Wind farm model state correction method based on gru neural network

Abstract

The application discloses a wind farm model state correction method based on a GRU neural network, and mainly relates to the technical field of wind power; the method comprises the following steps: S1, discretizing a wind farm in space and time, and defining a wind field boundary and an initial condition; S2, solving a model iteration algorithm, and defining a difference between a wind field data simulation value and an actual value; S3, fitting an improved model through a GRU deep neural network; under the premise that the high-fidelity CFD model with high calculation cost does not significantly increase the calculation amount, the application makes the corrected model state information more accurate, and improves the accuracy of the wind field simulation state value.

Description

Technical Field

[0001] This invention relates to the field of wind power technology, specifically a wind farm model state correction method based on GRU neural network. Background Technology

[0002] To study the impact of wake effects on wind farms, researchers have designed a wide variety of wind farm models. Existing wake models are mainly divided into three categories: one-dimensional linear models, two-dimensional viscous eddy current theory models based on axisymmetric assumptions, and complete three-dimensional CFD computational models. Common one-dimensional linear models include the Jensen wake model, Gaussian wake model, and Bastankhah wake model. These low-fidelity models are typically simple in design, consider few factors, and generally only describe the main airflow characteristics around the wind turbine. Low-fidelity models are static models that use simplifying assumptions to limit the necessary computation time; therefore, they can quickly provide flow rate estimates but are not suitable for practical analysis. Two-dimensional viscous eddy current theory models mainly include lift lines and vortex lattice methods. These models combine wake models to calculate the wake field. Complete three-dimensional CFD models, such as SOWFA, are high-fidelity models that capture as much airflow as possible in the wind farm and typically consider many factors that may affect the airflow, and include more complex wind turbine models. High-fidelity CFD models are generally much more accurate than low-fidelity models, but this comes at the cost of requiring more computing power and taking more time.

[0003] Because the mechanisms within the wake region of a wind farm are extremely complex, and the wind inflow itself possesses many characteristics such as high noise, non-stationarity, and strong nonlinearity, increasing the complexity of theoretical calculations solely for model accuracy during model derivation will increase computational costs in two ways: Firstly, the computational load of the theoretical model portion. Complete 3D CFD models, such as SOWFA, are high-fidelity models that capture as much airflow as possible in the wind farm and typically consider numerous factors that may affect airflow, including more complex turbine models. High-fidelity CFD models are usually much more accurate than low-fidelity models, but this comes at the cost of requiring more powerful computing capabilities and consuming more time. Secondly, this modeling method lacks universality from a computational cost perspective; different wind farm layouts require different high-fidelity CFD models.

[0004] Therefore, in order to take into account the ideal assumption of low computational cost, the accuracy of using model iteration to solve the wind farm state at the next moment needs to be further improved. Summary of the Invention

[0005] The purpose of this invention is to solve the problems existing in the prior art and provide a wind farm model state correction method based on GRU neural network. The method uses GRU deep neural network to fit the difference between the simulated value and the actual value of the wind farm model at the previous moment and the difference between the axial induction coefficient of the wind turbine and the simulated value and the actual value of the model at the next moment. Without significantly increasing the amount of computation compared with the high-fidelity CFD model with high computational cost, the corrected model state information is more accurate, thus improving the accuracy of the wind farm simulation state value.

[0006] To achieve the above objectives, the present invention employs the following technical solution:

[0007] A wind farm model state correction method based on GRU neural network includes the following steps:

[0008] S1. Discretize the wind farm in space and time, and define the wind farm boundary and initialization conditions;

[0009] S2. Solve the model iterative algorithm and define the difference between the simulated and actual wind field data;

[0010] S3. Improve the model by fitting it with a GRU deep neural network.

[0011] Preferably, in step S1, a wind farm model is constructed based on the Navier-Stokes equations, combined with the Prandtl mixed length turbulence model and the wind turbine actuation disk model, and the Navier-Stokes equations are discretized in space and time using the finite volume method and the mixed difference method.

[0012] Preferably, step S1 includes the following steps:

[0013] S11. The Navier-Stokes equations can be discretized spatially on an interlaced grid using the finite volume method and the hybrid finite difference method. Time discretization is then performed using an unconditionally stable implicit method, resulting in the interlaced grid and the following formula:

[0014]

[0015]

[0016] S12. Simplification: After processing the above equation by discretizing the nonlinear terms, approximating the convective mass flux, and approximating the central difference, we obtain the completely discrete Navier-Stokes equation.

[0017] S13. Define the wind field boundary and initialization conditions.

[0018] Preferably, in step S2, after the solver obtains the given current state and control action, it calculates the state at the next moment, and deploys high-precision sensors in the wind farm to obtain the true value, and defines the error by subtracting the simulated value output by the model.

[0019] Preferably, step S2 includes the following steps:

[0020] S21. Solve for each grid state, and calculate the state of the flow field at the next time step given the state of the flow field at the previous time step and the control action;

[0021] S22. Define the error based on the values ​​of the simulated state and the actual state: First, calculate the simulated state of all grids using step S21:

[0022]

[0023] Then, the measured accurate value of the real-time wind field status was obtained as follows:

[0024]

[0025] By comparing the simulated values ​​of the state of a cell in a wind farm with the accurate measured values, the difference between the measured and simulated values ​​at any time in a simulation can be obtained:

[0026]

[0027] Then at the previous time k-1, we have:

[0028]

[0029] Preferably, in step S3, the GRU neural network is used to fit the error value of the model with the real value of the wind farm related data based on the control input at the previous moment, thereby completing the correction of the simulated value derived from the model.

[0030] Preferably, step S3 includes the following steps:

[0031] S31, Wind farm parameter settings, simulated flow and control settings;

[0032] S32. Data preprocessing: Expand and flatten the nested sub-matrices, and perform the e operation in step S22. k Convert to a one-dimensional array:

[0033]

[0034] For a one-dimensional matrix, data standardization can be performed directly.

[0035] S33, GRU Neural Network Model Construction: Data sets were collected through simulation experiments; the root mean square error (RMSE) was used for error evaluation; and suitable hyperparameters of the model were selected through trial and error.

[0036] S34. Based on the model training results, fit the state error e at the previous moment. k-1 With control function w k-1 As input, with the state error e at this moment k As output, the correction results are applied to the wind farm model, thus obtaining the predicted state value of the grid in the wind field at a certain time k:

[0037]

[0038] In other words, by analyzing the state of historical data, the model achieves a more accurate prediction of the wind field state at a certain moment during model iteration.

[0039] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0040] This invention takes into account reducing computational costs in the model construction phase, simplifying the Navier-Stokes equations to a certain extent without sacrificing too much computational accuracy. However, this leaves room for improvement in the accuracy of the model's predictive control. Therefore, the core method of this invention addresses this issue to some extent by using a GRU neural network to fit simulated values ​​to actual values, thereby increasing the accuracy of the wind farm model. Attached Figure Description

[0041] Figure 1 This is a flowchart of the present invention;

[0042] Figure 2 Pseudocode for model iteration;

[0043] Figure 3 Schematic diagram of GRU neural network structure

[0044] Figure 4 It is an interlaced grid diagram. Detailed Implementation

[0045] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined in this application.

[0046] Example: As attached Figure 1-4As shown, the present invention describes a wind farm model state correction method based on a GRU neural network. This method uses a GRU deep neural network to fit the difference between the simulated and actual values ​​of the wind farm model at the previous moment, and the difference between the axial induction coefficient of the wind turbine and the simulated and actual values ​​at the next moment. Without significantly increasing the computational load compared to a high-fidelity CFD model with higher computational costs, this method makes the corrected model state information more accurate, thus improving the accuracy of the wind farm simulation state values.

[0047] Including the following steps:

[0048] S1. Discretize the wind farm spatially and temporally, define the wind farm boundary and initialization conditions, specifically including the following steps:

[0049] S11. The Navier-Stokes equations can be discretized spatially on an interlaced grid using the finite volume method and the hybrid finite difference method. Time discretization is then performed using an unconditionally stable implicit method, resulting in the following formula:

[0050]

[0051]

[0052] And as attached Figure 4 The interlaced grid shown.

[0053] S12. Simplification: After discretizing the nonlinear terms, approximating the convective mass flux, and approximating the central difference in the above equation, we obtain the completely discrete Navier-Stokes equation.

[0054] S13. Define the wind field boundary and initialization conditions.

[0055] S2. Solve the model iterative algorithm and define the difference between the simulated and actual values ​​of the wind field data, including the specific steps;

[0056] S21. Solve for each grid state. Given the state of the flow field at the previous time step and the control action, calculate the state of the flow field at the next time step, such as... Figure 2 As shown.

[0057] S22. Define the error based on the values ​​of the simulated and actual states: First, calculate the simulated state of all grids (including the simulated longitudinal wind speed, simulated lateral wind speed, and simulated pressure value for each grid) using step S21:

[0058]

[0059] By deploying high-precision lidar anemometers and barometers in the wind farm, wind speed and pressure data at various times can be accurately recorded for each grid in the wind farm. The accurate values ​​measured at time k are denoted as follows: transverse wind speed , longitudinal velocity , and pressure value , where is the total simulation time. This yields the accurate measured values ​​of the real-time wind farm conditions.

[0060]

[0061] By comparing the simulated values ​​of the state of a cell in a wind farm with the accurate measured values, the difference between the measured and simulated values ​​at any time in a simulation can be obtained:

[0062]

[0063] Then at the previous time k-1, we have:

[0064]

[0065] S3. Improve the model by fitting it with a GRU deep neural network, such as Figure 3 As shown, the specific steps include:

[0066] S31, Wind farm parameter settings, simulated flow and control settings;

[0067] S32. Data Preprocessing: The main difference between nested matrices and regular matrices in data processing lies in their data structures. A nested matrix is ​​a large matrix containing multiple sub-matrices of the same dimension. Therefore, we need to perform special data processing on errors, expanding and flattening the nested sub-matrices:

[0068] Convert to a one-dimensional array:

[0069] For a one-dimensional matrix w k =[β 1,k ,β 2,k ,β 3,k ,…,β N,k ] T The latter is relatively simple; standardized data processing can be performed directly.

[0070] S33. GRU Neural Network Model Construction: Data sets are collected through simulation experiments; to reflect the model's prediction accuracy, the root mean square error (RMSE) is used for error evaluation; suitable hyperparameters of the model are selected through trial and error to ensure that the neural network can achieve good results (without overfitting, underfitting, or other phenomena).

[0071] S34. Based on the model training results, fit the state error e at the previous moment. k-1With control function w k-1 As input, with the state error e at this moment k As output, the correction results are applied to the wind farm model, thus obtaining the predicted state value of the grid in the wind field at a certain time k:

[0072]

[0073] In this model, by analyzing the state of historical data, the present invention achieves a more accurate prediction of the state of the wind field at a certain moment in the model iteration, and the algorithm ends.

Claims

1. A wind farm model state correction method based on GRU neural network, characterized in that, Including the following steps: S1. Discretize the wind farm in space and time, and define the wind farm boundary and initialization conditions; S2. Solve the model iterative algorithm and define the difference between the simulated and actual wind field data; S3. Improve the model by fitting it with a GRU deep neural network; In step S2, after the solver obtains the given current state and control action, it calculates the state at the next moment and deploys high-precision sensors in the wind farm to obtain the real value. The difference between the real value and the simulated value output by the model is used to define the error. Step S2 includes the following steps: S21. Solve for each grid state, and calculate the state of the flow field at the next time step given the state of the flow field at the previous time step and the control action; S22. Define the error based on the values ​​of the simulated state and the actual state: First, calculate the simulated state of all grids using step S21: ； Then, the measured accurate value of the real-time wind field status was obtained as follows: ； By comparing the simulated values ​​of the state of a cell in a wind farm with the accurate measured values, the difference between the measured and simulated values ​​at any time in a simulation can be obtained: ； Then at the previous time k-1, we have: ； In step S3, the GRU neural network is used to fit the error value of the model with the real value of the wind farm related data based on the control input at the previous moment, thereby completing the correction of the simulated value derived from the model. Step S3 includes the following steps: S31, Wind farm parameter settings, simulated flow and control settings; S32. Data preprocessing: Expand and flatten the nested submatrices, performing the operations from step S22. Convert to a one-dimensional array: ； For a one-dimensional matrix, data standardization can be performed directly. S33, GRU Neural Network Model Construction: Data sets were collected through simulation experiments; the root mean square error (RMSE) was used for error evaluation; and suitable hyperparameters of the model were selected through trial and error. S34. Based on the model training results, fit the state error of the previous moment. With control As input, with the state error at this moment The correction results of the output GRU neural network model are applied to the wind farm model to obtain the state prediction value of the grid in the wind field at a certain time k: ； In other words, by analyzing the state of historical data, this model can make a more accurate prediction of the state of the wind field at a certain moment in the model iteration.

2. The wind farm model state correction method based on GRU neural network as described in claim 1, characterized in that: In step S1, a wind farm model is constructed based on the Navier-Stokes equations, combined with the Prandtl mixed length turbulence model and the wind turbine actuation disk model. The Navier-Stokes equations are then discretized in space and time using the finite volume method and the mixed difference method.

3. The wind farm model state correction method based on GRU neural network as described in claim 2, characterized in that: Step S1 includes the following steps: S11, the Navier-Stokes equations are discretized. Spatial discretization of the equations can be achieved on an interlaced grid using the finite volume method and the hybrid finite difference method. Temporal discretization is then performed using an unconditionally stable implicit method, resulting in the interlaced grid and the following formula: ； ； S12. Simplification: After discretizing the nonlinear terms, approximating the convective mass flux, and approximating the central difference in the above equation, we obtain the fully discrete Navier-Stokes equation. S13. Define the wind field boundary and initialization conditions.

Images