coefficient, the wind turbine can realize higher power generation efficiency while ensuring the safe operation of the unit (controlling the load within a reasonable range) under different wind conditions, and meet the grid scheduling demand.

[0081] The collaborative adjustment coefficients are iteratively updated through the Bayesian optimization framework, which can continuously optimize the control strategy of the wind turbine and output the optimal control action sequence to the main control system of the wind turbine. During the iteration process, the execution effect of the control strategy under the current collaborative adjustment coefficients is evaluated according to the actual operation feedback data of the wind turbine, such as the power generation of the unit, the load condition, the matching degree with the power grid, etc. Then, based on these evaluation results and prior knowledge, the collaborative adjustment coefficients are intelligently adjusted using the characteristics of the Bayesian optimization algorithm, and the most suitable collaborative adjustment coefficient combination for the current wind farm conditions and unit operating state is gradually found out, thereby realizing the optimal control of the wind turbine.

[0082] The implementation of the present application will be further described below in conjunction with Examples 1 to 5.

[0083] Example 1:

[0084] In constructing the wind farm dynamic analysis model, first, a historical wind farm data set under multiple meteorological conditions is collected. These meteorological conditions include different temperatures, humidities, air pressures, etc., because the change of meteorological conditions will have an impact on the wind farm. Multi-dimensional training samples containing wind speed fluctuation characteristics, wind direction shear gradient and unit response delay time are extracted from the historical wind farm data set. The wind speed fluctuation characteristics reflect the fluctuation of wind speed, the wind direction shear gradient reflects the degree of change of wind direction in space, and the unit response delay time takes into account the time required for the wind turbine to respond to the change of wind conditions.

[0085] The independent representation of the wind farm dominant mode and the random disturbance factor is extracted by performing hidden variable inference on the multi-dimensional training samples through the deep probabilistic graph model. The deep probabilistic graph model can mine the potential relationship in the data, and the wind farm dominant mode represents the main trend of the wind farm, and the random disturbance factor reflects some unpredictable interference factors.

[0086] The partial differential constraint condition of the wind farm evolution process is constructed in combination with the simplified form of the Navier-Stokes equation. The Navier-Stokes equation describes the motion law of fluid, and the simplified form reduces the computational complexity while ensuring a certain accuracy. When embedding the partial differential constraint condition, the Lie group symmetry analysis is performed on the hidden variable inference results to screen the physically consistent wind farm evolution path. The Lie group symmetry analysis can mathematically ensure that the wind farm evolution path conforms to the physical law. The characteristic evolution trajectories conforming to the fluid mechanics constraint are generated by sampling the stochastic differential equation. The convolution kernel of the spatio-temporal convolutional neural network is regularized and trained using these trajectory data to ensure that the model output conforms to the law of atmospheric motion. Assuming that in a certain wind farm, the related parameters in the simplified form of the Navier-Stokes equation are determined through multiple tests and data processing, and through the above series of operations, the wind farm dynamic analysis model can more accurately reflect the actual situation of the wind farm.

[0087] In this process, the Lie group symmetry analysis of the hidden variable inference results is involved, and the operation of screening the physically consistent wind field evolution path is involved. Although there is no specific formula to reflect this process, in actual operation, it is based on related mathematical theory and algorithm. The characteristic evolution trajectory conforming to the fluid mechanics constraint is generated by sampling the stochastic differential equation. This process can be understood by some general form of stochastic differential equation, such as the common Itô stochastic differential equation: dX t = μ(X t ,t)dt + σ(X t ,t)dW t , where X t represents the state of the random process at time t, μ(X t ,t) is the drift coefficient, which describes the average change rate of X t at time t, σ(X t ,t) is the diffusion coefficient, which reflects the volatility of X t , and dW t is the Wiener process increment, which represents random noise. In this embodiment, by adjusting the related parameters, the equation is used to generate the required characteristic evolution trajectory, which is used for subsequent regularization training of the convolution kernel of the spatio-temporal convolutional neural network.

[0088] The construction method of the deep probabilistic graph model includes:

[0089] A multi-dimensional training sample containing wind speed fluctuation characteristics, wind direction shear gradient and unit response delay time is constructed from historical wind field data sets under multiple meteorological conditions. The wind speed fluctuation characteristics reflect the fluctuation of wind speed, the wind direction shear gradient reflects the change of wind direction in space, and the unit response delay time considers the response time of the unit to the wind condition change.

[0090] When constructing the deep probabilistic graph model, the model structure includes multiple hidden layers and visible layers. The hidden layer is used to extract the potential features in the data, and the visible layer corresponds to the input multi-dimensional training sample data. The model is constructed based on the structure of the Bayesian network, and the dependence relationship between variables is represented by nodes and edges. For example, wind speed fluctuation characteristics, wind direction shear gradient and unit response delay time are nodes, and their mutual influence relationship is connected by edges.

[0091] In the model training stage, the maximum a posteriori estimation (MAP) method is adopted. Assuming that the multi-dimensional training sample set is D, the model parameter is θ, the prior distribution is P(θ), and the likelihood function is P(D|θ). According to the maximum a posteriori estimation, the goal is to maximize the posterior probability P(θ|D), according to the Bayes formula Since P(D) is constant during the training process, it is equivalent to maximizing P(D|0)P(0). In actual calculation, the logarithm of P(D|0)P(0) is taken to obtain lnP(D|0)+lnP(0), and the model parameter 0 is adjusted by optimizing the objective function.

[0092] Through the above construction and training process, the deep probabilistic graph model can perform hidden variable inference on the multi-dimensional training sample, and extract independent representations of the dominant mode of the wind field and the random disturbance factor.

[0093] Embodiment 2:

[0094] In constructing the multi-objective adaptive optimization model, first, the wind energy capture priority area is dynamically divided according to the spatio-temporal variation rate of the energy density distribution matrix. The energy distribution in the wind field is not fixed and will change in time and space dimensions. By analyzing the energy density distribution matrix data at different times and different spatial positions, the spatio-temporal variation rate is calculated. For example, by comparing the energy density values of adjacent time points and adjacent spatial regions, if the energy density of a certain region increases significantly in a short time and has obvious advantages compared with the surrounding region, then this region will be designated as a high-priority capture region; otherwise, if the energy density changes slowly or decreases, the priority is lower. The purpose of such division is to more efficiently utilize wind energy and preferentially capture energy in areas with rich and favorable changes in wind energy.

[0095] The benefit-risk trade-off score of the control strategy is calculated based on the mechanical fatigue accumulation index of the unit and the grid dispatch demand curve. First, the bearing vibration spectrum and gear box oil temperature time series data of the unit are obtained to construct a mechanical health state evaluation tensor. The bearing vibration spectrum can reflect the operating condition of the bearing. If there are abnormal frequency components in the vibration spectrum, it usually means that the bearing has hidden faults. The gear box oil temperature time series data reflects the working heat of the gear box. High oil temperature may cause problems such as increased wear of the gear box and decreased lubrication performance. Organize these data according to a specific tensor structure to form a mechanical health state evaluation tensor, which can comprehensively and comprehensively present the health status of the mechanical components of the unit.

[0096] The evaluation tensor is processed using a tensor decomposition algorithm to extract the principal component weights of the multi-dimensional features. The tensor decomposition algorithm can decompose a high-dimensional tensor into a combination of multiple low-dimensional tensors. Through this decomposition, key features hidden in the data can be mined, and the importance of each feature, i.e., the principal component weight, can be determined. For example, during the decomposition process, some features have larger weights, indicating that these features have a more significant impact on the mechanical health state of the unit, which needs to be considered in the subsequent trade-off score calculation.

[0097] The principal component weight is bilinearly pooled with the evaluation tensor to obtain a comprehensive trade-off score. The formula for calculating the comprehensive trade-off score is:

[0098]

[0099] In this formula, Γ represents the comprehensive trade-off score value, which is a key indicator for measuring the pros and cons of the control strategy. α k represents the fatigue damage coefficient of the kth mechanical component. Different types of mechanical components, such as blades and generator rotors, have different fatigue damage coefficients due to their different working principles and stress conditions. The larger the coefficient, the more likely it is that the corresponding component will experience fatigue damage under the current operating conditions, and the greater the potential threat to the overall performance of the unit. β k is the maintenance cost weight of the kth component. Components with high maintenance costs have relatively large weights. For example, the cost of replacing a generator is much higher than that of replacing a normal sensor, so the maintenance cost weight of the generator will be larger, and its impact on the result will be more prominent in the trade-off score. γ represents the unit base reliability constant, which is a fixed value determined based on factors such as the design, manufacturing process, and installation environment of the unit, and is used to reflect the reliability level of the unit base itself, providing a stable reference benchmark for the trade-off score. m represents the total number of mechanical component categories, covering all mechanical component categories that need to be considered in the wind turbine. Through comprehensive consideration of various components, the trade-off score can fully reflect the actual situation of the unit.

[0100] The trade-off score and priority area are nonlinearly weighted by the Sigmoid function to generate a unit-specific control threshold. The expression of the Sigmoid function is where x is the input value. This function has the property of mapping any real number to the interval (0, 1). In this embodiment, the trade-off score and priority area are inputted, and after processing by the Sigmoid function, they are reasonably integrated to obtain a control threshold between 0 and 1. This threshold is generated based on the characteristics of the unit itself and the current wind farm conditions, and has strong pertinence.

[0101] The switching conditions for multi-level regulation modes are triggered based on thresholds. When the calculated control threshold exceeds a certain preset value, it indicates that the current wind energy capture and turbine operation status need adjustment. At this time, the wind turbine is triggered to switch from one regulation mode to another. For example, when the threshold exceeds the preset high threshold, the turbine switches from low-power regulation mode to high-power regulation mode, increasing the blade pitch angle adjustment range and improving wind energy capture efficiency; when the threshold is lower than the preset low threshold, the turbine switches from high-power regulation mode to low-power regulation mode, reducing the blade pitch angle to reduce the turbine load and protect equipment safety. Through this threshold-based multi-level regulation mode switching mechanism, the wind turbine can maintain a relatively ideal operating state under different wind conditions, achieving efficient wind energy utilization and stable turbine operation.

[0102] Example 3:

[0103] When constructing the unit load-power generation efficiency balance equation, the first step is to collect unit operating state parameters and structural stress monitoring data to build a load-power mapping relationship dataset. Unit operating state parameters include wind speed, wind direction, and rotational speed, while structural stress monitoring data reflects the stress conditions of various unit components during operation. Let the wind speed be v, the wind direction be θ, the rotational speed be n, and the structural stress be σ. Multiple sets of data (v...) are collected... i ,θ i ,n i ,σ i ), i = 1, 2, ..., N, construct the dataset.

[0104] A fuzzy logic controller generates membership functions for blade aerodynamic loads and a fuzzy rule base for power generation efficiency. The fuzzy logic controller can fuzzify the blade aerodynamic loads and power generation efficiency based on input wind conditions and unit operating parameters. For example, for wind speed *v*, fuzzy sets such as "low wind speed," "medium wind speed," and "high wind speed" can be defined, and a corresponding membership function can be determined for each fuzzy set. Similarly, fuzzy sets and membership functions can be defined for power generation efficiency. Based on these fuzzy sets and membership functions, a fuzzy rule base for power generation efficiency is established, such as rules like "if the wind speed is low and the blade pitch angle is small, then the power generation efficiency is low."

[0105] A multi-objective genetic algorithm is used to construct a Pareto front search space and quantify the conflict strength between control parameters. The multi-objective genetic algorithm is an optimization algorithm that can find the optimal balance solution among multiple objectives. In this embodiment, the control parameters include yaw angle, blade pitch angle, etc., and the objectives are to maximize power generation efficiency and minimize unit load. Through the multi-objective genetic algorithm, a set of optimal control parameter combinations can be found, which form the Pareto front. In this process, the conflict strength between control parameters is quantified, for example, the adjustment of the yaw angle may affect the aerodynamic load on the blades, thereby affecting the power generation efficiency and unit load, and the degree of this influence is calculated to quantify the conflict strength.

[0106] The conflict strength and real-time wind field disturbance factor are input into a dynamic game network to generate a unit load-power generation efficiency balance equation. The real-time wind field disturbance factor reflects the uncertainty of the wind field, and the dynamic game network can simulate the interaction and game relationship between control parameters under different wind conditions. Through this network, an equation is generated that can balance the unit load and power generation efficiency, providing a scientific basis for unit control.

[0107] Embodiment 4:

[0108] After simulating the unit operation state based on the unit load-power generation efficiency balance equation, key parameter coupling nodes are identified according to the simulation results of the dynamic game network. Key parameter coupling nodes are those that have a greater impact on the unit operation state and are closely related to each other. For example, the blade pitch angle and the yaw angle may be a pair of key parameter coupling nodes, and their changes will simultaneously affect the load and power generation efficiency of the unit.

[0109] Game equilibrium strategies are configured for nodes in the control parameter instruction set. Game equilibrium strategies refer to finding an optimal parameter configuration scheme that allows the unit to achieve the best operating state under certain conditions while considering the mutual influence of each parameter. Based on this equilibrium strategy, a multi-unit cooperative control scheme is automatically generated, including a yaw system phase synchronization instruction and a pitch angle difference compensation strategy. The yaw system phase synchronization instruction can make the yaw systems of multiple wind turbines remain synchronized during operation, improving wind energy capture efficiency; the pitch angle difference compensation strategy adjusts the blade pitch angle according to the location and wind condition differences of different units to ensure efficient operation of each unit.

[0110] After configuring the game equilibrium strategy, the sensitivity index of the key parameter coupling node is monitored in real time. The sensitivity index is used to measure the sensitivity of the key parameter coupling node to external interference or parameter changes. If the sensitivity index exceeds the preset threshold, the simulated annealing optimization mechanism is triggered to re-plan the multi-unit collaborative time-space phase difference scheme. Simulated annealing optimization mechanism is a heuristic optimization algorithm that searches for the optimal solution within a certain range by simulating the physical annealing process. For example, when the sensitivity index of the key parameter coupling node of a certain wind turbine is too high, it means that the unit is too sensitive to wind changes, which may cause unstable operation. At this time, the simulated annealing optimization mechanism is started to adjust the multi-unit collaborative time-space phase difference scheme to find a better operation scheme to reduce the sensitivity index and improve the stability and efficiency of the unit. In the simulated annealing optimization mechanism, some temperature parameters and probability calculation formulas are used. For example, the probability formula for accepting new solutions is: where P represents the probability of accepting new solutions, ΔE represents the energy difference between the new solution and the current solution, and T represents the current temperature. As the temperature T gradually decreases, the probability of accepting worse solutions gradually decreases, and eventually converges to the vicinity of the global optimal solution.

[0111] Example 5:

[0112] When using the Bayesian optimization framework, first define the Gaussian process prior distribution function, which includes a balance factor that controls the balance between parameter exploration and utilization. The Gaussian process prior distribution function can describe the uncertainty and potential distribution of the control parameters. The balance factor is used to balance the exploration of new parameters and the utilization of existing good parameters in the search process.

[0113] A dynamic shrinkage model is established to define the confidence interval of the control parameters, and a overfitting penalty function is defined. The calculation formula of the dynamic shrinkage model is:

[0114]

[0115] In this formula, η represents the balance factor value, which is a dynamic value used to adjust the balance between control parameter exploration and utilization. λ(τ) represents the initial confidence of the τth dimension parameter, which reflects the degree of trust in the parameter at the initial stage, and the value generally ranges from 0 to 1, the larger the value, the higher the initial confidence. ν(τ) represents the shrinkage rate coefficient, which controls the shrinkage speed of the confidence interval, the larger the coefficient, the faster the confidence interval shrinks. δ(τ,t) represents the parameter activation state indicator function at time t, which takes the value of 1 when the parameter is activated at time t, otherwise it is 0. s represents the total number of parameter dimensions. Through this formula, the balance factor can be dynamically adjusted according to different parameter dimensions and time.

[0116] According to the wind field prediction error distribution, the information entropy gain value of each dimension parameter is calculated. The information entropy gain value can measure the amount of information contained in the parameter. By calculating the information entropy gain value, it can be determined which parameters are more important to the optimization result. The information entropy gain value is used as the adaptive adjustment coefficient of the acquisition function. The expected improvement acquisition function is used to select the optimal candidate parameter combination. In each iteration, the hyperparameters of the covariance matrix are dynamically adjusted according to the parameter uncertainty. As the number of iterations increases, the uncertainty range of the parameters is gradually reduced, and the accuracy of the search is improved. Finally, the optimal control action sequence that meets the global convergence condition is output. For example, after multiple iterations, when the variation range of the parameters is less than a certain preset value, and the optimization result no longer improves significantly, it is considered that the global convergence condition is met, and the output control action sequence is the optimal sequence, which can make the wind turbine reach the best operating state under the current wind field conditions. In the calculation process of the expected improvement acquisition function, some probability and statistical related formulas may be used, such as the formula for calculating the expectation: E(X) = ∑ i x i p i , where E(X) represents the expectation of the random variable X, x i represents the value of X, and p i represents the probability corresponding to the value x i . The expected improvement value is calculated by these formulas, the optimal candidate parameter combination is selected, and the optimization of the cooperative adjustment coefficient is realized.

[0117] It should be noted that in this paper, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that there is any such actual relationship or order between these entities or operations. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or equipment.

[0118] Although embodiments of the present application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A wind turbine control optimization method based on dynamic changes in wind speed and direction, characterized by, The method comprises the following steps: Receiving real-time wind speed time series data and three-dimensional wind direction vector field data of a target wind field, wherein the three-dimensional wind direction vector field data comprises laser radar wind measurement point cloud and atmospheric boundary layer turbulence spectrum; Constructing a wind field dynamic analysis model based on a spatio-temporal convolutional neural network, performing multi-scale feature extraction and fusion on real-time data, and generating a wind field energy density distribution matrix; According to the energy density distribution matrix, a multi-objective adaptive optimization model is constructed to generate a set of unit control parameter instructions, including yaw angle adjustment sequence and blade pitch angle optimization strategy; Based on the preset load-power balance equation of the unit, the dynamic game relationship of the unit operation state in the future preset time window is simulated, and the cooperative adjustment coefficient in the control parameter instruction set is optimized; Through the Bayesian optimization framework, the cooperative adjustment coefficient is iteratively updated, and the optimal control action sequence is output to the wind turbine main control system; The construction steps of the wind field dynamic analysis model include: collecting historical wind field data sets under multiple meteorological conditions, and constructing multi-dimensional training samples containing wind speed fluctuation characteristics, wind direction shear gradient and unit response delay time; Through the pre-constructed deep probability graph model, the multi-dimensional training samples are inferred for independent representation of the dominant mode of the wind field and the random disturbance factor; Combined with the simplified form of Navier-Stokes equation, the partial differential constraint condition of the wind field evolution process is constructed; The partial differential constraint condition is embedded into the residual module of the spatio-temporal convolutional neural network to generate the wind field dynamic analysis model supporting online updating; The multi-objective adaptive optimization model includes: dynamically dividing the wind energy capture priority area according to the spatio-temporal variation rate of the energy density distribution matrix; Based on the mechanical fatigue accumulation index of the unit and the grid dispatching demand curve, the benefit-risk trade-off score of the control strategy is calculated; The trade-off score and priority area are nonlinearly weighted by a Sigmoid function to generate a unit-specific control threshold; According to the threshold, the switching condition of the multi-level regulation mode is triggered.

2. The wind turbine control optimization method of claim 1, wherein, The construction steps of the load-power balance equation of the unit include: Collecting unit operation state parameters and structural stress monitoring data to construct a load-power mapping relationship data set; Through a fuzzy logic controller, a membership function of blade aerodynamic load and a fuzzy rule base of power generation efficiency are generated; Combined with a multi-objective genetic algorithm, a Pareto front search space is constructed to quantify the conflict intensity between control parameters; The conflict intensity and real-time wind field disturbance factor are input into a dynamic game network to generate the load-power balance equation of the unit.

3. The wind turbine control optimization method of claim 2, wherein, Further comprising: According to the simulation results of the dynamic game network, identify the key parameter coupling nodes; In the control parameter instruction set, configure the game equilibrium strategy for the nodes; Based on the equilibrium strategy, automatically generate a multi-unit cooperative control scheme, including yaw system phase synchronization instructions and pitch angle difference compensation strategies.

4. The wind turbine control optimization method of claim 1, wherein, The calculation of the benefit-risk trade-off score includes: Obtaining unit bearing vibration spectrum and gear box oil temperature time series data to construct a mechanical health state evaluation tensor; Through a tensor decomposition algorithm, the principal component weight of the multi-dimensional feature is extracted; The principal component weight is bilinearly pooled with the evaluation tensor to obtain a comprehensive trade-off score; The calculation formula of the comprehensive trade-off score is: ； wherein, represents a combined trade-off score value, represents the fatigue damage coefficient of a mechanical component, represents the maintenance cost weight of a component, represents a unit base reliability constant, represents the total number of mechanical component classifications.

5. The wind turbine control optimization method of claim 1, wherein, The embedding of the partial differential constraint condition includes: Perform Lie group symmetry analysis on the hidden variable inference result to screen physically consistent wind field evolution paths; Generate feature evolution trajectories that meet fluid mechanics constraints through stochastic differential equation sampling; Regularize the convolution kernel of the spatio-temporal convolutional neural network using trajectory data to ensure that the model output meets the atmospheric motion law.

6. The wind turbine control optimization method of claim 1, wherein, The execution of the Bayesian optimization framework includes: Define a Gaussian process prior distribution function that includes a balance factor that controls the exploration and utilization of parameters; Select the optimal candidate parameter combination through expected improvement acquisition functions; In each iteration, dynamically adjust the hyperparameters of the covariance matrix according to the parameter uncertainty; Output the optimal control action sequence that meets the global convergence condition.

7. The wind turbine control optimization method of claim 3, wherein, Also includes: After configuring the game equilibrium strategy, monitor the sensitivity index of the key parameter coupling node in real time; If the sensitivity index exceeds the preset threshold, trigger the simulated annealing optimization mechanism to re-plan the time-space phase difference scheme for multi-unit cooperation.

8. The wind turbine control optimization method of claim 6, wherein, The quantification of the balance factor includes: Establish a dynamic shrinkage model of the control parameter confidence interval and define an overfitting penalty function; According to the wind field prediction error distribution, calculate the information entropy gain value of each dimension parameter; Use the information entropy gain value as the adaptive adjustment coefficient of the acquisition function; The calculation formula of the dynamic shrinkage model is: ； wherein denotes the balancing factor value, denotes the initial confidence of the parameter, denotes the shrinkage rate coefficient, denotes the parameter activation status indicator function at time instant, denotes the total number of parameter dimensions.

Images