CN110765703B - A modeling method for wind farm aggregation characteristics - Google Patents

Abstract

The invention discloses a wind power plant aggregation characteristic modeling method, which comprises the following steps: the method comprises the following steps: collecting actual measurement operation data of a plurality of wind turbine generators of a wind power plant, and cleaning and normalizing the data; step two: establishing a wind power fluctuation measurement index; step three: establishing a smoothing effect measurement index of the aggregate output of the wind power plant, and obtaining a relational expression of the smoothing effect measurement index, the number N of the polymer units and the correlation coefficient of the power sequence among the units; step four: establishing a mapping model of the correlation between the multi-position point wind condition information and the power sequence among the units based on a convolutional neural network; step five: forming a model training sample, training a neural network model by using a root mean square error function index, and outputting a unit output correlation mapping result; step six: and realizing the wind power plant aggregation characteristic modeling based on the convolutional neural network.

Description

A modeling method for wind farm aggregation characteristics

technical field

The invention relates to the technical field of new energy power generation, in particular to a method for modeling aggregation characteristics of wind farms.

Background technique

The integration of large-scale wind power into the grid is one of the prominent features of the future development of the power system. As the proportion of wind power in the power supply of various countries in the world increases year by year, grid-connected wind power presents the characteristics of cluster development, weak grid access, and long-distance transmission. The inherent randomness and volatility of wind power have brought power balance to the power system. , reverse peak regulation, voltage stability, frequency stability and a series of problems, increasing the difficulty of power grid planning and scheduling. "Smoothing effect" is one of the common characteristics of large-scale wind power, which means that the aggregate output of wind power is smoother and less volatile than the individual output. Some studies have proposed a method of geographical diversification to smooth the fluctuation of wind power output. Therefore, it is especially important to understand and master the characteristics of wind power aggregation. It is the fluctuation law of wind power aggregate output and modeling it is the key basis to solve the above problems.

The difference in the distribution of wind resources is the key factor for the smoothing effect. Essentially, due to the spatial dispersion of wind turbines in the wind farm, the wind resources at different wind turbine locations in the wind farm show differences. The output of the unit changes with the wind speed and then through the synergistic decoupling effect, in some cases, the fluctuation of the other party is offset, thereby reducing the fluctuation of the wind power as a whole. Therefore, the volatility and smoothing effect of the aggregated power of the wind farm are closely related to the regional scale of the wind farm, the number of wind turbines in the wind farm, the wind condition of the wind farm, and the output correlation between the wind turbines in the wind farm, and these factors affect each other. Most of the existing researches on wind turbines at home and abroad describe and qualitatively analyze the law of wind power aggregation volatility and smoothing effect, and the following conclusions are mainly drawn: 1. The expansion of wind power scale (regional scale, number of units) can reduce wind power aggregation. Volatility, enhanced smoothing effect. 2. The stronger the output correlation between wind turbines, the worse the complementarity and the worse the smoothing effect, and vice versa. The few existing quantitative studies on the volatility of wind power aggregation only start from the observed historical data, and perform simple linear or nonlinear fitting of the functional relationship between it and various influencing factors. However, the power aggregation characteristic of a wind farm is a complex spatiotemporal coupling process affected by multiple variables such as the topography of the wind farm, the arrangement of units, and the wind conditions at multiple units. Linear or nonlinear fitting accurately describes the relationship between each influencing factor and the power aggregation characteristics of the wind farm.

Therefore, it is hoped that a modeling method for wind farm aggregation characteristics can solve the problems existing in the prior art.

SUMMARY OF THE INVENTION

The invention discloses a modeling method for wind farm aggregation characteristics, and the modeling method includes the following steps:

Step 1: Collect the measured operation data at the positions of multiple wind turbines in the wind farm, and clean and normalize the above data;

Step 2: Establish wind power volatility metrics;

Step 3: Establish the smoothing effect measurement index of the aggregated output of the wind farm, and obtain the relationship between the smoothing effect measurement index and the number of aggregated units N and the correlation coefficient of the power series between units by the analysis method of mathematical statistics;

Step 4: Establish a mapping model based on the convolutional neural network for the multi-location point wind condition information and the power sequence correlation between units, and model the number of units on the convolutional neural network according to the number of moments in the specified time period and the aggregation characteristics of the wind farm. Set the corresponding parameters;

Step 5: Use the measured wind speed series data and the measured wind direction series data of multiple units at the specified time scale as the model input, and use the output correlation coefficient between the two units as the output to form a model training sample, and use the root mean square error function. The index trains the neural network model, and outputs the unit output correlation mapping result;

Step 6: According to the unit output correlation mapping result in step 5 and the functional relationship between the smoothing effect measurement index in step 3 and the number of aggregated units N and the correlation coefficient of power series between units, calculate the smoothing effect measurement index s that characterizes the aggregation characteristics of the wind farm, Realize the modeling of wind farm aggregation characteristics based on convolutional neural network.

Preferably, the measured operation data in the first step includes: measured wind speed data, measured wind direction data, and measured power data.

Preferably, the aggregate output of the wind turbine cluster is the sum of the output of each single unit, as shown in formula (1):

where P _∑ (t) is the total power of the wind turbine cluster at time t; i is the number of wind turbines; N is the number of wind turbines; P _i (t) is the power of the ith wind turbine at time t;

In the second step, formulas (2) and (3) are used to express the standard deviation of the power sequence to measure the fluctuation of the output of a single wind turbine and the aggregate output of the wind turbine:

分别为第i号风电机组功率、聚合功率在相应时间统计尺度下采样值的均值；功率序列标准差反映了时间序列功率波动的程度，数值越小时间序列功率波动越小，平稳性越好。where σ _i and σ _∑ are the standard deviation of power and the standard deviation of aggregated power of the ith wind turbine, respectively; T is the statistical time scale;

are the mean values of the sampled values of the i-th wind turbine power and aggregate power under the corresponding time statistical scale; the standard deviation of the power series reflects the degree of power fluctuation of the time series. The smaller the value, the smaller the time series power fluctuation and the better the stability.

Preferably, the smoothing effect coefficient is the ratio of the standard deviation per unit value of the aggregate output of the wind turbines based on the installed capacity to the standard deviation per unit value of the output of a single unit, which is formula (4):

Among them, P _R is the rated power of the wind turbine, σ _i and σ _∑ are the standard deviation of power and the standard deviation of the aggregated power of the ith wind turbine, respectively, and N represents the number of aggregated units. The smaller the coefficient s, the smaller the output of each wind turbine. The stronger the complementarity between the two, the more obvious the smoothing effect;

Through the mathematical statistics in the third step, the relational formula (5) of the standard deviation of the aggregate power of the wind turbine and the standard deviation of the single-unit power is obtained:

In the formula, ri _,j is the correlation coefficient of the power sequence of the ith wind turbine and the jth wind turbine;

Assuming that the output standard deviation of each unit is the same, it is regarded as the average output standard deviation of a single unit, then the smoothing effect measurement index and the aggregated unit number N and the power sequence between units can be obtained from formulas (4) and (5). The correlation coefficient relationship of is formula (6):

Preferably, the convolutional neural network in the step 4 includes two convolution layers, two pooling layers, one flattening layer and one fully connected layer; the measured wind speed sequence data and the measured wind direction sequence data of multiple units pass through the first The convolutional layer and the ReLU activation layer generate a set of feature maps, which are then downsampled by non-overlapping maximum pooling, and then pass through the second convolutional layer and the ReLU activation layer to generate another set of feature maps, which are combined with the feature map. The flattening layer is connected to make the multi-dimensional tensor one-dimensional, and then connected to the fully connected layer, and finally activated by the sigmoid function as the output of the convolutional neural network.

Preferably, in the step 4, a batch normalization method is used to add a normalization layer after the output of some layers, and the output is normalized to the same distribution according to the characteristic values of the same batch, and the convolution operation process is represented by formula (7):

表示第L个卷积层中的第j个特征图，即该卷积层提取的风速分布特征，

是上一层的特征图集合，它同时也是这一层的输入，

为第L个卷积层中的第j个卷积核，*定义为卷积运算，b表示加性偏置值，f(x)表示激活函数，它将一个实数域上的值映射到一个有限域中；in,

Represents the jth feature map in the Lth convolutional layer, that is, the wind speed distribution feature extracted by this convolutional layer,

is the feature map set of the previous layer, which is also the input of this layer,

is the jth convolution kernel in the Lth convolutional layer, * is defined as the convolution operation, b represents the additive bias value, and f(x) represents the activation function, which maps a value in the real number domain to a in a finite field;

The operation process of the pooling layer is represented by formula (8):

表示第L个池化层中的第j个特征图，f(x)表示激活函数，

表示特征图的乘性偏置，down(x)为下采样函数，

表示特征图的加性偏置。in,

represents the jth feature map in the Lth pooling layer, f(x) represents the activation function,

Represents the multiplicative bias of the feature map, down(x) is the downsampling function,

Represents the additive bias of the feature map.

Preferably, the calculation formula of the output correlation coefficient between the units in the step 5 is formula (9):

分别为序列X、Y的均值，X_i、Y_i分别为序列X、Y第i个数的值。where σ _xy refers to the covariance of the two data series X and Y, σ _x and σ _y are the respective variances of X and Y, respectively,

are the mean values of the sequences X and Y respectively, and X _i and Y _i are the values of the i-th number of the sequences X and Y respectively.

The present invention proposes a method for modeling the aggregation characteristics of wind farms. The method for modeling the output aggregation characteristics of wind farms based on convolutional neural networks is derived from the deep learning method and is obtained by combining the existing analysis methods through mathematical statistics. Compared with traditional simple linear or nonlinear fitting methods, the power aggregation characteristics of wind farms are affected by the topography, unit arrangement and other factors of wind farms. The complex spatiotemporal coupling process affected by multi-variables such as wind conditions at multiple unit points, the specific mapping relationship is a complex nonlinear mapping, the present invention uses the convolutional neural network deep learning algorithm to simulate the flow field scene of the wind farm, and establishes the output of the wind farm The mapping model between the influencing factors of the aggregation characteristics and the aggregation characteristics of wind farms can more accurately describe the relationship between the influencing factors and the power aggregation characteristics of wind farms, so that it can more effectively cooperate with the planning and dispatching of the power system.

Description of drawings

Figure 1 is the flow chart of the modeling method of wind farm aggregation characteristics.

Figure 2 is a schematic structural diagram of a wind turbine output correlation mapping model based on a convolutional neural network.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions in the embodiments of the present invention will be described in more detail below with reference to the accompanying drawings in the embodiments of the present invention. Throughout the drawings, the same or similar reference numbers refer to the same or similar elements or elements having the same or similar functions. The described embodiments are some, but not all, of the embodiments of the present invention. The embodiments described below with reference to the accompanying drawings are exemplary, and are intended to explain the present invention and should not be construed as limiting the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

Taking the measured wind speed data, measured wind direction data, and measured power data of five wind turbines in a wind farm in northern China as a test example, the example collected the measured wind speed data and measured wind direction data from November 2016 to November 2017. , Measured power data, the data time resolution is 10 minutes. The first 70% data set of each month is selected as the training sample, and the second 30% data set of each month is selected as the test sample, and the aggregation characteristics of wind power with a time scale of 4 hours are modeled.

As shown in Figure 1, the modeling method of wind farm aggregation characteristics includes the following steps:

Step 1: Collect the measured operation data at the locations of multiple wind turbines in the wind farm, and clean and preprocess the above data; the measured operation data includes the measured wind speed data, the measured wind direction data, and the measured power data, and the preprocessing includes data cleaning and normalization. unified treatment;

Step 2: Establish wind power volatility metrics. The aggregate output of the wind turbine cluster is the sum of the output of each single unit, such as formula (1):

where P _∑ (t) is the total power of the wind turbine cluster at time t; i is the number of wind turbines; N is the number of wind turbines; P _i (t) is the power of the ith wind turbine at time t;

Formulas (2) and (3) are used to express the standard deviation of the power series to measure the volatility of the output of a single wind turbine and the aggregate output of wind turbines:

分别为第i号风电机组功率、聚合功率在相应时间统计尺度下采样值的均值；功率序列标准差反映了时间序列功率波动的程度，数值越小时间序列功率波动越小，平稳性越好。where σ _i and σ _∑ are the standard deviation of power and the standard deviation of aggregated power of the ith wind turbine, respectively; T is the statistical time scale;

are the mean values of the sampled values of the i-th wind turbine power and aggregate power under the corresponding time statistical scale; the standard deviation of the power series reflects the degree of power fluctuation of the time series. The smaller the value, the smaller the time series power fluctuation and the better the stability.

Step 3: Establish the smoothing effect measurement index of the aggregate output of the wind farm, and define the smoothing effect coefficient as the ratio of the standard deviation per unit value of the wind turbine aggregate output based on the installed capacity to the standard deviation per unit value of the output of a single unit, that is, the formula ( 4):

Among them, P _R is the rated power of the wind turbine, σ _i and σ _∑ are the standard deviation of power and the standard deviation of the aggregated power of the ith wind turbine, respectively, and N represents the number of aggregated units. The smaller the coefficient s, the smaller the output of each wind turbine. The stronger the complementarity between the two, the more obvious the smoothing effect;

Through the mathematical statistics in the third step, the relational formula (5) of the standard deviation of the aggregate power of the wind turbine and the standard deviation of the single-unit power is obtained:

In the formula, ri _,j is the correlation coefficient of the power sequence of the ith wind turbine and the jth wind turbine;

Assuming that the output standard deviation of each unit is the same, it is regarded as the average output standard deviation of a single unit, then the smoothing effect measurement index and the aggregated unit number N and the power sequence between units can be obtained from formulas (4) and (5). The correlation coefficient relationship of is formula (6):

Step 4: Establish a mapping model based on the convolutional neural network based on the multi-location point wind condition information and the power sequence correlation between the units. The convolutional neural network includes two convolutional layers, two pooling layers, and one expansion layer. A flat layer, and a fully connected layer, the structure of which is shown in Figure 2. The measured wind speed sequence data and the measured wind direction sequence data of the 5 units pass through the first convolution layer and the ReLU activation layer to generate a set of feature maps, which are then down-sampled through non-overlapping max pooling, and then passed through the second convolution layer. And the ReLU activation layer to generate another set of feature maps, connect this set of feature maps to the flattening layer to make the multi-dimensional tensor one-dimensional, and then connect to the fully connected layer, and finally activate the sigmoid function as the output of the convolutional neural network, In view of the uneven distribution of each layer and the "gradient dispersion" problem, this scheme adopts the batch normalization method (Batch Normalization-BN) to reduce this effect, that is, a normalization layer is added after the output of some layers, and the output is based on the same The eigenvalues of the batches are normalized to the same distribution;

The convolution operation process is represented by formula (7):

表示第L个卷积层中的第j个特征图，即该卷积层提取的风速分布特征，

是上一层的特征图集合，它同时也是这一层的输入，

为第L个卷积层中的第j个卷积核，*定义为卷积运算，b表示加性偏置值，f(x)表示激活函数，它将一个实数域上的值映射到一个有限域中；in,

Represents the jth feature map in the Lth convolutional layer, that is, the wind speed distribution feature extracted by this convolutional layer,

is the feature map set of the previous layer, which is also the input of this layer,

is the jth convolution kernel in the Lth convolutional layer, * is defined as the convolution operation, b represents the additive bias value, and f(x) represents the activation function, which maps a value in the real number domain to a in a finite field;

The operation process of the pooling layer is represented by formula (8):

表示第L个池化层中的第j个特征图，f(x)表示激活函数，

表示特征图的乘性偏置，down(x)为下采样函数，

表示特征图的加性偏置。in,

represents the jth feature map in the Lth pooling layer, f(x) represents the activation function,

Represents the multiplicative bias of the feature map, down(x) is the downsampling function,

Represents the additive bias of the feature map.

The network structure parameters of the output correlation mapping model based on convolutional neural network are shown in Table 1:

Table 1 Network structure parameters of output correlation mapping model based on convolutional neural network

Step 5: Use the measured wind speed sequence data and the measured wind direction sequence data of multiple units at a specified time scale (such as 4 hours) as the model input, and use the output correlation coefficient between the units as the output to form a model training sample. The root mean square error function index trains the neural network model and outputs the mapping result. The formula for calculating the output correlation coefficient between two units is formula (9):

分别为序列X、Y的均值，X_i、Y_i分别为序列X、Y第i个数的值。where σ _xy refers to the covariance of the two data series X and Y, σ _x and σ _y are the respective variances of X and Y, respectively,

are the mean values of the sequences X and Y respectively, and X _i and Y _i are the values of the i-th number of the sequences X and Y respectively.

After training and prediction, the error between the model mapping result and the actual unit output correlation coefficient is shown in Table 2.

Table 2 The error of the model mapping result and the actual unit output correlation coefficient

Step 6: According to the unit output correlation mapping result in step 5 and formula (6), calculate the smoothing effect measurement index s representing the aggregation characteristics of the wind farm, and realize the modeling of the aggregation characteristics of the wind farm based on the convolutional neural network. In this embodiment The mean absolute error and root mean square error of the smoothing effect system s calculated from the mapping results and the real smoothing effect coefficient are 0.1386 and 0.1781, respectively.

Finally, it should be pointed out that the above embodiments are only used to illustrate the technical solutions of the present invention, but not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: it is still possible to modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements to some of the technical features; and these Modifications or substitutions do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims (7)

1. A method for modeling wind farm aggregation characteristics, characterized in that the modeling method comprises the following steps: Step 1: Collect the measured operation data at the positions of multiple wind turbines in the wind farm, and clean and normalize the above data; Step 2: Establish wind power volatility metrics; Step 3: Establish the smoothing effect measurement index of the aggregated output of the wind farm, and obtain the relationship between the smoothing effect measurement index and the number of aggregated units N and the correlation coefficient of the power series between units by the analysis method of mathematical statistics; Step 4: Establish a mapping model based on the convolutional neural network for the multi-location point wind condition information and the power sequence correlation between units, and model the number of units on the convolutional neural network according to the number of moments in the specified time period and the aggregation characteristics of the wind farm. Set the corresponding parameters; Step 5: Use the measured wind speed series data and the measured wind direction series data of multiple units at the specified time scale as the model input, and use the output correlation coefficient between the two units as the output to form a model training sample, and use the root mean square error function. The index trains the neural network model, and outputs the unit output correlation mapping result; Step 6: According to the unit output correlation mapping result in step 5 and the functional relationship between the smoothing effect measurement index in step 3 and the number of aggregated units N and the correlation coefficient of power series between units, calculate the smoothing effect measurement index s that characterizes the aggregation characteristics of the wind farm, Realize the modeling of wind farm aggregation characteristics based on convolutional neural network. 2 . The method for modeling the aggregation characteristics of a wind farm according to claim 1 , wherein the measured operation data in step 1 includes: measured wind speed data, measured wind direction data, and measured power data. 3 . 3. The method for modeling the aggregated characteristics of a wind farm according to claim 1, wherein the aggregated output of the wind turbine cluster is the sum of the output of each single unit, as in formula (1):

where P _∑ (t) is the total power of the wind turbine cluster at time t; i is the number of wind turbines; N is the number of wind turbines; P _i (t) is the power of the ith wind turbine at time t; In the second step, formulas (2) and (3) are used to express the standard deviation of the power sequence to measure the fluctuation of the output of a single wind turbine and the aggregate output of the wind turbine:

where σ _i and σ _∑ are the standard deviation of power and the standard deviation of aggregated power of the ith wind turbine, respectively; T is the statistical time scale;

are the mean values of the sampled values of the i-th wind turbine power and aggregate power under the corresponding time statistical scale; the standard deviation of the power series reflects the degree of power fluctuation of the time series. The smaller the value, the smaller the time series power fluctuation and the better the stability. 4. The method for modeling wind farm aggregation characteristics according to claim 1, wherein the smoothing effect coefficient is the standard deviation per unit value of the aggregate output of wind turbines with the installed capacity as the base value and the standard deviation per unit output of a single unit The ratio of values is equation (4):

Among them, P _R is the rated power of the wind turbine, σ _i and σ _∑ are the standard deviation of power and the standard deviation of the aggregated power of the ith wind turbine, respectively, and N represents the number of aggregated units. The smaller the coefficient s, the smaller the output of each wind turbine. The stronger the complementarity between the two, the more obvious the smoothing effect; Through the mathematical statistics in the third step, the relational formula (5) of the standard deviation of the aggregate power of the wind turbine and the standard deviation of the single-unit power is obtained:

In the formula, ri _,j is the correlation coefficient of the power sequence of the ith wind turbine and the jth wind turbine; Assuming that the output standard deviation of each unit is the same, it is regarded as the average output standard deviation of a single unit, then the smoothing effect measurement index and the aggregated unit number N and the power sequence between units can be obtained from formulas (4) and (5). The correlation coefficient relationship of is formula (6):

5. The method for modeling wind farm aggregation characteristics according to claim 1, wherein the convolutional neural network in step 4 comprises two convolutional layers, two pooling layers, a flattening layer and a full Connection layer; the measured wind speed sequence data and the measured wind direction sequence data of multiple units pass through the first convolution layer and the ReLU activation layer to generate a set of feature maps, and then perform downsampling through non-overlapping maximum pooling, and then pass through the second convolution layer. And the ReLU activation layer to generate another set of feature maps, connect this set of feature maps with the flattening layer to make the multi-dimensional tensor one-dimensional, and then connect with the fully connected layer, and finally activate the sigmoid function as the output of the convolutional neural network. 6 . The method for modeling aggregation characteristics of wind farms according to claim 5 , wherein in the step 4, a batch normalization method is used to add a layer of normalization layer after the output of some layers, and the output is based on the characteristics of the same batch. 7 . The values are normalized to the same distribution, and the convolution operation process is represented by formula (7):

in,

Represents the jth feature map in the Lth convolutional layer, that is, the wind speed distribution feature extracted by this convolutional layer,

is the feature map set of the previous layer, which is also the input of this layer,

is the jth convolution kernel in the Lth convolutional layer, * is defined as the convolution operation, b represents the additive bias value, and f(x) represents the activation function, which maps a value in the real number domain to a in a finite field; The operation process of the pooling layer is represented by formula (8):

in,

represents the jth feature map in the Lth pooling layer, f(x) represents the activation function,

Represents the multiplicative bias of the feature map, down(x) is the downsampling function,

Represents the additive bias of the feature map. 7. The method for modeling the aggregation characteristics of a wind farm according to claim 1, wherein the calculation formula of the output correlation coefficient between two units in the step 5 is formula (9):

where σ _xy refers to the covariance of the two data series X and Y, σ _x and σ _y are the respective variances of X and Y, respectively,

are the mean values of the sequences X and Y respectively, and X _i and Y _i are the values of the i-th number of the sequences X and Y respectively.

Images