CN106991285A - A kind of short-term wind speed multistep forecasting method and device - Google Patents

Abstract

The invention discloses a short-term wind speed multi-step prediction method and device, comprising: using a heuristic segmentation algorithm to divide the original wind speed time series into a plurality of stable subsequences; using adaptive variable mode decomposition technology to decompose each stable subsequence It is a series of sub-patterns with limited bandwidth; through the robust online extreme learning machine based on the forgetting factor of COOK distance, a basic prediction model is established for each sub-pattern, and the parameters of the prediction model are adjusted by using a cross-cut optimization algorithm with an adaptive mutation mechanism Optimization; using the online ensemble learning method and ordered aggregation technology, the final prediction value is obtained by weighted aggregation of multiple basic prediction models; it can be seen that the model is updated according to the change of wind speed through the robust online extreme learning machine based on the forgetting factor of COOK distance, Using ordered aggregation for online integrated learning technology OEOA weighted aggregation of multiple basic prediction models to obtain the final prediction value, which greatly improves the accuracy of multi-step prediction.

Description

A short-term wind speed multi-step prediction method and device

technical field

The present invention relates to the technical field of wind speed prediction, and more specifically, relates to a short-term wind speed multi-step prediction method and device.

Background technique

At present, short-term wind speed prediction is of great significance for wind farm management and power system operation, but the randomness, fluctuation and intermittency of wind speed make wind speed prediction more difficult. The methods proposed in the field of wind speed prediction in recent years are often complex in structure and computationally intensive, and most of the researches are in offline mode, which cannot track the change of wind speed in real time.

Therefore, how to track the change of wind speed in real time is a problem to be solved by those skilled in the art.

Contents of the invention

The purpose of the present invention is to provide a short-term wind speed multi-step prediction method and device to realize real-time tracking of wind speed changes and improve wind speed prediction accuracy.

In order to achieve the above object, the embodiment of the present invention provides the following technical solutions:

A short-term wind speed multi-step forecasting method, comprising:

Using the heuristic segmentation algorithm BGA to segment the original wind speed time series into multiple stationary subsequences;

Use the adaptive variable mode decomposition technique SAVMD to decompose each stationary subsequence into a series of submodes with limited bandwidth;

Through the robust online extreme learning machine λ _CDFF OS-ORELM based on the forgetting factor of COOK distance, a basic prediction model is established for each series of sub-patterns, and the cross-cutting optimization algorithm CSO-SAM with an adaptive mutation mechanism is used to predict the model Carry out parameter optimization;

Using online ensemble learning and ordered aggregation technology OEOA, the final prediction value is obtained by weighted aggregation of multiple λ _CDFF OS-ORELM basic prediction models.

Wherein, the original wind speed time series is divided into multiple stationary subsequences by using the heuristic segmentation algorithm BGA, including:

Determine the left average value and right average value of other target points in the original wind speed time series except the starting point and the end point, and determine the statistical value of each target point according to the left average value and right average value of each target point;

Calculate the statistical significance of each target point according to the statistical value of each target point, use the target point whose statistical significance is greater than a predetermined threshold as a segmentation point, and use the segmentation point to segment the original wind speed time series to form Multiple stationary subsequences.

Wherein, after using the segmentation point to segment the original wind speed time series, it also includes:

Judging whether the divided left and right parts of the sequence are both greater than the minimum segmentation length; if not, cancel the segmentation of the original wind speed time series.

Wherein, the adaptive variable mode decomposition technique SAVMD is used to decompose each stationary subsequence into a series of limited bandwidth submodes, including:

Utilize the empirical mode decomposition technique EMD to determine the modal number K that decomposes the target stationary subsequence, and decompose the target stationary subsequence into K modes by the variable mode decomposition technique VDM;

The fluctuation parameters of each mode are determined by the detrending fluctuation analysis method DFA, and the target stationary subsequence is denoised and reconstructed according to the fluctuation parameters of each mode.

Among them, through the robust online extreme learning machine λ _CDFF OS-ORELM based on the forgetting factor of COOK distance, a basic prediction model is established for each series of sub-patterns, and a crossover optimization algorithm CSO- SAM optimizes the parameters of the prediction model, including:

Utilize the initial training data set to create an initial prediction model, determine the number of hidden layer nodes of the initial prediction model by a 10-fold cross-validation method, and calculate model parameters; the model parameters include a hidden layer output matrix and initial output weights;

Calculate the COOK distance, determine the forgetting factor λ according to the COOK distance, update the model parameters through the forgetting factor λ, and optimize the model parameters through the crossover optimization algorithm CSO-SAM with an adaptive mutation mechanism to obtain the basic prediction model ; The basic forecasting model includes a plurality of forecasting models corresponding to the stationary subsequence.

Wherein, the online integrated learning and ordered aggregation technology OEOA is used to obtain the final predicted value through weighted aggregation of multiple λ _CDFF OS-ORELM basic prediction models, including:

Update the root mean square error of multiple prediction models corresponding to each stationary subsequence, and select a predetermined number of models as the optimal model set according to the update results;

Assign weights to each prediction model in the optimal model set, calculate the prediction value of each stationary subsequence, and determine the final prediction value according to the prediction value of each stationary subsequence.

Wherein, after determining the final predicted value according to the predicted value of each stationary subsequence, it includes:

After obtaining the wind speed time series at the next moment, update the initial training data set according to the real value at this moment to obtain an updated training data set;

According to the real value at this moment, calculate the prediction error of each prediction model in the optimal model set, if there is a prediction model with a prediction error greater than the error threshold, then establish a training model through the updated training data set, replace A prediction model with a prediction error greater than the error threshold in the optimal model set.

A short-term wind speed multi-step forecasting device, comprising:

The smooth subsequence segmentation module is used for utilizing the heuristic segmentation algorithm BGA to divide the original wind speed time series into a plurality of stationary subsequences;

Decomposition module for decomposing each stationary subsequence into a series of limited-bandwidth sub-modes using the adaptive variable mode decomposition technique SAVMD;

Predictive model building module for building a basic predictive model for each series of sub-patterns by a robust online extreme learning machine λ _CDFF OS-ORELM based on the COOK distance-based forgetting factor, and adopting cross-cutting optimization with an adaptive mutation mechanism The algorithm CSO-SAM optimizes the parameters of the prediction model;

The prediction module is used to obtain the final prediction value through weighted aggregation of multiple λ _CDFF OS-ORELM basic prediction models by using online ensemble learning and ordered aggregation technology OEOA.

Wherein, the stable subsequence segmentation module includes:

Statistical value determination unit, used to determine the left average value and right average value of other target points in the original wind speed time series except the starting point and the end point, and determine each target according to the left average value and right average value of each target point point statistics;

The segmentation unit is used to calculate the statistical significance of each target point according to the statistical value of each target point, and use the target point whose statistical significance is greater than a predetermined threshold as a segmentation point, and use the segmentation point to calculate the original wind speed time The sequence is divided into multiple stationary subsequences.

Wherein, the predictive model building module includes:

The initial prediction model determination unit is used to create an initial prediction model by using the initial training data set, determine the number of hidden layer nodes of the initial prediction model by the 10-fold cross-validation method, and calculate the model parameters; the model parameters include the hidden layer output matrix and the initial output weight;

The basic predictive model determination unit is used to calculate the COOK distance, determine the forgetting factor λ according to the COOK distance, update the model parameters through the forgetting factor λ, and use the crossover optimization algorithm CSO-SAM with an adaptive mutation mechanism to model the model The parameters are optimized to obtain a basic forecasting model; the basic forecasting model includes multiple forecasting models corresponding to the stationary subsequences.

It can be seen from the above scheme that a short-term wind speed multi-step forecasting method provided by the embodiment of the present invention includes: using the heuristic segmentation algorithm BGA to divide the original wind speed time series into multiple stationary subsequences; using the adaptive variable mode decomposition technology SAVMD Decompose each stationary subsequence into a series of limited-bandwidth subpatterns; through the robust online extreme learning machine λ _CDFF OS-ORELM based on the forgetting factor of COOK distance, a basic prediction model is established for each series of subpatterns, and a band The crossover optimization algorithm CSO-SAM with an adaptive mutation mechanism optimizes the parameters of the prediction model; the final prediction value is obtained by weighted aggregation of multiple λ _CDFF OS-ORELM basic prediction models by using online integrated learning and ordered aggregation technology OEOA.

It can be seen that in this scheme, the robust online extreme learning machine based on the COOK distance forgetting factor can be used to update the model according to the change of wind speed, and the online ensemble learning technology OEOA is used for orderly aggregation, and multiple λ _CDFF OS- The ORELM weighted aggregation obtains the final prediction value, which greatly improves the multi-step prediction accuracy; the invention also discloses a short-term wind speed multi-step prediction device, which can also achieve the above technical effects.

Description of drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only These are some embodiments of the present invention. Those skilled in the art can also obtain other drawings based on these drawings without creative work.

Fig. 1 is a schematic flow chart of a short-term wind speed multi-step prediction method disclosed in an embodiment of the present invention;

FIG. 2 is a schematic diagram of a denoising process disclosed in an embodiment of the present invention;

Fig. 3 is a schematic flow chart of a short-term wind speed multi-step forecasting method disclosed in an embodiment of the present invention;

Fig. 4 is a schematic diagram of a wind speed time series disclosed by an embodiment of the present invention;

Fig. 5 is a schematic diagram of another wind speed time series disclosed by the embodiment of the present invention;

FIG. 6 is a schematic diagram of another wind speed time series disclosed in an embodiment of the present invention;

Fig. 7 is a schematic diagram of another wind speed time series disclosed by the embodiment of the present invention;

FIG. 8 is a schematic diagram of a prediction result disclosed in an embodiment of the present invention;

Fig. 9 is a schematic diagram of another prediction result disclosed in the embodiment of the present invention;

Figure 10 is a schematic diagram of another prediction result disclosed in the embodiment of the present invention

Fig. 11 is a schematic structural diagram of a short-term wind speed multi-step forecasting device disclosed in an embodiment of the present invention.

detailed description

The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some, not all, embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without creative efforts fall within the protection scope of the present invention.

The embodiment of the invention discloses a short-term wind speed multi-step prediction method and device, so as to realize real-time tracking of wind speed changes and improve wind speed prediction accuracy.

Referring to Fig. 1, a kind of short-term wind speed multi-step prediction method that the embodiment of the present invention provides, comprises:

S101. Using the heuristic segmentation algorithm BGA to segment the original wind speed time series into multiple stationary subsequences;

Wherein, the original wind speed time series is divided into multiple stationary subsequences by using the heuristic segmentation algorithm BGA, including:

Determine the left average value and right average value of other target points in the original wind speed time series except the starting point and the end point, and determine the statistical value of each target point according to the left average value and right average value of each target point;

Calculate the statistical significance of each target point according to the statistical value of each target point, use the target point whose statistical significance is greater than a predetermined threshold as a segmentation point, and use the segmentation point to segment the original wind speed time series to form Multiple stationary subsequences.

Wherein, after using the segmentation point to segment the original wind speed time series, it also includes: judging whether the divided left and right parts of the sequence are both greater than the minimum segmentation length; if not, canceling the time series of the original wind speed sequence segmentation.

Specifically, in this embodiment, the specific steps of using the BGA (Bernaola Galvan algorithm) algorithm to divide the wind speed time series into several stationary subsequences are:

11) For a wind speed time series X _N ={x ₁ ,...,x _N } containing N points, except for the start point and the end point, calculate the left part and right part of the remaining points from left to right The average value u _l and u _r of the t test is used to quantify the difference between the two parts of the mean value of the left and right parts of point i with the statistical value t _i of the t test, calculated as formula (1):

t _i ＝|(μ _l -μ _r )/s _D |.......................... ................(1)

Among them, the pooled variance s _l and s _r are the standard deviations of the time series on the left and right of the calculation point, respectively, where N _l and N _r are the number of time series on the left and right sides of the calculation point.

12) Calculate the statistical significance P(t _max ) of the maximum value t _max in t, as shown in formula (2):

P(t _max )＝Prob(t≤t _max ).......................... ......(2)

Among them, P(t _max ) represents the probability that the t value obtained in the random process is less than t _max .

13) If the significance P(t _max ) is greater than the selected critical value, then at this point the wind speed time series is divided into two subsequences with significant differences in mean values, otherwise no division; in this embodiment, P(t _max ) can take 95%.

14) In order to ensure the validity of the statistics, check whether the length of the left and right subsequences formed after segmentation is less than the minimum segmentation length (l ₀ ), if the length is less than l ₀ , no longer segment it, otherwise follow the above method Continue to divide the subsequence of P(τ)≥P ₀ , P(τ) can be obtained by Monte Carlo simulation.

After the segmentation is completed, the non-stationary wind speed time series is divided into multiple stationary subsequences with different mean values, so that the mean value difference of adjacent parts can be maximized, and the stationarity of each subsequence ensures higher prediction accuracy.

S102. Decompose each stationary subsequence into a series of bandwidth-limited subpatterns by using the adaptive variable pattern decomposition technique SAVMD;

Wherein, the adaptive variable mode decomposition technique SAVMD is used to decompose each stationary subsequence into a series of limited bandwidth submodes, including:

Utilize the empirical mode decomposition technique EMD to determine the modal number K that decomposes the target stationary subsequence, and decompose the target stationary subsequence into K modes by the variable mode decomposition technique VDM;

The fluctuation parameters of each mode are determined by the detrending fluctuation analysis method DFA, and the target stationary subsequence is denoised and reconstructed according to the fluctuation parameters of each mode.

Specifically, in this embodiment, the specific steps of using SAVMD (self-adaptive variational modedecomposition) technology to decompose the wind speed subsequence into a series of submodes with limited bandwidth are:

21) Use empirical mode decomposition (EMD) to determine the decomposed mode number K, and then decompose the wind speed time subsequence signal x(t) through VMD preset scale, and decompose the subsequence into K center frequencies is a sub-signal of w _k , thus K modal functions u _k are obtained, and the variable constraints are:

In the formula, {u _k }={u ₁ ,u ₂ ,...,u _k } are the modal functions; {w _k }={w ₁ ,w ₂ ,..., w _k } is each center frequency; is the sum of all modal functions; δ is the Dirac distribution, and "*" means convolution.

22) For the minimization problem in formula (3), the Lagrangian operator is solved as follows:

In the formula, a is the balance parameter of data fidelity, and the above formula is solved by the alternate direction method of multipliers (ADMN), and the iterative format of the optimal value of each parameter obtained is as follows:

in is the complex frequency domain form of u _k (t), n is the number of iterations, ifft(·) represents the inverse Fourier transform, Represents the real part of the solution signal.

23) By eliminating the fluctuation parameter h _urt of the detrended fluctuation analysis (DFA) to detect the wind speed time sub-mode that contains noise or is superimposed with a polynomial trend signal, the h _urt index is calculated as follows:

For a wind speed time series {x(i),i=1,2,...,N} with N points, calculate its cumulative dispersion:

Separately divide y(k) into equal lengths, divide the sequence into n non-overlapping intervals with length N _n =N/n, and use l-degree polynomial to fit each interval. In this paper, l=2, that is,

y _n (k)＝a _n k ² +b _n k+c _n .................. ..(9)

Calculate the mean value and square root of n equal-length intervals, and calculate the DFA fluctuation function as follows:

Take the double logarithm for the DFA fluctuation function and n, h _urt is the scatter plot in the double logarithmic coordinates (ln(n), ln(F(n))), and use the least square method to fit the data points, The slope of the straight line is the _urt exponent, which can be obtained by formula (11):

ln(F(n))＝ _hourt ln(n)+c...........................(11 )

Define the critical value of the slope θ = h _urt +0.25 (in this paper, h _urt = 0.5), when the h _urt parameter is less than θ, define the internal model function of the wind speed time subsequence as a noise signal as an independent input; when h _urt When the exponent is greater than θ, it is defined as a pure wind speed subsequence internal model function, and it is reconstructed:

Among them, h _urtj represents the fluctuation parameter of the internal model function of the jth wind speed time subsequence, where 1≤j≤k.

In summary, the implementation process of SAVMD-DFA is: firstly use EMD to determine the number K of VMD modes, then calculate the _hurt index of each mode through DFA to distinguish the random components of the signal, and perform Refactoring to finally determine the number of modes, the implementation process is shown in Figure 2.

S103. Through the robust online extreme learning machine λ _CDFF OS-ORELM based on the forgetting factor of COOK distance, establish a basic prediction model for each series of sub-patterns, and use the crossover optimization algorithm CSO-SAM with an adaptive mutation mechanism to Predictive model for parameter optimization;

Among them, through the robust online extreme learning machine λ _CDFF OS-ORELM based on the forgetting factor of COOK distance, a basic prediction model is established for each series of sub-patterns, and a crossover optimization algorithm CSO- SAM optimizes the parameters of the prediction model, including:

Utilize the initial training data set to create an initial prediction model, determine the number of hidden layer nodes of the initial prediction model by a 10-fold cross-validation method, and calculate model parameters; the model parameters include a hidden layer output matrix and initial output weights;

Calculate the COOK distance, determine the forgetting factor λ according to the COOK distance, update the model parameters through the forgetting factor λ, and optimize the model parameters through the crossover optimization algorithm CSO-SAM with an adaptive mutation mechanism to obtain the basic prediction model ; The basic forecasting model includes a plurality of forecasting models corresponding to the stationary subsequence.

Specifically, in this embodiment, λ _CDFF OS-ORELM (online sequential ORELM with forgetting factor based on Cook's distance) is used to establish a basic prediction model for each series of sub-patterns, and CSO-SAM (crisscross optimization algorithm with Self-Adaptive Mutation) algorithm for parameter optimization specific steps are:

31) Initialization phase.

The model utilizes the initial training dataset to create the initial ORELM model. The number of hidden layer nodes L _opt of the model is determined by 10-fold cross-validation method (10-fold cross-validation), setting the controller k=0, and calculating the initial hidden layer output matrix H ₀

Then the initial output weight is

in

32) Online sequence learning phase. When the k+1th sample of the training set D arrives, first calculate the hidden layer output layer matrix

Use the recursive least squares method to update the output weights, namely

in is the prediction error, Γ _k+1 is the gain matrix, obtained by formula (17)

33) Calculate the time-varying Cook distance for new observations

m is the length of the output parameter β, is the consistent estimator of ^σ2 , calculated as

Where M _k is the effective number of samples, given M ₀ =0.

34) The forgetting factor λ _k+1 can gradually eliminate old data and ensure the use of the latest data to build models. In this paper, the forgetting factor based on Cook’s distance (λ _CDFF ) is used to update the model parameters. When the k+1th sample arrives , find the forgetting factor based on Cook's distance:

C _k = νD _k ................................................ ........(twenty one)

S _k ＝P(χ _ν >C _k )(0<S _k <1).......................... ...(twenty two)

λ _k+1 ＝λ _min +(1-λ _min )S _k+1 .......................... ......(twenty three)

In the formula, λ _min is the minimum value of the forgetting factor, S _k+1 is the survival function, which uses mathematical statistics tools to convert the Cook distance into a forgetting factor, where C _k obeys the chi-square distribution with ν as the degree of freedom:

35) The input weights and hidden node deviations of the above basic prediction models need to be optimized using the CSO-SAM algorithm. The specific steps are as follows:

Assuming that the parent particles X(i) and X(j) cross horizontally in the d-th dimension, the calculation formula is as follows:

Among them, r ₁ and r ₂ are random numbers evenly distributed in [0,1], c ₁ and c ₂ are random numbers evenly distributed in [-1,1], MS _hc (i,d) and MS _hc ( j,d) is the offspring generated by the intersection of X(i,d) and X(j,d).

36) Take the updated population after the horizontal crossover is performed as the parent population of the vertical crossover, assuming that the particle X(i) performs vertical crossover in the d ₁ and d ₂ dimensions, the calculation formula is as follows:

Among them, r is a random number uniformly distributed in [0,1]; D is the total number of dimensions of particles; M is the population size. Compete the mean solution generated by horizontal crossover and vertical crossover with parent particles to get the optimal population.

37) Automatically adjust p _vc through the variable σ ² in the population fitness value, where the expression of σ ² is as follows:

Among them, f _i is the fitness value of particle i, f _avg is the average value of all fitness at this time, and n is the number of particles;

38) Finally, the linear expression of the vertical intersection probability p _vc can be obtained by using formula (27):

Among them, p _vcmax and p _vcmin are the maximum and minimum values of the vertical intersection probability p _vc ;

39) The training error of the λ _CDFF OS-ORELM model will be used as the optimization goal of the CSO-SAM algorithm to determine the optimal input weight and bias of ORELM, namely

In the formula, with are the actual value and predicted value of the pth training sample, respectively; N is the number of training samples.

S104. Using the online ensemble learning and ordered aggregation technology OEOA, the final predicted value is obtained by weighted aggregation of multiple λ _CDFF OS-ORELM basic prediction models.

Among them, using online ensemble learning and ordered aggregation technology OEOA, the final prediction value obtained by weighted aggregation of multiple λ _CDFF OS-ORELM basic prediction models includes:

Update the root mean square error of multiple prediction models corresponding to each stationary subsequence, and select a predetermined number of models as the optimal model set according to the update results;

Assign weights to each prediction model in the optimal model set, calculate the prediction value of each stationary subsequence, and determine the final prediction value according to the prediction value of each stationary subsequence.

Wherein, after determining the final predicted value according to the predicted value of each stationary subsequence, it includes:

After obtaining the wind speed time series at the next moment, update the initial training data set according to the real value at this moment to obtain an updated training data set;

According to the real value at this moment, calculate the prediction error of each prediction model in the optimal model set, if there is a prediction model with a prediction error greater than the error threshold, then establish a training model through the updated training data set, replace A prediction model with a prediction error greater than the error threshold in the optimal model set.

Specifically, in this embodiment, the specific steps for obtaining wind speed point prediction by weighting using the λ _CDFF OS-ORELM basic prediction model and OEOA (onlineensemble using ordered aggregation) technology are as follows:

41) Multi-model online selection, OEOA uses the initial training data set D ₀ to establish M _max models f _m (m=1,...,M _max ), forming a model set ε. For each model, the training root mean square error is updated using Equation (29) during ensemble learning.

in, is the root mean square error of the mth model training at time t, and l _m is the number of times used in the process of predicting each sample data using the model f _m .

Sort in ascending order according to the mean square error obtained from each model training, adaptively select the first B models of ε as the optimal subset, and predict the tth (t=T ₀ +1,...,T) sample data x _t Corresponding wind speed forecast value. When the t+1th new sample arrives, the true value of the tth sample is added to the training data set, and the old data sample removed from the training dataset, i.e.

42) Model ensemble learning, OEOA selects the best B models from the model set ε for ensemble learning to predict new samples. In the process of ensemble learning, it is necessary to assign weights to the B models, and the model weights are obtained by calculating the mean square error, and the calculation formula is shown in (31).

in, For the m-th model training root mean square error at time t-1, The root mean square error vectors are trained for the B optimal models, and median(E ^t-1 ) represents the median of the vector E ^t-1 .

Finally, the predicted value of the new sample data is

43) Model update, before predicting the samples at time t+1, first use the real value of the sample data at time t to verify the prediction error of the model, and determine whether to perform the model set ε at this time by setting the error threshold α renew. when satisfied

When , a new training model f _new is established according to the new training set D ^t formed by formula (30), and f _new is used to replace the model with the largest ε prediction error. When formula (33) is not satisfied, keep the model set ε unchanged.

It can be seen that in this scheme, considering the non-stationarity of the original wind speed time series, the original wind speed is divided into several stationary wind speed sub-sequences by using BGA, and each sub-sequence is decomposed and reconstructed by SAVMD to form a series of sub-patterns with similar frequencies. Then use λ _CDFF OS-ORELM-OEOA to model and predict each sub-mode, and model parameters are dynamically adjusted through CSO-SAM to obtain a short-term wind speed multi-step prediction model. Compared with other prediction methods, this prediction method shows good prediction performance under different signal decomposition techniques (SAVMD and EEMD (ensemble empirical mode decomposition)), the specific performance is as follows:

1) Using SAVMD technology to decompose the original wind speed time subsequence, compared with VMD technology, SAVMD can adaptively determine the number of sub-signals;

2) CSO-SAM has a strong search ability and can determine the optimal parameters of the model to improve the prediction accuracy;

3) The online ensemble learning prediction model uses the forgetting factor based on Cook distance to track wind speed changes in real time to adjust model parameters; and uses OEOA technology to select the optimal sub-model, and obtains the final prediction value through weighted aggregation to further improve the wind speed prediction accuracy.

Specifically, refer to Fig. 3. Fig. 3 is a flow chart of realizing a short-term wind speed multi-step prediction method based on online robust extreme learning machine adaptive integrated learning provided by this embodiment; in order to describe this scheme in detail, an application is provided The concrete embodiment of this scheme:

The measured wind speed time series of a wind farm in 2004 provided by NREL is selected as the research object. In order to reduce the influence of seasons on wind speed prediction, April, July, October and January are selected as the typical months of each season, and different prediction models are established respectively. . The sampling frequency of wind speed data is 10 minutes, and the number of sample points in four months is 4320, 4464, 4464 and 4464 respectively. The time series of wind speed is shown in Figure 4-7. One-step and five-step forecasts. Using BGA to divide the original wind speed time series into multiple subsequences, the mean and variance are shown in Table 1.

Table 1 Segmentation results of spring wind speed time series in 2004

The last 100 points of each wind speed subseries are used as the validation set, and the rest are used as the training set to train the wind speed model. Compare the prediction results of this prediction model with the prediction results of λ _CDFF OS-ORELM-OEOA prediction model, λ _CDFF OS-ORELM prediction model, BPNN, offline-ELM, offline-ORELM and persistence model (persistence model, PM), At the same time, the above prediction models based on SAVMD and EEMD were established respectively, and the prediction results were compared.

In order to qualitatively evaluate the performance of the prediction model, the mean absolute error (mean absolute error, MAE), mean absolute percentage error (mean absolute percentage error, MAPE) and root mean square error (root mean-squared error, RMSE) are introduced as the prediction model. Performance evaluation indicators:

In the above formula, with are the actual and predicted values of wind speed at the tth sample point, respectively; N is the number of test samples.

Taking the wind speed time series in April 2004 as an example, the original wind speed is divided into nine subsequences by using BGA. The mean and variance of each subsequence wind speed are shown in Table 1. The fifth subsequence (S5) is selected for analysis, the first 551 sample points are used as the training data set, and the last 100 points are used as the verification sample set.

The parameters of the λ _CDFF OS-ORELM-OEOA model are set as follows: the number of initial training sets T ₀ =20; the error threshold ρ=0.04; the minimum value of the forgetting factor λ _min =0.6; M _min =10; M _max =15.

The parameters of CSO-SAM are set as follows: the number of populations is 20; the maximum number of iterations is 100; according to empirical values, the maximum and minimum values of the vertical crossover probability are P _vmax = 0.8 and P _vmin = 0.2, respectively.

SAVMD decomposes the original sub-wind speed time series of S5 into 5 sub-signals, and EEMD decomposes S5 into 8 sub-signals, establishes a prediction model for each sub-signal separately according to the above prediction method, and finally sums the predicted values of each sub-signal to obtain The final wind speed forecast. Table 2 shows the multi-step error evaluation indexes of S5 without signal decomposition, and the multi-step model error evaluation indexes after SAVMD and EEMD decomposition are shown in Tables 3 and 4, respectively.

Table 2 Prediction results of wind speed without signal decomposition

Table 3 Wind speed prediction results based on SAVMD technology

Table 4 Wind speed prediction results based on EEMD technology

Comparing Table 2-Table 4, we can see that:

(1) After the wind speed time series is processed by SAVMD and EEMD technology, the prediction accuracy of various prediction models has been greatly improved. Taking the method in this paper as an example, after using the signal decomposition technology, the single-step prediction E _MAPE has increased by 17.07% and 9.76% respectively;

(2) Among all the forecasting models, the multi-step forecasting performance of the model proposed in this paper is better than other models, among which the PM model has the worst forecasting performance.

Further analysis from the above data shows that:

(1) CSO-SAM has a strong ability to optimize the parameters of the ORELM model;

(2) Compared with the traditional offline model, the online prediction model in this paper can change the parameters in real time to adapt to the change of wind speed, so as to ensure the strong robustness of the prediction model;

(3) The forgetting factor (λ _CDFF ) based on Cook's distance weakens the influence of old data samples during the entire online prediction process, and uses new data samples to update model parameters to improve the accuracy of the prediction model.

Similarly, other subsequences formed after BGA cutting in Table 1 can be predicted by the above method, and the prediction results are shown in Figures 8-10. From the error evaluation indicators in Figure 8-10, it can be seen that the prediction accuracy of the subsequence segmented by the BGA algorithm is higher than that of the complete sequence, which indicates the effectiveness of using BGA in data preprocessing. Applying the λ _CDFF OS-ORELM-OEOA model in this paper to the four seasons in 2004, the single-step error evaluation indicators of the model are shown in Table 5.

Table 5 The prediction results of different seasons obtained by the method in this paper

It can be seen from Table 5 that the single-step error evaluation index values in each season are not much different, and the prediction performance in spring is slightly worse than that in other seasons, indicating that the model in this paper can reflect the wind speed change caused by seasonal changes and reflects the seasonal characteristics of wind speed prediction .

The short-term wind speed multi-step forecasting device provided by the embodiment of the present invention is introduced below. The short-term wind speed multi-step forecasting device described below and the short-term wind speed multi-step forecasting method described above can be referred to each other.

Referring to Figure 11, a short-term wind speed multi-step forecasting device provided by an embodiment of the present invention includes:

Stationary subsequence segmentation module 100, for utilizing the heuristic segmentation algorithm BGA to divide the original wind speed time series into a plurality of stationary subsequences;

The decomposition module 200 is used to decompose each stationary subsequence into a series of limited-bandwidth sub-patterns by using the adaptive variable mode decomposition technique SAVMD;

The prediction model building module 300 is used to establish a basic prediction model for each series of sub-patterns through the robust online extreme learning machine λ _CDFF OS-ORELM based on the forgetting factor of COOK distance, and adopts a crossover with an adaptive mutation mechanism The optimization algorithm CSO-SAM optimizes the parameters of the prediction model;

The prediction module 400 is configured to use the online ensemble learning and ordered aggregation technology OEOA to obtain the final prediction value through weighted aggregation of multiple λ _CDFF OS-ORELM basic prediction models.

Based on the foregoing embodiments, the stable subsequence segmentation module 100 includes:

Statistical value determination unit, used to determine the left average value and right average value of other target points in the original wind speed time series except the starting point and the end point, and determine each target according to the left average value and right average value of each target point point statistics;

The segmentation unit is used to calculate the statistical significance of each target point according to the statistical value of each target point, and use the target point whose statistical significance is greater than a predetermined threshold as a segmentation point, and use the segmentation point to calculate the original wind speed time The sequence is divided into multiple stationary subsequences.

Wherein, the stable subsequence segmentation module 100 also includes:

A judging module, configured to judge whether the divided left and right sequences are both greater than the minimum segmentation length;

The segmentation unit is configured to segment the original wind speed time series when both the divided left and right part sequences are greater than the minimum segmentation length; otherwise, cancel the division of the original wind speed time series.

Based on the foregoing embodiments, the decomposition module 200 includes:

The mode decomposition module is used to determine the modal number K for decomposing the target stationary subsequence by using the empirical mode decomposition technique EMD, and decompose the target stationary subsequence into K modes by the variable mode decomposition technique VDM;

The denoising and reconstruction module is used to determine the fluctuation parameters of each mode by DFA, and perform denoising and reconstruction on the target stationary subsequence according to the fluctuation parameters of each mode.

Based on the foregoing embodiments, the prediction model building module 300 includes:

The initial prediction model determination unit is used to create an initial prediction model by using the initial training data set, determine the number of hidden layer nodes of the initial prediction model by the 10-fold cross-validation method, and calculate the model parameters; the model parameters include the hidden layer output matrix and the initial output weight;

The basic prediction model determination unit is used to calculate the COOK distance, determine the forgetting factor λ according to the COOK distance, update the model parameters through the forgetting factor λ, and use the crossover optimization algorithm CSO-SAM with an adaptive mutation mechanism to update the model The parameters are optimized to obtain a basic forecasting model; the basic forecasting model includes multiple forecasting models corresponding to the stationary subsequences.

Based on the foregoing embodiments, the prediction module 400 includes:

The optimal model set determination module is used to update the root mean square error of multiple prediction models corresponding to each stationary subsequence, and select a predetermined number of models as the optimal model set according to the update results;

The final predictive value determination module is configured to assign weights to each predictive model in the optimal model set, calculate the predictive value of each stationary subsequence, and determine the final predictive value according to the predictive value of each stationary subsequence.

Wherein, the prediction module 400 includes:

The update module is used to update the initial training data set according to the real value at this moment after obtaining the wind speed time series at the next moment, so as to obtain the updated training data set;

The optimal model update module is used to calculate the prediction error of each prediction model in the optimal model set according to the real value at this moment. If there is a prediction model with a prediction error greater than the error threshold, the updated A training model is established in the training data set, and a prediction model whose prediction error is greater than the error threshold in the optimal model set is replaced.

Each embodiment in this specification is described in a progressive manner, each embodiment focuses on the difference from other embodiments, and the same and similar parts of each embodiment can be referred to each other.

The above description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention will not be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims (10)

1. A short-term wind speed multi-step forecasting method is characterized in that, comprising: Using the heuristic segmentation algorithm BGA to segment the original wind speed time series into multiple stationary subsequences; Use the adaptive variable mode decomposition technique SAVMD to decompose each stationary subsequence into a series of submodes with limited bandwidth; Through the robust online extreme learning machine λ _CDFF OS-ORELM based on the forgetting factor of COOK distance, a basic prediction model is established for each series of sub-patterns, and the cross-cutting optimization algorithm CSO-SAM with an adaptive mutation mechanism is used to predict the model Carry out parameter optimization; Using online ensemble learning and ordered aggregation technology OEOA, the final prediction value is obtained by weighted aggregation of multiple λ _CDFF OS-ORELM basic prediction models. 2. the short-term wind speed multi-step prediction method according to claim 1, is characterized in that, described utilizes heuristic segmentation algorithm BGA to divide original wind speed time series into a plurality of stable subsequences, comprising: Determine the left average value and right average value of other target points in the original wind speed time series except the starting point and the end point, and determine the statistical value of each target point according to the left average value and right average value of each target point; Calculate the statistical significance of each target point according to the statistical value of each target point, use the target point whose statistical significance is greater than a predetermined threshold as a segmentation point, and use the segmentation point to segment the original wind speed time series to form Multiple stationary subsequences. 3. short-term wind speed multi-step forecasting method according to claim 2, is characterized in that, described utilizing described segmentation point, after described original wind speed time series is segmented, also comprises: Judging whether the divided left and right parts of the sequence are both greater than the minimum segmentation length; if not, cancel the segmentation of the original wind speed time series. 4. short-term wind speed multi-step forecasting method according to claim 1, is characterized in that, described utilizing self-adaptive variable mode decomposition technique SAVMD to decompose each stable subsequence into a series of limited-bandwidth submodes, comprising: Utilize the empirical mode decomposition technique EMD to determine the modal number K that decomposes the target stationary subsequence, and decompose the target stationary subsequence into K modes by the variable mode decomposition technique VDM; The fluctuation parameters of each mode are determined by the detrending fluctuation analysis method DFA, and the target stationary subsequence is denoised and reconstructed according to the fluctuation parameters of each mode. 5. short-term wind speed multi-step prediction method according to claim 1, is characterized in that, described by the robust online extreme learning machine λ _CDFF OS-ORELM of the forgetting factor based on COOK distance, establishes basic for each series sub-pattern Forecast model, and use the cross-cross optimization algorithm CSO-SAM with adaptive mutation mechanism to optimize the parameters of the forecast model, including: Utilize the initial training data set to create an initial prediction model, determine the number of hidden layer nodes of the initial prediction model by a 10-fold cross-validation method, and calculate model parameters; the model parameters include a hidden layer output matrix and initial output weights; Calculate the COOK distance, determine the forgetting factor λ according to the COOK distance, update the model parameters through the forgetting factor λ, and optimize the model parameters through the crossover optimization algorithm CSO-SAM with an adaptive mutation mechanism to obtain the basic prediction model ; The basic forecasting model includes a plurality of forecasting models corresponding to the stationary subsequence. 6. the short-term wind speed multi-step forecasting method according to claim 5, is characterized in that, described utilizes online integrated learning and ordered aggregation technology OEOA, obtains final forecast by weighted aggregation of a plurality of λ _CDFF OS-ORELM basic forecasting models values, including: Update the root mean square error of multiple prediction models corresponding to each stationary subsequence, and select a predetermined number of models as the optimal model set according to the update results; Assign weights to each prediction model in the optimal model set, calculate the prediction value of each stationary subsequence, and determine the final prediction value according to the prediction value of each stationary subsequence. 7. short-term wind speed multi-step prediction method according to claim 6, is characterized in that, after the described predicted value according to each smooth subsequence determines final predicted value, comprises: After obtaining the wind speed time series at the next moment, update the initial training data set according to the real value at this moment to obtain an updated training data set; According to the real value at this moment, calculate the prediction error of each prediction model in the optimal model set, if there is a prediction model with a prediction error greater than the error threshold, then establish a training model through the updated training data set, replace A prediction model with a prediction error greater than the error threshold in the optimal model set. 8. A short-term wind speed multi-step forecasting device, characterized in that it comprises: The smooth subsequence segmentation module is used for utilizing the heuristic segmentation algorithm BGA to divide the original wind speed time series into a plurality of stationary subsequences; Decomposition module for decomposing each stationary subsequence into a series of limited-bandwidth sub-modes using the adaptive variable mode decomposition technique SAVMD; Predictive model building module for building a basic predictive model for each series of sub-patterns by a robust online extreme learning machine λ _CDFF OS-ORELM based on the COOK distance-based forgetting factor, and adopting cross-cutting optimization with an adaptive mutation mechanism The algorithm CSO-SAM optimizes the parameters of the prediction model; The prediction module is used to obtain the final prediction value through weighted aggregation of multiple λ _CDFF OS-ORELM basic prediction models by using online ensemble learning and ordered aggregation technology OEOA. 9. short-term wind speed multi-step forecasting device according to claim 8, is characterized in that, described stable subsequence segmentation module comprises: Statistical value determination unit, used to determine the left average value and right average value of other target points in the original wind speed time series except the starting point and the end point, and determine each target according to the left average value and right average value of each target point point statistics; The segmentation unit is used to calculate the statistical significance of each target point according to the statistical value of each target point, and use the target point whose statistical significance is greater than a predetermined threshold as a segmentation point, and use the segmentation point to calculate the original wind speed time The sequence is divided into multiple stationary subsequences. 10. short-term wind speed multi-step forecasting device according to claim 8, is characterized in that, described prediction model building module comprises: The initial prediction model determination unit is used to create an initial prediction model by using the initial training data set, determine the number of hidden layer nodes of the initial prediction model by the 10-fold cross-validation method, and calculate the model parameters; the model parameters include the hidden layer output matrix and the initial output weight; The basic prediction model determination unit is used to calculate the COOK distance, determine the forgetting factor λ according to the COOK distance, update the model parameters through the forgetting factor λ, and use the crossover optimization algorithm CSO-SAM with an adaptive mutation mechanism to model the model The parameters are optimized to obtain a basic forecasting model; the basic forecasting model includes multiple forecasting models corresponding to the stationary subsequences.