CN114386324A - Ultra-short-term wind power segmented prediction method based on turning period identification - Google Patents

Abstract

The invention relates to an ultra-short-term wind power sectional prediction method based on turning time interval identification, which extracts a time sequence trend by utilizing a mobile averaging method; smoothing an exponential moving average line (EMA) by adopting a Gaussian window method, and calculating the time sequence change rate alpha at each moment; the window adjustment strategy based on the local time sequence characteristics adaptively adjusts the time window width; based on an inflection point detection strategy of a double-timing sliding window, introducing alpha as one of criteria, and extracting and dividing a turning weather sudden change time period; adopting improved GRU algorithm point prediction for a turning section time sequence, and combining an improved Attention mechanism of a CRS algorithm; probability prediction is carried out on the gentle section time sequence, an empirical distribution estimation method is adopted to establish a time sequence mode-power prediction error probability density distribution model, and wind power probability prediction is carried out on the basis of a variable bandwidth kernel density estimation method; and combining the point prediction and the probability prediction to obtain a final prediction result by time sequence segmentation prediction. Compared with the prior art, the method has the advantages of improving the model operation efficiency and the like.

Description

A segmented forecasting method of ultra-short-term wind power based on turning point identification

technical field

The invention relates to the technical field of ultra-short-term wind power prediction of high-concentration wind farms, in particular to a sub-section prediction method of ultra-short-term wind power based on turning period identification.

Background technique

In recent years, with the development of large-scale and high-concentration wind power, the proportion of wind power connected to the power grid is increasing. But with the frequent occurrence of extreme weather, the large wind speed fluctuations caused by it can lead to potential disaster, especially in large-capacity wind farms. In order to formulate effective prevention and control strategies in time, it is particularly important to predict wind power in extreme weather in advance. The impact of extreme weather on wind farms is intuitively reflected in the significant changes in wind speed in a short period of time. The extreme weather power periods are characterized by large and violent power fluctuations on the time series scale. Identifying and extracting extreme weather power periods has become the primary task.

Under extreme meteorological conditions represented by transitional weather, wind power fluctuates violently in a short period of time. Existing ultra-short-term wind power single-value forecasting methods lack a forecasting model for extreme weather, resulting in poor forecast accuracy and stability; secondly, for gentle power periods, traditional single-value forecasting methods have higher prediction accuracy, and for extreme weather power periods , the probabilistic prediction method has better prediction performance because it can quantify the prediction error. The existing method lacks the optimal strategy of prediction method based on extreme weather period, and cannot combine the advantages of single-value prediction and probability prediction in the whole period. Therefore, it is necessary to further search for ultra-short-term wind power prediction methods suitable for the "complex meteorological model" for forecasting under the traditional "simple weather model". In particular, it should be considered to combine the transitional weather characteristics and build an adaptive power mutation identification mechanism to further improve the generalization ability of the model.

At present, the research on the sudden change of wind power under transitional weather is still limited to the description and prediction of wind power climbing events. The main methods used in the existing technology are: iterative optimization of the parameters of the probability generation model through the genetic algorithm of the multi-objective fitness function. , obtain a large number of prediction scenarios, and evaluate the prediction method through the probability features of climbing events excavated in the scene capture band, but the model still needs to improve the robustness in extreme weather; based on the event detection framework, data-driven algorithms are used to improve However, the power ramp event in this model still uses the traditional evaluation index, and lacks flexible response ability to interference factors such as pseudo-inflection points; through the continuous adjustment of the integrated learning method, probabilistic prediction is generated to quantify the uncertain factors of the prediction, but this method lacks The accurate detection and identification method of wind power ramping events does not highlight the accuracy improvement of the model under extreme weather. In summary, preliminary research results have been achieved in the detection, identification and advance prediction of wind power sudden change periods, but there is still room for improvement in refined identification of sudden change periods.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a segmented prediction method for ultra-short-term wind power based on the identification of turning points in order to overcome the above-mentioned defects of the prior art.

The object of the present invention can be realized through the following technical solutions:

A method for segmented forecasting of ultra-short-term wind power based on identification of turning period, comprising the following steps:

1) Extract the time series trend, obtain the EMA curve representing the short-term development trend of the original wind power data of the wind farm, and use the Gaussian window method for smoothing to obtain the change rate α at each time;

2) Based on the EMA curve obtained in step 1), use the local feature difference of the EMA curve to formulate a window adjustment strategy, and set the detection threshold ε. If the fluctuation of the distribution difference between the window and its previous window is less than the detection threshold ε, the window width will be expanded to speed up. Detection speed, otherwise, reduce the window width to improve detection accuracy;

3) based on the window adjustment strategy of the local feature difference formulated in step 2), use the α obtained in step 1) as one of the criteria, and mark the position where the mean value in the two windows has a minimum value as an inflection point;

4) Improve the traditional power mutation period criteria, merge adjacent same-trend mutation periods, and completely extract the turning weather mutation period;

5) According to the extraction result of the adaptive turning period as the division basis, the time sequence is divided into turning sections and smooth sections;

6) Using point prediction for the flat segment, using GRU as the original algorithm for point prediction, introducing an improved Attention mechanism combined with the CRS algorithm, assigning different weights to the transition feature vectors of the neural network model, and then transferring the attention weights to the GRU layer, Output the training results of the GRU neural network, read the training loss curve and error curve, observe the longitudinal distance between the loss curves of the training set and the validation set during the convergence process, and visually evaluate the convergence performance of the network prediction results based on the absolute error of the training set and the validation set;

7) Probabilistic prediction is used for the turning section, and the time series mode-adaptive bandwidth kernel density estimation method is used for probability prediction;

8) Combining steps 6) and 7) to complete the ultra-short-term wind power prediction based on the transition period, and obtain the predicted power.

Further, in step 1), the time series trend is extracted by using the moving average method.

Further, in step 2), the specific steps of formulating the window adjustment strategy by utilizing the local feature difference of the EMA curve include:

21) Divide the original power sequence into several segments, and perform turning point detection on each segment;

其中Vs_i为待检测数据的第i个窗口数据数据分布的均值波动，Ds_i为差值波动；22) In turning point detection, define diff _i to measure the fluctuation of the distribution difference between the i-th window and the previous window, denoted as

Wherein Vs _i is the mean fluctuation of the data distribution of the i-th window data of the data to be detected, and Ds _i is the difference fluctuation;

23) Set the threshold ε, if the value of diff _i is less than or equal to the threshold ε, expand the sliding window width W to increase the detection speed; if the value of diff _i is greater than ε, reduce the sliding window width W to improve the detection accuracy.

Further, in step 3), double fixed-time sliding windows are used to detect inflection points. The specific contents are:

First, the change rate α obtained in step 1) is introduced as one of the criteria, and the inflection point is set to satisfy the condition α=0; based on the EMA curve, two closely connected sliding windows are established, and the data in the two windows are updated frame by frame. , mark the power value at the junction point when the mean difference in the two windows reaches the minimum as the inflection point, and repeat the above steps to obtain the corresponding time series trend inflection point set T _ip .

Further, in step 4), the expression of the improved mutation period criterion is:

为拐点集T_ip中第j点功率值；

为拐点集T_ip中第j+1点功率值；

为拐点集T_ip中经过第j点时刻；

为拐点集T_ip中经过第j+1点的时刻；λ为转折时段突变幅度阈值；β为转折时段突变速率阈值；合并相邻同趋势的突变时段，以完整提取转折时段。In the formula,

is the power value of the jth point in the inflection point set _Tip ;

is the power value of the j+1th point in the inflection point set T _ip ;

is the moment passing through the jth point in the inflection point set T _ip ;

is the time passing through the j+1th point in the inflection point set T _ip ; λ is the threshold of mutation amplitude in the turning period; β is the threshold of mutation rate in the turning period; the adjacent mutation periods with the same trend are merged to completely extract the turning period.

Further, in step 6), the mathematical expression of GRU neural network is:

z _t =σ(W _z ·[h _t-1 ,x _t ])

r _t =σ(W _r ·[h _t-1 ,x _t ])

为输入和过去隐层状态的汇总，h_t为隐藏层输出，W_z，W_r，W为可训练参数矩阵。where: z _t is the update gate, r _t is the reset gate, X _t is the current input,

is the summary of the input and the past hidden layer state, h _t is the hidden layer output, W _z , W _r , and W are the trainable parameter matrix.

The improved Attention mechanism combined with the CRS algorithm is introduced, and the specific steps to assign different weights to the transition feature vectors of the neural network model include:

61) Provide the weight W of the attention layer, and the specific calculation steps of the weight are as follows:

611) Calculate similarity M(Q, K) for given task query vector Q and attention variable K;

612) Perform Softmax operation on the obtained similarity and normalize to obtain the normalized similarity η _i :

613) For the weights obtained by the above calculations, weighted summation is performed on all the obtained weights to obtain the Attention vector;

62) Convert the provided weight W of the attention layer into binary code W ^B , the subset _Wi is the attention weight, transmit the subset to the GRU neural network, and the GRU neural network generates the corresponding prediction error in the network according to the prediction error in the network. loss value;

并对其子集组合进行反复循环；63) According to the loss of W ^B , select the optimal attention weight subset W _i ^B and

And repeat the cycle of its subset combination;

64) Rebuild a new attention weight

Further, in step 7), the concrete steps of probability prediction using time series mode-adaptive bandwidth kernel density estimation method include:

71) Divide the time series mode based on the characteristics of the power period, and divide it into five categories: sharp rise, sharp fall, slow rise, slow fall and oscillation;

72) The empirical distribution estimation method is used to establish the time series mode-wind power prediction error probability density distribution model for each category under different weather conditions.

In step 71), with α as the division basis, the weather types are divided, the probability density distribution of power prediction errors under the corresponding time series pattern characteristics is calculated, and the distribution is visually reflected through the boxplot; the asymptotic integral mean square error method is used to obtain The optimal window width is substituted into the estimation function, and the probability density distribution curves under each time series feature are respectively fitted to provide a basis for the presentation of interval prediction results.

Compared with the prior art, the ultra-short-term wind power segment prediction method based on the identification of the turning period provided by the present invention at least includes the following beneficial effects:

1) The method of the present invention proposes a power time series trend discrimination method based on moving average iteration to fully consider the historical wind power time series characteristics, which is conducive to improving the power mutation period under transition weather conditions, in view of the changeable wind power scene under transitional weather conditions. the accuracy of trend descriptions;

2) Aiming at the problem of insufficient extraction of the turning period of the power sequence, the present invention proposes a sliding window width adjustment strategy based on the difference of local timing characteristics to improve the extraction integrity of the turning period and significantly improve the effectiveness of the algorithm;

3) Aiming at the difference of power time series characteristics in meteorological mode, a segmented power prediction method based on point prediction-probability interval prediction based on time series characteristic matching is proposed. The point prediction adopts the improved GRU algorithm, which reduces the amount of training. The variable bandwidth kernel density estimation method improves the prediction performance.

Description of drawings

1 is a schematic flowchart of a method for segmented prediction of ultra-short-term wind power based on identification of a turning point in an embodiment;

2 is a schematic diagram of a window adjustment strategy in an embodiment;

Fig. 3 is the schematic flow chart of the improved Attention mechanism combined with CRS in the embodiment;

FIG. 4 shows the distribution of power prediction errors in each timing mode in the embodiment.

Detailed ways

The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are some, but not all, embodiments of 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 work shall fall within the protection scope of the present invention.

Example

The invention relates to an ultra-short-term wind power subsection prediction method based on turning period identification. The method extracts the time series trend based on the moving average method, realizes the turning period division through an adaptive window adjustment method, and adopts different forecasting methods considering different characteristic periods. , formulate a point forecasting-probabilistic forecasting subsection forecasting strategy, and make accurate forecasts for full-time wind power including transitional weather.

The main principles of the ultra-short-term wind power segment prediction model based on transitional weather established by the present invention are:

In terms of the extraction of turning weather periods, extreme weather power periods are characterized by large and violent fluctuations in power on the time series scale. In order to improve the prediction accuracy of wind power under transitional weather, it is the primary task to identify and extract extreme weather power periods, and to achieve accurate and complete extraction of sudden changes as much as possible. Due to the variety of meteorological patterns in a short period of time under transitional weather, the traditional time series feature extraction of the original power time series is prone to misjudgment of the trend. Therefore, the moving average iteration power time series trend discrimination method is used to improve the accuracy of the time period trend description. In order to fully extract the turning period of the power sequence, the traditional slope extraction algorithm represented by the fixed-time sliding window is difficult to accurately and completely extract the turning power period. .

In terms of forecasting methods, the current single-value forecasting methods lack forecasting models for extreme weather, resulting in poor forecasting accuracy and stability. For the moderate power period, the traditional single-value prediction method has higher prediction accuracy, and for the extreme weather power period, the probabilistic prediction method has better prediction performance due to quantifying the prediction error. Therefore, the present invention proposes to use a segmented prediction strategy to make accurate predictions of wind power in the whole period including transitional weather, that is, point prediction is used for the time sequence of the smooth segment, and probability prediction is used for the time sequence of the turning segment.

LSTM is an RNN model. In addition to processing the input data at the current time point, each unit in the RNN also processes the output of the previous unit, and finally outputs a single prediction. The basic RNN model only processes the output of the previous unit, so that the output of the distant unit is processed several times, and the influence gradually disappears. In view of the above problems, the present invention proposes to use the GRU algorithm as the original algorithm for point prediction. GRU is a simplified variant of the LSTM network and belongs to the gated recurrent neural network. The update gate in GRU is composed of the forget gate and the input gate in the LSTM network, and the model architecture is simpler, which reduces the amount of computation and training time while ensuring the prediction accuracy of the model. Based on the existing model, an Attention mechanism combined with the CRS algorithm is introduced to guide the weight distribution between time series.

Traditional wind power interval forecasts often only consider the distribution of forecast errors at different power levels, ignoring the impact of sudden changes in power caused by weather type changes on forecast errors. In view of the above problems, the present invention divides the time sequence mode based on the power period characteristics, and divides it into five types of modes, such as sharp rise, sharp fall, slow rise, slow fall, and oscillation. Combined with the distribution characteristics of wind power prediction errors under different weather types, the prediction performance of the model is improved, and the empirical distribution estimation method is used to establish the time series mode-wind power power prediction error probability density distribution model for each category under different weather types.

Based on the above principles and design ideas, as shown in FIG. 1 , the ultra-short-term wind power segmented prediction method based on the identification of the turning period of the present invention specifically includes the following steps:

Step 1: Extract the time series trend. First, obtain the EMA curve representing the short-term development trend of the original data (original wind power data of the wind farm); secondly, the Gaussian window method is used for smoothing to obtain the change rate α at each moment. The specific operations are as follows:

11) Find the EMA curve

The EMA curve adopts the method of statistical processing to perform a weighted average of the original data, and then the connected curve is used to observe the changing trend of the data in the future.

Obtain the N-day smoothed moving average Y _N of t, and Y _N-1 is the N-1 day smoothed moving average, that is, the EMA curve is calculated as follows:

12) Find α

Before the rising or falling trend of the time series has not changed, the time series slope can directly reflect the change trend of the time series, and the indicator of the change trend of the power time series can be found by tracking the slope of the time series. For the EMA curve obtained above, the curve is first smoothed by the Gaussian window method, and then the initial power mutation sensitivity factor α is calculated from the rate of change at the current moment. The calculation formula is as follows:

Among them, Y _smooth,N (t) is the EMA value before smoothing at time t. Y _smooth,N-1 (t-Δt) is the smoothed EMA value after the time series rising or falling trend changes Δt.

13) Trend extraction

When the raw power is above the short-term moving average, α>0, and when the raw power is below the short-term moving average, α<0. Then, according to the aggregation and separation between the short-term moving average and the original power, and further combining the timing characteristics of the moving average itself, the image can intuitively judge the high and low points of the predicted object.

In step 2, an adaptive time window turning period division method is proposed to quickly identify the turning power period. Firstly, according to the EMA curve obtained in the previous step, use the local characteristic difference of the EMA curve to calculate the distribution difference between adjacent windows, and then set the detection threshold. If the width is larger than this value, the window width will be reduced to improve the detection accuracy.

Most power mutation detection algorithms use the difference between the distribution of the local data to be detected and the standard data distribution to determine whether there is a sudden change in power time series. The window adjustment strategy based on time series features can speed up the extraction of turning points.

The specific steps of the window adjustment method based on local feature differences are as follows:

A) Introduce the original sliding window model, first divide the original power time series into several segments, and then perform turning point detection on each segment;

B) In turning point detection, calculate the mean fluctuation Vs _i and the difference fluctuation Ds _i of the data distribution of the i-th window data of the data to be detected, and the expressions are as follows:

In the formula, U _i represents the data in the ith window, U represents the entire original power time series, and var and std represent the variance and standard deviation, respectively. max(U) and min(U) represent the maximum and minimum values of the entire original power sequence, respectively.

C) Calculate the distribution difference diff _i between the ith window and the previous window.

D) As shown in FIG. 2 , for the distribution difference diff _i obtained by the above formula, a threshold ε=0.2 is set. If the value of diff _i is less than or equal to ε, it belongs to the same data feature, and the sliding window width W is enlarged; if the value of diff _i is greater than ε, it is in a turning period, the sliding window width W is reduced, and finally the purpose of adaptive window adjustment is achieved. .

The adaptive window adjustment method is used to identify the power mutation period, so as to further extract the inflection point set, which greatly increases the detection speed and improves the detection accuracy.

Step 3. Use the double fixed-time sliding window method to detect the inflection point in the trend. First, based on the above-mentioned window adjustment strategy for local feature differences, α is introduced as one of the criteria to mark the minimum value of the mean value in the two windows. The position of is the inflection point; then the traditional power mutation period criterion is improved, the adjacent same trend mutation periods are merged, and the turning weather mutation period is completely extracted.

The specific steps for extracting the inflection point set in the trend are as follows:

A) Under the oscillating output period, there will be a misjudgment of the inflection point. The method of the present invention introduces α as one of the criteria, that is, the inflection point must satisfy the condition α=0;

B) on the basis of the EMA curve, set up 2 sliding windows that are closely connected, update the data in the 2 windows frame by frame, the present invention takes the sliding window mean value as the benchmark of the difference comparison, and the calculation method of the _inflection point detection score S is as follows:

In the formula: X _1,i is all power data in the previous window, X _2,i is all power data in the next window. When the mean difference in the two windows reaches the minimum, the power value at the marked junction point is the inflection point, and the above steps are repeated to finally obtain the corresponding time series trend inflection point set T _ip .

Then, the traditional power period criterion is improved, and the improved mutation period criterion is as follows:

为拐点集T_ip中第j点功率值；

为拐点集T_ip中第j+1点功率值；

为拐点集T_ip中经过第j点时刻；

为拐点集T_ip中经过第j+1点的时刻；λ为转折时段突变幅度阈值；β为转折时段突变速率阈值；不仅考虑突变时段幅值的变化，且考虑突变速率，从而排除伪拐点的存在。In the formula,

is the power value of the jth point in the inflection point set _Tip ;

is the power value of the j+1th point in the inflection point set T _ip ;

is the moment passing through the jth point in the inflection point set T _ip ;

is the moment when the inflection point set T _ip passes through the j+1th point; λ is the threshold of the mutation amplitude in the turning period; β is the threshold of the mutation rate in the turning period; not only the change of the amplitude of the mutation period, but also the mutation rate is considered, so as to exclude the false inflection point. exist.

Finally, the adjacent mutation periods with the same trend are merged, and the power transition period is completely extracted.

Step 4. According to the extraction result of the adaptive turning period as the division basis, the time series is divided into the turning section and the flat section, the point prediction based on the GRU algorithm is used for the flat section, and the time series mode-adaptive bandwidth kernel density estimation method is used for the turning section. Probability predict.

Step 5. Input the time series features extracted by the GRU network into the improved Attention mechanism of the CRS algorithm, assign different weights to the transition feature vectors of the neural network model, and then input the transition feature vectors managed by the weights to the GRU layer by time step. Output the training results of the improved GRU neural network, read the training loss curve and error curve, observe the longitudinal distance between the loss curves of the training set and the validation set during the convergence process, and visually evaluate the convergence performance of the network prediction results based on the absolute errors of the training set and the validation set.

The following are convergence conditions represented by three common fitting states:

1) When the loss curve of the training set has almost no decline, it is an underfitting state and a non-converging state;

2) When the loss curve of the training set continues to decrease, and the loss curve of the validation set does not decrease at a certain time, it is in the state of overfitting, which is in the state of convergence but not perfect convergence;

3) When there is no obvious gap between the loss curves of the training set and the validation set, it is in a state of perfect fitting and perfect convergence.

Further, the present invention improves and optimizes the traditional GRU model, combined with the Attention mechanism of the CRS algorithm, the specific contents are as follows:

Due to the large number of features of the input model, in order to highlight more critical influencing factors and help the model to make more accurate judgments, the present invention proposes an improved Attention mechanism, which assigns different weights to the transition feature vectors of the neural network model. In the traditional Attention mechanism, the information carried by the content input to the network first will be overwritten by the information input later, and the semantic vector may not fully represent the information of the entire sequence. Therefore, in view of the above shortcomings, the present invention proposes an improved Attention mechanism (Improved attention mechanism) combined with the CRS (Competitive random search) algorithm, which makes up for the shortage of the network's attention to the characteristics of different related factors on the same time scale, and improves the network The level of attention to various related factors.

CRS is used to generate optimal parameter combinations in the attention layer. The operation process of the CRS is introduced in Fig. 3, and the CRS consists of four parts "I, II, III, IV".

并对其子集组合进行反复循环。最终，在“IV”中重建了一个新的注意力权重

"I" provides the weights W of the attention layer; then through the conversion to binary code in "II", the subset _Wi is the attention weights, which are transmitted to the GRU neural network, where corresponding prediction errors are generated according to the network loss value. Then, according to the loss of ^WB in "III", the optimal attention weight subsets _Wi ^B and

And iteratively loops its subset combination. Finally, a new attention weight is reconstructed in "IV"

The detailed steps of CRS are as follows:

1) Randomly generate an attention weight set of length M=n (n is the model input feature dimension),

2) Input the subset Wi into the attention layer and convert W into _binary code:

来计算预测误差：

3) According to the true value y and predicted value of the GRU model

to calculate the prediction error:

每个子集由二进制字符串组成，并均匀地划分为n段。相应地，W_i ^B和

由W_i ^B＝(F_i ¹,F_i ²,...F_i ⁿ)和

表示。F_i ¹和

分别是W_i ^B和

的一部分。4) According to the error feedback, select the optimal attention weight subset _Wi ^B and

Each subset consists of binary strings and is evenly divided into n segments. Correspondingly, _Wi ^B and

By Wi ^B =(F _i ¹ ,F _i ² ,...F _{i n} ⁾ and

express. F _i ¹ and

are W _i ^B and

a part of.

例如，选择两者的第n-1段F_i ^n-1和

然而所选分段的数目并不固定。5) _Randomly extract part of Wi ^B and

For example, select the _n -1th segment of both ^Fin-1 and

However, the number of selected segments is not fixed.

的遗传重组。F_i ^n-1和

由长度为6的二进制码表示，两者在相应的6个索引上随机交换，得到重组段

6) Get _Fin ^-1 and

genetic recombination. _Fin ^-1 and

Represented by a binary code of length 6, the two are randomly exchanged at the corresponding 6 indices to obtain a reorganized segment

的基因型。例如，0被反转为1。然后

取代了在W_i ^B中相应的F_i ^n-1，形成新的

插入到W^B中。7) Simulated a gene mutation and reversed

genotype. For example, 0 is inverted to 1. Then

replaces the corresponding _Fin ^-1 _in Wi ^B , forming a new

Insert into ^WB .

8) W ^B is decoded to obtain an updated set of attention weights: W'=(W' ₁ ,W' ₂ ,...,W' _k ,...,W' _M ).

9) Repeat steps 2) to 8) until the preset time K is reached.

Further, the mathematical expression of the GRU neural network is:

z _t =σ(W _z ·[h _t-1 ,x _t ])

r _t =σ(W _r ·[h _t-1 ,x _t ])

输入和过去隐层状态的汇总，h_t为隐藏层输出，W_z，W_r，W为可训练参数矩阵。where: z _t updates the gate, r _t resets the gate, X _t is the current input,

A summary of the input and past hidden layer states, h _t is the hidden layer output, W _z , W _r , W are the trainable parameter matrix.

Step 7. The specific steps of the probability prediction method using the time series mode-adaptive bandwidth kernel density estimation method for the time period of sudden change are as follows:

1) Divide the time series mode based on the characteristics of the power period, and divide it into five modes: sharp rise, sharp fall, slow rise, slow fall, and oscillation;

2) The empirical distribution estimation method is used to establish the time series mode-wind power prediction error probability density distribution model for each category under different weather conditions. The details are as follows:

21) Assuming that the overall variable X has a set of sample observations x ₁ , x ₂ ,...x _p ... x _m with a capacity of m, rearrange them in the order from small to large to obtain order statistics x ₁ ′, x ₂ ′,…x _p ′…x _m ′ For any real number x, its empirical distribution expression is:

22) Calculate the power prediction error probability distribution according to the corresponding time series pattern characteristics, and obtain the boxplot based on the error distribution of various weather patterns. For example, as shown in Figure 4, the left and right borders of the square are respectively the corresponding points of the 50% quantile, the scale line at the center position is the median of this type of pattern error, and the outer points are outliers (outliers). The power prediction error distribution in the mode;

23) From the original prediction error probability distribution f _error from the empirical distribution, the formula for estimating and calculating f _e kernel density is as follows:

In the formula, N is the number of samples of a set of data; h is the window width, also known as the smooth parameter; K(u) is the kernel function, u=h ^-1 (ee _i ); e _i is the ith in the prediction error data. sample value.

表达式中，高斯核函数具体表达式如下：24) Select Gaussian kernel function to substitute for estimation

In the expression, the specific expression of the Gaussian kernel function is as follows:

Substitute into the expression:

和真实的概率密度函数f_e(x)两者存在的差异，表达式如下：25) First, the integral mean square error is introduced to judge the estimated probability density function

The difference between the real probability density function f _e (x) and the real probability density function f e (x) is expressed as follows:

MISE表示为主项AMISE的和，当h→0，nh→∞时，定义AMISE表达式为：in

MISE is expressed as the sum of the main term AMISE. When h→0, nh→∞, the AMISE expression is defined as:

可得最优窗宽h_x表达式如下：In the formula, AMISE is an expression about the window width h, and the optimal h value is reached when AMISE takes the minimum value, that is,

The optimal window width h _x can be expressed as follows:

3) Fit the probability density distribution curve under each time series feature respectively, and provide a theoretical basis for the presentation of interval prediction results.

Traditional interval forecasting only considers historical power levels to classify errors, but considering the physical dynamic process of time series pattern classification can improve the prediction accuracy of interval forecasting. The invention proposes a short-term interval prediction method based on time series pattern classification and Monte Carlo method, which has the best effect, and can obtain interval coverage rates greater than preset confidence levels under different confidence levels.

Step 8: Establish a point prediction-probability prediction subsection prediction model, and quantitatively analyze the result accuracy of point prediction, probability prediction and subsection prediction point prediction and probability prediction in the whole period.

The method of the present invention considers full-time wind power forecasting including transitional weather. Firstly, a power time series trend discrimination method based on moving average iteration is proposed, and the Gaussian window method is used to smooth the EMA curve to extract the turning weather trend. Secondly, a window adjustment strategy based on local time series characteristics is proposed to adaptively adjust the time window width. The inflection point detection strategy of double fixed-time sliding window, combined with the criterion to obtain the inflection point set, extract and divide the transitional weather mutation period. The proposed time-series segmented prediction algorithm for point prediction and probability interval prediction performs power prediction for the whole time period: the point prediction of the improved GRU algorithm is used for the time sequence of the turning segment, and the improved Attention mechanism combined with the CRS algorithm is proposed, and the transition feature vector of the neural network model is assigned to Different weights; probabilistic prediction is used for the time series of the gentle segment, the time series mode-power prediction error probability density distribution model is established by the empirical distribution estimation method, and the wind power probability prediction is performed based on the variable bandwidth kernel density estimation method. The model not only fits the typical characteristics of abrupt weather, but also significantly improves the generalization ability and prediction performance of full-time power prediction.

The above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited to this. Any person familiar with the technical field can easily think of various equivalents within the technical scope disclosed by the present invention. Modifications or substitutions should be included within the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims (10)

1. A method for segmented forecasting of ultra-short-term wind power based on turning period identification, is characterized in that, comprises the following steps: 1) Extract the time series trend, obtain the EMA curve representing the short-term development trend of the original wind power data of the wind farm, and use the Gaussian window method for smoothing to obtain the change rate α at each time; 2) Based on the EMA curve obtained in step 1), use the local characteristic difference of the EMA curve to formulate a window adjustment strategy, and set the detection threshold ε. If the fluctuation of the distribution difference between the window and its previous window is less than the detection threshold ε, the window width will be expanded to speed up. Detection speed, otherwise, reduce the window width to improve detection accuracy; 3) based on the window adjustment strategy of the local feature difference formulated in step 2), use the α obtained in step 1) as one of the criteria, and mark the position where the mean value in the two windows has a minimum value as an inflection point; 4) Improve the traditional power mutation period criterion, merge adjacent same-trend mutation periods, and completely extract the turning weather mutation period; 5) According to the extraction result of the adaptive turning period as the division basis, the time sequence is divided into turning sections and smooth sections; 6) Using point prediction for the flat segment, using GRU as the original algorithm for point prediction, introducing an improved Attention mechanism combined with the CRS algorithm, assigning different weights to the transition feature vectors of the neural network model, and then transferring the attention weights to the GRU layer, Output the training results of the GRU neural network, read the training loss curve and error curve, observe the longitudinal distance between the loss curves of the training set and the validation set during the convergence process, and visually evaluate the convergence performance of the network prediction results based on the absolute error of the training set and the validation set; 7) Probabilistic prediction is used for the turning section, and the time series mode-adaptive bandwidth kernel density estimation method is used for probability prediction; 8) Combining step 6) and step 7) to complete the ultra-short-term wind power prediction based on the transition period, and obtain the predicted power. 2 . The ultra-short-term wind power segment prediction method based on the identification of the turning period according to claim 1 , wherein, in step 1), the time series trend is extracted by using the moving average method. 3 . 3. the ultra-short-term wind power segmented prediction method based on turning period identification according to claim 1, is characterized in that, in step 2), utilizes EMA curve local characteristic difference to formulate the concrete step of window adjustment strategy comprises: 21) Divide the original power sequence into several segments, and perform turning point detection on each segment; 22) In turning point detection, define diff _i to measure the fluctuation of the distribution difference between the i-th window and the previous window, denoted as

Wherein Vs _i is the mean fluctuation of the data distribution of the i-th window data of the data to be detected, and Ds _i is the difference fluctuation; 23) Set the threshold ε, if the value of diff _i is less than or equal to the threshold ε, expand the sliding window width W to increase the detection speed; if the value of diff _i is greater than ε, reduce the sliding window width W to improve the detection accuracy. 4. The ultra-short-term wind power segmented prediction method based on turning period identification according to claim 1, characterized in that, in step 3), double fixed-time sliding windows are used to detect inflection points. 5. The ultra-short-term wind power segmented prediction method based on turning period identification according to claim 4, is characterized in that, the concrete content that adopts double fixed time sliding window to carry out inflection point detection is: First, the change rate α obtained in step 1) is introduced as one of the criteria, and the inflection point is set to satisfy the condition α=0; based on the EMA curve, two closely connected sliding windows are established, and the data in the two windows are updated frame by frame. , mark the power value at the junction point when the mean difference in the two windows reaches the minimum as the inflection point, and repeat the above steps to obtain the corresponding time series trend inflection point set T _ip . 6. The ultra-short-term wind power segmented prediction method based on the identification of the turning period according to claim 5, is characterized in that, in step 4), the expression of the improved sudden change period criterion is:

In the formula,

is the power value of the jth point in the inflection point set _Tip ;

is the power value of the j+1th point in the inflection point set T _ip ;

is the moment passing through the jth point in the inflection point set T _ip ;

is the time passing through the j+1th point in the inflection point set T _ip ; λ is the threshold of mutation amplitude in the turning period; β is the threshold of mutation rate in the turning period; the adjacent mutation periods with the same trend are merged to completely extract the turning period. 7. The ultra-short-term wind power segmented prediction method based on turning period identification according to claim 1, is characterized in that, in step 6), the mathematical expression of GRU neural network is: z _t =σ(W _z ·[h _t-1 ,x _t ]) r _t =σ(W _r ·[h _t-1 ,x _t ])

where: z _t is the update gate, r _t is the reset gate, X _t is the current input,

is the summary of the input and the past hidden layer state, h _t is the hidden layer output, W _z , W _r , and W are the trainable parameter matrix. 8. The ultra-short-term wind power segmented prediction method based on the identification of turning period according to claim 7, it is characterized in that, in step 6), introduce the improved Attention mechanism combined with CRS algorithm, the transition feature vector of neural network model is introduced. The specific steps for assigning different weights include: 61) Provide the weight W of the attention layer, and the specific calculation steps of the weight are as follows: 611) Calculate similarity M(Q, K) for given task query vector Q and attention variable K; 612) Perform Softmax operation on the obtained similarity and normalize to obtain the normalized similarity η _i :

613) For the weights obtained by the above calculations, weighted summation is performed on all the obtained weights to obtain the Attention vector; 62) Convert the provided weight W of the attention layer into binary code W ^B , the subset _Wi is the attention weight, transmit the subset to the GRU neural network, and the GRU neural network generates the corresponding prediction error in the network according to the prediction error in the network. loss value; 63) According to the loss of W ^B , select the optimal attention weight subset W _i ^B and

And repeat the cycle of its subset combination; 64) Rebuild a new attention weight

9. The ultra-short-term wind power segmented prediction method based on the identification of turning period according to claim 1, characterized in that, in step 7), the concrete steps of adopting time series mode-adaptive bandwidth kernel density estimation method probability prediction include: : 71) Divide the time series mode based on the characteristics of the power period, and divide it into five categories: sharp rise, sharp fall, slow rise, slow fall and oscillation; 72) The empirical distribution estimation method is used to establish the time series mode-wind power prediction error probability density distribution model for each category under different weather conditions. 10. The ultra-short-term wind power segmented prediction method based on the identification of the turning period according to claim 9, characterized in that, in step 71), with α as the division basis, the weather types are divided, and the corresponding time series pattern characteristics are calculated The probability density distribution of the power prediction error under , and the distribution situation is directly reflected by the box plot; the optimal window width is obtained by the asymptotic integral mean square error method, and it is substituted into the estimation function to fit the probability density distribution curve under each time series feature respectively. , which provides a basis for the presentation of interval prediction results.

Images