CN107169612A

CN107169612A - The prediction of wind turbine active power and error revising method based on neutral net

Info

Publication number: CN107169612A
Application number: CN201710473103.3A
Authority: CN
Inventors: 彭丽霞; 杜杰; 孙泓川; 王雷; 陆金桂; 曹家; 曹一家; 朱伟军; 曾刚
Original assignee: Nanjing University of Information Science and Technology
Current assignee: Nanjing University of Information Science and Technology
Priority date: 2017-06-21
Filing date: 2017-06-21
Publication date: 2017-09-15

Abstract

The invention discloses a kind of wind turbine active power prediction based on neutral net and error revising method, Wavelet spectral analysis method is used first, extract the harmonic compoment sequence implied in wind turbine active power time series and isolated residual sequence, then harmonic compoment sequence and residual sequence are predicted using neural network model respectively, wherein harmonic compoment sequence uses the BP neural network corrected based on particle cluster algorithm optimization and overlay error to be predicted, the RBF neural that residual sequence is optimized using particle cluster algorithm and overlay error is corrected is predicted, it can be obtained final wind turbine active power by predicting the outcome for harmonic compoment sequence and residual sequence and predicted the outcome.The present invention is realized to carrying out fine forecast per typhoon motor active power in wind power plant, so as to effectively improve the short-term forecast level of exerting oneself of whole wind power plant.

Description

The prediction of wind turbine active power and error revising method based on neutral net

Technical field

The invention belongs to wind energy forecast field, and in particular to it is a kind of based on neutral net wind turbine active power prediction and Error revising method.

Background technology

In effectively wind energy is connected to the grid, it is extremely necessary and closes for the progress accurate forecast of exerting oneself of wind power plant Key, among these, the short-period forecast of 0 to 6 hour is for power network Real-Time Scheduling, it is ensured that mains frequency, power and balance of voltage etc. The technical parameter for being related to power grid security is significant.

Wind energy is as a kind of reproducible clean energy resource, and with installation scaleable, wind power generator group reliability is high, make The advantages of valency is low, operation maintenance is simple.Announced according to 2 months 2015 National Energy Boards《Wind Power Generation Industry monitors situation within 2014》, By the end of the year 2014, China's wind-powered electricity generation adds up installed capacity and has reached 96,370,000 kilowatts, accounts for the 7% of whole capacities of installed generator, accounts for The 27% of global wind-powered electricity generation installation.The kilowatt hour of wind-powered electricity generation electricity volume 153,400,000,000 in 2014, accounts for the 2.78% of whole generated energy.2014 National Energy Board's issue in December《Energy development Strategic Action Plan (2014-2020)》, it is contemplated that to the year two thousand twenty, wind-powered electricity generation installation will Reach 200,000,000 kilowatts.So far, wind-powered electricity generation turns into the third-largest main force's power supply of China after thermoelectricity and water power.With installation Capacity is continuously increased, and the electric problem of abandoning of wind-powered electricity generation is more protruded always, is counted according to National Energy Board, wind-powered electricity generation amount is abandoned in the whole nation within 2012 About 20,000,000,000 kilowatt hours, averagely abandon wind rate and reach 17%；The kilowatt hour of wind-powered electricity generation amount about 15,000,000,000 is abandoned in the whole nation in 2013, is averagely abandoned wind rate and is reached To 10%, newest statistics is shown, by by the end of September, 2014, and wind-powered electricity generation abandons the kilowatt hour of wind-powered electricity generation amount 8,600,000,000, averagely abandons wind rate 7.5%. A major reason for causing wind-powered electricity generation to abandon electricity is that the intermittence of wind causes the fluctuation and unstability of wind-powered electricity generation to have impact on wind-powered electricity generation Quality, electricity is abandoned in order to ensure the safety of power network in vain.Based on this, National Energy Board issued in 2011《Wind power is pre- Survey forecast management Tentative Measures》, it is desirable to all wind parks being incorporated into the power networks of China should be set up before 1 day January in 2012 and blow Electric prediction system and generation schedule declaration work mechanism simultaneously start trial operation, report and submit wind power prediction to forecast as requested As a result.

The active power forecast common method of current wind power plant includes physical method and statistical method, and physical method refers to root The timing of high-spatial and temporal resolution, fixed point, quantitative numerical weather prediction model wind are obtained according to the numerical weather prediction model that becomes more meticulous Power predicts output result, while running actual conditions according to wind electric field blower, considers various wind turbine power generation influence factors, builds Vertical prediction physical model of exerting oneself, carries out output of wind electric field prediction.Physical method does not need substantial amounts of measurement data, but requires to big The physical characteristic of gas and the characteristic of wind power plant wind turbine have accurate mathematical description, and these equation solutions are difficult, required data sea Amount, computationally intensive, the calculating time is long, and the difficulty for obtaining data from meteorological department is big, costly, therefore in short-term wind-electricity In active power forecast, statistical method is still commonly used.At present, statistical method is adopted mostly according to the historical summary of wind power plant anemometer tower With the methods such as lasting method, Random time sequence method, Kalman filtering method, neural network, SVMs using wind power plant as One entirety is predicted, and such a method disadvantage is that wind power plant is influenceed by landform, turbulent flow and wind turbine operating mode etc., must Larger prediction error will be caused, this is unrelated with specific forecasting procedure.With e measurement technology and computer computation ability Improve, the active power to separate unit generator cabin become more meticulous carries out forecast and is possibly realized.

The content of the invention

In view of the above-mentioned problems, the present invention proposes a kind of wind turbine active power prediction based on neutral net and error revising Method, by carrying out fine forecast to the active power in wind power plant per typhoon motor, so as to effectively improve whole wind power plant Short-term forecast level of exerting oneself.

Above-mentioned technical purpose is realized, above-mentioned technique effect is reached, the present invention is achieved through the following technical solutions：

It is a kind of based on neutral net wind turbine active power prediction and error revising method, it is characterised in that including with Lower step：

(1) the crude sampling active power time series p={ p (i), i=1,2 ..., N } of wind turbine is read in, wherein N is Original wind turbine active power sampled point number；P is adjusted to the average active power time series p ' by forecast space requirement ={ p ' (j), j=1,2 ..., M }, wherein M are by the wind turbine average active power sequence after forecast space requirement adjustment Sampled point number, p ' average value isOrder

(2) multi-scale wavelet power spectrum analysis method is used, the harmonic compoment sequence { P in P is extracted₁,P₂,…,P_k,…, P_K, wherein K is the number of the harmonic compoment sequence in P, P_k={ P_k(1),P_k(2),…,P_k(M) }, thus P=P₁+P₂+…+P_K + R, wherein R=P-P₁-P₂-…-P_K, it is that the residual sequence after harmonic compoment sequence is rejected in P；

(3) to the harmonic compoment sequence { P in P₁,P₂,…,P_k,…,P_K, particle group optimizing and overlay error is respectively adopted The BP neural network corrected is predicted, and sets prediction step as l, then each harmonic compoment sequence { P₁,P₂,…,P_k,…,P_K Predict the outcome forWherein

(4) the RBF nerve nets that the first-order difference sequence D to residual sequence R is corrected using particle group optimizing and overlay error Network is predicted, and predicting the outcome for residual sequence is obtained by difference inverse operation, set prediction step as l, then residual sequence R Predict the outcome as Y_R={ Y_R(1),Y_R(2),...,Y_R(l)}；

(5) willIt is added with each periodic signal sequence, predicting the outcome for residual sequence, obtains the final Y that predicts the outcome,

Further, the detailed process of the BP neural network of the particle group optimizing employed in the step (3) is：

(1) according to Kolmogorov theorems, 3 layers of BP neural network model are set up, it is hidden if input layer number is I Neuron number containing layer is H, and output layer neuron number is O；Wherein, H=2*I+1, O=1；

(2) parameter for needing to optimize is determined, including：The input layer number I and the length of training set of BP neural network L is spent, in addition to：One group objects W=(w (1), w (2) ..., w (q)), q=I*H+H*O+H+O, wherein, w (1)~w (I*H) is The input layer of BP neural network is to the link weights of hidden layer neuron, and w (I*H+1)~w (I*H+H*O) is BP neural network Hidden layer is to the link weights of output layer neuron, and w (I*H+H*O+1)~w (I*H+H*O+H) is BP neural network hidden layer god Threshold value through member, w (I*H+H*O+H+1)~w (I*H+H*O+H+O) is the threshold value of BP neural network output layer neuron；

(3) population is initializedWherein Q₁For the sum of particle, i-th of particle is X_i=(I_i, W_i,L_i), particle rapidity is V_i=(v_I_i,v_W_i,v_L_i), wherein I_i、W_i、L_iFor parameter I, W, L, mono- group alternatively solves；

(4) to each particle X in colony_i=(I_i,W_i,L_i) determine parameter, construction BP neural network training set it is defeated Enter and output matrix, wherein for harmonic compoment sequence P_kAnd BP neural network input layer number I_iInitially set up matrix Z₁And Z₂, wherein：

For neural metwork training collection length L, Z to be optimized₁In last L_iArrange the input matrix I as training set_train, Z₂In last L_iArrange the output matrix O as training set_train；It regard forecast step-length l as test step-length, Z₁In last l row make For the input matrix I of test set_test, Z₂In last l arrange output matrix O as test set_test；Constructed according to training set BP neural network to the error sum of squares of test set analog result as its fitness value, with the minimum optimization direction of fitness value The quality of each particle, record particle X are judged as evaluation criterion_iCurrent individual extreme value is P_best(i) P in colony, is taken_best(i) Optimal individual is used as overall extreme value G_best；

(5) each particle X in colony_i, its position and speed are updated respectively；

In formula：ω is inertia weight, c₁、c₂For acceleration factor, g is current iteration number of times, and r₁、r₂To be distributed in [0, 1] random number；

(6) target function value of each particle now is recalculated, P is updated_bestAnd G (i)_best；

(7) judge whether to reach maximum iteration, terminate optimization process if meeting, obtain BP neural network and most preferably join Number I_best、W_best(w_best(1),w_best(2),...,w_best(q))、L_best, otherwise return to step (4).

Further, the detailed process of the overlay error correction method employed in the step (3) is：

(1) first by the BP neural network parameter I of optimization_best、W_best(w_best(1),w_best(2),...,w_best(q))、 L_bestConstructing neural network training set Z₃With test set Z₄And BP neural network link weights and threshold value are initialized, wherein：

w_best(1)~w_best(I*H) the initial of weights is linked to hidden layer neuron for the input layer of BP neural network Value, w_best(I*H+1)~w_best(I*H+H*O) the first of weights is linked to output layer neuron for the hidden layer of BP neural network Initial value, w_best(I*H+H*O+1)~w_best(I*H+H*O+H) for BP neural network hidden layer neuron threshold value initial value, w_best(I*H+H*O+H+1)~w_best(I*H+H*O+H+O) for BP neural network output layer neuron threshold value initial value；

(2) defined nucleotide sequence P_kIt is c=in sampled point i rate of change | p_k(i)-p_k(i-1) |, for prediction step l, calculate Z₄The maximum rate of change of last l step-lengths is c_max, the l step predictions that just can be iterated afterwards, during prediction, with c_maxMake The output valve of neutral net during for jth step-ahead predictionWith previous predicted value or actual valueMaximum change Rate, ifThenIt is used as the predicted value of jth step-length；If otherwiseThenIfThen

Further, the detailed process of the particle group optimizing RBF neural employed in the step (4) is：

(1) determine to need Optimal Parameters, including：RBF neural input layer number I and training set length L；

(2) population is initializedWherein Q₂For the sum of particle, i-th of particle is X_i=(I_i, L_i), particle rapidity isWherein I_i,L_iFor parameter I, L, mono- group alternatively solves；

(3) each particle X in colony_i(I_i,L_i) determine parameter, construct RBF neural training set input And output matrix, wherein for residual sequence R and RBF neural input layer number I_iInitially set up matrix Z₅And Z₆, Wherein：

For neural metwork training collection length L, Z to be optimized₅In last L_iArrange the input matrix I as training set_train, Z₆In last L_iArrange the output matrix O as training set_train；It regard forecast step-length l as test step-length, Z₅In last l row make For the input matrix I of test set_test, Z₆In last l arrange output matrix O as test set_test；Constructed according to training set RBF neural to the error sum of squares of test set analog result as its fitness value, with the minimum optimization side of fitness value To the quality that each particle is judged as evaluation criterion, record particle X_iCurrent individual extreme value is P_best(i) P in colony, is taken_best (i) optimal individual is used as overall extreme value G_best；

(4) each particle X in colony_i, its position and speed are updated respectively；

(5) target function value of each particle now is recalculated, P is updated_bestAnd G (i)_best；

(6) judge whether to reach maximum iteration, terminate optimization process if meeting, obtain RBF neural optimal Parameter I_bestAnd L_best, otherwise return to step (3).

Further, the detailed process of the overlay error correction method employed in the step (4) is：

First by the RBF neural parameter I of optimization_bestAnd L_bestConstructing neural network training set Z₇With test set Z₈And Initialize BP neural network and link weights and threshold value, wherein：

(2) defined nucleotide sequence R is c=in sampled point i rate of change | R (i)-R (i-1) |, for prediction step l, calculate Z₈ The maximum rate of change of last l step-lengths is c_max, the l step predictions that just can be iterated afterwards, during prediction, with c_maxAs The output valve Y of neutral net during jth step-ahead prediction_R(j) with previous predicted value or actual value Y_R(j-1) maximum rate of change, if |Y_R(j)-Y_R(j-1)|<c_max, then Y_R(j) as the predicted value of jth step-length；If otherwise Y_R(j)>Y_R(j-1), then Y_R(j)=Y_R (j-1)+c_maxIf, Y_R(j)<Y_R(j-1), then Y_R(j)=Y_R(j-1)-c_max。

Further, inertia weight ω=0.5, acceleration factor c₁=c₂=1.49445.

Beneficial effects of the present invention：

(1) the harmonic compoment sequence for the wind turbine active power extracted through multiscale analysis is due to regular strong, Er Qie Proportion is larger in original work(power sequence, therefore has established the basis of degree of precision prediction；Eliminate harmonic compoment sequence Residual sequence afterwards is little due to the proportion in entirety, therefore its predicated error is relatively limited, therefore proposed by the invention By original active power sequence through multiscale analysis, harmonic compoment sequence and residual sequence are decomposed into, and then to each sequence point The thinking not being predicted can greatly improve overall prediction effect.

(2) because synoptic process change is continuous, and not mutated, this is showed extremely in harmonic compoment sequence Protrude, the although smooth smooth-going of harmonic compoment sequence, but the amplitude and phase of wherein sequence are changed over time substantially, and this can influence god Study and Generalization Capability through network, cause to predict value mutation, precision of prediction declines, continuity of the application based on Changes in weather Feature, by maximum rate of change index come the behavior of constrained forecast value mutation, through experiment, this technology can be effectively improved nerve net The catastrophic behavior of network predicted value, improves precision of prediction, and the method is also applied in the error revising of residual sequence simultaneously.

(3) influence differed for prediction performance is selected for neural network structure, the present invention proposes have for wind turbine The characteristics of harmonic compoment sequence that work(power sequence is extracted and the residual sequence of separation, BP neural network and RBF god is respectively adopted Through network, for the structural parameters of neutral net, training set scale carries out excellent method using particle cluster algorithm, significantly improves god Generalization Capability through network, finally improves precision of prediction.

Brief description of the drawings

Fig. 1 is flow chart of the method for the present invention；

Fig. 2 is minute rank wind turbine active power sequence；

Fig. 3 is sequence P Wavelet spectral analysis result figure；

Fig. 4 is the residual sequence figure of the harmonic compoment sequence that sequence P is extracted and separation；

Fig. 5 (a) is the one-step prediction result of the inventive method；

Fig. 5 (b) predicts the outcome for two steps of the inventive method；

Fig. 5 (c) predicts the outcome for three steps of the inventive method；；

Fig. 6 (a) is that maximum rate of change index is directed to harmonic compoment sequence P₅Predict the outcome contrast；

Fig. 6 (b) is that maximum rate of change index predicts the outcome contrast for residual error R；

Fig. 7 (a) is to set up particle cluster algorithm optimization RBF neural one for minute rank wind turbine active power sequence Step predicts the outcome；

Fig. 7 (b) is to set up particle cluster algorithm optimization RBF neural two for minute rank wind turbine active power sequence Step predicts the outcome；

Fig. 7 (c) is to set up particle cluster algorithm optimization RBF neural three for minute rank wind turbine active power sequence Step predicts the outcome；

Fig. 8 is minute rank wind turbine active power sequence db3 wavelet decomposition results in contrast experiment；

Fig. 9 (a) is to fix wavelet decomposition-RBF neural one-step prediction knot of RBF neural parameter in contrast experiment Really；

Fig. 9 (b) is to fix wavelet decomposition-RBF neural one-step prediction knot of RBF neural parameter in contrast experiment Really；

Fig. 9 (c) is to fix wavelet decomposition-RBF neural one-step prediction knot of RBF neural parameter in contrast experiment Really；

Figure 10 (a) is the wavelet decomposition-RBF god in contrast experiment using particle cluster algorithm optimization RBF neural parameter Through network one-step prediction result；

Figure 10 (b) is the wavelet decomposition-RBF god in contrast experiment using particle cluster algorithm optimization RBF neural parameter Predicted the outcome through the step of network two；

Figure 10 (c) is the wavelet decomposition-RBF god in contrast experiment using particle cluster algorithm optimization RBF neural parameter Predicted the outcome through the step of network three.

Embodiment

In order to make the purpose , technical scheme and advantage of the present invention be clearer, with reference to embodiments, to the present invention It is further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, it is not used to Limit the present invention.

The application principle of the present invention is explained in detail below in conjunction with the accompanying drawings.

As shown in figure 1, a kind of prediction of wind turbine active power and error revising method based on neutral net, including it is following Step：

S1：The crude sampling active power time series p={ p (i), i=1,2 ..., N } of wind turbine is read in, wherein N is Original wind turbine active power sampled point number；P is adjusted to the average active power time series p ' by forecast space requirement ={ p ' (j), j=1,2 ..., M }, wherein M are by the wind turbine average active power sequence after forecast space requirement adjustment Sampled point number, p ' average value isOrder

S2：Using multi-scale wavelet power spectrum analysis method, the harmonic compoment sequence { P in P is extracted₁,P₂,…,P_k,…, P_K, wherein K is the number of the harmonic compoment sequence in P, P_k={ P_k(1),P_k(2),…,P_k(M) }, thus P=P₁+P₂+…+P_K + R, wherein R=P-P₁-P₂-…-P_K, it is that the residual sequence after harmonic compoment sequence is rejected in P；

The use multi-scale wavelet power spectrum analysis method, extracts the harmonic compoment sequence { P in P₁,P₂,…,P_k,…, P_K, detailed process is：

Assuming that a discrete-time series x_n, wherein n=1 ..., N, common N number of sampled point, sampling time interval δ t=1 are answered Morlet wavelet transformations are used, the harmonic compoment of the time series is analyzed, the corresponding time series of each harmonic compoment band is extracted.

2.1 determine analytical cycle

The cycle T of wavelet transformation_j(Fig. 3 abscissa values) and wavelet analysis mesoscale parameter s_jIt is relevant, it is considered to which that mother Morlet is small Ripple center cyclophysis, herein T_j=s_j.The selection of scale parameter is：s_j=2^jδj, wherein δ j=1/4, j=0,1 ..., J, altogether The maximum of J+1 yardstick, wherein J is no more than J_max=4log₂(N), J=48 here.

2.2 determine global wavelet transformation spectrum

In the n-th sampled point, s_jThe corresponding local wavelet transformation spectrum W of scale parameter_n(s_j) be：

Wherein ψ^*() is ψ () conjugate function, to the n-th sampled point, scale parameter s_jMorlet small echo mother wavelet letters Number is：

To W_n(s_j) mould | W_n(s_j) | integrated along whole sampling interval, obtain scale parameter s_jCorresponding global wavelet transformation SpectrumI.e.：

The present invention is using the global wavelet transformation spectrum standardized(Fig. 3 solid lines), wherein σ²For x_nSide Difference.

2.3 global wavelet transformation spectrum significance tests

The maximum corresponding cycle of global wavelet transformation spectrum curve is generally set to the major cycle, but whether significantly, To pass through significance test.Here global wavelet transformation spectrum obtained as above and red noise spectrum are compared, judge that it shows Work property, wherein red noise spectrum Q_kIt is expressed as：

Wherein α is x_nTime series falls behind the auto-correlation coefficient of a sampled point, k=0,1 ..., N/2.

Assuming that global wavelet transformation spectrum is a certain acyclic process spectrum, then its defer to the ratio between with red noise spectrum by from Removed by spending νDistribution：

The wherein free degreeγ is the decorrelation factor, to Morlet small echos, γ=2.32.Herein 0.05 significance is taken, whenWhen, the global wavelet transformation spectrum corresponding cycle be it is significant, its InFor Fig. 3 dotted lines.

2.4 extract the corresponding time series of harmonic compoment band

Extract certain specific period band i.e. [T₁,T₂] corresponding time series x '_n, specific period band [T is known by (1)₁, T₂] correspondence Scale parameter beFor Morlet small echos, extractThe corresponding time series of scale parameter is to this Scale parameter band correspondenceSummed, i.e.,：

Wherein, whereinFor W_n(s_j) real part, for Morlet small echos, ψ₀(0)=π^-1/4, C_δ=0.776.

S3：To the harmonic compoment sequence { P in P₁,P₂,…,P_k,…,P_K, particle group optimizing and overlay error is respectively adopted The BP neural network corrected is predicted, and sets prediction step as l, then each harmonic compoment sequence { P₁,P₂,…,P_k,…,P_K Predict the outcome forWherein

The detailed process of the BP neural network of particle group optimizing employed in the step S3 is：

(2) parameter for needing to optimize is determined, including：The input layer number I and the length of training set of BP neural network L is spent, in addition to：One group objects W=(w (1), w (2) ..., w (q)), q=I*H+H*O+H+O, wherein, w (1)~w (I*H) is The input layer of BP neural network is to the link weights of hidden layer neuron, and w (I*H+1)~w (I*H+H*O) is BP nerves The hidden layer neuron of network is to the link weights of output layer neuron, and w (I*H+H*O+1)~w (I*H+H*O+H) is BP nerves The threshold value of network hidden layer neuron, w (I*H+H*O+H+1)~w (I*H+H*O+H+O) is BP neural network output layer neuron Threshold value；

The detailed process of overlay error correction method employed in the step S3 is

The RBF neural that S4 is corrected to residual sequence R first-order difference sequence D using particle group optimizing and overlay error It is predicted, predicting the outcome for residual sequence is obtained by difference inverse operation, set prediction step as l, then residual sequence R's is pre- Survey result is Y_R={ Y_R(1),Y_R(2),...,Y_R(l)}；

The detailed process of particle group optimizing RBF neural employed in the step S4 is：

The detailed process of overlay error correction method employed in the step (4) is：

Inertia weight ω=0.5, acceleration factor c₁=c₂=1.49445.

Specific test case：

The flow chart shown by Fig. 1, take China's wind power plant typhoon motor from 5 days 10 October in 2015 when 07 point 53 Second starts the original active power time series of the second rank of collection, because the ultra-short term for minute rank that this example shows is pre- Report, therefore pressed first by forecast space requirementOriginal active power time series is adjusted to obtain The average active power time series of minute rank is obtained, as shown in Fig. 2 wherein p (i) is the original active power that wind turbine is gathered Time series (second level, but sampling interval disunity), p ' (j) is the average active power time sequence of the minute rank after adjustment Row, t and z refer to the sequence number of this minute terminal corresponding sampled point in original work(power time series respectively.This test is real Example takes preceding 2150 points in p ' (j) to be training data, and development is 1 step, 2 steps and the 3 step prognostic experiments of 50 step-lengths by a definite date, and with Percentage ratio error MAPE is the validity of this algorithm of standard testing：

Wherein, Y (i) and p ' (i) are respectively prediction wind turbine active power value and sampled value, and l is prediction step.

The wind turbine active power sequence for rejecting the minute rank after average is designated as P, and Fig. 3 show the small echo work(for P Rate analysis of spectrum result, the red noise measuring line using 5% significance is threshold value, during the average minute active power of the wind turbine Between sequence have 4096,2048,1218,609 and 256 be extreme point totally 5 harmonic compoments, take its extreme point each side First periodic point less than red noise measuring line, constitutes half period zones, and this half period zones is harmonic compoment band, by taking this example as an example, is deposited In [2896.3,4871], [1722.2,2435.5], [861.1,1722.2], [430.5,724.1] and [215.3,304.4] altogether 5 harmonic compoment bands, according to wavelet reconstruction method, extract the corresponding time series P of this 5 half period zones₁、P₂、P₃、P₄And P₅, and Corresponding residual sequence R is obtained, thus P=P₁+P₂+P₃+P₄+P₅+ R, is shown in Fig. 4.It can be seen that, the regularity of 5 harmonic compoment signals By force, it is expected to the prediction of degree of precision；On the other hand, although the predicated error for residual error is inevitable, but is computed, residual error R Energy (variance) accounting P energy (variance) be 40.80%, decline notable, therefore, the predicated error for residual error will be much Less than the error being directly predicted for P.

Although neutral net has powerful nonlinear fitting ability and quick learning ability, how to select appropriate Neural network model, determines that structure, training set and the test set of neutral net still gather mainly by artificial experience or examination, its is pervasive Property is poor.Found by the analysis of the harmonic compoment signal to extraction, particularly P₃、P₄And P₅Sequence, although with obvious week Phase Variation Features, and sequence smoothes out, but the amplitude and phase of wherein sequence change over time substantially, and this can influence neutral net Study and Generalization Capability, cause to predict value mutation, precision of prediction declines, so on the one hand present invention selection fault-tolerant ability is stronger BP neural network, on the other hand continuity Characteristics based on Changes in weather, by introducing maximum rate of change index come constrained forecast The behavior of value mutation；And the fluctuation around 0 axle is presented in sequence of the residual signals after first-order difference, more suitable for RBF nerves Network, therefore the present invention is to the harmonic compoment signal P of extraction₁~P₅Using based on particle cluster algorithm optimization, simultaneously overlay error is corrected BP neural network be predicted, and for residual signals R then using being optimized based on particle cluster algorithm and overlay error is corrected RBF neural is predicted, through experiment, and this technology can be effectively improved the catastrophic behavior of neural network prediction value, improves pre- Survey precision.

To P₁~P₅Using the BP neural network model optimized based on particle cluster algorithm, the model of input layer number is taken Enclose for [5,14], the length of training set is [50,2100], the scope of neural network weight and threshold value is [- 3,3], population kind Group's scale is 50, iteration 30 times.For R then using the RBF neural optimized based on particle cluster algorithm, input layer is taken The scope of number is [5,20], and the length of training set is [50,2100], and population population scale is 50, iteration 30 times, the institute of table 1 When being shown as carrying out the prediction of 3 steps, for harmonic compoment sequence P₁~P₅With residual error R input layer number I and training set length The optimum results of two parameters of L, for P₁~P₅The BP neural network weights of foundation and the optimum results of threshold value are excessive due to parameter Without listing one by one.

Table 1

This test case has carried out 1 step, 2 steps and the 3 step prognostic experiments that total prediction step is 50, predicts the outcome such as Fig. 5 (a) shown in-(c), table 2 counts for predicated error.

Table 2

	1 step is predicted	2 steps are predicted	3 steps are predicted
				MAPE	0.0635	0.0730	0.1248

Contrast experiment 1

Influence of the maximum rate of change index for precision of prediction is tested, because maximum rate of change index is aobvious both for each Write periodic sequence P₁~P₅Or residual sequence R is individually configured, therefore this sentences P₅It is illustrated with exemplified by R.Fig. 6 (a) It show for P₅Total prediction step for 50 1 step prediction contrast experiment, discovery take after maximum rate of change index, its MAPE Error is down to 0.0032 by 0.0093, and the flatness of prediction curve is guaranteed, and precision of prediction improves 190.63%；Fig. 6 (b) 1 step that the total prediction step showing for R is 50 predicts contrast experiment, and discovery is taken after maximum rate of change index, its MAPE errors are down to 0.9926 by 1.0268, and precision of prediction improves 3.45%

Contrast experiment 2

For the wind turbine active power sequence of minute rank, first difference computing is directly carried out, afterwards using population Optimization RBF neural network model is modeled, and the scope for taking input layer number is [5,20], and the length of training set is [50,2100], population population scale is 50, iteration 30 times, when table 3 show progress 3 steps prediction, for input layer Number I and training set length two parameters of L optimum results：

Table 3

Input layer number (I)	Training set length (L)
		12	169

This test case has carried out 1 step, 2 steps and the 3 step prognostic experiments that total prediction step is 50, predicts the outcome such as Fig. 7 (a) shown in-(c), table 4 counts for predicated error.

Table 4

	1 step is predicted	2 steps are predicted	3 steps are predicted
				MAPE	0.2727	0.4825	0.4715

The average MAPE errors of its 1-3 steps add 369.46% compared with table 2, and what the original wind turbine of this description of test was gathered has Work(power data is violent due to fluctuation, regular poor, is directly used in neural net model establishing effect bad.

Contrast experiment 3

In addition, for the validity of verification method, being compared present invention is alternatively directed to a kind of conventional wavelet analysis method Relatively test, the method for conventional wavelet analysis is that former sequence is decomposed into relatively smooth low frequency trend sequence and high-frequency fluctuation sequence Row, are individually predicted to each sequence afterwards, by predicted value carry out it is cumulative be it is final predict the outcome, existing literature has been demonstrate,proved It is well many that bright such a method ratio is predicted effect individually for original series.

Contrast experiment carries out 3 layers of decomposition to P (i) using db3 small echos, obtains low frequency trend sequence A3 and high-frequency fluctuation sequence D1, D2, D3, as shown in Figure 8.Each sequence is individually predicted using RBF neural afterwards, in order to verify institute of the present invention The particle group optimizing RBF of proposition validity, the present invention tests fixed RBF parameters and particle cluster algorithm optimization RBF parameters two The method of kind：

1) preset parameter：It is 20 to choose input layer number, that is, takes the active power data of preceding 20 sampled points to use In the active power data of prediction thereafter, training set length is 500, and its total prediction step is missed for 50 1 step, 2 steps and 3 steps prediction It is poor as shown in table 5, predict the outcome and see shown in Fig. 9 (a)-(c), it is seen then that the average MAPE errors of its 1-3 steps are added compared with table 2 131.15%.

Table 5

	1 step is predicted	2 steps are predicted	3 steps are predicted
				MAPE	0.1829	0.1927	0.2284

2) particle group optimizing RBF parameters：The scope for taking input layer number is [10,20], and the length of training set is [50,2100], population population scale is 50, iteration 30 times, when carrying out the prediction of 3 steps, RBF neural optimum results such as table 6 It is shown, predict the outcome as shown in Figure 8：

Table 6

	A3	D1	D2	D3
					Input layer number	15	18	15	17
Training set length	164	113	129	111

Its total prediction step is as shown in table 7 for 50 1 step, 2 steps and 3 step predicated errors, predicts the outcome and sees Figure 10 (a)-(c) It is shown, it is seen then that the average MAPE errors of its 1-3 steps add 18.33% compared with table 2.

Table 7

	1 step is predicted	2 steps are predicted	3 steps are predicted
				MAPE	0.0847	0.1029	0.1216

The general principle and principal character and advantages of the present invention of the present invention has been shown and described above.The technology of the industry Personnel are it should be appreciated that the present invention is not limited to the above embodiments, and the simply explanation described in above-described embodiment and specification is originally The principle of invention, without departing from the spirit and scope of the present invention, various changes and modifications of the present invention are possible, these changes Change and improvement all fall within the protetion scope of the claimed invention.The claimed scope of the invention by appended claims and its Equivalent thereof.

Claims

1. a kind of prediction of wind turbine active power and error revising method based on neutral net, it is characterised in that including following Step：

(1) the crude sampling active power time series p={ p (i), i=1,2 ..., N } of wind turbine is read in, wherein N is original Wind turbine active power sampled point number；P is adjusted to average active power time series p '={ p ' by forecast space requirement (j), j=1,2 ..., M }, wherein M is the sampled point by the wind turbine average active power sequence after forecast space requirement adjustment Number, p ' average value isOrder

(2) multi-scale wavelet power spectrum analysis method is used, the harmonic compoment sequence { P in P is extracted₁,P₂,…,P_k,…,P_K, its Middle K is the number of the harmonic compoment sequence in P, P_k={ P_k(1),P_k(2),…,P_k(M) }, thus P=P₁+P₂+…+P_K+ R, its Middle R=P-P₁-P₂-…-P_K, it is that the residual sequence after harmonic compoment sequence is rejected in P；

(3) to the harmonic compoment sequence { P in P₁,P₂,…,P_k,…,P_K, particle group optimizing is respectively adopted and overlay error is corrected BP neural network be predicted, set prediction step as l, then each harmonic compoment sequence { P₁,P₂,…,P_k,…,P_KPrediction As a result it isWherein

(4) RBF neural that the first-order difference sequence D to residual sequence R is corrected using particle group optimizing and overlay error is entered Row prediction, obtains predicting the outcome for residual sequence by difference inverse operation, sets prediction step as l, then residual sequence R prediction As a result it is Y_R={ Y_R(1),Y_R(2),...,Y_R(l)}；

(5) willIt is added with each harmonic compoment sequence, predicting the outcome for residual sequence, obtains the final Y that predicts the outcome,

2. a kind of prediction of wind turbine active power and error revising method based on neutral net according to claim 1, It is characterized in that：The detailed process of the BP neural network of particle group optimizing employed in the step (3) is：

(1) according to Kolmogorov theorems, 3 layers of BP neural network model are set up, if input layer number is I, hidden layer Neuron number is H, and output layer neuron number is O；Wherein, H=2*I+1, O=1；

(2) parameter for needing to optimize is determined, including：The input layer number I of the BP neural network and length L of training set, Also include：One group objects W=(w (1), w (2) ..., w (q)), q=I*H+H*O+H+O, wherein, w (1)~w (I*H) is BP god Input layer through network is to the link weights of hidden layer neuron, and w (I*H+1)~w (I*H+H*O) is the implicit of BP neural network Layer is to the link weights of output layer neuron, and w (I*H+H*O+1)~w (I*H+H*O+H) is BP neural network hidden layer neuron Threshold value, w (I*H+H*O+H+1)~w (I*H+H*O+H+O) be BP neural network output layer neuron threshold value；

(3) population is initializedWherein Q₁For the sum of particle, i-th of particle is X_i=(I_i,W_i,L_i), Particle rapidity is V_i=(v_I_i,v_W_i,v_L_i), wherein I_i、W_i、L_iFor parameter I, W, L, mono- group alternatively solves；

(4) to each particle X in colony_i=(I_i,W_i,L_i) determine parameter, construct BP neural network training set input and Output matrix, wherein for harmonic compoment sequence P_kAnd BP neural network input layer number I_iInitially set up matrix Z₁With Z₂, wherein：

For neural metwork training collection length L, Z to be optimized₁In last L_iArrange the input matrix I as training set_train, Z₂In Last L_iArrange the output matrix O as training set_train；It regard forecast step-length l as test step-length, Z₁In last l row conducts The input matrix I of test set_test, Z₂In last l arrange output matrix O as test set_test；The BP constructed according to training set Neutral net, as its fitness value, is made to the error sum of squares of test set analog result with the minimum optimization direction of fitness value The quality of each particle, record particle X are judged for evaluation criterion_iCurrent individual extreme value is P_best(i) P in colony, is taken_best(i) most Excellent individual is used as overall extreme value G_best；

<mrow> <msubsup> <mi>V</mi> <mi>i</mi> <mrow> <mi>g</mi> <mo>+</mo> <mn>1</mn> </mrow> </msubsup> <mo>=</mo> <msubsup> <mi>&omega;V</mi> <mi>i</mi> <mi>g</mi> </msubsup> <mo>+</mo> <msub> <mi>c</mi> <mn>1</mn> </msub> <msub> <mi>r</mi> <mn>1</mn> </msub> <mrow> <mo>(</mo> <msub> <mi>P</mi> <mrow> <mi>b</mi> <mi>e</mi> <mi>s</mi> <mi>t</mi> </mrow> </msub> <mo>(</mo> <mi>i</mi> <mo>)</mo> <mo>-</mo> <msubsup> <mi>X</mi> <mi>i</mi> <mi>g</mi> </msubsup> <mo>)</mo> </mrow> <mo>+</mo> <msub> <mi>c</mi> <mn>2</mn> </msub> <msub> <mi>r</mi> <mn>2</mn> </msub> <mrow> <mo>(</mo> <msub> <mi>G</mi> <mrow> <mi>b</mi> <mi>e</mi> <mi>s</mi> <mi>t</mi> </mrow> </msub> <mo>-</mo> <msubsup> <mi>X</mi> <mi>i</mi> <mi>g</mi> </msubsup> <mo>)</mo> </mrow> <mo>,</mo> </mrow>

In formula：ω is inertia weight, c₁、c₂For acceleration factor, g is current iteration number of times, and r₁、r₂To be distributed in [0,1] Random number；

(7) judge whether to reach maximum iteration, terminate optimization process if meeting, obtain BP neural network optimal parameter I_best、W_best(w_best(1),w_best(2),...,w_best(q))、L_best, otherwise return to step (4).

3. a kind of prediction of wind turbine active power and error revising method based on neutral net according to claim 2, It is characterized in that：The detailed process of overlay error correction method employed in the step (3) is：

(1) first by the BP neural network parameter I of optimization_best、W_best(w_best(1),w_best(2),...,w_best(q))、L_bestStructure Make neural metwork training collection Z₃With test set Z₄And BP neural network link weights and threshold value are initialized, wherein：

w_best(1)~w_best(I*H) for BP neural network input layer to hidden layer neuron link weights initial value, w_best(I*H+1)~w_best(I*H+H*O) the initial of weights is linked to output layer neuron for the hidden layer of BP neural network Value, w_best(I*H+H*O+1)~w_best(I*H+H*O+H) for BP neural network hidden layer neuron threshold value initial value, w_best (I*H+H*O+H+1)~w_best(I*H+H*O+H+O) for BP neural network output layer neuron threshold value initial value；

(2) defined nucleotide sequence P_kIt is c=in sampled point i rate of change | p_k(i)-p_k(i-1) |, for prediction step l, calculate Z₄Finally The maximum rate of change of l step-lengths is c_max, the l step predictions that just can be iterated afterwards, during prediction, with c_maxIt is used as jth The output valve of neutral net during step-ahead predictionWith previous predicted value or actual valueMaximum rate of change, ifThenIt is used as the predicted value of jth step-length；If otherwiseThenIfThen

4. a kind of prediction of wind turbine active power and error revising method based on neutral net according to claim 1, It is characterized in that：The detailed process of particle group optimizing RBF neural employed in the step (4) is：

(2) population is initializedWherein Q₂For the sum of particle, i-th of particle is X_i=(I_i,L_i), grain Sub- speed isWherein I_i,L_iFor parameter I, L, mono- group alternatively solves；

(3) each particle X in colony_i(I_i,L_i) parameter that determines, construct the input of RBF neural training set and defeated Go out matrix, wherein for residual sequence R and RBF neural input layer number I_iInitially set up matrix Z₅And Z₆, its In：

For neural metwork training collection length L, Z to be optimized₅In last L_iArrange the input matrix I as training set_train, Z₆In Last L_iArrange the output matrix O as training set_train；It regard forecast step-length l as test step-length, Z₅In last l row conducts The input matrix I of test set_test, Z₆In last l arrange output matrix O as test set_test；The RBF constructed according to training set Neutral net, as its fitness value, is made to the error sum of squares of test set analog result with the minimum optimization direction of fitness value The quality of each particle, record particle X are judged for evaluation criterion_iCurrent individual extreme value is P_best(i) P in colony, is taken_best(i) most Excellent individual is used as overall extreme value G_best；

5. a kind of prediction of wind turbine active power and error revising method based on neutral net according to claim 4, It is characterized in that：The detailed process of overlay error correction method employed in the step (4) is：

(1) first by the RBF neural parameter I of optimization_bestAnd L_bestConstructing neural network training set Z₇With test set Z₈And just Beginningization RBF neural links weights and threshold value, wherein：

(2) defined nucleotide sequence R is c=in sampled point i rate of change | R (i)-R (i-1) |, for prediction step l, calculate Z₈Last l The maximum rate of change of step-length is c_max, the l step predictions that just can be iterated afterwards, during prediction, with c_maxWalked as jth The output valve Y of neutral net during long prediction_R(j) with previous predicted value or actual value Y_R(j-1) maximum rate of change, if | Y_R (j)-Y_R(j-1)|<c_max, then Y_R(j) as the predicted value of jth step-length；If otherwise Y_R(j)>Y_R(j-1), then Y_R(j)=Y_R(j- 1)+c_maxIf, Y_R(j)<Y_R(j-1), then Y_R(j)=Y_R(j-1)-c_max。

6. a kind of prediction of wind turbine active power and error revising side based on neutral net according to claim 2 or 4 Method, it is characterised in that：Inertia weight ω=0.5, acceleration factor c₁=c₂=1.49445.