CN116757101B - Cabin wind speed correction method and system based on mechanism model and neural network - Google Patents

Abstract

According to the cabin wind speed correction method and system based on the mechanism model and the neural network, wind speed data and power data measured in an SCADA system are selected, the SCADA wind speed is divided into wind speed data sets of different working conditions according to rated wind speed, and theoretical wind speed is obtained through calculation of an empirical formula; calculating high-low frequency residual errors of the real wind speed and the theoretical wind speed by wavelet transformation, establishing a relation between the SCADA wind speed and the high-low frequency residual errors by using a neural network, inputting SCADA wind speed data from an actual running wind power plant into the trained neural network, and obtaining corresponding high-low frequency residual errors; and (3) performing one-to-one correspondence on the SCADA wind speed data and the residual data, and performing linear addition to obtain a real wind speed correction value. The method and the system have the characteristics of high calculation speed, high precision and good generalization, and can be widely used for correcting the wind speed of wind driven generators of different types.

Description

A cabin wind speed correction method and system based on mechanism model and neural network

Technical field

This application relates to the fields of data processing and data transmission, and more specifically, to a cabin wind speed correction method and system based on a mechanism model and a neural network.

Background technique

With the development of wind power technology, it is necessary to obtain accurate wind speed data when conducting economic benefit assessment, power generation assessment, fleet layout, and operation control in wind farms. Since the real wind speed is not easy to measure directly, the nacelle wind speed recorded by the Supervisory Control And Data AcquiSition System (SCADA) is usually used as an important parameter for analysis when calculating wind turbines. However, due to factors such as blade rotation and airflow distortion caused by the nacelle, the wind speed value measured by SCADA often cannot directly reflect the true wind speed of the impeller and must be corrected before use. This correction to the wind speed is called the nacelle transfer function.

At present, the methods for correcting SCADA wind speed are mainly divided into theoretical calculation correction methods and function fitting methods. Among them, the theoretical calculation correction method is mainly based on aerodynamic theory, using multiple parameters of the wind turbine and the operating data recorded by SCADA to achieve cabin wind speed correction through theoretical calculations. However, due to its large number of parameters and complex internal mechanism, this method The calculation accuracy of the class method is insufficient and cannot cover different aircraft types; the function fitting method corrects the cabin wind speed by fitting the direct functional relationship between the wind speed of the wind tower and the wind speed of the cabin. This method is simple and easy to use, but it relies heavily on the wind tower data. However, the configuration site of the wind measurement tower is demanding and expensive. Usually, only one typical wind turbine is selected to equip it with a wind measurement tower for characteristic evaluation, which cannot meet the needs of all wind turbines in the entire site. Therefore, the application cost of this method is high and the application cost is high. Scenes are limited.

Contents of the invention

In order to overcome the shortcomings of the existing technology, the present invention provides a nacelle wind speed correction method and system that uses neural network, theoretical characteristics, model evaluation and correction and other technologies and considers wind power operating conditions. Specifically, a method and system based on a mechanism model are proposed and neural network cabin wind speed correction method and system. This invention measures the real wind speed by installing a laser radar at the impeller, and then combines the physical mechanism with the data model. It has the characteristics of fast calculation speed, high accuracy and good generalization, and can be widely used to measure the wind speed of different types of wind turbines. Correction.

In the first aspect, the present invention provides a cabin wind speed correction method based on a mechanism model and a neural network, which includes the following steps:

Select the wind speed data and power data measured in the SCADA system, divide the wind speed data and power data according to preset rules, and obtain a pitch-changing operating condition data set and a non-pitch operating condition data set;

Wherein, the dividing according to the preset rules includes: determining whether the wind speed data is less than the rated wind speed data; if so, determining that the wind speed data is the wind speed under non-pitch operating conditions; if not, determining that the wind speed data is Wind speed during pitch operation;

The resolution of wind speed data and power data in the SCADA system is 1s level;

Convert the pitch-changing operating condition data set and the non-pitch operating condition data set into average data with a resolution of 30s as a feature data set;

Obtain the air density information and sweep area information, calculate the air density information, sweep area information, wind speed data and power data according to the preset empirical formula to obtain the theoretical wind speed under the corresponding working conditions. ;

Among them, the preset empirical formula is:

;

in is the power data,/> is the wind speed data,/> is the air density information,/> It is the sweep area information;

Get true wind speed , and then use wavelet transform to convert the theoretical wind speed/> and the actual wind speed/> Decompose them separately to obtain the corresponding high-frequency data and low-frequency data, and obtain the corresponding high-frequency residuals/> and low-frequency residual/> , convert the high-frequency residual/> and the low frequency residual/> As the target data set, the wind speed data is a feature data set;

Among them, the true wind speed measured by lidar , and then use wavelet transform to convert the real wind speed/> and theoretical wind speed/> Decompose them separately to obtain the corresponding high-frequency data/> ,/> and low frequency data/> ,/> , then // and/> Subtract the high-frequency residual of the real wind speed high-frequency data and the theoretical wind speed high-frequency data/> , will/> and/> Subtract to obtain the low-frequency residual of the real wind speed low-frequency data and the theoretical wind speed low-frequency data/> , Specifically, after obtaining the real wind speed data, the real wind speed data is decomposed into high-frequency data and low-frequency data through wavelet transformation, and the corresponding time points are recorded, and then the theoretical wind speed is decomposed into high-frequency data and low-frequency data according to the time point of the real wind speed;

Use the characteristic data set as input and the target data set as output, use a neural network for training, use an optimization method to determine the time step, and establish a wind speed-residual prediction model based on operating conditions and high and low frequency data;

Input SCADA wind speed data of a certain time step into the trained wind speed-residual prediction model, and obtain high-frequency residual prediction value data and low-frequency residual prediction value data corresponding to the time step;

The SCADA wind speed data is linearly added to the high-frequency residual prediction value data and the low-frequency residual prediction value data to obtain the final true wind speed correction value.

Optional, where the real wind speed in front of the wind turbine hub is measured through lidar ;

Measuring cabin wind speed through SCADA system .

Optionally, the feature data set is used as input, the target data set is used as output, a neural network is used for training, an optimization method is used to determine the time step, and the wind speed- Residual prediction models include:

Use neural networks for training to establish wind speed-residual prediction models under two working conditions, a total of four: one is a high-frequency residual prediction model in which SCADA wind speed is used as input and high-frequency residuals are output under pitch working conditions; The second is a low-frequency residual prediction model in which SCADA wind speed is used as input under pitch changing conditions and low-frequency residuals are used as output. The third is a high-frequency residual prediction model in which SCADA wind speed is used as input under non-pitching conditions and high-frequency residuals are used as output. ; The fourth is a low-frequency residual prediction model using SCADA wind speed as input and low-frequency residual as output under non-pitch operating conditions.

Optionally, in the process of establishing the wind speed-residual prediction model, gamma detection technology is used to optimize the training time step and prediction time step, and the layer-by-layer random jumping method is used to optimize the parameters of each layer. Combination solution.

Optionally, the layer-by-layer random jumping method includes the following steps:

1) Determine the parameters, jump range and jump step size of random jumps:

Determine the parameters that need to be optimized as the second layer LSTM unit number, dropout value, and activation function; the fourth layer fully connected layer unit number and activation function; set the range and jump step size of each parameter; the activation function is the sigmod function, tanh The function and relu function are randomly selected, and the parameter value calculation formula is as follows:

;

Calculates the value for the current round,/> Calculated value for the previous round,/> is the minimum value of the range,/> is the maximum value of the range, is the jump step size,/> is a function that randomly selects integers within a given range,/> ;

2) In the first round, the number of units, dropout value, mapping dimension and activation function of the selected layer are randomly initialized within a given range, and then a certain amount of wind speed and residual data sets are divided into training sets and test sets for training and testing. And calculate the accuracy, get the accuracy value of the corresponding parameter, record the parameter combination and result; then add a random positive or negative integer to the number of units and dropout value of the selected layer multiplied by the jump step size to get the next set of parameter values. , where the parameter values cannot exceed the given range, and the activation function is also randomly selected and executed repeatedly for five rounds;

3) Calculate the accuracy of five groups of random parameters, select the parameter combination with the highest accuracy, and then execute the parameters of the fourth layer according to step 2). Also calculate the accuracy of the five groups of parameters and select the parameter combination with the highest accuracy, and finally get the optimal Parameter combination.

Optionally, after obtaining the SCADA wind speed data and judging the working conditions based on the SCADA wind speed data, input it into the corresponding high-frequency residual prediction model and low-frequency residual prediction model to obtain the high-frequency residual of the corresponding time step. and low-frequency residual/> , and finally the cabin wind speed/> Add the high-frequency residuals of the corresponding time steps/> and low-frequency residual/> Get corrected true wind speed/> ,Right now

.

In a second aspect, the present invention provides a cabin wind speed correction system based on a mechanism model and a neural network, which is used to implement a cabin wind speed correction method based on a mechanism model and a neural network described in the first aspect, including:

A division module, used to select wind speed data and power data measured in the SCADA system, divide the wind speed data and power data according to preset rules, and obtain a pitch-changing operating condition data set and a non-pitch operating condition data set;

The calculation module is used to obtain air density information and sweep area information, and calculate the air density information, sweep area information, wind speed data and power data according to a preset empirical formula to obtain the corresponding work The theoretical wind speed under ;

Decomposition module for obtaining true wind speed , and then use wavelet transform to convert the theoretical wind speed/> and the actual wind speed/> Decompose them separately to obtain the corresponding high-frequency data and low-frequency data, and obtain the corresponding high-frequency residuals/> and low-frequency residual/> , convert the high-frequency residual/> and the low frequency residual/> As the target data set, the wind speed data is a feature data set;

Establish a module for taking the characteristic data set as input and the target data set as output, using a neural network for training, using an optimization method to determine the time step, and establishing wind speed-residual predictions based on working conditions and high and low frequency data. Model;

The input module is used to input SCADA wind speed data of a certain time step into the trained wind speed-residual prediction model to obtain high-frequency residual prediction value data and low-frequency residual prediction value data corresponding to the time step;

The output module is used to linearly add the SCADA wind speed data to the high-frequency residual prediction value data and the low-frequency residual prediction value data to obtain the final true wind speed correction value.

The invention discloses a cabin wind speed correction method and system based on a mechanism model and a neural network. The wind speed data and power data measured in the SCADA system are selected, and the SCADA wind speed is divided into different working condition wind speed data sets according to the rated wind speed. Through experience The theoretical wind speed is calculated using the formula; wavelet transform is used to calculate the high- and low-frequency residuals between the real wind speed and the theoretical wind speed, and a neural network is used to establish the relationship between SCADA wind speed and high- and low-frequency residuals. The SCADA wind speed data from the actual operating wind farm is input into the trained neural network. network to obtain the corresponding high and low frequency residuals; correspond the SCADA wind speed data and the residual data one-to-one, and linearly add them to obtain the true wind speed correction value. The invention has the characteristics of fast calculation speed, high accuracy and good generalization, and can be widely used for wind speed correction of different types of wind turbines.

Description of the drawings

Figure 1 shows a schematic flow chart of the steps of a cabin wind speed correction method based on a mechanism model and a neural network according to the present invention;

Figure 2 shows the architecture diagram of a cabin wind speed correction method based on a mechanism model and neural network according to the present invention;

Figure 3 shows a neural network structure diagram of the present invention;

Figure 4 shows a schematic module diagram of a cabin wind speed correction system based on a mechanism model and a neural network according to the present invention.

Detailed ways

In order to more clearly understand the above objects, features and advantages of the present invention, the present invention will be described in further detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that, as long as there is no conflict, the embodiments of the present application and the features in the embodiments can be combined with each other.

Many specific details are set forth in the following description in order to fully understand the present invention. However, the present invention can also be implemented in other ways different from those described here. Therefore, the protection scope of the present invention is not limited by the specific details disclosed below. Limitations of Examples.

Figures 1 and 2 show a schematic flowchart and architecture diagram of the steps of a cabin wind speed correction method based on a mechanism model and a neural network according to the present invention. A cabin wind speed correction method 100 based on a mechanism model and a neural network includes:

S102: Select the wind speed data and power data measured in the SCADA system, divide the wind speed data and power data according to the preset rules, and obtain a pitch-changing operating condition data set and a non-pitch operating condition data set;

This step is used to establish the original data set, select the wind speed and power data in the SCADA system, and divide the wind speed and power data into the wind speed data set under changing pitch conditions and the wind speed data set under unchanging pitch conditions according to whether the wind speed reaches the rated wind speed. Specifically, of:

When dividing wind speed data into working conditions, you first need to determine the fan model, then determine the corresponding rated wind speed value when the rated power is reached, and then use the rated wind speed as the basis for judgment. Taking a 2MW direct-drive wind turbine in a wind farm as an example, it is determined that the rated power is 2000kw and the rated wind speed is 12m/s.

Optionally, use lidar to measure the true wind speed in front of the turbine hub. ;

Measuring cabin wind speed through SCADA system ,

Further, the wind speed of the fan recorded by the SCADA system Taking the rated wind speed as the boundary, determine whether the wind speed data is less than the rated wind speed data;

If so, it is determined that the wind speed data is the wind speed under non-pitch operating conditions;

If not, it is determined that the wind speed data is the wind speed under the pitching condition.

Optionally, the resolution of the wind speed data and power data in the SCADA system is 1s level; the pitch operating condition data set and the non-pitch operating condition data set are averagely converted into data with a resolution of 30s level. as a feature dataset.

S104: Obtain the air density information and sweep area information, and calculate the air density information, sweep area information, wind speed data and power data according to the preset empirical formula to obtain the Theoretical wind speed ;

This step is used to process the data set and input SCADA wind speed, power, environmental and aircraft model parameters into the empirical formula to obtain the corresponding theoretical wind speed. . Specifically, air density information and sweep area information are obtained, and the air density information, sweep area information, wind speed data and power data are calculated according to a preset empirical formula to obtain the corresponding theoretical wind speed/ > ;

Among them, when calculating the theoretical wind speed When , the empirical formula used is as follows:

;

in is the generator output power,/> is the wind speed recorded by SCADA,/> is the air density,/> is the swept area.

The derivation process of this empirical formula is as follows:

According to aerodynamic theory, when the air mass moves at a speed Flow through area/> When , the power of air movement is expressed as:

(1)

In formula (1), is the air density,/> is the swept area.

Power coefficient Usually defined as:

(2)

In formula (2), is the mechanical power of the wind turbine.

according to Equation (2), the relationship between fan mechanical power and power coefficient can be written as:

(3)

in: ,

is the tip speed ratio,/> is the blade radius,/> is the generator speed.

Considering that the mechanical power of the wind turbine is difficult to measure directly, the actual measured power is the generator power in the SCADA system. The above equation (3) can be rewritten as:

(4)

in, is the energy conversion rate.

According to aerodynamic theory, the airflow passing through the fan blades can be divided into three categories. One category represents the wind speed in front of the blades. ; One type represents the wind speed when passing through the blades/> ; One type represents the wind speed behind the blade/> . Since anemometers are usually installed near wind turbine blades, SCADA wind speed can be considered/> . According to aerodynamic theory, the aerodynamic energy absorbed by fan blades can be expressed as:

(5)

(6)

Substituting formula (6) into the above formula (5) we get:

(7)

Ignoring the difference between the mechanical power of the fan and the power of the generator, and substituting equation (4) into equation (7) we get:

(8)

From formula (4) and formula (8), we can get:

(9)

The empirical formulas for the actual wind speed and SCADA wind speed are finally obtained as follows:

(10)

in is the generator output power,/> is the wind speed recorded by SCADA.

S106: Get the real wind speed , and then use wavelet transform to convert the theoretical wind speed/> and the actual wind speed/> Decompose them separately to obtain the corresponding high-frequency data and low-frequency data, and obtain the corresponding high-frequency residuals/> and low-frequency residual/> , convert the high-frequency residual/> and the low frequency residual/> As the target data set, the wind speed data is a feature data set;

In this step, the true wind speed measured by lidar , and then use wavelet transform to convert the real wind speed/> and theoretical wind speed/> Decompose them separately to obtain the corresponding high-frequency data/> ,/> and low frequency data/> ,/> , then // and/> Subtract the high-frequency residual of the real wind speed high-frequency data and the theoretical wind speed high-frequency data/> , will/> and/> Subtract to obtain the low-frequency residual of the real wind speed low-frequency data and the theoretical wind speed low-frequency data/> .

Specifically, after obtaining the real wind speed data, the real wind speed data is decomposed into high-frequency data and low-frequency data through wavelet transformation, and the corresponding time points are recorded. Then the theoretical wind speed is decomposed into high-frequency data and low-frequency data according to the time point of the real wind speed. Wavelet The transformation formula is shown in formula (11):

(11)

In formula (11), represents the mother wavelet function,/> represent the stretching factor and the translation factor respectively. /> Two conditions must be met:

First: ,/> ;

second: ;

in Yes/> The Fourier transform of . Signal at this time/> The continuous wavelet transform of is calculated by formula (12):

(12)

in, is the wavelet coefficient,/> Yes/> of conjugation.

To describe a time series The wavelet transform discretizes the continuous wavelet transform. Let //> in formula (11) , then equation (11) becomes:

(13)

sequentially The calculation formula of discrete wavelet coefficient is:

(14)

Therefore, the inverse transformation formula for reconstructing the original signal from the wavelet coefficients is:

(15)

S108: Use the feature data set as input and the target data set as output, use neural networks for training, use optimization methods to determine the time step, and establish a wind speed-residual prediction model based on working conditions and high and low frequency data;

Figure 3 shows a neural network structure diagram of the present invention. The neural network used is a long short-term memory recurrent neural network (Self-Attention-LSTM) based on the self-attention mechanism. The neural network has a five-layer structure. The first The layer is the input layer, which inputs the feature data set. The second layer is the LSTM layer. The data of the first layer is calculated through LSTM neurons in batches to obtain the corresponding hidden state as the input of the next layer; the third layer is self-attention. In the force layer, the hidden state of the second layer input is mapped and reorganized using the self-attention mechanism. The dimension after Q, K, and V mapping is 10; the fourth layer is the fully connected layer (Dense); the fifth layer is the output layer. Output the value finally calculated by the model. The weights and bias parameters of this neural network are the relationship model between wind speed and residual. If the wind speed is input into this neural network, the residual value can be obtained.

Optionally, in one embodiment, the feature data set is used as input and the target data set is used as output, a neural network is used for training, and an optimization method is used to determine the time step, based on working conditions and high and low frequency data. , establishing a wind speed-residual prediction model includes:

Use neural networks for training to establish wind speed-residual prediction models under two working conditions, a total of four: one is a high-frequency residual prediction model in which SCADA wind speed is used as input and high-frequency residuals are output under pitch working conditions; The second is a low-frequency residual prediction model in which SCADA wind speed is used as input under pitch changing conditions and low-frequency residuals are used as output. The third is a high-frequency residual prediction model in which SCADA wind speed is used as input under non-pitching conditions and high-frequency residuals are used as output. ; The fourth is a low-frequency residual prediction model using SCADA wind speed as input and low-frequency residual as output under non-pitch operating conditions.

In the process of establishing the wind speed-residual prediction model, gamma detection technology is used to optimize the training time step and prediction time step, and the layer-by-layer random jumping method is used to optimize and combine the parameters of each layer.

Specifically, for different wind turbines, the time step of the input and output is determined through the gamma detection algorithm, and the parameters of each layer are determined using the layer-by-layer random jumping method, but the overall structure of the neural network will not change.

Gamma detection directly estimates the minimum mean square error present in the nonlinear model through observation data, and in this way the parameters of the input data can be determined.

Suppose the form of sample data is ,/> Indicates the first/> Line input, each/> May contain multiple features, use/> Indicates which feature,/> Indicates the first/> output. Assuming that the vector contains useful factors that affect the output, the relationship between and can be expressed as:

(16)

is a smooth unknown regression function,/> represents a noisy random variable. Since any constant deviation can be contained in the unknown function, the expectation/> , variance/> .

If two points and/> (/> ) are very close in the input space, then their corresponding output/> and/> The distance in the output space should also be very close, if not, then we think this difference is caused by noise/> Caused. /> The function calculates the mean of the squared value of the proximity distance of each point in the input space, as shown in Equation (17):

(17)

in, for/> of/> Close to the point, i.e./> is closest/> of/> points,/> for/> of/> Close to the point, i.e./> is closest/> of/> points,/> Represents the Euclidean distance, in the output space, the corresponding /> The function is as follows:

(18)

Gamma detection calculates gamma statistics by drawing a linear regression line on the points ,Calculated as follows:

(19)

In equation (19), the vertical intercept (when is zero) described by/> The value of the gamma statistic represented. if/> is very small, then the regression function/> Exists, output value/> Depends a lot on the input variables/> , there is a close correlation between input and output. if/> Very large, indicating that the input has nothing to do with the output.

Since the value of the gamma statistic is affected by the numerical size of the sample, consider the statistic Standardize the results. Statistics/> defined as:

(20)

In formula (20), Indicates output/> Variance. Close to 0/> The value indicates the output/> Highly predictable, close to 1/> The value indicates the output/> is random, with the input variable/> Nothing to do.

Relevant instructions for selecting embedding dimensions for gamma detection: The value of the corresponding gamma statistic when M is chosen for the embedding dimension. The smaller the gamma statistic, the more reasonable the corresponding embedding dimension, that is to say, the statistic/> The closer the value is to 0, the more reasonable the corresponding embedding dimension is. Therefore, gamma detection can be performed separately according to the increasing embedding dimension, using statistics/> Make judgments and choose appropriate embedding dimensions.

In the process of determining the parameters of each layer using the layer-by-layer random jumping method, the layer-by-layer random jumping method includes the following steps:

1) Determine the parameters, jump range and jump step size of random jumps:

Determine the parameters that need to be optimized as the second layer LSTM unit number, dropout value, and activation function; the fourth layer fully connected layer unit number and activation function; set the range and jump step size of each parameter; the activation function is the sigmod function, tanh The function and relu function are randomly selected, and the parameter value calculation formula is as follows:

(twenty one)

In the formula, Calculates the value for the current round,/> Calculated value for the previous round,/> is the minimum value of the range,/> is the maximum value of the range,/> is the jump step size,/> is a function that randomly selects integers within a given range,/> ;

2) In the first round, the number of units, dropout value, mapping dimension and activation function of the selected layer are randomly initialized within a given range, and then a certain amount of wind speed and residual data sets are divided into training sets and test sets for training and testing. And calculate the accuracy, get the accuracy value of the corresponding parameter, record the parameter combination and result; then add a random positive or negative integer to the number of units and dropout value of the selected layer multiplied by the jump step size to get the next set of parameter values. , where the parameter values cannot exceed the given range, and the activation function is also randomly selected and executed repeatedly for five rounds;

3) Calculate the accuracy of five groups of random parameters, select the parameter combination with the highest accuracy, and then execute the parameters of the fourth layer according to step 2). Also calculate the accuracy of the five groups of parameters and select the parameter combination with the highest accuracy, and finally get the optimal Parameter combination.

S110: Input SCADA wind speed data of a certain time step into the trained wind speed-residual prediction model, and obtain high-frequency residual prediction value data and low-frequency residual prediction value data corresponding to the time step;

According to the input time step and resolution during training, the SCADA second-level wind speed is extracted every 10 to 20 times the resolution times the time step. The SCADA wind speed is divided into the pitch speed and the non-pitch wind speed according to whether it reaches the rated wind speed, and then the average is calculated according to the set resolution, and the reduced pitch wind speed is input into the training system. The high-frequency and low-frequency residual prediction model of the pitch operating condition is used to obtain the high-frequency residual prediction value and the low-frequency residual prediction value of the wind speed under the pitch operating condition; the wind speed of the unvaried pitch condition after the resolution is reduced is input to the trained unvaried wind speed. The high-frequency and low-frequency residual prediction model of the propeller operating condition is used to obtain the high-frequency residual prediction value and the low-frequency residual prediction value of the wind speed under the unchanging propeller operating condition.

S112: Linearly add the SCADA wind speed data, high-frequency residual prediction value data and low-frequency residual prediction value data to obtain the final true wind speed correction value.

The high-frequency residual prediction values of wind speed under pitching conditions are obtained respectively. and low-frequency residual prediction value/> , and the high-frequency residual prediction value and low-frequency residual prediction value of the wind speed in the un-pitch operating condition, and then the SCADA wind speed in the pitch operating condition, the high-frequency residual prediction value of the wind speed in the pitch operating condition and the low-frequency residual prediction value of the wind speed in the pitch operating condition. The difference prediction values are linearly added to obtain the final true pitch wind speed correction value; the SCADA un-pitch operating condition wind speed, the un-pitch operating condition wind speed high-frequency residual prediction value and the un-pitch operating condition wind speed low-frequency residual prediction value are linearly added Add together to get the final true unpitched wind speed correction value.

That is, the cabin wind speed Add the high-frequency residuals of the corresponding time steps/> and low-frequency residual/> Get corrected true wind speed/> ,Right now

. (twenty two)

According to this embodiment, the invention discloses a cabin wind speed correction method based on a mechanism model and a neural network. The wind speed data and power data measured in the SCADA system are selected, and the SCADA wind speed is divided into different working condition wind speed data sets according to the rated wind speed. , the theoretical wind speed is calculated through empirical formulas; the wavelet transform is used to calculate the high and low frequency residuals of the real wind speed and the theoretical wind speed, the neural network is used to establish the relationship between SCADA wind speed and the high and low frequency residuals, and the SCADA wind speed data from the actual operating wind farm is input. The trained neural network obtains the corresponding high and low frequency residuals; the SCADA wind speed data and the residual data are corresponded one-to-one, and the true wind speed correction value is obtained by linear addition. The invention has the characteristics of fast calculation speed, high accuracy and good generalization, and can be widely used for wind speed correction of different types of wind turbines.

Figure 4 shows a schematic module diagram of a cabin wind speed correction system based on a mechanism model and a neural network according to the present invention. A cabin wind speed correction system 200 based on a mechanism model and a neural network is used to implement the aforementioned cabin wind speed correction method based on a mechanism model and a neural network, including:

The dividing module 210 is used to select the wind speed data and power data measured in the SCADA system, divide the wind speed data and power data according to the preset rules, and obtain a pitch-changing operating condition data set and a non-pitch operating condition data set;

The calculation module 220 is used to obtain air density information and sweep area information, and calculate the air density information, sweep area information, wind speed data and power data according to preset empirical formulas to obtain the corresponding Theoretical wind speed under working conditions ;

Decomposition module 230, used to obtain the true wind speed , and then use wavelet transform to convert the theoretical wind speed/> and the actual wind speed/> Decompose them separately to obtain the corresponding high-frequency data and low-frequency data, and obtain the corresponding high-frequency residuals/> and low-frequency residual/> , convert the high-frequency residual/> and the low frequency residual/> As the target data set, the wind speed data is a feature data set;

The establishment module 240 is used to take the characteristic data set as input and the target data set as output, use neural network for training, use optimization method to determine time step, and establish wind speed-residual error according to working conditions and high and low frequency data. Predictive models;

The input module 250 is used to input SCADA wind speed data of a certain time step into the trained wind speed-residual prediction model to obtain high-frequency residual prediction value data and low-frequency residual prediction value data corresponding to the time step;

The output module 260 is used to linearly add the SCADA wind speed data to the high-frequency residual prediction value data and the low-frequency residual prediction value data to obtain the final true wind speed correction value.

Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working process of the electronic device of the present invention can be referred to the corresponding processes in the foregoing embodiments, as well as the same parts and beneficial effects as the foregoing embodiments. , which will not be described in detail here.

According to this embodiment, the invention discloses a cabin wind speed correction system based on a mechanism model and a neural network. The wind speed data and power data measured in the SCADA system are selected, and the SCADA wind speed is divided into different working condition wind speed data sets according to the rated wind speed. , the theoretical wind speed is calculated through empirical formulas; the wavelet transform is used to calculate the high and low frequency residuals of the real wind speed and the theoretical wind speed, the neural network is used to establish the relationship between SCADA wind speed and the high and low frequency residuals, and the SCADA wind speed data from the actual operating wind farm is input. The trained neural network obtains the corresponding high and low frequency residuals; the SCADA wind speed data and the residual data are corresponded one-to-one, and the true wind speed correction value is obtained by linear addition. The invention has the characteristics of fast calculation speed, high accuracy and good generalization, and can be widely used for wind speed correction of different types of wind turbines.

In the several embodiments provided in this application, it should be understood that the disclosed methods and systems can be implemented in other ways. The device embodiments described above are only illustrative. For example, the division of modules is only a logical function division. In actual implementation, there may be other division methods, such as: multiple modules or components can be combined, or can be integrated into another system, or some features can be ignored, or not implemented. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed may be through some interfaces, and the indirect coupling or communication connection of the devices or units may be electrical, mechanical, or other forms. of.

The modules described above as separate components may or may not be physically separated, and the components shown as modules may or may not be physical units; they may be located in one place or distributed to multiple network units; Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

In addition, each functional unit in each embodiment of the present invention can be all integrated into one processing unit, or each unit can be separately used as a unit, or two or more units can be integrated into one unit; the above-mentioned integration The unit can be implemented in the form of hardware or in the form of hardware plus software functional units.

Those of ordinary skill in the art can understand that all or part of the steps to implement the above method embodiments can be completed through hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, the execution includes: The steps of the above method embodiment; and the aforementioned storage media include: mobile storage devices, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disks or optical disks, etc. The medium on which program code is stored.

Alternatively, if the above-mentioned integrated unit of the present invention is implemented in the form of a software function module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of the present invention can be embodied in the form of software products in essence or those that contribute to the existing technology. The computer software products are stored in a storage medium and include a number of instructions to A computer device (which may be a personal computer, a server, a network device, etc.) is caused to execute all or part of the methods described in various embodiments of the present invention. The aforementioned storage media include: mobile storage devices, ROM, RAM, magnetic disks or optical disks and other media that can store program codes.

Although some specific embodiments of the invention have been described in detail by way of examples, those skilled in the art will understand that the above examples are for illustration only and are not intended to limit the scope of the invention. It will also be understood by those skilled in the art that various modifications may be made to the embodiments without departing from the scope and spirit of the invention. The scope of the invention is defined by the appended claims.

Claims (7)

1. The cabin wind speed correction method based on the mechanism model and the neural network is characterized by comprising the following steps of:

selecting wind speed data and power data measured in an SCADA system, and dividing the wind speed data and the power data according to a preset rule to obtain a variable-pitch working condition data set and an unchanged-pitch working condition data set;

wherein the dividing according to the preset rule comprises: judging whether the wind speed data is smaller than rated wind speed data or not, if yes, judging that the wind speed data is the wind speed under the condition of no pitch change; if not, judging that the wind speed data is the wind speed under the variable pitch working condition;

the resolution ratio of wind speed data and power data in the SCADA system is 1s level;

the variable-pitch working condition data set and the non-variable-pitch working condition data set are averagely converted into data with the resolution of 30s level to be used as a characteristic data set;

acquiring air density information and wind sweeping area information, and calculating the air density information, the wind sweeping area information, the wind speed data and the power data according to a preset empirical formula to obtain a theoretical wind speed under corresponding working conditions；

Wherein, the preset empirical formula is:

；

wherein the method comprises the steps ofIs power data, +.>Is wind speed data>Is air density information, ++>Is wind sweeping area information;

obtaining the true wind speedThe theoretical wind speed is then converted by wavelet>And the true wind speed +.>Respectively decomposing to obtain corresponding high-frequency data and low-frequency data, and obtaining corresponding high-frequency residual ∈>And low frequency residual->The high frequency residual error is->And said low frequency residual->As a target dataset, the wind speed data is a feature dataset;

wherein, the real wind speed is measured by a laser radarThen the real wind speed is converted by wavelet>And theoretical wind speed->Respectively decomposing to obtain corresponding high frequency data ∈>、/>And low frequency data->、/>Then ∈>And->Subtracting to obtain high-frequency residual error of real wind speed high-frequency data and theoretical wind speed high-frequency data>Will->And->Subtracting to obtain a low-frequency residual error of the real wind speed low-frequency data and the theoretical wind speed low-frequency data>Specifically, real wind speed data are obtained, then the real wind speed data are decomposed into high-frequency data and low-frequency data through wavelet transformation, corresponding time points are recorded, and then theoretical wind speed is decomposed into the high-frequency data and the low-frequency data according to the time points of the real wind speed;

taking the characteristic data set as input, taking the target data set as output, training by using a neural network, determining a time step by using an optimization method, and establishing a wind speed-residual prediction model according to working conditions and high-low frequency data;

inputting SCADA wind speed data of a certain time step into a trained wind speed-residual error prediction model to obtain high-frequency residual error prediction value data and low-frequency residual error prediction value data of a corresponding time step;

and linearly adding the SCADA wind speed data with the high-frequency residual error predicted value data and the low-frequency residual error predicted value data to obtain a final real wind speed correction value.

2. The nacelle wind speed correction method based on a mechanism model and a neural network of claim 1, wherein,

real wind speed in front of fan hub is measured through laser radar；

Measuring nacelle wind speed by SCADA system。

3. A nacelle wind speed correction method based on a mechanism model and a neural network as recited in claim 1, wherein,

the characteristic data set is taken as input, the target data set is taken as output, the neural network is used for training, the time step length is determined by an optimization method, and the wind speed-residual prediction model is built according to working conditions and high-low frequency data and comprises the following steps:

training by using a neural network, and establishing wind speed-residual error prediction models under two working conditions, wherein the total number of the wind speed-residual error prediction models is four: the method comprises the steps that firstly, SCADA wind speed is used as input under a variable pitch working condition, and high-frequency residual errors are used as output high-frequency residual error prediction models; second, SCADA wind speed is used as input under the variable pitch working condition, and low-frequency residual is used as output low-frequency residual prediction model; thirdly, under the condition of no pitch change, the SCADA wind speed is taken as input, and the high-frequency residual is taken as an output high-frequency residual prediction model; and fourthly, taking the SCADA wind speed under the condition of no pitch variation as an input, and taking the low-frequency residual as an output low-frequency residual prediction model.

4. The cabin wind speed correction method based on the mechanism model and the neural network according to claim 1, wherein in the process of establishing a wind speed-residual error prediction model, a gamma detection technology is used for carrying out optimization solution on a training time step and a prediction time step, and a layer-by-layer random jump method is used for carrying out optimization combined solution on parameters of each layer.

5. The nacelle wind speed correction method based on a mechanism model and a neural network according to claim 4, wherein the layer-by-layer random jump method comprises the steps of:

1) Determining parameters of random jitter, jitter range and jitter step length:

determining parameters to be optimized as the number of LSTM units of the second layer, a dropout value and an activation function; the fourth layer of full connection layer unit number and activation function; setting the range and the jumping step length of each parameter; the activation function is selected randomly for sigmod function, tanh function and relu function, and the parameter value calculation formula is as follows:

；

calculating a value for the current wheel,/->Calculating a value for the previous round,/->Is the minimum value of the range>For the maximum value of the range>For jumping step length +.>For randomly selecting a function of integers within a given range, < >>；

2) In the first round, randomly initializing the unit number, dropout value, mapping dimension and activation function of the selected layer in a given range, dividing a certain amount of wind speed and residual error data set into a training set and a testing set for training test and calculating precision, obtaining the precision value of the corresponding parameter, and recording the parameter combination and result; then adding a random positive integer or a negative integer to the unit number and the dropout value of the selected layer to multiply the jumping step length to obtain a next group of parameter values, wherein the parameter values cannot exceed a given range, the activation function is also randomly selected, and five rounds of repeated execution are performed;

3) Calculating the precision of five groups of random parameters, selecting the parameter combination with the highest precision, then executing the parameters of the fourth layer according to the step 2), likewise calculating the precision of the five groups of parameters, then selecting the parameter combination with the highest precision, and finally obtaining the optimal parameter combination.

6. A nacelle wind speed correction method based on a mechanism model and a neural network according to claim 3,

acquiring SCADA wind speed data, judging working conditions according to the SCADA wind speed data, and inputting the SCADA wind speed data into a corresponding high-frequency residual error prediction model and a corresponding low-frequency residual error prediction model to obtain a high-frequency residual error of a corresponding time stepAnd low frequency residual->Finally, cabin wind speed->Adding the high frequency residual of said corresponding time step +.>And low frequency residual->Obtaining corrected true wind speed +.>I.e.

。

7. Cabin wind speed correction system based on a mechanism model and a neural network for implementing a cabin wind speed correction method based on a mechanism model and a neural network as claimed in any one of claims 1-6, comprising:

the division module is used for selecting wind speed data and power data measured in the SCADA system, dividing the wind speed data and the power data according to a preset rule, and obtaining a variable-pitch working condition data set and an unchangeable-pitch working condition data set;

the calculation module is used for acquiring air density information and wind sweeping area information to be obtainedThe air density information, the wind sweeping area information, the wind speed data and the power data are calculated according to a preset empirical formula to obtain a theoretical wind speed under corresponding working conditions；

The decomposition module is used for obtaining the real wind speedThe theoretical wind speed is then converted by wavelet>And the true wind speed +.>Respectively decomposing to obtain corresponding high-frequency data and low-frequency data, and obtaining corresponding high-frequency residual ∈>And low frequency residualThe high frequency residual error is->And said low frequency residual->As a target dataset, the wind speed data is a feature dataset;

the building module is used for taking the characteristic data set as input, taking the target data set as output, training by using a neural network, determining a time step by using an optimization method, and building a wind speed-residual prediction model according to working conditions and high-low frequency data;

the input module is used for inputting SCADA wind speed data with a certain time step into the trained wind speed-residual error prediction model to obtain high-frequency residual error prediction value data and low-frequency residual error prediction value data corresponding to the time step;

and the output module is used for carrying out linear addition on the SCADA wind speed data, the high-frequency residual error predicted value data and the low-frequency residual error predicted value data to obtain a final real wind speed correction value.