CN101788692A - Short-time wind speed forecasting method based on neural network - Google Patents

Abstract

The invention discloses a short-time wind speed forecasting method based on a neural network. The method comprises the following steps of: (1) recording the moving average observing values of the wind speed, wind direction, temperature and air pressure on the same district once at the interval of 10 minutes, and ordering the observed data in a time sequence from front to back to obtain original wind speed data; (2) calculating the data of the adjacent time according to the time sequence, and generating an original wind speed added value sequence; (3) entering the original airspeed added value into a BP artificial neural network to construct a wind speed added value neural network model, calculating and counting an original wind speed added value tendency by utilizing the BP artificial neural network, training a RP artificial neural network by respectively utilizing the original wind speed added value and the original wind speed added value tendency as the original data, and obtaining a wind speed added value predicting value and a wind speed added value error tendency; (4) adding the wind speed added value predicting value into statistical noise to reduce and generate a wind speed predicting value; (5) carrying out moving filtering on the wind speed predicting value; and (6) obtaining the wind speed predicting value 4 hours in advance.

Description

Short-time wind speed forecasting method based on neural network

Technical field:

The present invention relates to short-term forecasting wind speed field, is the prediction that utilizes (24 hours) in short-term wind speed that artificial neural network combines with time series method specifically.

Background technology

Wind-power electricity generation is a kind of cleaning, non-pollution of renewable energy generation mode.Developing regenerative resource energetically is the important component part of China's energy development strategy, and to melt the regenerative resource development of sending out as a kind of large-scale commercial applications very rapid for wind energy in recent years.Wind-powered electricity generation has characteristics such as intermittent and undulatory property, and this is bigger to the safe and stable operation influence that wind-powered electricity generation inserts electrical network.Especially in recent years the fast development of China's wind-powered electricity generation cause, urgent especially to effective solution of this problem.Empirical evidence both domestic and external is predicted exerting oneself of wind-power electricity generation can reduce wind power-generating grid-connected adverse effect to a great extent.

Wind energy turbine set being done the forecasting wind speed of short-term, obtained the predicted value of wind-power electricity generation power again by the wind powertrace, is one of effective way of carrying out the wind-power electricity generation power prediction.The cube of wind energy and wind speed is directly proportional, and the variation of wind energy is more violent when the wind speed acute variation.The effectively accurately prediction of wind speed is the basis of wind power reasonable prediction.

At present, with regard to world wide, the achievement in research of wind energy turbine set short-term wind speed is also unsatisfactory.The wind speed forecasting method of finding wind energy turbine set by literature search can be divided into statistical model method and meteorological model predictions method.

Wind speed and wind power that Yang Xiuyuan etc. utilize time series method and neuroid method to provide and put forward previous observation time section forecast.Yang Xiuyuan, Xiao Yang, Chen Shuyong.Wind farm wind velocity and generated output forecasting research.Proceedings of the CSEE, 2005,25 (11), 1-5.Mention in the literary composition and use time series modeling, draw the basic parameter of wind speed characteristics, and select the Forecasting Methodology of the input variable of neural network, and utilize tumbling-type weights adjustment means to improve precision of prediction with parameter.The prediction of wind speed is had certain effect to this method but predicated error is bigger, can not use in the middle of the practical application preferably.

Xiao Yongshan, Wang Weiqing, Huo Xiaoping.Wind farm wind velocity time series forecasting research based on neural network.Power-saving technology, 2007,25 (142) 106-175.Utilize Time Serial Neural Network that the wind speed of wind energy turbine set was predicted in advance in one hour in the literary composition.Wu Guochang, Xiao Yang, Weng Shasha. wind-powered electricity generation factory short-term forecasting wind speed is inquired into. Jilin electric power, 2005 (6)., 21-24.In the literary composition wind speed forecasting model that utilizes time series method to provide has been carried out instance analysis.But these methods can only be predicted the wind farm wind velocity that shifts to an earlier date a hour, predict poor in timeliness, can not satisfy the traffic control needs of electric system fully.Fourth is bright, Zhang Lijun, Wu Yichun.Predicting wind speed of wind farm model based on time series analysis.Electric Power Automation Equipment, 2005,25 (8), 32-34.

Summary of the invention

The purpose of this method is to propose a kind of predicting the outcome of (4 hours in advance per 10 minutes) wind speed in a short time.Be directed to the ageing weak point of wind energy turbine set Short-term Forecast result of study, artificial neural network there is no and takes into full account filtering, and utilizes method such as increment corrections to forecasting that the result corrects etc.The prediction period of this method increases the wind farm wind velocity forecast for 4 hours in advance.In order to improve precision of prediction and to keep the stable of precision of prediction, in artificial neural network, add filtering; And the method that has proposed to utilize increment to proofread and correct improves precision of prediction.Solved neural network weight and passed in time and become inapplicable gradually, problem such as the prediction effect deviation is big.The wind speed result of this method prediction has anastomose property preferably with the contrast of actual measurement air speed data.Sequential neural network method has improved the precision of forecasting wind speed effectively.

Technical scheme of the present invention is achieved in that the employing multilayer feedforward neural network, i.e. the BP neural network.Constitute by an input layer, two hidden layers and an output layer.The neuron of interlayer carries out unidirectional connection, and neuron is separate in the layer.The hidden layer function adopts the Sigmoid function.The training process of network promptly is the process that weights are adjusted.

Because the amplitude of variation of wind speed on room and time is bigger, predicts that directly wind speed will cause great error.Used filtering wind speed incremental forecasting method in the current neural network forecasting wind speed.The method avoids directly predicting air speed value, and revises the prediction wind speed that generates reality from the prediction of increment.

Among current short-term forecasting, need to predict the wind speed in the future 4 hours.The actual observation wind series is spaced apart 10 minutes, and predicting 4 hours is exactly prediction 24 air speed value in the future.In order to predict wind speed accurately, according to historical wind-resources correlation of data is analyzed, drawing the correlation distance of the wind comes from is 13 ~ 20 days, and correlativity therebetween accounts for leading factor.When carrying out forecasting wind speed, the sample point of being taked is about 16 days before first future position, and the timed sample sequence data are 2304 to be advisable.After sample is determined substantially, can carry out the noise that sample space is taken out in pre-service, time series is carried out filtering, can reach good trend prediction effect.To the sample data processing of rising in value, promptly the adjacent moment air speed value is poor.This increment sequence has certain pseudo-periodicity, and this characteristic is applied to output result's the effect preferably of having corrected.

Description of drawings

The probability density of Fig. 1 wind speed

The original wind speed accumulative total of Fig. 2 is sewed and mend function

The probability density of Fig. 3 wind speed increment

Fig. 4 increment cumulative distribution function

Fig. 5 increment distributes and normal distribution is compared

Fig. 6 wind speed forecast in 4 hours in advance and measured result comparison diagram;

Fig. 7 wind speed forecast in 4 hours in advance process flow diagram.

Embodiment

Data acquisition: utilize NRG anemometer tower observation data as sample data, observation data comprises per 10 minutes wind speed, wind direction, temperature and air pressure once.Become continuous time series with surveying the wind data arrangement, it is carried out filter preprocessing.

Prediction is proofreaied and correct in increment: because the amplitude of variation of wind speed on room and time is bigger, predict that directly wind speed will cause great error.Used filtering wind speed incremental forecasting method in the current neural network forecasting wind speed.The method avoids directly predicting air speed value, and revises the prediction wind speed that generates reality from the prediction of increment.

The prediction fundamental assumption

(1) there is certain pseudoperiod in the supposition wind speed

(2) also there is certain pseudoperiod in supposition wind speed increment

(3) there is pseudoperiod in the supposition prediction error value

(4) the supposition cycle is 144 integral multiple

Wind speed profile and wind increment distribute

Fig. 1 ~ Fig. 5 has provided wind speed profile and wind speed increment function.

As seen from Figure 1, wind speed mainly concentrates in the middle of 5 ~ 10, the wind speed 17 ~ 20 that high wind speed Duan Youyi section is comparatively mild.Can be got by Fig. 2, prediction can reflect correctly that mainly the correct prediction of the wind speed more than 4 meters just can reach the accuracy more than 90%.As seen from Figure 3, the distribution function of speed increment almost is symmetrical about x=0, and probability density concentrates between-1 to 1, and pulsation of the wind speed that this explanation is most can not surpass 1 scope.Neural network algorithm can be predicted the time series that wind speed fluctuates fully between 1.Fig. 4 increment cumulative distribution function also has further elaboration to this conclusion.Fig. 5 illustrates that the increment cumulative distribution disobeys normal distribution, but very approaching.

Utilize the wind speed increment that network is trained based on above-mentioned research conclusion, increment is predicted, then predict wind speed.

Calculate back one wind speed and this increment of wind speed constantly constantly, obtain the time series that constitutes by original wind speed increment.Analyze this increment seasonal effect in time series trend.

Set up mathematical model: the method that adopts the BP artificial neural network to combine with time series is set up model.BP neural network (Back-propagation Neural Network) is meant the multilayer feedforward neural network based on error backpropagation algorithm, adopts the training patterns that the tutor is arranged.It is that D.E.Rumelhart and J.L.McCelland and research group thereof are research in 1986 and design.Multilayer feedforward neural network has following characteristics:

(1) can approach any Nonlinear Mapping with arbitrary accuracy, bring a kind of new nonlinear representation aids for the modeling of complication system;

(2) it can be learnt and the self-adaptation unknown message, can change the control effect by the connection value of revising network if system has taken place to change;

(3) distributed information storage and Processing Structure have certain fault-tolerance, and therefore the system that constructs has robustness preferably;

(4) structural model of many inputs, many outputs is fit to handle challenge.

The mathematical model of BP algorithm is the optimum solution problem of finding the solution with minor function:

Wherein, x is a training sample,

Be the actual output of network, y _k(t) be the desired output of network, ω _IjBe the weights of input layer i to hidden layer node j, v _JkBe the weights of hidden layer node j to output layer node k, θ _jBe the threshold value at hidden layer node j place, γ _tBe the threshold value at output node t place, f (x) is an activation function.Realize that global error function E descends by gradient on curved surface, adopt the gradient rule, ask E to the connection weight of output layer and hidden layer and the negative gradient of threshold value:

$\left\{\begin{array}{c}-\frac{\∂E}{\∂{\mathrm{\ω}}_{\mathrm{ij}}}=\underset{k=1}{\overset{N}{\mathrm{\Σ}}}(-\frac{\∂E_k}{\∂{\mathrm{\ω}}_{\mathrm{ij}}})\\ -\frac{\∂E}{\∂{\mathrm{\θ}}_j}=\underset{k=1}{\overset{N}{\mathrm{\Σ}}}(-\frac{\∂E_k}{\∂{\mathrm{\θ}}_j})\\ -\frac{\∂E}{\∂v_{\mathrm{jk}}}=\underset{k=1}{\overset{N}{\mathrm{\Σ}}}(-\frac{\∂E_k}{\∂v_{\mathrm{jk}}})\\ -\frac{\∂E}{\∂{\mathrm{\γ}}_t}=\underset{k=1}{\overset{N}{\mathrm{\Σ}}}(-\frac{\∂E_k}{\∂{\mathrm{\γ}}_t})\end{array}\right.$

By gradient decline principle, i.e. the variation of connection weight and threshold value is proportional to negative gradient, so have:

Wherein 0＜η＜1 is learning rate, i=1, and 2 ..., n, t=1,2 ..., n, j=1,2 ..., p.

Adjusted network connection weights and threshold value are as follows:

$\left\{\begin{array}{c}{\mathrm{\ω}}_{\mathrm{ij}}(l+1)={\mathrm{\ω}}_{\mathrm{ij}}\left(l\right)+\mathrm{\Δ}{\mathrm{\ω}}_{\mathrm{ij}}\\ v_{\mathrm{jt}}(l+1)=v_{\mathrm{jt}}\left(l\right)+\mathrm{\Δ}{v}_{\mathrm{jt}}\\ {\mathrm{\θ}}_j(l+1)={\mathrm{\θ}}_j\left(l\right)+{\mathrm{\Δ\θ}}_j\\ {\mathrm{\γ}}_t(l+1)={\mathrm{\γ}}_t\left(l\right)+{\mathrm{\Δ\γ}}_t\end{array}\right.$

Wherein, l represents frequency of training.When neural network is predicted for time series, promptly adopt neural network to go to approach a time series or a seasonal effect in time series distortion, with m before the seasonal effect in time series value (Y (s-1), Y (s-2), ..., Y (s-m)) remove prediction in the future t value (Y (s), Y (s+1), ..., Y (s+t-1))

With a structure be: the neural network of m-p-t is come match or approximating function

T(Y(s)，Y(s+1)，...，Y(s+t-1))＝F(Y(s-1)，Y(s-2)，...，Y(s-m))

T (Y (s), Y (s+1) ..., Y (s+t-1)) when t equals 1, be exactly the one-step prediction of neural network, be exactly

Y(s)＝F(Y(s-1)，Y(s-2)，...，Y(s-m))

When t greater than 1 the time, just be referred to as multi-step prediction, at this moment, t input is corresponding to t predicted value.If predict next step, current predicted value is joined in the sample space carry out the next prediction in a step as actual value.Here the method that is adopted is a kind of forecast method one by one, also can be directly via the needed unknown quantity of the disposable prediction of the network that trains for a plurality of predictions.

Training pattern: utilize the time series that constitutes by original wind speed increment to the model training.Provide 24 wind speed incremental forecasting values.

Predicted value output: 24 wind speed incremental forecasting values are added statistical noise be reduced to 24 forecasting wind speed values.It is carried out The disposal of gentle filter.Output is the forecasting wind speed value of 4 hours (per 10 minutes one time 24 points) in advance.Predict the outcome and the measured result comparison diagram is seen Fig. 6.

Above content is to further describing that the present invention did in conjunction with concrete preferred implementation; can not assert that the specific embodiment of the present invention only limits to this; for the general technical staff of the technical field of the invention; without departing from the inventive concept of the premise; can also make some simple deduction or replace, all should be considered as belonging to the present invention and determine scope of patent protection by claims of being submitted to.

Claims (4)

1. based on the short-time wind speed forecasting method of neural network, it is characterized in that this method is according to following steps:

(1) every wind speed, wind direction, temperature and the air pressure of 10 minutes record areals, the observation data that continuous acquisition acquisition in 4 hours is made up of 24 groups of data is arranged the original air speed data of acquisition with described observation data according to time sequencing from front to back; Described original air speed data is carried out filter preprocessing;

(2) according to time sequencing, calculate the data of adjacent moment, generate original wind speed increment;

(3) described original wind speed increment input BP artificial neural network is obtained wind speed increment neural network model; Utilize the original wind speed increment of BP artificial neural network counting statistics trend; The BP artificial neural network is trained as raw data with described original wind speed increment and original wind speed increment trend respectively; Obtain wind speed increment predicted value and wind speed increment error trend;

(4) utilize the reduction of described wind speed increment predicted value to generate the forecasting wind speed value;

(5) to forecasting wind speed value smothing filtering;

(6) obtain 4 hours forecasting wind speed values in advance.

2. according to claim 1 based on the short-time wind speed forecasting method of neural network, it is characterized in that: original air speed data carries out filter preprocessing and is meant that the time series that original wind speed is constituted carries out Fourier filtering in the described step (1), removes the noise of original wind series; Described Fourier transform promptly is to have converted original reluctant time-domain signal to be easy to analyze frequency-region signal, can also utilize inverse fourier transform to convert these frequency-region signals to time-domain signal at last;

Be provided with limit sequence x (n), n=0,1,2 ..., N-1

Its discrete Fourier transformation is:

$X\left(k\right)=\mathrm{DFT}\left[x\left(n\right)\right]=\underset{n=0}{\overset{N-1}{\mathrm{\Σ}}}x\left(n\right)e^{-j\frac{2\mathrm{\π}}{N}\mathrm{kn}},0\≤k\≤N-1$

Its discrete inversefouriertransform is:

$x\left(n\right)=\mathrm{IDFT}[X\left(k\right)e^{j\frac{2\mathrm{\π}}{N}\mathrm{kn}},0\≤n\≤N-1.$

3. according to claim 1 based on the short-time wind speed forecasting method of neural network, it is characterized in that: described step (2) be meant adjacent wind speed do poor, promptly in the sequence X adjacent two do poorly, constitute new sequence Y;

Y(s)＝X(t)-X(t-1)，0≤s≤N-1

；

4. according to claim 1 based on the short-time wind speed forecasting method of neural network, it is characterized in that: described step (3) is meant

Wherein, x is a training sample,

Be the actual output of network, y _k(t) be the desired output of network, ω _IjBe the weights of input layer i to hidden layer node j, v _JkBe the weights of hidden layer node j to output layer node k, θ _jBe the threshold value at hidden layer node j place, γ _tBe the threshold value at output node t place, f (x) is an activation function;

Realize that global error function E descends by gradient on curved surface, adopt the gradient rule, ask E to the connection weight of output layer and hidden layer and the negative gradient of threshold value:

$\left\{\begin{array}{c}-\frac{\∂E}{\∂{\mathrm{\ω}}_{\mathrm{ij}}}=\underset{k=1}{\overset{N}{\mathrm{\Σ}}}(-\frac{\∂E_k}{\∂{\mathrm{\ω}}_{\mathrm{ij}}})\\ -\frac{\∂E}{\∂{\mathrm{\θ}}_j}=\underset{k=1}{\overset{N}{\mathrm{\Σ}}}(-\frac{\∂E_k}{\∂{\mathrm{\θ}}_j})\\ -\frac{\∂E}{\∂v_{\mathrm{jk}}}=\underset{k=1}{\overset{N}{\mathrm{\Σ}}}(-\frac{\∂E_k}{\∂v_{\mathrm{jk}}})\\ -\frac{\∂E}{\∂{\mathrm{\γ}}_t}=\underset{k=1}{\overset{N}{\mathrm{\Σ}}}(-\frac{\∂E_k}{\∂{\mathrm{\γ}}_t})\end{array}\right.$

By gradient decline principle, i.e. the variation of connection weight and threshold value is proportional to negative gradient, so have:

Wherein 0＜η＜1 is learning rate, i=1, and 2 ..., n, t=1,2 ..., n, j=1,2 ..., p;

Adjusted network connection weights and threshold value are as follows:

$\left\{\begin{array}{c}{\mathrm{\ω}}_{\mathrm{ij}}(l+1)={\mathrm{\ω}}_{\mathrm{ij}}\left(l\right)+\mathrm{\Δ}{\mathrm{\ω}}_{\mathrm{ij}}\\ v_{\mathrm{jt}}(l+1)=v_{\mathrm{jt}}\left(l\right)+\mathrm{\Δ}{v}_{\mathrm{jt}}\\ {\mathrm{\θ}}_j(l+1)={\mathrm{\θ}}_j\left(l\right)+\mathrm{\Δ}{\mathrm{\θ}}_j\\ {\mathrm{\γ}}_t(l+1)={\mathrm{\γ}}_t\left(l\right)+\mathrm{\Δ}{\mathrm{\γ}}_t\end{array}\right.$

Wherein, l represents frequency of training; When neural network is predicted for time series, promptly adopt neural network to go to approach a time series or a seasonal effect in time series distortion, with m value (Y (s-1) before the seasonal effect in time series, Y (s-2) .., Y (s-m)) remove prediction t value (Y (s) in the future, Y (s+1) ..., Y (s+t-1))

With a structure be: the neural network of m-p-t is come match or approximating function

T (Y (s), Y (s+1) .., Y (s+t-1))=F (Y (s-1), Y (s-2) ..., Y (s-m)) T (Y (s), Y (s+1) ..., Y (s+t-1)) when t equals 1, be exactly the one-step prediction of neural network, be exactly

Y(s)＝F(Y(s-1)，Y(s-2)，...，Y(s-m))

When t greater than 1 the time, just be referred to as multi-step prediction, at this moment, t input is corresponding to t predicted value; If predict next step, current predicted value is joined in the sample space carry out the next prediction in a step as actual value; Here the method that is adopted is a kind of forecast method one by one, also can be directly via the needed unknown quantity of the disposable prediction of the network that trains for a plurality of predictions.

Images