CN105741192A - Short-term wind speed combined forecasting method for wind turbine cabin of wind power plant - Google Patents

Abstract

The invention discloses a short-term wind speed combined forecasting method for a wind turbine cabin of a wind power plant. The method comprises the steps of analyzing the similarity of wind speed subsequences of a forecasted wind turbine cabin and all wind turbine cabins within a certain time period day by day by adopting a dynamic time warping method and a correlation coefficient method separately, extracting wind speed data of a plurality of subsequences with the most similar evolution, separately establishing a generalized regression neural network sub-model forecasting unit based on the dynamic time warping method and the correlation coefficient method, and globally optimizing specific parameters of each sub-model by adopting a particle swarm optimization, wherein the mean of the forecasting results of the two sub-models is used as a final forecasting result of the combined forecasting method. The method realizes fine forecasting of the cabin wind speed of each wind turbine in the wind power plant, thereby effectively improving the short-term output forecasting level of the whole wind power plant.

Description

A kind of wind energy turbine set wind turbine cabin short-term wind speed combining prediction method

Technical field

The invention belongs to wind energy turbine set wind turbine technical field, particularly to a kind of wind energy turbine set wind turbine cabin short-term wind speed combining prediction method.

Background technology

In effectively wind energy being connected to the grid, wind energy turbine set exerted oneself that to carry out accurate forecast be extremely necessary and crucial, this is wherein, the short-period forecast of 0 to 6 hour is for electrical network Real-Time Scheduling, it is ensured that the technical parameter that mains frequency, power and the balance of voltage etc. relate to power grid security is significant.

Wind energy, as a kind of reproducible clean energy resource, has the advantages such as scaleable, wind power generator group reliability is high, cost is low, operation maintenance is simple of installing.According to " Wind Power Generation Industry monitoring situation in 2014 " that in February, 2015, National Energy Board announced, by the end of the year 2014, the accumulative installed capacity of China's wind-powered electricity generation has reached 96,370,000 kilowatts, accounts for the 7% of whole capacity of installed generator, accounts for the 27% of whole world wind-powered electricity generation installation.Wind-powered electricity generation electricity volume 153,400,000,000 kilowatt hour in 2014, accounts for the 2.78% of whole generated energy.In December, 2014 National Energy Board issues " energy development Strategic Action Plan (2014-2020) ", it is contemplated that to the year two thousand twenty, wind-powered electricity generation installation is up to 200,000,000 kilowatts.So far, wind-powered electricity generation has become the third-largest main force of China power supply after thermoelectricity and water power.Along with being continuously increased of installed capacity, the electricity problem of abandoning of wind-powered electricity generation is always comparatively prominent, adds up according to National Energy Board, and within 2012, wind-powered electricity generation amount about 20,000,000,000 kilowatt hour is abandoned in the whole nation, on average abandons wind rate and reaches 17%；Within 2013, wind-powered electricity generation amount about 15,000,000,000 kilowatt hour is abandoned in the whole nation, on average abandons wind rate and reaches 10%, and up-to-date statistics shows, by by the end of September, 2014, wind-powered electricity generation abandons wind-powered electricity generation amount 8,600,000,000 kilowatt hour, on average abandons wind rate 7.5%.Cause the major reason that wind-powered electricity generation abandons electricity to be in that the intermittence of wind causes that the undulatory property of wind-powered electricity generation and unstability have impact on wind-powered electricity generation quality, abandon electricity in vain to ensure the safety of electrical network.Based on this, National Energy Board issued " wind farm power prediction forecast management Tentative Measures " in 2011, require that all wind parks being incorporated into the power networks of China should be set up wind-powered electricity generation prediction system and generation schedule declaration work mechanism before 1 day January in 2012 and start trail run, report and submit wind power prediction forecast result as requested.

The wind speed forecast common method of current wind energy turbine set includes physical method and statistical method, physical method refers to and obtains the timing of high-spatial and temporal resolution, fixed point, quantitative numerical weather prediction model wind-force prediction output result according to the numerical weather prediction model that becomes more meticulous, run practical situation according to wind electric field blower simultaneously, consider various wind turbine power generation influence factor, foundation is exerted oneself prediction physical model, carries out output of wind electric field prediction.Physical method does not need substantial amounts of measurement data, but require that the physical characteristic of air and the characteristic of wind energy turbine set are had mathematical description accurately, these equation solution difficulties, required information magnanimity, computationally intensive, calculate the time long, and from the difficulty of meteorological department's acquisition data greatly, costly, therefore in short-term wind-electricity field wind speed forecasts, still conventional statistical method.At present, statistical method historical summary according to wind energy turbine set anemometer tower mostly, adopt methods such as continuing method, Random time sequence method, Kalman filtering method, neural network, support vector machine.Rely solely on the disadvantage that anemometer tower data carry out forecasting and be in that wind energy turbine set is subject to landform, turbulent flow etc. to affect, the wind speed in cabin, wind turbine place and anemometer tower place wind speed would be likely to occur obvious difference, therefore only exerting oneself of whole wind energy turbine set is forecast with the wind speed of measuring of anemometer tower, will causing bigger prediction error, this is unrelated with concrete forecasting procedure.Along with the raising of the technology of measurement and computer computation ability, the wind speed to separate unit electromotor cabin become more meticulous carries out forecast and is possibly realized.

Summary of the invention

In order to solve the technical problem that above-mentioned background technology proposes, it is desirable to provide a kind of wind energy turbine set wind turbine cabin short-term wind speed combining prediction method, by the wind speed of typhoon motor every in wind energy turbine set is carried out fine forecast, the forecast level thus the short-term being effectively improved whole wind energy turbine set is exerted oneself.

In order to realize above-mentioned technical purpose, the technical scheme is that

A kind of wind energy turbine set wind turbine cabin short-term wind speed combining prediction method, it is characterised in that comprise the following steps:

(1) set wind turbine and sample air speed data number as m every day, read in the crude sampling wind speed collection in all wind turbine cabins in wind energy turbine set be v (#i, j), i=1,2 ..., M；J=1,2 ..., N}, wherein (#i, j) represents the wind turbine #i nacelle wind speed value at sampled point j to v, and M is the sum of wind turbine, and N is total number of sample points, then always sampling natural law is

Treating the wind turbine #p, 1≤p≤M of forecast nacelle wind speed, forecast step-length is L_f, it is respectively adopted dynamic time warping and Pearson correlation coefficient method carries out sampling wind speed similarity measurement: the length of similarity system design is L, namely take wind turbine #p to start to retrodict from sampled point N the sampling air speed data { v (#p of L length, N-L+1), v (#p, N-L+2), ..., v (#p, N) }, the cabin sampling wind speed of corresponding period T every day period the last period of wind turbine D days all with in wind energy turbine set, similarity system design is carried out respectively by dynamic time warping and Pearson correlation coefficient method, respectively the result of two kinds of similarity system design methods is ranked up from high to low by similarity measurement result；

(2) result of the two kinds of similarity system design methods obtained according to step (1), constructs two GRNN and forecasts submodels, and structure the two GRNN forecasts submodel it needs to be determined that training set P, test set T and tri-parameters of smoothing factor S；

Construct two simulation experiment collection and determine that two GRNN forecast the optimized parameter of submodel respectively, the input of the building method of described simulation experiment collection: training set P is for according to dynamic time warping or Pearson correlation coefficient method, for the wind series length L carrying out similarity system design, acquired with { v (#p, N-L+1), v (#p, N-L+2) ..., v (#p, N) } the wind series collection Q that constitutes of front I wind series the most similar_P, training set P be output as v (#p, N-L+1), v (#p, N-L+2) ..., v (#p, N)；Test set T is identical with training set P；

With v (#p, N-L+1), v (#p, N-L+2) ..., and v (#p, N) } self simulation error is standard, and L, I and S are optimized, the optimized parameter of the GRNN submodel set up for dynamic time warping is L_DTW、I_DTWAnd S_DTW, the optimized parameter for the GRNN submodel of Pearson correlation coefficient method foundation is L_PCC、I_PCCAnd S_PCC；

(3) the GRNN submodel set up with dynamic time warping, by L_DTW、I_DTWAnd S_DTWSet up the training set of GRNNTest setAnd smoothing factor S_DTW, wherein, training setInput for according to dynamic time warping, for the length L of the wind series carrying out similarity system design_DTW, acquired with { v (#p, N-L_DTW+1),v(#p,N-L_DTW+ 2) ..., v (#p, N) } the most similar front I_DTWThe wind series collection that individual wind series is constitutedTraining setIt is output as { v (#p, N-L_DTW+1),v(#p,N-L_DTW+2),…,v(#p,N)}；Test setInput beIn each wind turbine forIn L after each sampling wind series_fThe wind series collection that individual sampling wind series is constituted, test setBeing output as wind turbine #p from sampled point N, step-length is L_fForecast wind series F_DTW, wherein $F_{DTW}\{{\hat{v}}_{DTW}(\#p,N+1),{\hat{v}}_{DTW}(\#p,N+2),...,{\hat{v}}_{DTW}(\#p,N+L_f)\};$

With the GRNN submodel that Pearson correlation coefficient method is set up, by L_PCC、I_PCCAnd S_PCCSet up the training set of GRNNTest setAnd smoothing factor S_PCC, wherein, training setInput for according to Pearson correlation coefficient method, for the length L of the wind series carrying out similarity system design_PCC, acquired with { v (#p, N-L_PCC+1),v(#p,N-L_PCC+ 2) ..., v (#p, N) } the most similar front I_PCCThe wind series collection that individual wind series is constitutedTraining setIt is output as { v (#p, N-L_PCC+1),v(#p,N-L_PCC+2),…,v(#p,N)}；Test setInput beIn each wind turbine forIn L after each sampling wind series_fThe wind series collection that individual sampling wind series is constituted, test setBeing output as wind turbine #p from sampled point N, step-length is L_fForecast wind series F_PCC, wherein $F_{PCC}=\{{\hat{v}}_{PCC}(\#p,N+1),{\hat{v}}_{PCC}(\#p,N+2),...,{\hat{v}}_{PCC}(\#p,N+L_f)\};$

(4) wind turbine #p from sampled point N, step-length be L_fForecast wind series result be F=0.5 (F_DTW+F_PCC)。

Further, the definition of " T in period the last period of D day " in step (1):

If wind series sampling periods was less than one week, then before taking D day, total data carries out similarity measurement, and the data of the last week or identical season year by year of otherwise taking D day carry out similarity measurement.

Further, in step (2), adopt particle cluster algorithm that L, I and S are optimized.

Further, the detailed process that L, I and S are optimized by particle cluster algorithm is adopted:

A () initializes population X=(X₁,X₂,...,X_W), wherein W is the sum of particle, and i-th particle is X_i=(L_i,I_i,S_i), particle rapidity is V_i=(v_L_i,v_I_i,v_S_i), wherein L_i,I_i,S_iOne group of alternative solution for parameter L, I and S；

B () is to each particle X in colony_iThe GRNN parameter determined, modelling structural experiment collection, according to { v (#p, N-L_i+1),v(#p,N-L_i+ 2) ..., v (#p, N) } self simulation error calculates its fitness value, minimum for the quality optimizing direction and passing judgment on as evaluation criterion each particle using fitness value, records particle X_iCurrent individual extreme value is P_best (i), takes individuality optimum for P_best (i) in colony as overall extreme value G_best；

Each particle X in (c) colony_i, respectively its position and speed are updated；

$V^{k+1}={\mathrm{\ωV}}^k+c_1{r}_1(P\_best(i)-X_i^k)+c_2{r}_2(G\_best-X_i^k),$

$X_i^{k+1}=X_i^k+V^{k+1}$

In formula: ω is inertia weight, c₁、c₂For acceleration factor, and r₁、r₂For being distributed in the random number of [0,1]；

D () recalculates each particle target function value now, update P_best (i) and G_best；

E () judges whether to reach maximum iteration time, then terminate optimization process as met, otherwise return step (c).

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

Adopt the beneficial effect that technique scheme is brought:

(1) present invention is in close relations and present the feature of 24 hours weak mechanical periodicity from the diurnal variation of wind speed and underlying surface surface temperature, adopt dynamic time warping and Pearson correlation coefficient method that with the wind speed subsequence of all wind turbine cabins corresponding period day by day, forecast wind turbine cabin is carried out similarity analysis, the air speed data extracting the most like some subsequences that develop builds forecasting model, this thought not only follows the weak periodic feature of wind speed, and from wind energy turbine set all wind turbines historical series, carry out similarity measurement be more conducive to search the sequence the most similar to treating forecast wind turbine wind speed evolution, finally can be effectively improved forecast precision.It is simultaneously introduced DTW and two kinds of wind speed similarity system design methods of Pearson correlation coefficient method, particularly DTW method, its essence be research two time series drawns and shrink after Similarity Problem, carry out DTW calculating two wind speed sequential can Length discrepancy, meet the actual condition of measuring wind speed；Meanwhile, two wind speed sequential carrying out nonlinear mapping on a timeline that stretch/shrink, more meet plesiomorphism but moment and the actual condition of the inconsistent wind speed of generation amplitude occur, therefore its theoretical basis is particularly suitable for the process of wind.Experiments show that, this system can be effectively improved the forecast precision of wind turbine cabin short-term wind speed.

(2) present invention adopts generalized regression nerve networks as forecasting model, its fault-tolerance with very strong non-linear mapping capability and flexible network structure and height and robustness, when sample data is less, the value of forecasting is also better, simultaneously need to the parameter regulated only only has smoothing factor S, therefore the applied environment of native system it is particularly suitable for, and when constructing generalized regression nerve networks, particle cluster algorithm is adopted to carry out global optimizing for the parameter of neutral net, improve the system general applicability for different wind energy turbine set wind turbine nacelle wind speed data, further increase forecast precision.

Accompanying drawing explanation

Fig. 1 is the method flow diagram of the present invention；

Fig. 2 is dynamic time warping flow chart；

Fig. 3 is GRNN submodel parameter optimization schematic diagram in the present invention；

Fig. 4 is GRNN submodel forecast schematic diagram in the present invention；

Fig. 5 is wind turbine #1～#4 cabin acquired original air speed data schematic diagram；

Fig. 6 is wind turbine #1 cabin acquired original air speed data Morlet Wavelet Spectrum schematic diagram.

Detailed description of the invention

Below with reference to accompanying drawing, technical scheme is described in detail.

The method flow diagram of the present invention, comprises the following steps that as shown in Figure 1.

The first step, data similarity judges:

False wind motor air speed data number of sampling every day is m, read in the crude sampling wind speed collection in all wind turbine cabins in wind energy turbine set be v (#i, j), i=1,2 ..., M；J=1,2 ..., N}, wherein (#i, j) represents the wind turbine #i nacelle wind speed value at sampled point j to v, and M is the sum of wind turbine, and N is total number of sample points, then always sampling natural law is(For downward rounding operation), usual N is also for forecast starting point.Treating the wind turbine #p (1≤p≤M) of forecast nacelle wind speed, forecast step-length is L_f, it is respectively adopted dynamic time warping (DynamicTimeWarping, it is called for short DTW) and Pearson correlation coefficient method (Pearson ' sCorrelationCoefficient, it is called for short PCC) carry out sampling wind speed similarity measurement: the length of similarity system design is L, namely take wind turbine #p to start to retrodict from sampled point N the sampling air speed data of L length, i.e. { v (#p, N-L+1), v (#p, N-L+2), ..., v (#p, N) }, the cabin sampling wind speed of corresponding period T every day period the last period of wind turbine D days all with in wind energy turbine set, similarity system design is carried out respectively by DTW and PCC method, and result is ranked up from high to low by similarity measurement result.

The definition of " T in period the last period of D day " above: if wind series sampling periods was less than one week, then before taking D day, total data carries out similarity measurement, the data of the last week or identical season year by year of otherwise taking D day carry out similarity measurement.

Wherein, 2 kinds of methods of wind speed evolution similarity system design are carried out:

1, dynamic time warping, as illustrated in fig. 2, it is assumed that two wind series X={x₁,x₂,…,x_N, Y={y₁,y₂,…,y_N, wherein N is the total number of wind series.Initialization sequence distance matrix d_NN, wherein d_NNEach element be:I, j=1,2 ..., N.At matrix d_NNIn, the set of matrix element one group adjacent is called crooked route, is designated as W={w₁,w₂,...,w_K, the kth element w of W_k=(i, j)_k, this paths meets following condition: (a) N≤K < 2N-1；(b)w₁=(1,1), w_K=(N, N)；C () is for w_k=(i, j), w_k-1=(i', j'), meets 0≤i-i'≤1,0≤j-j'≤1.On this basis,DTW algorithm can be attributed to utilization Dynamic Programming Idea find one from d (1,1) to d the shortest path D of (N, N), its state transition equation is: $\{\begin{array}{c}D(1,1)=d(1,1)\\ D(i,j)=\min \{D(i-1,j-1),D(i,j-1),D(i-1,j)\}+d(i,j)\end{array},$ The more little explanation sequence X of DTW (X, Y), Y flexible rear similarity on a timeline is more high.

2, correlation coefficient process, two wind series X={x₁,x₂,…,x_N, Y={y₁,y₂,…,y_N, wherein N is the total number of sequence.Sequence X, the correlation coefficient of Y is:Wherein,WithRespectively sequence X, the meansigma methods of Y.The more big explanation sequence X of R (X, Y), the linear dependence of Y is more strong.

Second step, GRNN submodel parameter optimization:

The present invention uses generalized regression nerve networks (GeneralizedRegressionNeuralNetwork, GRNN) forecast wind turbine nacelle wind speed, press DTW method and PCC method respectively according to step 1) result of air speed data similarity system design, construct two GRNN forecast submodels.Based on GRNN algorithm, structure GRNN is it needs to be determined that training set P, test set T and smoothing factor S totally 3 parameters, and return value is the neutral net NET constructed；Two simulation experiment collection of structure are to determine the optimized parameter of two submodels herein, the make of simulation experiment collection is: the input of training set P is for according to DTW method or PCC method, for the wind series length L specifically carrying out similarity system design, acquired with { v (#p, N-L+1), v (#p, N-L+2), ..., v (#p, N) } the wind series collection Q that constitutes of front I wind series the most similar_P, training set P be output as v (#p, N-L+1), v (#p, N-L+2) ..., v (#p, N)；Test set T is identical with training set P；By to { v (#p, N-L+1), v (#p, N-L+2) ..., v (#p, N) } simulation error of self is standard, adopting particle cluster algorithm (ParticleSwarmOptimization is called for short PSO) that L, I and S are optimized, the optimized parameter for the submodel of DTW method foundation is L_DTW、I_DTWAnd S_DTW, the optimized parameter for the submodel of PCC method foundation is L_PCC、I_PCCAnd S_PCC；

The result schematic diagram of this step as shown in Figure 3, for wind turbine #D, the optimum results integrated by modelling structural experiment samples the wind series of G2 sampling interval of day and wind turbine #C first samples the wind series of G3 sampling interval of day as the sample wind series of G1 sampling interval of day, wind turbine #B second of wind turbine #A the 4th, and it is minimum that G1, G2 and G3 are simulated time error for H1；

Wherein, the concrete steps that L, I and S are optimized by particle cluster algorithm are adopted:

(2-1) population X=(X is initialized₁,X₂,...,X_W), wherein W is the sum of particle, and i-th particle is X_i=(L_i,I_i,S_i), particle rapidity is V_i=(v_L_i,v_I_i,v_S_i) wherein L_i,I_i,S_iOne group of alternative solution for parameter L, I and S；

(2-2) to each particle X in colony_iThe GRNN parameter determined, modelling structural experiment collection, according to { v (#p, N-L_i+1),v(#p,N-L_i+ 2) ..., v (#p, N) } self simulation error calculates its fitness value, minimum for the quality optimizing direction and passing judgment on as evaluation criterion each particle using fitness value, records particle X_iCurrent individual extreme value is P_best (i), takes individuality optimum for P_best (i) in colony as overall extreme value G_best；

(2-3) each particle X in colony_i, respectively its position and speed are updated；

$V^{k+1}={\mathrm{\&omega;V}}^k+c_1{r}_1(P\_best(i)-X_i^k)+c_2{r}_2(G\_best-X_i^k)$

$X_i^{k+1}=X_i^k+V^{k+1}$

In above formula, ω is inertia weight, desirable ω=0.5；c₁、c₂For acceleration factor, desirable c₁=c₂=1.49445；And r₁、r₂For being distributed in the random number of [0,1]；

(2-4) recalculate each particle target function value now, update P_best (i) and G_best；

(2-5) with maximum iteration time for according to judging whether to meet the condition of convergence, then terminating optimization process as met, otherwise return step (2-3).

3rd step, GRNN submodel forecasts:

With the submodel that DTW method is set up, by L_DTW、I_DTWAnd S_DTWSet up the training set of GRNNTest setAnd smoothing factor S_DTW, wherein: training setInput be according to DTW method, for the length L of the wind series carrying out similarity system design_DTW, acquired with { v (#p, N-L_DTW+1),v(#p,N-L_DTW+ 2) ..., v (#p, N) } the most similar front I_DTWThe wind series collection that individual wind series is constitutedTraining setIt is output as { v (#p, N-L_DTW+1),v(#p,N-L_DTW+2),…,v(#p,N)}；Test setInput beIn each wind turbine forIn L after each sampling wind series_fThe wind series collection that individual sampling wind series is constituted, test setBeing output as wind turbine #p from sampled point N, step-length is L_fForecast wind series F_DTW, wherein $F_{DTW}\{{\hat{v}}_{DTW}(\#p,N+1),{\hat{v}}_{DTW}(\#p,N+2),...,$ ${\hat{v}}_{DTW}(\#p,N+L_f)\}.$

With the submodel that PCC method is set up, by L_PCC、I_PCCAnd S_PCCSet up the training set of GRNNTest setAnd smoothing factor S_PCC, wherein: training setInput be according to PCC method, for the length L of the wind series carrying out similarity system design_PCC, acquired with { v (#p, N-L_PCC+1),v(#p,N-L_PCC+ 2) ..., v (#p, N) } the most similar front I_PCCThe wind series collection that individual wind series is constitutedTraining setIt is output as { v (#p, N-L_PCC+1),v(#p,N-L_PCC+2),…,v(#p,N)}；Test setInput beIn each wind turbine forIn L after each sampling wind series_fThe wind series collection that individual sampling wind series is constituted, test setBeing output as wind turbine #p from sampled point N, step-length is L_fForecast wind series F_PCC, wherein $F_{PCC}=\{{\hat{v}}_{PCC}(\#p,N+1),{\hat{v}}_{PCC}(\#p,N+2),...,$ ${\hat{v}}_{PCC}(\#p,N+L_f)\};$

The result schematic diagram of this step as shown in Figure 4, for L after wind turbine #D, G1, G2 and G3_fThe wind series that wind series G1 ', the G2 ' of individual sampling wind series and G3 ' are constituted integrates the input as test set, is output as the wind speed predicted value H1 ' of #D, and forecast length is L_f。

4th step, combining prediction:

Wind turbine #p is from sampled point N, and step-length is L_fForecast wind series result be F=0.5 (F_DTW+F_PCC)。

Concrete test case:

Certain wind energy turbine set has 274 typhoon motors, Fig. 5 is the wind turbine #1～#4 of this wind energy turbine set SCADA system collection crude sampling air speed data of continuous 5 days in 23 days 7 October in 2008, each air speed data is every 10 minutes interior meansigma methodss, there is air speed data 144 every day, air speed data totally 720 in figure.As can be seen from the figure the strong nonlinearity of wind speed and randomness, for wind turbine #1, sampling air speed data maximum within this time period is about 18m/s, minima is about 0.4m/s, to the 170th sampled point near the 90th sampled point, in 14 hours, about 0.4m/s is suddenly down in wind speed change from about 13m/s, skyrocketing again afterwards to about 15m/s, change is acutely.In addition, it is also shown in the weak periodicity that wind speed presents with diurnal variation, the Morlet Wavelet Spectrum that front 600 the sampling air speed datas of wind turbine #1 are done, result shows that this sequence comprises accurate 144 sampled points (corresponding 24 hours) cycle, and only 24 hours periods are by 95% red noise credit assigned line in this sequence, as shown in Figure 6, the weak periodicity that wind speed presents with diurnal variation is also the starting point of this patent.

The application encloses 2 test cases, and example 1 gives wind turbine #7 from the 620th sampled point, arranges a new forecast starting point every 6 sampled points, and common mode intends 11 forecast starting points, and forecast step-length is the modeling process of 1；Example 2 gives #7～#37 totally 31 typhoon motor, and for 11 forecast starting points of experiment 1 simulation, forecast step-length is when being 1～6, the provided method of this patent and BP neutral net extrapolation, GRNN neutral net extrapolation, ARIMA time series method comparative result.Judgment of error standard is mean square error MSE (MeanSquaredError) and mean absolute error MAE (MeanAbsoluteError), $MSE=\sqrt{\frac{1}{n}\underset{i=1}{\overset{n}{\&Sigma;}}{(Y_i-y_i)}^2},MAE=\frac{1}{n}\underset{i=1}{\overset{n}{\&Sigma;}}|Y_i-y_i|,$ Wherein, Y_iAnd y_iRespectively step-length is actual value during i and predictive value, and n is total prediction step.

Example 1

For the collection air speed data of #7 wind turbine, Simulation prediction starting point N={620,626,632,638,644,650,656,662,668,674,680} and forecast step-length L_f=1, construct two simulation experiment collection to determine the optimized parameter of two submodels, the make of simulation experiment collection is: the submodel set up with DTW method or PCC method, for the wind series length L carrying out similarity system design, acquired with { v (#7, N-L+1), v (#7, N-L+2) ..., v (#7, N) } the wind series collection Q that constitutes of front I wind series the most similar_P, training set P be output as v (#7, N-L+1), v (#7, N-L+2) ..., v (#7, N)；Test set T is identical with training set P；By to { v (#7, N-L+1), v (#7, N-L+2) ..., v (#7, N) } simulation error of self is standard, makes I ∈ [2,6], L ∈ [4,36], S ∈ [0.1,0.5], adopt particle cluster algorithm that L, I and S are optimized, maximum iteration time 30 times, for the submodel called after DTW-PSO-GRNN that DTW method is set up, its optimized parameter is L_DTW、I_DTWAnd S_DTW；For the submodel called after PCC-PSO-GRNN that PCC method is set up, its optimized parameter is L_PCC、I_PCCAnd S_PCC, the results are shown in Table 1, final built-up pattern called after COM-PSO-GRNN, and a step of forecasting result and error thereof that according to it, each each forecast starting point is done by two submodels of optimized parameter foundation and the built-up pattern of correspondence thereof are listed in table 2.Consolidated statement 1, table 2 are visible, and two submodels after optimizing in table 1, the wind series length for carrying out similarity system design is all less, L_DTWValue between 5～12, average is 7.7, L_PCCValue between 5～20, average is 9.5, wind series for nonlinearity is described, the sequence length carrying out similarity system design is unsuitable long, also indicates that GRNN neutral net still has study and generalization ability preferably for shorter training set data simultaneously；In table 1, the smoothing factor of the GRNN of two submodels, all close to 0.1, illustrates that the smoothing factor of GRNN is more little simultaneously, and its generalization ability is more strong；The forecast result of table 2 shows, the error MSE of DTW-PSO-GRNN relatively PCC-PSO-GRNN reduces 7.33%, the precision of built-up pattern COM-PSO-GRNN is affected by PCC-PSO-GRNN, precision is between two submodels, MSE relatively PCC-PSO-GRNN improves 3.53%, and relatively DTW-PSO-GRNN reduces 3.67%.

Table 1

Table 2

Example 2

In order to verify the universality of the application, adopt #7～#37 totally 31 typhoon motor, 11 forecast starting points for experiment 1 simulation, when forecast step-length is 1～6, to the provided method of this patent, comparing including two submodels and built-up pattern and BP neutral net extrapolation, GRNN neutral net extrapolation and ARIMA time series method, the results are shown in Table 3.From table 3, no matter for MSE or MAE error criterion, the forecast precision of COM-PSO-GRNN is all the highest, and precision respectively GRNN neutral net extrapolation, BP neutral net extrapolation and ARIMA time series method from high to low in 3 kinds of Extrapolation method, illustrate:

1) it is feasible for carrying out the forecast of short-term wind speed based on similarity principle, and its effect is better than the method based on extrapolation；

2) DTW-PSO-GRNN submodel is better than PCC-PSO-GRNN submodel, what reason was in that dynamic time warping substantially weighs is wind speed sequential non-linear similarity on a timeline, and correlation coefficient process is a kind of linear correlation rule, and wind speed is strong nonlinearity；

3) the precision acceptor model accuracy impact of built-up pattern, the precision of built-up pattern may not be the highest, but its stationarity and overall forecast effect are often optimum；

4) basic reason that the error of 3 kinds of Extrapolation method is bigger is in that the conduction of error, and along with the increase error of forecast step-length sharply increases, this is wherein, the forecast precision of two kinds of neural net methods will far above ARIMA time series method, wind series for nonlinearity is described, the superiority of neural network model, GRNN neutral net extrapolation forecast precision is higher than BP neutral net extrapolation simultaneously, illustrate that the Generalization Capability of GRNN neutral net is better than BP neutral net, GRNN neutral net is adopted to forecast the reasonability of basic methods as wind speed herein thus also demonstrating.

Table 3

Above example is only the technological thought that the present invention is described, it is impossible to limits protection scope of the present invention, every technological thought proposed according to the present invention, any change done on technical scheme basis with this, each falls within scope.

Claims (5)

1. a wind energy turbine set wind turbine cabin short-term wind speed combining prediction method, it is characterised in that comprise the following steps:

(1) set wind turbine and sample air speed data number as m every day, read in the crude sampling wind speed collection in all wind turbine cabins in wind energy turbine set be v (#i, j), i=1,2 ..., M；J=1,2 ..., N}, wherein (#i, j) represents the wind turbine #i nacelle wind speed value at sampled point j to v, and M is the sum of wind turbine, and N is total number of sample points, then always sampling natural law is

Treating the wind turbine #p, 1≤p≤M of forecast nacelle wind speed, forecast step-length is L_f, it is respectively adopted dynamic time warping and Pearson correlation coefficient method carries out sampling wind speed similarity measurement: the length of similarity system design is L, namely take wind turbine #p to start to retrodict from sampled point N the sampling air speed data { v (#p of L length, N-L+1), v (#p, N-L+2), ..., v (#p, N) }, the cabin sampling wind speed of corresponding period T every day period the last period of wind turbine D days all with in wind energy turbine set, similarity system design is carried out respectively by dynamic time warping and Pearson correlation coefficient method, respectively the result of two kinds of similarity system design methods is ranked up from high to low by similarity measurement result；

(2) result of the two kinds of similarity system design methods obtained according to step (1), constructs two GRNN and forecasts submodels, and structure the two GRNN forecasts submodel it needs to be determined that training set P, test set T and tri-parameters of smoothing factor S；

Construct two simulation experiment collection and determine that two GRNN forecast the optimized parameter of submodel respectively, the input of the building method of described simulation experiment collection: training set P is for according to dynamic time warping or Pearson correlation coefficient method, for the wind series length L carrying out similarity system design, acquired with { v (#p, N-L+1), v (#p, N-L+2) ..., v (#p, N) } the wind series collection Q that constitutes of front I wind series the most similar_P, training set P be output as v (#p, N-L+1), v (#p, N-L+2) ..., v (#p, N)；Test set T is identical with training set P；

With v (#p, N-L+1), v (#p, N-L+2) ..., and v (#p, N) } self simulation error is standard, and L, I and S are optimized, the optimized parameter of the GRNN submodel set up for dynamic time warping is L_DTW、I_DTWAnd S_DTW, the optimized parameter for the GRNN submodel of Pearson correlation coefficient method foundation is L_PCC、I_PCCAnd S_PCC；

(3) the GRNN submodel set up with dynamic time warping, by L_DTW、I_DTWAnd S_DTWSet up the training set of GRNNTest setAnd smoothing factor S_DTW, wherein, training setInput for according to dynamic time warping, for the length L of the wind series carrying out similarity system design_DTW, acquired with { v (#p, N-L_DTW+1),v(#p,N-L_DTW+ 2) ..., v (#p, N) } the most similar front I_DTWThe wind series collection that individual wind series is constitutedTraining setIt is output as { v (#p, N-L_DTW+1),v(#p,N-L_DTW+2),…,v(#p,N)}；Test setInput beIn each wind turbine forIn L after each sampling wind series_fThe wind series collection that individual sampling wind series is constituted, test setBeing output as wind turbine #p from sampled point N, step-length is L_fForecast wind series F_DTW, wherein $F_{DTW}=\{{\hat{v}}_{DTW}(\#p,N+1),{\hat{v}}_{DTW}(\#p,N+2),...,{\hat{v}}_{DTW}(\#p,N+L_f)\};$

With the GRNN submodel that Pearson correlation coefficient method is set up, by L_PCC、I_PCCAnd S_PCCSet up the training set of GRNNTest setAnd smoothing factor S_PCC, wherein, training setInput for according to Pearson correlation coefficient method, for the length L of the wind series carrying out similarity system design_PCC, acquired with { v (#p, N-L_PCC+1),v(#p,N-L_PCC+ 2) ..., v (#p, N) } the most similar front I_PCCThe wind series collection that individual wind series is constitutedTraining setIt is output as { v (#p, N-L_PCC+1),v(#p,N-L_PCC+2),…,v(#p,N)}；Test setInput beIn each wind turbine forIn L after each sampling wind series_fThe wind series collection that individual sampling wind series is constituted, test setBeing output as wind turbine #p from sampled point N, step-length is L_fForecast wind series F_PCC, wherein $F_{PCC}=\{{\hat{v}}_{PCC}(\#p,N+1),{\hat{v}}_{PCC}(\#p,N+2),...,{\hat{v}}_{PCC}(\#p,N+L_f)\};$

(4) wind turbine #p from sampled point N, step-length be L_fForecast wind series result be F=0.5 (F_DTW+F_PCC)。

2. a kind of wind energy turbine set wind turbine cabin short-term wind speed combining prediction method according to claim 1, it is characterised in that: in step (1), " T in period the last period of D day " is defined as,

If wind series sampling periods was less than one week, then before taking D day, total data carries out similarity measurement, and the data of the last week or identical season year by year of otherwise taking D day carry out similarity measurement.

3. a kind of wind energy turbine set wind turbine cabin short-term wind speed combining prediction method according to claim 1, it is characterised in that: in step (2), adopt particle cluster algorithm that L, I and S are optimized.

4. a kind of wind energy turbine set wind turbine cabin short-term wind speed combining prediction method according to claim 3, it is characterised in that: step (2) adopts the detailed process that L, I and S are optimized by particle cluster algorithm:

A () initializes population X=(X₁,X₂,...,X_W), wherein W is the sum of particle, and i-th particle is X_i=(L_i,I_i,S_i), particle rapidity is V_i=(v_L_i,v_I_i,v_S_i), wherein L_i,I_i,S_iOne group of alternative solution for parameter L, I and S；

B () is to each particle X in colony_iThe GRNN parameter determined, modelling structural experiment collection, according to { v (#p, N-L_i+1),v(#p,N-L_i+ 2) ..., v (#p, N) } self simulation error calculates its fitness value, minimum for the quality optimizing direction and passing judgment on as evaluation criterion each particle using fitness value, records particle X_iCurrent individual extreme value is P_best (i), takes individuality optimum for P_best (i) in colony as overall extreme value G_best；

Each particle X in (c) colony_i, respectively its position and speed are updated；

$V^{k+1}={\mathrm{\&omega;V}}^k+c_1{r}_1(P\_best(i)-X_i^k)+c_2{r}_2(G\_best-X_i^k),$

$X_i^{k+1}=X_i^k+V^{k+1}$

In formula: ω is inertia weight, c₁、c₂For acceleration factor, and r₁、r₂For being distributed in the random number of [0,1]；

D () recalculates each particle target function value now, update P_best (i) and G_best；

E () judges whether to reach maximum iteration time, then terminate optimization process as met, otherwise return step (c).

5. state a kind of wind energy turbine set wind turbine cabin short-term wind speed combining prediction method according to claim 4, it is characterised in that: inertia weight ω=0.5, acceleration factor c₁=c₂=1.49445.