Wind farm wake and power prediction method based on convolutional neural network model

Abstract

The present invention discloses a method for predicting the wake and power of a wind farm based on a convolutional neural network model, comprising the following steps: 1) constructing a multi-wind turbine wake superposition prediction model based on a convolutional neural network, training and verifying the model by generating a wake flow field data set through analytical wake model and numerical simulation; 2) predicting the wake velocity field and turbulence field and total power generation of the wind farm wake according to the multi-wind turbine wake superposition prediction model. The present invention can achieve a wake prediction accuracy close to that of numerical simulation while ensuring efficient calculation by adopting deep learning model technology, and can be quickly deployed in various types of wind farms by relying on data-driven model training methods; secondly, the prediction performance can be further improved through measured data of the wind farm, avoiding traditional parameter identification steps, and making the application more efficient.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power generation, and in particular to a method for predicting the wake and power of a wind farm based on a convolutional neural network model. Background Art

[0002] As one of the main renewable energy sources in the global energy market, wind energy has seen rapid growth in recent years. In wind farms, the wake effect is a key factor causing power loss in downstream wind turbines, and accurate prediction of wind farm wakes can significantly improve the power generation efficiency of large-scale wind farms. Therefore, wake prediction technology is of great significance for improving the economic benefits of wind farms. Wind turbine wake refers to a feather-like area in the downstream area of ​​the unit, characterized by reduced wind speed and increased turbulence intensity. Under conditions where the wind direction is completely aligned with the wind turbine fleet, the wind speed is lower than the rated wind speed, or the layout is tight, the wake loss may be as high as 70%.

[0003] The two mainstream methods for predicting wind turbine wakes are analytical wake models and numerical simulations. The analytical wake model is easy to use and deploy, but it cannot accurately capture the complex turbulence changes in the wake. High-fidelity numerical simulations can significantly improve the accuracy of wake predictions, but the computational cost is extremely high when solving large-scale wind farm environmental problems. Summary of the invention

[0004] In view of the above-mentioned deficiencies in the prior art, the technical problem to be solved by the present invention is: to provide a wind farm wake and power prediction method based on a convolutional neural network model, which solves the problem that the analytical wake model method in the prior art is convenient for prediction but has low accuracy, while the numerical simulation method has high accuracy but high computational cost, and it is difficult to have both.

[0005] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0006] A method for predicting wind farm wake and power based on a convolutional neural network model comprises the following steps:

[0007] 1) A multi-wind turbine wake superposition prediction model is constructed based on a convolutional neural network, and the model is trained and verified by generating a training data set through analytical wake model and numerical simulation;

[0008] 2) Predicting the wake velocity field, turbulence field and total power generation of the wind farm wake based on the multi-wind turbine wake superposition prediction model.

[0009] As an optimization, step 1) includes the following steps:

[0010] 101) A multi-wind turbine wake superposition prediction model is constructed using a repeatedly stacked residual network basic module;

[0011] 102) Pre-training stage: The wake flow field dataset generated by the analytical wake model is used to conduct preliminary training on the multi-wind turbine wake superposition prediction model;

[0012] 103) Fine-tuning stage: The training data set generated by the numerical simulation method is used to further train and verify the multi-wind turbine wake superposition prediction model;

[0013] 104) After the pre-training stage and the fine-tuning stage, the final multi-wind turbine wake superposition prediction model is obtained.

[0014] As an optimization, in step 101), the input for constructing the multi-wind turbine wake superposition prediction model is the single wake prediction result of each wind turbine considered separately, including wind speed and turbulence distribution, which is obtained through a single wind turbine wake prediction model.

[0015] As an optimization, the training data set is a wake wind speed and turbulence data set generated under different incoming flows, yaw angles and wind turbine layouts.

[0016] As an optimization, in step 102), the wind turbine layout used to generate the training data set is randomly generated under the constraint of the minimum spacing between wind turbines.

[0017] As an optimization, in step 103), the fan layout used to generate the training data set is that a certain number of fans are arranged in fixed positions.

[0018] As an optimization, step 2) includes the following steps:

[0019] 201) Set the standard reference wind direction and rotate the wind turbine coordinates in the wind farm so that the wind turbines are arranged in order from upstream to downstream of the wind direction;

[0020] 202) According to the rotated coordinate system, the single wake field is calculated starting from the most upstream wind turbine in sequence, that is, the incoming flow is input into the single wind turbine wake prediction model to obtain the single wake field of the most upstream wind turbine, and then starting from the second wind turbine, the single wake field results calculated by each wind turbine according to the single wind turbine wake prediction model and the single wake field result of the previous wind turbine are input into the multi-wind turbine wake superposition prediction model to obtain the multi-wake field prediction results, and the calculation is accumulated to the most downstream wind turbine in sequence, so as to obtain the final wake field result;

[0021] 203) According to the wake field result, the average inflow wind speed of each wind turbine is calculated by obtaining wind speed information at the coordinate points at the cross section of each wind turbine impeller;

[0022] 204) According to the inflow wind speed of each wind turbine, the power of each wind turbine is predicted according to the preset power curve and the power generation power of the entire wind farm is obtained.

[0023] Compared with the prior art, this application has the following beneficial effects:

[0024] The present invention aims to solve the problem of rapid prediction of flow field and power in wind farms under different incoming flow conditions. Based on the generative adversarial network and Transformer model in deep learning technology, a wake calculation method that combines the advantages of analytical wake model and numerical simulation method is proposed. By adopting deep learning model technology, the wake prediction accuracy close to that of numerical simulation can be achieved while ensuring efficient calculation. Relying on data-driven model training methods, it can be quickly deployed in various types of wind farms; secondly, the prediction performance can be further improved through the measured data of the wind farm, avoiding the traditional parameter identification steps, and the application is more efficient. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 A flowchart of the multi-wind turbine wake superposition prediction model of the present invention for predicting the wake of a wind farm;

[0026] Figure 2 A schematic diagram of a wind turbine layout for generating a wake flow field data set for a multi-wind turbine wake superposition prediction model pre-training stage using a method of analyzing the wake model in the present invention; wherein D represents the wind turbine hub diameter;

[0027] Figure 3 A schematic diagram of a wind turbine layout for generating a wake flow field data set for the fine-tuning stage of a multi-wind turbine wake superposition prediction model using a numerical simulation method in the present invention; wherein D represents the diameter of the wind turbine hub;

[0028] Figure 4 A schematic diagram of a specific wind turbine layout for multi-wake prediction in an embodiment of the present invention;

[0029] Figure 5 For Figure 4 The Y-axis wake prediction data of the wind farm with wind turbine layout;

[0030] Figure 6 For Figure 4 The z-axis wake prediction data of the wind farm with wind turbine layout;

[0031] Figure 7 For Figure 4 Cross-sectional wake cloud diagram of wind farm with wind turbine layout;

[0032] Figure 8 Another specific wind turbine layout schematic diagram for multi-wake prediction in an embodiment of the present invention;

[0033] Fig. 9 For Figure 8 The Y-axis wake prediction data of the wind farm with wind turbine layout;

[0034] Fig.10 For Figure 8The z-axis wake prediction data of the wind farm with wind turbine layout;

[0035] Fig.11 For Figure 8 Cross-sectional wake cloud diagram of a wind farm with wind turbine layout. DETAILED DESCRIPTION

[0036] The present invention will be further described in detail below in conjunction with the accompanying drawings.

[0037] For specific implementation: see Figure 1-11 ,

[0038] In order to achieve the dual advantages of high efficiency and accuracy, the present invention proposes a wake prediction framework based on machine learning technology. The machine learning-based model will be trained and tested using a large number of wake data sets generated by CFD simulations. Once well-trained, the machine learning-based wake prediction model can effectively characterize the wake distribution of a wind farm, with a calculation speed comparable to that of a traditional analytical wake model, while maintaining good accuracy equal to that of a high-fidelity numerical simulation. The framework is expected to be applicable to the modeling of single or multiple wakes under various operating conditions, and further deployed in application problems of wind farms, such as power prediction, layout optimization, or intelligent control.

[0039] Specifically, the present invention proposes a wind farm wake and power prediction method based on a convolutional neural network model, comprising the following steps: Figure 1 As shown,

[0040] 1) A multi-wind turbine wake superposition prediction model is constructed based on a convolutional neural network, and the model is trained and verified by generating a training data set through analytical wake model and numerical simulation.

[0041] Specifically, step 1) includes the following steps:

[0042] 101) A multi-fan wake superposition prediction model is constructed using a repeatedly stacked residual network basic module; the input for constructing the multi-fan wake superposition prediction model is the single wake prediction result of each fan considered separately, including wind speed and turbulence distribution, obtained through a single fan wake prediction model. The single fan wake prediction model can be constructed based on a transformer model and a conditional generative adversarial network, and the model is trained and verified by generating a training data set through an analytical wake model and numerical simulation.

[0043] 102) Pre-training stage: The wake flow field dataset generated by the analytical wake model is used to conduct preliminary training on the multi-wind turbine wake superposition prediction model; the training dataset is a wake wind speed and turbulence dataset generated under different incoming flows, yaw angles and wind turbine layouts. The inflow conditions include different inflow profiles (hub height wind speed 4~15m / s and turbulence 5%~16%), yaw angles (-30~30 degrees), etc., and this dataset is used to conduct preliminary training on the single wind turbine wake prediction model; the parameter randomness of the analytical wake model is also considered in data generation to improve the generalization ability of the model in the pre-training stage; the wind turbine layout adopts a random generation method within a specific area, and ensures that the minimum spacing between wind turbines is maintained, such as Figure 2 shown.

[0044] 103) Fine-tuning stage: The training data set generated by the numerical simulation method is used to further train and verify the multi-wind turbine wake superposition prediction model; the wind turbine layout adopts a certain amount of fixed-position simulation schemes, such as Figure 3 shown.

[0045] 104) After the pre-training stage and the fine-tuning stage, the final multi-wind turbine wake superposition prediction model is obtained.

[0046] 2) Predicting the wake field and power of the wind farm wake according to the multi-wind turbine wake superposition prediction model.

[0047] Specifically, step 2) includes the following steps:

[0048] 201) Set the standard reference wind direction and rotate the wind turbine coordinates in the wind farm so that the wind turbines are arranged in sequence from upstream to downstream of the wind direction.

[0049] 202) According to the rotated coordinate system, the single wake field is calculated from the most upstream wind turbine in sequence, that is, the incoming flow is input into the single wind turbine wake prediction model to obtain the single wake field of the most upstream wind turbine, and then starting from the second wind turbine, the single wake field results calculated by each wind turbine according to the single wind turbine wake prediction model and the single wake field results of the previous wind turbine are input into the multi-wind turbine wake superposition prediction model to obtain the multi-wake field prediction results, and then accumulated to the most downstream wind turbine in sequence to obtain the final wake field result. That is, the single wake field is calculated from the most upstream wind turbine in sequence, that is, the incoming flow is input into the single wake prediction model to obtain the wake field considering only the most upstream wind turbine, and then the wake fields of the second, third, fourth, etc. are calculated, wherein starting from the second wind turbine, after each calculation of the single wake field, the current result and the wake field obtained by the previous calculation need to be input into the multi-wake prediction model to obtain the multi-wake prediction results, and so on, and the final wake field result can be obtained by repeatedly accumulating to the last wind turbine.

[0050] 203) According to the wake field result, the average inflow wind speed of each wind turbine is calculated by obtaining wind speed information at the coordinate points at the cross section of each wind turbine impeller;

[0051] 204) According to the inflow wind speed of each wind turbine, the wake velocity field, turbulence field and total power generation are predicted according to the corresponding relationship curve.

[0052] The invention predicts the wake of multiple wind turbines by preparing the prediction results of a single wake field under the influence of a single wind turbine in advance, and then continues to predict the wake field results of each wind turbine according to the Figure 1 The prediction process of the multi-wake prediction model shown in the figure can obtain the wake field information of multiple wake superposition after combining the input.

[0053] Specifically, use Figure 4 When the wind turbine layout is shown, there are two operating parameters, one of which is the hub height wind speed, turbulence, and yaw angle of 0°, and the other is the hub height wind speed, turbulence, and yaw angle of 10°. The prediction results of the multi-wake prediction model of the present invention are as follows: Figure 5-Figure 7 As shown in the figure, the upper part is the first operating condition parameters, and the lower part is the second operating condition parameters. The cross sections at 9D, 11D, 13D, 15D, 17D and 20D away from the wind farm are respectively intercepted along the x-axis direction. The wake wind speed in the y-axis direction is as follows Figure 5 As shown in the figure, the horizontal axis represents 1-predicted wind speed / hub height wind speed, the vertical axis represents the ratio of the distance from the x-axis along the y-axis to the hub diameter, the red line and the green line represent the prediction results of the multi-wind turbine wake superposition prediction model after the pre-training stage and the fine-tuning stage, × represents the data of the analytical wake model proposed by Ishihara, and ○ represents the data of the Reynolds stress turbulence and brake disc coupling model. The wake wind speed in the z-axis direction is as follows Figure 6 As shown, the ordinate represents the ratio of height to fan hub diameter, and the others are Figure 5 The wake cloud diagram of its cross section is as follows: Figure 7 As shown by Figure 7 The wake distribution data at different cross-sectional locations can be seen more intuitively.

[0054] Use Figure 8 When the wind turbine layout is shown, the operating parameters are still two, one of which is the hub height wind speed, turbulence, and yaw angle of 0°, and the other is the hub height wind speed, turbulence, and yaw angle of 10°. The prediction results of the multi-wake prediction model of the present invention are as follows Figure 9-11 As shown in the figure, the upper part is the first operating condition parameters, and the lower part is the second operating condition parameters. The cross sections at 9D, 11D, 13D, 15D, 17D and 20D away from the wind farm are respectively intercepted along the x-axis direction. The wake wind speed in the y-axis direction is as follows Fig. 9As shown in the figure, the horizontal axis represents 1-predicted wind speed / hub height wind speed, the vertical axis represents the ratio of the distance from the x-axis along the y-axis to the hub diameter, the red line and the green line represent the prediction results of the multi-wind turbine wake superposition prediction model after the pre-training stage and the fine-tuning stage, × represents the data of the analytical wake model proposed by Ishihara, and ○ represents the data of the Reynolds stress turbulence and brake disc coupling model. The wake wind speed in the z-axis direction is as follows Fig.10 As shown, the ordinate represents the ratio of height to fan hub diameter, and the others are Figure 5 The wake cloud diagram of its cross section is as follows: Fig.11 As shown by Fig.11 The wake distribution data at different cross-sectional locations can be seen more intuitively.

[0055] The present invention aims to solve the problem of rapid prediction of flow field and power in wind farms under different incoming flow conditions, and proposes a wake calculation method based on a deep learning framework that combines the advantages of both analytical wake models and numerical simulation methods. By adopting deep learning model technology, the wake prediction accuracy close to that of numerical simulation can be achieved while ensuring efficient calculation. Relying on data-driven model training methods, it can be quickly deployed in various types of wind farms; secondly, the prediction performance can be further improved through the measured data of the wind farm, avoiding the traditional parameter identification steps, and the application is more efficient.

[0056] Although the embodiments of the present invention have been shown and described, it is apparent to those skilled in the art that various changes, modifications, substitutions and variations may be made to these embodiments without departing from the principles and basis of the present invention. The scope of the present invention is defined by the appended claims and their equivalents. Therefore, the embodiments of the present invention are merely illustrative examples of the present invention. The embodiments of the present invention do not constitute limitations on the present invention from any point of view.

Claims

1. A wind farm wake and power prediction method based on a convolutional neural network model, characterized in that: The steps include: 1) A multi-wind turbine wake superposition prediction model is constructed based on a convolutional neural network. The wake flow field dataset is generated through empirical analytical wake model and numerical simulation to train and verify the model. Specifically, 101) A multi-wind turbine wake superposition prediction model is constructed using a repeatedly stacked residual network basic module; wherein, the input for constructing the multi-wind turbine wake superposition prediction model is the single wake prediction result of each wind turbine considered separately, including wind speed and turbulence intensity distribution, obtained through a single wind turbine wake prediction model; the single wind turbine wake prediction model is constructed based on the transformer model and conditional generative adversarial network, and the model is trained and verified by generating a training data set through analytical wake model and numerical simulation; 102) Pre-training stage: The wake flow field dataset generated by the analytical wake model is used to conduct preliminary training on the multi-wind turbine wake superposition prediction model; 103) Fine-tuning stage: The training data set generated by the numerical simulation method is used to further train and verify the multi-wind turbine wake superposition prediction model; 104) After the pre-training stage and the fine-tuning stage, the final multi-wind turbine wake superposition prediction model is obtained; The training data set is a wake wind speed and turbulence data set generated under different incoming flows, yaw angles and wind turbine layouts; 2) Predicting the wake velocity field, turbulence field and total power generation of the wind farm wake according to the multi-wind turbine wake superposition prediction model; specifically, 201) Set the standard reference wind direction and rotate the wind turbine coordinates in the wind farm so that the wind turbines are arranged in order from upstream to downstream of the wind direction; 202) According to the rotated coordinate system, the single wake field is calculated starting from the most upstream wind turbine in sequence, that is, the incoming flow is input into the single wind turbine wake prediction model to obtain the single wake field of the most upstream wind turbine, and then starting from the second wind turbine, the single wake field results calculated by each wind turbine according to the single wind turbine wake prediction model and the single wake field result of the previous wind turbine are input into the multi-wind turbine wake superposition prediction model to obtain the multi-wake field prediction results, and the calculation is accumulated to the most downstream wind turbine in sequence, so as to obtain the final wake field result; 203) According to the wake field result, the average inflow wind speed of each wind turbine is calculated by obtaining wind speed information at the coordinate points at the cross section of each wind turbine impeller; 204) According to the inflow wind speed of each wind turbine, the power of each wind turbine is predicted according to the preset power curve and the power generation power of the entire wind farm is obtained.

2. The method for predicting wind farm wake and power based on a convolutional neural network model according to claim 1, characterized in that: In step 102), the wind turbine layout used to generate the training data set is randomly generated under the constraint of the minimum spacing between wind turbines.

3. The method for predicting wind farm wake and power based on a convolutional neural network model according to claim 1, characterized in that: In step 103), the fan layout used to generate the training data set is that a certain number of fans are arranged in fixed positions.

Images