A neural network-based fan mechanism modeling and optimization method and system

Abstract

The application discloses a fan mechanism modeling and optimization method and system based on a neural network, and the method comprises the following steps: a model of a wind energy utilization coefficient and a thrust coefficient of a fan is established, and steady-state simulation is performed to obtain a two-dimensional relationship table model of the wind energy utilization coefficient and the thrust coefficient, and a fan aerodynamics model is established based on the two-dimensional relationship table model; a fan transmission chain system model and a fan blade and tower coupling model are respectively established based on nonlinear aerodynamic torque T r and nonlinear air thrust F t in the fan aerodynamics model, so that a complete fan mechanism model is obtained; actual operation data of the fan are collected, deviations of the fan mechanism model and the actual operation data of the fan are analyzed based on a neural network, and correction is performed through the neural network, so that an optimized fan mechanism model is obtained. The application can improve the precision of fan modeling.

Description

Technical Field

[0001] This invention relates to the field of wind turbine modeling technology, specifically to a method and system for wind turbine mechanism modeling and optimization based on neural networks. Background Technology

[0002] Because wind turbine operation involves a complex nonlinear model, deviations in turbine parameters and neglect of some turbine flexibility characteristics occur during the modeling process, leading to discrepancies between the modeled wind turbine mechanism and the actual turbine model. In practical applications, the wind energy utilization coefficient and thrust coefficient lack functional expressions and are typically obtained in real-time by looking up tables (relationship tables with pitch angle and tip speed ratio). However, commonly used fitting methods (polynomial fitting, trigonometric function fitting, etc.) struggle to achieve multivariate fitting of the wind energy utilization coefficient and thrust coefficient, resulting in a less accurate final wind turbine mechanism model. Furthermore, the inaccuracies in the wind turbine mechanism modeling process pose a significant challenge to designing high-precision control strategies based on the wind turbine model.

[0003] In the prior art, Chinese Patent Publication No. CN119982370A establishes a nonlinear dynamic model of the wind turbine, predicts and controls the wind turbine using this nonlinear model, and designs a state observer to estimate unmeasurable states. However, the model used in this patent document is relatively general, and the process of optimizing the modeled nonlinear dynamic model is not described. Those skilled in the art have found in practical applications that the nonlinear dynamic model disclosed in this patent document suffers from inaccurate parameter measurements, structural simplification assumptions, and insufficient modeling of complex aerodynamic couplings, making it difficult to directly use for accurate wind turbine modeling and prediction. Therefore, there is an urgent need for a modeling method that can improve the accuracy of wind turbine modeling and address the shortcomings of insufficient accuracy in existing modeling techniques. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method and system for wind turbine mechanism modeling and optimization based on neural networks, in order to improve the accuracy of wind turbine modeling.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0006] A method for modeling and optimizing wind turbine mechanisms based on neural networks, comprising the following steps:

[0007] S1. Establish a model of the wind energy utilization coefficient and thrust coefficient of the wind turbine and perform steady-state simulation to obtain a two-dimensional relationship table model of the wind energy utilization coefficient and thrust coefficient, and establish a wind turbine aerodynamic model based on the two-dimensional relationship table model.

[0008] S2, based on the nonlinear aerodynamic torque in the aforementioned wind turbine aerodynamic model and nonlinear air thrust A complete wind turbine mechanism model was obtained by establishing separate models of the wind turbine drive chain system and the wind turbine blade and tower coupling model.

[0009] S3: Collect actual operating data of the wind turbine, analyze the deviation between the wind turbine mechanism model and the actual operating data of the wind turbine based on the neural network, and correct it through the neural network to obtain the optimized wind turbine mechanism model;

[0010] Step S3 includes:

[0011] A neural network is introduced to correct the errors in the unrepresented and inaccurate parts of the wind turbine mechanism model, resulting in an optimized model.

[0012] The optimized model is trained by backpropagation based on a truncated time window, and the optimized wind turbine mechanism model is obtained when the training is completed.

[0013] Furthermore, the expression for the aerodynamic model of the wind turbine is as follows:

[0014] ;

[0015] In the above formula, air density, The radius of the fan rotor. Relative wind speed, The wind energy utilization coefficient, For thrust coefficient, For the fan rotor speed, The pitch angle is the propeller angle. For the tip speed ratio, It is a nonlinear aerodynamic torque. It is a nonlinear air thrust.

[0016] Furthermore, the expression for the wind turbine drivetrain system model is as follows:

[0017] ;

[0018] In the above formula, The torsion angle of the drive shaft. for The first derivative, For the fan rotor speed, for The first derivative, For generator speed, for The first derivative, This refers to the gearbox transmission ratio. The equivalent rotational inertia on the rotor side. The equivalent moment of inertia on the generator side. This is the equivalent torsional stiffness coefficient of the transmission chain system. This is the equivalent damping coefficient of the transmission chain system. This represents the generator torque.

[0019] Furthermore, the expression for the coupled model of the wind turbine blades and tower is as follows:

[0020]

[0021] ;

[0022] In the above formula, Total blade mass For the equivalent mass of the tower, This is the equivalent damping coefficient of the tower. This is the equivalent stiffness coefficient of the tower. This is the distance from the center of gravity of the blade to the center of gravity at the top of the tower. This is the distance from the center of thrust to the center of gravity at the top of the tower. For the angle of the blade's waving, The angular velocity of the blade flapping. The acceleration of the blade flapping angle, For the forward and backward displacement of the tower, The forward and backward speed of the tower. The acceleration of the tower's forward and backward motion. This is the blade damping coefficient. This is the blade stiffness coefficient. This is the total mass of the blades. For the tower mass, For cabin quality, For the quality of the wheel hub, Mass of a single blade The structural damping ratio, The frequency of the tower's forward and backward movement.

[0023] Further, in step S2, a state-space model of the wind turbine is established by combining the wind turbine aerodynamic model, the wind turbine transmission chain system model, and the wind turbine blade and tower coupling model. The expression of the state-space model is as follows:

[0024]

[0025] , , , , , , , ,

[0026]

[0027] ;

[0028] In the above formula, x represents the various state variables of the fan. For model input, The torsion angle of the drive shaft. For the fan rotor speed, For generator speed, For the forward and backward displacement of the tower, The forward and backward speed of the tower. For the angle of the blade's waving, The angular velocity of the blade flapping. For generator torque, This refers to the gearbox transmission ratio. This is the equivalent torsional stiffness coefficient of the transmission chain system. The equivalent rotational inertia on the rotor side. This is the equivalent damping coefficient of the transmission chain system. The equivalent moment of inertia on the generator side. This is the equivalent stiffness coefficient of the tower. For the equivalent mass of the tower, This is the equivalent damping coefficient of the tower. This is the blade stiffness coefficient. This is the distance from the center of gravity of the blade to the center of gravity at the top of the tower. This is the blade damping coefficient. Total blade mass This is the distance from the center of thrust to the center of gravity at the top of the tower;

[0029] Fourth-order Runge-Kutta discretization of the state-space model yields:

[0030]

[0031]

[0032] ;

[0033] In the above formula, The discretization control cycle for the wind turbine model. Represents the state variables of the wind turbine at time k+1. Let represent the system matrix and input matrix of the discretized wind turbine dynamics mechanism model, respectively, and I represent the identity matrix.

[0034] Furthermore, in step S3, a neural network is introduced. After correcting for errors in the unrepresented and inaccurate parts of the wind turbine mechanism model, the expression for the optimized model is obtained as follows:

[0035]

[0036] ;

[0037] In the above formula, This represents a wind turbine mechanism model. This represents a neural network correction model. For matrix , Parameters to be optimized The parameters to be optimized in the neural network This represents the actual operating status of the wind turbine. For wind speed, The pitch angle is the propeller angle. This represents the generator torque.

[0038] Furthermore, the backpropagation training of the optimized model based on the truncated time window specifically includes the following steps:

[0039] T1: Obtain the timing data of the actual operation of the wind turbine to obtain the state sequence. and input sequence :

[0040] ;

[0041] In the above formula, For the amount of data in each state, For each input data volume;

[0042] T2, at the Step, extract Substitute the data from the time window into the optimization model:

[0043] ;

[0044] In the above formula, k+1, k+2, ..., k+τ represent the data corresponding to time points k+1, k+2, ..., k+τ, respectively. For hyperparameters;

[0045] T3, calculate the prediction error according to the following formula. :

[0046] ;

[0047] T4 performs backpropagation and parameter optimization for each time window to achieve truncated training. The overall loss function is constructed by jointly weighting temporal and feature data, as shown in the following expression:

[0048]

[0049]

[0050]

[0051]

[0052] ;

[0053] In the above formula, Representing feature dimension, As time weight, These are parameters used to control the initial weights and growth rate. For feature weights, for The There are two elements, where T is the total duration of data collection for the actual operation of the wind turbine, and t is the time point. The regularization coefficient is . For the joint weighted loss function, For regular expression constraints, This is the total loss function.

[0054] Furthermore, the neural network is a three-layer neural network, including an input layer, an intermediate layer, and an output layer, wherein the intermediate layer uses the sigmoid activation function, the output layer uses the linear activation function, and the input and output of the hidden layer are:

[0055] ;

[0056] In the above formula, As input to the neural network, for The Middle One element, For the input of the hidden layer, This is the output of the i-th node in the hidden layer. Input the number of nodes. This represents the number of hidden layer nodes. The weights of the hidden layers, For the bias of the hidden layer;

[0057] The input and output of the output layer are:

[0058] ;

[0059] In the above formula, For the input of the output layer, For the weights of the output layer, For the bias of the output layer, This is the output of the l-th node in the output layer. This represents the number of nodes in the output layer.

[0060] A neural network-based wind turbine mechanism modeling and optimization system includes an interconnected microprocessor and a memory, wherein the microprocessor is programmed or configured to execute the steps of a neural network-based wind turbine mechanism modeling and optimization method.

[0061] A computer-readable storage medium storing a computer program / instructions programmed or configured to execute, via a processor, the steps of a neural network-based wind turbine mechanism modeling and optimization method.

[0062] Compared with the prior art, the advantages of the present invention are as follows:

[0063] This invention first establishes a complete wind turbine mechanism model, then collects a large amount of wind turbine operation data, and uses a neural network to analyze and correct the deviation between the wind turbine mechanism model and the actual wind turbine operation data. This can compensate for the model's insufficient accuracy under complex nonlinearity and environmental disturbances, correct nonlinear errors that are difficult to analyze in the model, and improve the accuracy of wind turbine modeling. The neural network has self-learning and nonlinear approximation capabilities, making the model highly consistent with the measured results, significantly improving the prediction accuracy and engineering applicability of the wind turbine mechanism model. Attached Figure Description

[0064] Figure 1 This is a flowchart of a wind turbine mechanism modeling and optimization method based on neural networks, according to an embodiment of the present invention. Detailed Implementation

[0065] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0066] like Figure 1 As shown, the wind turbine mechanism modeling and optimization method based on neural networks in this embodiment includes the following steps:

[0067] S1. Establish models of the wind turbine's wind energy utilization coefficient and thrust coefficient, and perform steady-state simulations to obtain a two-dimensional relationship table model of the wind energy utilization coefficient and thrust coefficient (for example, the wind energy utilization coefficient can be obtained experimentally). and thrust coefficient A two-dimensional data table), and an aerodynamic model of the wind turbine is established based on the two-dimensional relationship table model;

[0068] S2, based on the nonlinear aerodynamic torque in the aforementioned wind turbine aerodynamic model and nonlinear air thrust A complete wind turbine mechanism model was obtained by establishing separate models of the wind turbine drive chain system and the wind turbine blade and tower coupling model.

[0069] S3: Collect actual operating data of the wind turbine, analyze the deviation between the wind turbine mechanism model and the actual operating data of the wind turbine based on the neural network, and correct it through the neural network to obtain the optimized wind turbine mechanism model;

[0070] Step S3 includes:

[0071] A neural network is introduced to correct the errors in the unrepresented and inaccurate parts of the wind turbine mechanism model, resulting in an optimized model.

[0072] The optimized model is trained by backpropagation based on a truncated time window, and the optimized wind turbine mechanism model is obtained when the training is completed.

[0073] As can be understood, this embodiment first establishes a complete wind turbine mechanism model, then collects a large amount of wind turbine operation data, and uses a neural network to analyze and correct the deviation between the wind turbine mechanism model and the actual wind turbine operation data. This can compensate for the model's insufficient accuracy under complex nonlinearity and environmental disturbances, correct nonlinear errors in the model that are difficult to analyze, and improve the accuracy of wind turbine modeling. The neural network has self-learning and nonlinear approximation capabilities, making the model highly consistent with the measured results, significantly improving the prediction accuracy and engineering applicability of the wind turbine mechanism model.

[0074] In this embodiment, the expression for the wind turbine aerodynamic model is:

[0075] ;

[0076] In the above formula, Air density (unit: kg / m3). The radius of the fan rotor (unit: m). Relative wind speed (unit: m / s) The displacement of the tower section is measured in meters (m). The forward and backward velocity of the tower (unit: m / s). The wind energy utilization coefficient, For thrust coefficient, The speed of the fan rotor (unit: rad / s). The pitch angle is expressed in deg. For the tip speed ratio, It is a nonlinear aerodynamic torque. It is a nonlinear air thrust.

[0077] The wind turbine drivetrain system consists of a low-speed shaft, a gearbox, and a high-speed shaft. Considering the flexibility of the torsion spring on the low-speed shaft, the high-speed shaft is modeled as a rigid body. In this embodiment, the expression for the wind turbine drivetrain system model is:

[0078] ;

[0079] In the above formula, The torsion angle of the drive shaft (unit: rad). for The first derivative, For the fan rotor speed, for The first derivative, Generator speed (unit: rad / s). for The first derivative, This refers to the gearbox transmission ratio. The equivalent moment of inertia on the rotor side (including rotor, hub, and main shaft, unit: kg·m) 2 ), Equivalent moment of inertia on the generator side (including gearbox, high-speed shaft and generator, unit: kg·m) 2 ), The equivalent torsional stiffness coefficient of the transmission chain system (unit: Nm / rad). The equivalent damping coefficient of the transmission chain system (unit: Nms / rad). Generator torque (unit: Nm).

[0080] In actual wind turbine operation, the blades and tower interact and influence each other. Therefore, analyzing the blades and tower as a unified system during modeling can more accurately reflect the dynamic behavior of the wind turbine. In this embodiment, the expression for the coupled model of the wind turbine blades and tower is as follows:

[0081] ;

[0082] In the above formula, Total blade mass (unit: kg). Equivalent mass of the tower (unit: kg). The equivalent damping coefficient of the tower (unit: Nms / rad). The equivalent stiffness coefficient of the tower (unit: Nm / rad). This is the distance from the center of gravity of the blade to the center of gravity at the top of the tower (unit: m). This is the distance from the center of thrust to the center of gravity at the top of the tower (unit: m). The blade flapping angle (unit: rad). The angular velocity of the blade flapping. The acceleration of the blade flapping angle, For the forward and backward displacement of the tower, The forward and backward speed of the tower. The acceleration of the tower's forward and backward motion. The blade damping coefficient (unit: Nms / rad). This is the blade stiffness coefficient (unit: Nm / rad).

[0083] in and The empirical calculation formula is:

[0084] ;

[0085] In the above formula, Total mass of the blades (unit: kg). The tower mass is expressed in kg. Cabin mass (unit: kg) Wheel hub mass (unit: kg) Mass of a single leaf (unit: kg). The structural damping ratio, The frequency of the tower's forward and backward movement (unit: Hz).

[0086] In this embodiment, in step S2, a state-space model of the wind turbine is established by combining the wind turbine aerodynamic model, the wind turbine transmission chain system model, and the wind turbine blade and tower coupling model, so that... , , , , , , , , The expression for the state-space model is as follows:

[0087]

[0088]

[0089] ;

[0090] Applying a high-precision fourth-order Runge-Kutta discretization method to the above equation, we obtain:

[0091]

[0092]

[0093] ;

[0094] In the above formula, The discretization control cycle for the wind turbine model. Represents the state variables of the wind turbine at time k+1. Let represent the system matrix and input matrix of the discretized wind turbine dynamics mechanism model, respectively, and I represent the identity matrix.

[0095] In this embodiment, a neural network is introduced in step S3. After correcting for errors in the unrepresented and inaccurate parts of the wind turbine mechanism model, the expression for the optimized model is obtained as follows:

[0096]

[0097] ;

[0098] In the above formula, This represents a wind turbine mechanism model. This represents a neural network correction model. For matrix , Parameters to be optimized The parameters to be optimized in the neural network This represents the actual operating status of the wind turbine. For wind speed, The pitch angle is the propeller angle. The generator torque; where the neural network It consists of three linear layers, with the activation function between the linear layers being the Sigmoid function.

[0099] In a specific application embodiment, step S3 is a combined method of mechanism system identification and neural network modeling, which is used to balance physical interpretability and nonlinear modeling capability in wind turbine state prediction. Its core includes a mechanism modeling part and a data-driven compensation part.

[0100] The mechanism section is based on the defined state-space model equations of the wind turbine. Considering the possibility of inaccuracies in mechanism modeling, such as inaccurate measurements of important parameters like component moment of inertia, radius, and coefficients, a matrix is ​​used. It will also participate in the optimization. The optimization process does not involve adjusting the matrix. Instead of optimizing all elements, this approach combines mechanistic knowledge to fix elements that are necessarily zero or known constants, setting only elements that may physically have non-zero values ​​as parameters to be optimized. This "partially sparse optimization" method can significantly reduce the degrees of freedom, avoid the overfitting risk caused by full parameter free fitting, and maintain the physical rationality of the mechanistic modeling.

[0101] Data-driven component: Considering the wind field disturbances, aerodynamic nonlinearities, and higher-order effects not modeled in the actual operation, this embodiment introduces a neural network to compensate for residuals based on the mechanistic equations, resulting in the following overall optimized model:

[0102] ;

[0103] in For matrix , Parameters to be optimized These are the parameters to be optimized in the neural network. The specific optimization form is as follows:

[0104] .

[0105] In this embodiment, the backpropagation training of the optimized model based on the truncated time window specifically includes the following steps:

[0106] T1: Obtain the timing data of the actual operation of the wind turbine to obtain the state sequence. and input sequence :

[0107] ;

[0108] In the above formula, For the amount of data in each state, For each input data volume;

[0109] T2, at the Step, extract Substitute the data from the time window into the optimization model:

[0110] ;

[0111] In the above formula, k+1, k+2, ..., k+τ represent the data corresponding to time points k+1, k+2, ..., k+τ, respectively. These are hyperparameters (selected flexibly according to downstream tasks);

[0112] T3, calculate the prediction error according to the following formula. :

[0113] ;

[0114] T4 performs backpropagation and parameter optimization for each time window to achieve truncated training. The overall loss function is constructed by jointly weighting temporal and feature data, as shown in the following expression:

[0115]

[0116]

[0117] (Time weighting: Considering that in the long-term prediction of wind turbine operating status, the error accumulation effect is more significant the later the time step, the time weight is set to a linearly increasing form.)

[0118]

[0119] ;

[0120] In the above formula, Representing feature dimension, As time weight, These are parameters used to control the initial weights and growth rate, thereby emphasizing the model's accuracy requirements in long-term prediction. Characteristic weights are used to measure the different dimensions and focuses of wind turbine monitoring data, such as rotor angular velocity and tower top sway velocity, which are particularly crucial for load estimation and safety control. Different variables are subjected to fuzzy normalization, while key features are given higher weights. for The There are two elements, where T is the total duration of data collection for the actual operation of the wind turbine, and t is the time point. The regularization coefficient is . For the joint weighted loss function, For regular expression constraints, This is the total loss function.

[0121] It is understandable that, unlike conventional full-sequence backpropagation, this embodiment avoids the gradient explosion and vanishing problems in long-sequence training by truncating the time window, significantly improving the training stability of long-term wind turbine modeling; at the same time, it shortens the computation path of a single training session and improves computational efficiency. During the modeling process, the prediction difficulty and importance of different feature variables vary at different time steps. If the traditional mean squared error (MSE) is directly used as the loss function, it cannot reflect the "long-term prediction error accumulation risk" and the "importance of key feature indicators to the safe operation of wind turbines" in time-series prediction, resulting in insufficient attention to key prediction tasks during model training. Therefore, this embodiment introduces joint weights of time series and features on the basis of mean squared error and constructs a joint weighted loss function. Since the mechanistic parameters are highly sensitive to system performance, different optimization settings are used in training than in conventional models, that is, for the mechanistic parameters... Use parameters far lower than those of a neural network Learning rate: ,in Take the lowest value (e.g.) (at a certain magnitude) to avoid excessive fluctuations in mechanistic parameters during training; Maintain the learning rate of the conventional model. Adjust the mechanistic parameters in the loss function. Apply levels far higher than normal Regularization constraints, achieved by increasing the regularization coefficient To further suppress the deviation of the mechanism parameters from the physically reasonable range, a total loss function is constructed. The above optimization strategy differs from the conventional approach of "uniform learning rate and ordinary regularization". It can effectively reduce the training instability of mechanism parameters, ensure their physical interpretability, and at the same time allow the neural network to undertake the nonlinear compensation task, thereby improving the overall stability and reliability of prediction.

[0122] In this embodiment, the neural network is a three-layer neural network, including an input layer, an intermediate layer, and an output layer. The intermediate layer uses the sigmoid activation function, the output layer uses a linear activation function, and the input and output of the hidden layer are:

[0123] ;

[0124] In the above formula, As input to the neural network, for The Middle One element, For the input of the hidden layer, This is the output of the i-th node in the hidden layer. Input the number of nodes. This represents the number of hidden layer nodes. The weights of the hidden layers, For the bias of the hidden layer;

[0125] The input and output of the output layer are:

[0126] ;

[0127] In the above formula, For the input of the output layer, For the weights of the output layer, For the bias of the output layer, This is the output of the l-th node in the output layer. This represents the number of nodes in the output layer.

[0128] Compared with the prior art, the present invention has the following advantages:

[0129] The modeling method proposed in this invention, which integrates mechanistic model and data-driven model, accurately corrects inaccurate parts of the wind turbine mechanistic model by introducing a data-driven parameter identification mechanism. At the same time, it adopts a neural network structure to adaptively compensate for the unmodeled dynamic parts of the mechanistic model. This significantly improves the accuracy and robustness of the model while maintaining the interpretability of the mechanism, effectively copes with model deviations and external disturbances in wind turbine operation, and achieves high-precision description and prediction of complex nonlinear dynamic characteristics.

[0130] This invention establishes a state-space model based on various state variables of the wind turbine. During the state-space model establishment phase, multiple parameters are incorporated to accurately describe the multivariable coupled dynamic characteristics of the wind turbine system, comprehensively reflecting the dynamic responses of the structure, aerodynamics, and transmission chain. The state-space model facilitates system stability analysis and controller design, enables fault diagnosis and performance optimization, improves the safety and control accuracy of wind turbine operation, and facilitates subsequent implementation through neural network optimization steps. Together with neural network correction, it achieves high-precision modeling of the mechanistic model.

[0131] A neural network-based wind turbine mechanism modeling and optimization system includes an interconnected microprocessor and a memory, wherein the microprocessor is programmed or configured to execute the steps of a neural network-based wind turbine mechanism modeling and optimization method.

[0132] A computer-readable storage medium storing a computer program / instructions programmed or configured to execute, via a processor, the steps of a neural network-based wind turbine mechanism modeling and optimization method.

[0133] The system and medium of the present invention, corresponding to the methods described above, also have the advantages described above.

[0134] The present invention can implement all or part of the processes in the methods of the above embodiments, or it can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, it can implement the steps of the above method embodiments. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form. Computer-readable media include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. The memory is used to store computer programs and / or modules. The processor implements various functions by running or executing the computer programs and / or modules stored in the memory, and by calling data stored in the memory. The memory may include high-speed random access memory, as well as non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart media cards (SMC), secure digital (SD) cards, flash cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0135] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A neural network based fan mechanism modeling and optimization method, characterized in that, Including the following steps: S1. Establish a model of the wind energy utilization coefficient and thrust coefficient of the wind turbine and perform steady-state simulation to obtain a two-dimensional relationship table model of the wind energy utilization coefficient and thrust coefficient, and establish a wind turbine aerodynamic model based on the two-dimensional relationship table model. S2, a nonlinear aerodynamic torque in the wind turbine aerodynamic model and a nonlinear air thrust A wind turbine transmission chain system model and a wind turbine blade and tower coupling model are respectively established to obtain a complete wind turbine mechanism model. S3: Collect actual operating data of the wind turbine, analyze the deviation between the wind turbine mechanism model and the actual operating data of the wind turbine based on the neural network, and correct it through the neural network to obtain the optimized wind turbine mechanism model; Step S3 includes: A neural network is introduced to correct the errors in the unrepresented and inaccurate parts of the wind turbine mechanism model, resulting in an optimized model. The optimized model is trained by backpropagation based on a truncated time window, and the optimized wind turbine mechanism model is obtained when the training is completed. The expression for the optimization model is: ； In the above formula, represents a fan mechanism model, represents a neural network correction model, is a matrix A, B to be optimized parameters, is a neural network to be optimized parameters, is the actual running state of the fan, is the wind speed, is the pitch angle, is the generator torque.

2. The neural network based fan mechanism modeling and optimization method according to claim 1, wherein, The expression for the aerodynamic model of the wind turbine is: ； in the above formulae, for the air density, for the fan rotor radius, for the relative wind speed, for the wind energy utilization coefficient, for the thrust coefficient, for the fan rotor speed, for the pitch angle, for the tip speed ratio, for the nonlinear aerodynamic torque, for the nonlinear air thrust.

3. The neural network based fan mechanism modeling and optimization method of claim 1, wherein, The expression for the wind turbine drive train system model is: ； in the above formula, is the torsion angle of the transmission shaft, is the first derivative of is the first derivative of is the rotor speed of the fan, is the first derivative of is the first derivative of is the generator speed, is the first derivative of is the first derivative of is the gear box transmission ratio, is the equivalent moment of inertia on the rotor side, is the equivalent moment of inertia on the generator side, is the equivalent torsional stiffness coefficient of the drive train system, is the equivalent damping coefficient of the drive train system, is the generator torque.

4. The neural network based fan mechanism modeling and optimization method of claim 1, wherein, The expression for the coupled model of the wind turbine blades and tower is as follows: ； in the above formula, is the total blade mass, is the tower tube equivalent mass, is the tower tube equivalent damping coefficient, is the tower tube equivalent stiffness coefficient, is the distance from the blade center of gravity to the tower top center of gravity, is the distance from the thrust center to the tower top center of gravity, is the blade flapping angle, is the blade flapping angular velocity, is the blade flapping angular acceleration, is the tower tube fore-aft displacement, is the tower tube fore-aft movement velocity, is the tower tube fore-aft movement acceleration, is the blade damping coefficient, is the blade stiffness coefficient, is the total blade mass, is the tower tube mass, is the nacelle mass, is the hub mass, is the single blade mass, is the structural damping ratio, is the tower tube fore-aft movement frequency.

5. The neural network based fan mechanism modeling and optimization method of claim 1, wherein, In step S2, a state-space model of the wind turbine is established by combining the wind turbine aerodynamic model, the wind turbine transmission chain system model, and the coupled model of the wind turbine blades and tower. The expression of the state-space model is as follows: ， ， ， ， ， ， ， ， ； In the above formula, x is each state quantity of the fan, is a model input quantity, is a transmission shaft torsion angle, is a fan rotor speed, is a generator rotational speed, is a tower displacement before and after, is a tower movement speed before and after, is a blade flapping angle, is a blade flapping angular velocity, is a generator torque, is a gearbox transmission ratio, is an equivalent torsion stiffness coefficient of a transmission chain system, is an equivalent rotational inertia on the rotor side, is an equivalent damping coefficient of a transmission chain system, is an equivalent rotational inertia on the generator side, is an equivalent stiffness coefficient of a tower, is an equivalent mass of a tower, is an equivalent damping coefficient of a tower, is a blade stiffness coefficient, is a distance from a blade center of gravity to a tower top center of gravity, is a blade damping coefficient, is a total blade mass, is a distance from a thrust center to a tower top center of gravity; Fourth-order Runge-Kutta discretization of the state-space model yields: ； In the above formula, is the control period of the fan model discretization, denotes each state quantity of the fan at time k+1, denote the system matrix and the input matrix of the discretized fan dynamics mechanism model, respectively, and I denotes the identity matrix.

6. The neural network based fan mechanism modeling and optimization method of claim 1, wherein, The backpropagation training of the optimized model based on the truncated time window specifically includes the following steps: T1, obtain the time sequence data of the actual operation of the fan, and obtain a state sequence and an input sequence : ； In the above formulae, is the amount of data for each state, is the amount of data for each input; T2, in the first Step, extract data of the time window, into the optimization model: ； In the above formula, k+1, k+2, …, k+τ respectively represent data corresponding to the k+1, k+2, …, k+τ time points, is a hyperparameter; T3, the prediction error is calculated as follows : ； T4 performs backpropagation and parameter optimization for each time window to achieve truncated training. The overall loss function is constructed by jointly weighting temporal and feature data, as shown in the following expression: ； In the above formula, Representing feature dimension, As time weight, These are parameters used to control the initial weights and growth rate. For feature weights, for The There are two elements, where T is the total duration of data collection for the actual operation of the wind turbine, and t is the time point. The regularization coefficient is . For the joint weighted loss function, For regular expression constraints, This is the total loss function.

7. The neural network based fan mechanism modeling and optimization method according to any one of claims 1-6, characterized in that, The neural network is a three-layer neural network, including an input layer, an intermediate layer, and an output layer. The intermediate layer uses the sigmoid activation function, the output layer uses a linear activation function, and the input and output of the hidden layer are: ； In the above formula, is the input of the neural network, is the input of the neural network, is the input of the neural network, is the input of the neural network, is the input of the neural network, is the input of the neural network, is the input of the neural network, is the input of the neural network, is the input of the neural network, is the input of the neural network, The input and output of the output layer are: ； In the above formula, For the input of the output layer, For the weights of the output layer, For the bias of the output layer, This is the output of the l-th node in the output layer. This represents the number of nodes in the output layer.

8. A wind turbine mechanism modeling and optimization system based on neural networks, comprising a microprocessor and a memory interconnected, characterized in that, The microprocessor is programmed or configured to perform the steps of the neural network-based wind turbine mechanism modeling and optimization method according to any one of claims 1 to 7.

9. A computer-readable storage medium storing a computer program / instructions, characterized in that, The computer program / instructions are programmed or configured to execute the steps of the neural network-based wind turbine mechanism modeling and optimization method according to any one of claims 1 to 7 via a processor.

Images