CN105512745A - Wind power section prediction method based on particle swarm-BP neural network - Google Patents

Abstract

The invention relates to a wind power section prediction method based on a particle swarm-BP neural network, and the method comprises the steps: employing the BP neural network to achieve the prediction of a wind power section; enabling historical data of the wind power to serve as an input vector of a prediction model, wherein output values {Y1, Y2} respectively represent the upper and lower limits of a wind power prediction output section at a future time node; forming a two-dimension flying particle in a PSO (Particle Swarm Optimization) through the weight value and a threshold value of the BP neural network; adjusting the network weight value and the threshold value to meet the requirements for the reliability of the prediction model; and searching the optimal particle according to a given optimization target function. According to the invention, the method introduces an index for describing section information, builds a comprehensive optimization target of wind power section prediction, carries out the parameter optimization of the BP neural network through employing a the PSO, achieves the short-term section prediction of wind power through employing the multi-output characteristic of the BP neural network, and provides a basis for power grid decision. The technical scheme of the invention is shown in the description.

Description

Based on the wind power interval prediction method of population-BP neural network

Technical field

The invention belongs to technical field of power systems, relate to a kind of wind power interval prediction method.

Background technology

In field of new energy generation, wind energy earns widespread respect as the high-efficiency cleaning energy of technology maturation and widely applies, and has good development prospect.Along with the progress of correlation technique and the support of national policy, China's wind-power electricity generation was obtaining develop rapidly in recent years, was that the access of Large Scale Wind Farm Integration or middle-size and small-size blower fan all achieves remarkable achievement.But wind-power electricity generation has randomness, undulatory property and intermittent natural quality, the grid-connected meeting of large-scale wind power brings adverse effect to power grid security and the quality of power supply, very large challenge is proposed to dispatching of power netwoks and system cloud gray model, if system call is unreasonable or auxiliary plan imperfection, be easy to occur power supply trouble; Even if middle-size and small-size blower fan also also exists the problem affecting the quality of power supply and energy management.Therefore, from security, the economy point of operation of power networks, short-term forecasting is carried out to wind power very necessary.But wind power forecasting method main is at present fixed-point value prediction really on timing node, is difficult to the uncertainty describing wind-power electricity generation, brings many uncertainties, increase the operation risk containing blower fan electrical network to dispatching of power netwoks.And interval prediction is carried out to wind power, the uncertainty of wind-powered electricity generation can be described to a certain extent, for electrical network decision-making provides foundation, have more practical application meaning.

Summary of the invention

The object of the invention is the above-mentioned deficiency improving prior art, a kind of short-term interval prediction method of wind power is provided, the present invention introduces the index describing block information, set up the complex optimum target of wind power interval prediction, particle cluster algorithm is adopted to carry out parameter optimization to BP neural network, the multi output characteristic of BP neural network is utilized to realize the short-term interval prediction of wind power, for electrical network decision-making provides foundation.Technical scheme of the present invention is as follows:

A kind of wind power interval prediction method based on population-BP neural network, the method adopts BP neural network to carry out wind power interval prediction, using the input vector of the historical data of wind power as forecast model, output valve { Y ₁, Y ₂represent the interval bound that future time node wind power prediction exports respectively, by the flight particle that the dimension in the weights and threshold constituent particle group PSO algorithm of BP neural network is 2, by adjustment network weight and threshold value to meet the reliability requirement of forecast model, according to given optimization object function search optimal particle, specific as follows:

If normalized wind power time series is X=[x ₁, x ₂..., x _n], U _iwith L _ifor the upper and lower bound of t=i moment interval prediction, N _tfor choosing the sample size for computation interval parameter, fiducial probability is θ=100 (1-α) %, ω is the flight particle in PSO algorithm, the i.e. parameter vector of BP neural network weight and threshold value, for realizing describing the rationalization in wind power prediction interval, adopt the index describing block information to form the optimization object function of wind power interval prediction, i.e. the fitness function of PSO algorithm, each index is as follows:

1) mulching measures

Mulching measures refers to that actual value falls into the probability in forecast interval, is N for capacity _tforecast sample collection, actual value x _ifall into forecast interval [L _i, U _i] probability should be not less than θ=100 (1-α) %, the mulching measures PICP computing formula of forecast interval is:

$PICP=\frac{1}{N_t}\underset{i=1}{\overset{N_t}{\Σ}}{\mathrm{\ξ}}_i$

${\mathrm{\ξ}}_i=\left\{\begin{array}{cc}0& x_i\∉\[L_i,U_i\]\\ 1& x_i\∈\[L_i,U_i\]\end{array}\right.$

2) interval average bandwidth

Consider the average bandwidth PINAW of forecast interval, reduce the complete risk of mistake that the bound because of forecast interval is brought close to its ultimate value, for normalized wind power time series, PINAW is calculated by following formula:

$PINAW=\frac{1}{N_t}\underset{i=1}{\overset{N_t}{\Σ}}(U_i-L_i)$

3) mean center error

For the distribution consistency degree of assessment actual value in forecast interval, consider the mean center error PIACE of forecast interval, the deviation namely between interval intermediate value and real output value, computing formula is:

$PIACE=\sqrt{\underset{i=1}{\overset{N_t}{\Σ}}{(x_i-(U_i+L_i)/2)}^2}$

4) optimization object function

The requirement of wind power interval prediction is prediction mulching measures high as far as possible, has again interval average bandwidth little as far as possible and mean center error simultaneously, sets up wind power interval prediction parametric synthesis optimization object function:

F(ω)＝PINAW·(1+k·λ(PICP)·e ^PIACE)

$\mathrm{\&lambda;}\left(PICP\right)=\left\{\begin{array}{cc}0& PICP\&GreaterEqual;\mathrm{\&theta;}\\ 1& PICP<\mathrm{\&theta;}\end{array}\right.$

Wherein k is nonnegative constant, is called punishment parameter,

Wind power interval prediction problem is converted into the following optimization problem containing constraint:

Min:F(ω)

s.t.θ≤PICP(ω)≤100％

Preferably, modified particle swarm optiziation is adopted to carry out optimizing to the weights and threshold of BP neural network, improve one's methods as follows: on the basis of primary particle speed, increase multiple search speed component, single particle is made to have multiple speed varied in size, the speed that amplitude is larger meets the global optimizing requirement of particle, the speed that amplitude is less can make particle have better local search ability, in order to maintain the diversity of particle; For improving the search capability of algorithm, the weights W of being fixed by numerical value is set as the dynamic value of successively decreasing according to sequential.

Accompanying drawing explanation

Figure 1B P neural network topology structure

Fig. 2 wind power interval model process flow diagram

Fig. 3 wind speed raw data

Fig. 4 blower fan output power-wind speed curve

Fig. 5 wind power raw data

Fig. 6 wind-powered electricity generation short-term interval prediction result (fiducial probability 80%)

Embodiment

First technical scheme of the present invention is described in detail below.

(1) BP neural network and improve PSO algorithm

BP neural network is a kind of multilayer feedforward neural network, has the feature of forward transfer signal, reverse propagated error, and input signal successively to go forward one by one process from input layer through hidden layer, and the neuronic coverage of every one deck is only limitted to lower one deck neuron.Once output layer fails to obtain expected result, then network adjusts weights and threshold automatically according to its output error, thus makes predicting the outcome of output constantly approach expectation value, and Fig. 1 is the network topology structure of BP neural network.Wherein, { X ₁, X ₂..., X _nthe input quantity of BP neural network, { Y ₁, Y ₂..., Y _mthe prediction output valve of network, ω _ijand ω _jkfor the weights of BP neural network.Adopt BP neural network to carry out wind power interval prediction, need the input vector of the historical data of wind power as forecast model, output valve { Y ₁, Y ₂represent the interval bound that future time node wind power prediction exports respectively, by adjustment network weight and threshold value to meet the reliability requirement of forecast model.

For improving the performance of BP neural network wind power interval prediction method, particle cluster algorithm is adopted to carry out optimizing to the weights and threshold of BP neural network.In particle cluster algorithm, a potential solution of each particle representing optimized problem, determine its feature by position, speed and fitness value three indexs, fitness value determined by fitness function.Particle according to the flying experience dynamic conditioning of self and companion, thus realizes individual can the optimizing of solution space.

If the search volume dimension of optimization problem is D dimension, the number that group comprises flight particle is n, ω=(ω ₁, ω ₂..., ω _n), i-th particle represents a D dimensional vector ω _i=(ω _i1, ω _i2..., ω _iD), the vector of particle is made up of the speed of particle position in space, particle self and history optimal location three part of individuality:

Current location: x _i=(x _i1, x _i2..., x _iD);

Optimal location: p _i=(p _i1, p _i2..., p _iD);

Particle rapidity: v _i=(v _i1, v _i2..., v _iD);

Meanwhile, the global extremum of note population is P _g=(P _g1, P _g2..., P _gD).

For each particle, its Position And Velocity upgrades according to formula (1) and formula (2):

$\begin{array}{c}{v}_{id}^{k+1}=W\&CenterDot;v_{id}^k+c_1\&CenterDot;{\mathrm{rand}}_1\&CenterDot;\left(p_{id}^k-x_{id}^k\right)\\ +c_2\&CenterDot;{\mathrm{rand}}_2\&CenterDot;\left(p_{gd}-x_{id}^k\right)\end{array}---\left(1\right)$

$x_{id}^{k+1}=x_{id}^k+v_{id}^{k+1}---\left(2\right)$

Wherein d=1,2 ..., D; I=1,2 ..., N, c ₁and c ₂for nonnegative constant, rand () is the random number in [0,1], and W is inertia weight, determines the impact of particle previous experience on present speed.

In wind power prediction, the weights and threshold dimension formed in PSO algorithm of BP neural network is the flight particle of 2, according to given optimization object function search optimal particle.Because the renewal speed of particle does not have adaptive time-varying characteristics, the search that becomes more meticulous cannot be realized, cause particle cluster algorithm to exist search precision is not high, be easy to be absorbed in local optimum shortcoming, need make improvements.

In order to keep particle diversity spatially, setting single particle has the speed varied in size, the speed that amplitude is larger meets the global optimizing requirement of particle, the speed that amplitude is less can make particle have better local search ability, so formula (1) and formula (2) are deformed into following formula (3):

$\left\{\begin{array}{c}{v}_{id}^{k+1}=W\&CenterDot;v_{id}^k+c_1\&CenterDot;{\mathrm{rand}}_1\&CenterDot;\left(p_{id}^k-x_{id}^k\right)+\\ c_2\&CenterDot;{\mathrm{rand}}_2\&CenterDot;\left(p_{gd}-x_{id}^k\right)\\ x_{id}^{k+1}=x_{id}^k+v_{id}^{k+1}\\ v_{id}^m=a\left(m\right)v_{id}^0,m=1,2,...,j\\ x_{id}^m=x_{id}^0+v_{id}^m,m=1,2,...,j\end{array}\right.---\left(3\right)$

Wherein be called the datum velocity component that particle i ties up at d, be called the search speed component that particle i ties up at d, 1≤j≤k+1; be called the reference position component that particle i ties up at d, be called the searching position component that particle i ties up at d; A (m) is called velocity variation coefficient, and in order to determine the relation between two speed, its value can be determined by formula (4):

$a\left(m\right)=\left\{\begin{array}{cc}m& v_{id}^0<v_{min}\\ m/j& v_{id}^0\&GreaterEqual;v_{\max }\\ 1\&PlusMinus;m/j& v_{min}<v_{id}^0<v_{max}\end{array}\right.---\left(4\right)$

Inertia weight W affects the search performance of particle cluster algorithm, and the ability of searching optimum of PSO algorithm is strengthened along with the increase of W value, and Local Search optimizing ability is then contrary.For improving the search capability of algorithm, the weights W of being fixed by numerical value is set as the dynamic value of successively decreasing according to sequential, and the value of W changes by generation according to formula (5):

$W\left(t\right)=W_{max}-\frac{W_{max}-W_{min}}{t_{\max }}\&CenterDot;t---\left(5\right)$

By above-mentioned improvement, improve the traversal performance of particle on search volume, ensure that the diversity of particle flight, dynamic inertia weight achieves the balance of particle swarm optimization algorithm between global search and Local Search preferably, improves the performance of algorithm from speed of convergence and global optimum two aspect.

(2) wind power interval prediction Optimality Criteria and Forecasting Methodology

The target of wind power interval prediction is the reliability ensureing to predict the outcome, and meets decision requirements on this basis as far as possible, improves engineering practicability.If the time series of wind power is X=[x ₁, x ₂..., x _n], U _iwith L _ifor the bound of t=i moment interval prediction, N _tfor choosing the sample size for computation interval parameter, the fiducial probability flight particle that to be θ=100 (1-α) %, ω be in PSO algorithm, the i.e. parameter vector of BP neural network weight and threshold value.For realizing describing the rationalization in wind power prediction interval, the index describing block information is adopted to form the optimization object function of wind power interval prediction, i.e. the fitness function of PSO algorithm.

1) mulching measures

Mulching measures refers to that actual value falls into the probability in forecast interval, is N for capacity _tforecast sample collection, actual value x _ifall into forecast interval [L _i, U _i] probability should be not less than θ=100 (1-α) %.The mulching measures PICP (predictedintervalcoverageprobability) of forecast interval is calculated by formula (6):

$PICP=\frac{1}{N_t}\underset{i=1}{\overset{N_t}{\&Sigma;}}{\mathrm{\&xi;}}_i---\left(6\right)$

${\mathrm{\&xi;}}_i=\left\{\begin{array}{cc}0& x_i\&NotElement;\&lsqb;L_i,U_i\&rsqb;\\ 1& x_i\&Element;\&lsqb;L_i,U_i\&rsqb;\end{array}\right.---\left(7\right)$

5) interval average bandwidth

Different from point forecast, the degree of accuracy predicted the outcome merely is not pursued in the interval prediction of wind power, the more important thing is as electrical network decision-making provides foundation, and it is in theory all acceptable for namely meeting predicting the outcome of certain mulching measures requirement.In interpretational criteria, consider the average bandwidth PINAW (PInormalizedaveragewidth) of forecast interval for this reason, reduce the complete risk of mistake brought close to its ultimate value because of the bound of forecast interval, for normalized wind power time series, PINAW through type (8) calculates.

$PINAW=\frac{1}{N_t}\underset{i=1}{\overset{N_t}{\&Sigma;}}(U_i-L_i)---\left(8\right)$

6) mean center error

In order to assess the distribution consistency degree of actual value in forecast interval, consider the mean center error PIACE (PIaveragecentererror) of forecast interval, the deviation namely between interval intermediate value and real output value, computing formula is:

$PIACE=\sqrt{\underset{i=1}{\overset{N_t}{\&Sigma;}}{(x_i-(U_i+L_i)/2)}^2}---\left(9\right)$

7) optimization object function

The requirement of wind power interval prediction is prediction mulching measures high as far as possible, there are again interval average bandwidth little as far as possible and mean center error simultaneously, i.e. maxPICP (ω), minPINAW (ω), minPIACE (ω).This is a multi-objective optimization question, needs comprehensive indices, sets up wind power interval prediction parametric synthesis optimization object function:

$\begin{array}{c}F\left(\mathrm{\&omega;}\right)=\\ PINAW\&CenterDot;(1+k\&CenterDot;\mathrm{\&lambda;}(PICP)\&CenterDot;e^{PIACE})\end{array}---\left(10\right)$

$\mathrm{\&lambda;}\left(PICP\right)=\left\{\begin{array}{cc}0& PICP\&GreaterEqual;\mathrm{\&theta;}\\ 1& PICP<\mathrm{\&theta;}\end{array}\right.---\left(11\right)$

Wherein k is nonnegative constant, is called punishment parameter.

By the foundation of complex optimum function, wind power interval prediction problem can be converted into the following optimization problem containing constraint:

Min:F(ω)(12)

s.t.θ≤PICP(ω)≤100％

The step setting up wind power interval prediction model is as follows:

(1) choose wind power historical data, set up training dataset and test data set, and original wind power data are normalized.

(2) parameter initialization of forecast model, comprises the initial parameter vector ω of the BP neural network of weights and threshold, the population scale of particle cluster algorithm, iterations, the constant interval of inertia weight, and random initializtion particle position and particle rapidity.

(3) forward prediction of neural network is carried out based on BP neural network parameter vector ω, with past [t-m, t-m+1, t] the wind power data in moment are as input quantity, prediction obtains the bound in t+1 moment wind power interval, and calculates fitness function F (ω) according to formula (10).

(4) obtain by improve PSO algorithm search the particle obtaining optimal fitness function value, be the optimal parameter vector of BP neural network.

(5) utilize the forecast model containing optimum network parameter test data set to be carried out to the interval prediction of wind power, error analysis is carried out to result.

The Structure and Process of wind power interval prediction model is shown in Fig. 2.

The present invention for the measured data of key lab of the intelligent grid Ministry of Education of University Of Tianjin, checking the present invention put forward the validity of algorithm in wind power interval prediction.A hour measured data for the axial fan hub place wind speed in January, 2013 is shown in Fig. 3.Using the physical model that the AOC15/50 blower fan of AtlanticOrient production calculates as wind power, its wind speed-powertrace is shown in Fig. 4, and mathematic(al) representation is shown in formula (13), and the raw data calculating wind power is shown in Fig. 5.

$P=\left\{\begin{array}{cc}0& v<v_{ci},v>v_{co}\\ 0.0444v^3-1.6114v^2+24.272v-87.417& v_{ci}\&le;v\&le;v_{cr}\\ 0.025v^2-1.525v+82.5& v_{cr}<v\&le;v_{co}\end{array}\right.---\left(13\right)$

In formula, P represents the output power of blower fan, and v represents the ambient wind velocity at axial fan hub place, v _cifor the threshold wind velocity of blower fan, v _crfor the nominal operation wind speed of blower fan, v _cofor the cut-off wind speed of blower fan.

The population scale of setting particle cluster algorithm is 20, iterations is 100, and random initializtion particle position and particle rapidity, inertia weight successively decreases with iteration algebraic linear by 0.9 to 0.4.

Selected fiducial probability is 80% and 90% respectively, test set data are adopted to carry out the interval prediction of wind power short-term, respectively with the present invention propose improvement population-BP neural network and basic BP neural network carry out wind power interval prediction, predict the outcome and see Fig. 6, forecast interval parameters is:

From predicting the outcome, the BP neural network interval prediction model based on improve PSO algorithm optimization can realize the short term power interval prediction of wind-powered electricity generation.By the Optimum search of improve PSO algorithm, the interval average bandwidth predicted the outcome and mean center error have clear improvement.With regard to predicting the outcome, the prediction accuracy of PSO-BP model is higher, and when the real output of blower fan is less, when changing relative continuous and stable, bound and the real power value of forecast interval are very close, interval smaller bandwidth; And for having the time period of random fluctuation or Characteristics of Mutation, although bound can follow the tracks of the generating variation tendency of distributed generation system, the broader bandwidth of forecast interval, prediction accuracy is slightly poor.

Can find out according to the interval parameter predicted the outcome, when fiducial probability is larger, the average bandwidth of forecast interval is wider, this is because higher confidence level means higher mulching measures, in order to ensure predicting the outcome, there is higher completeness, the corresponding increase of mean center error.

Claims (2)

1., based on a wind power interval prediction method for population-BP neural network, the method adopts BP neural network to carry out wind power interval prediction, using the input vector of the historical data of wind power as forecast model, and output valve { Y ₁, Y ₂represent the interval bound that future time node wind power prediction exports respectively, by the flight particle that the dimension in the weights and threshold constituent particle group PSO algorithm of BP neural network is 2, by adjustment network weight and threshold value to meet the reliability requirement of forecast model, according to given optimization object function search optimal particle, specific as follows:

If normalized wind power time series is X=[x ₁, x ₂..., x _n], U _iwith L _ifor the upper and lower bound of t=i moment interval prediction, N _tfor choosing the sample size for computation interval parameter, fiducial probability is θ=100 (1-α) %, ω is the flight particle in PSO algorithm, the i.e. parameter vector of BP neural network weight and threshold value, for realizing describing the rationalization in wind power prediction interval, adopt the index describing block information to form the optimization object function of wind power interval prediction, i.e. the fitness function of PSO algorithm, each index is as follows:

1) mulching measures

Mulching measures refers to that actual value falls into the probability in forecast interval, is N for capacity _tforecast sample collection, actual value x _ifall into forecast interval [L _i,u _i] probability should be not less than θ=100 (1-α) %, the mulching measures PICP computing formula of forecast interval is:

2) interval average bandwidth

Consider the average bandwidth PINAW of forecast interval, reduce the complete risk of mistake that the bound because of forecast interval is brought close to its ultimate value, for normalized wind power time series, PINAW is calculated by following formula:

3) mean center error

For the distribution consistency degree of assessment actual value in forecast interval, consider the mean center error PIACE of forecast interval, the deviation namely between interval intermediate value and real output value, computing formula is:

4) optimization object function

The requirement of wind power interval prediction is prediction mulching measures high as far as possible, has again interval average bandwidth little as far as possible and mean center error simultaneously, sets up wind power interval prediction parametric synthesis optimization object function:

F(ω)＝PINAW·(1+k·λ(PICP)·e ^PIACE)

Wherein k is nonnegative constant, is called punishment parameter,

Wind power interval prediction problem is converted into the following optimization problem containing constraint:

Min:F(ω)

s.t.θ≤PICP(ω)≤100％。

2. the wind power interval prediction method based on population-BP neural network according to claim 1, it is characterized in that, modified particle swarm optiziation is adopted to carry out optimizing to the weights and threshold of BP neural network, improve one's methods as follows: on the basis of primary particle speed, increase multiple search speed component, single particle is made to have multiple speed varied in size, the speed that amplitude is larger meets the global optimizing requirement of particle, the speed that amplitude is less can make particle have better local search ability, in order to maintain the diversity of particle; For improving the search capability of algorithm, the weights W of being fixed by numerical value is set as the dynamic value of successively decreasing according to sequential.