CN105824987A - Wind field characteristic statistical distributing model building method based on genetic algorithm - Google Patents

Abstract

The invention provides a wind field characteristic statistical distributing model building method based on genetic algorithm. The implementation process is as follows: A, processing original wind speed and wind direction data; B, estimating parameters in a wind speed probability density function; C, estimating parameters in a wind direction probability density function; D, estimating parameters in a related coefficient probability density function; E, obtaining a wind speed and wind direction joint distribution function.

Description

A kind of Characteristics of Wind Field statistical distribution pattern method for building up based on genetic algorithm

Technical field

The present invention relates to the fields such as monitoring structural health conditions, Structural Wind Engineering, statistical mathematics modeling, be specially and carry out wind speed and direction Joint Distribution parameter estimation and probability Distribution Model foundation based on genetic algorithm.

Background technology

Along with bridge constantly develops to more Long span direction, it is considered to wind has become more and more important to bridge beam action.Research is at bridge site under the conditions of various possible wind fields, and the static(al) effect of bridge structure and dynamic response, can be the design of newly building bridge, construction offer wind resistance scheme.When studying wind to bridge beam action, need to set about accounting for from architectural characteristic, wind characteristic and the wind of bridge in terms of structural interaction three, and wherein analysis to near-earth natural wind characteristic is the most basic and important link.Bridge on-site near-earth wind characteristic is by the basic foundation of Wind-resistance of Bridges design and checking computations.When research wind is to the effect of engineering structure, we will be generally substantially that random natural wind is divided into two kinds of compositions: the average wind represented with average speed and average are the fluctuating wind of zero.

Wherein, the wind speed of average wind is the major parameter describing wind load, but wind speed is stochastic variable, needs to add up by the method for theory of probability, i.e. wind speed probabilistic model, the probability distribution of wind speed is closely related to Structural analysis and design.Up to now, existing a large amount of probabilistic models, for approximating the distribution of wind speed, mainly have Gumbel distribution, Weibull distribution, Gamma distribution, logarithm normal distribution etc..Above-mentioned probability distribution belongs to unimodal model, but the distribution of actual wind speed is the most complicated, uses single distribution form cannot fully describe the multi-modal statistical property of wind speed.Finite mixtures distribution function is made up of a series of distribution function weighted superposition with different statistical property, the distribution function being made up of different statistical properties by estimating the unknown distribution parameter in limited mixed distribution function to simulate.

In addition to wind speed, wind direction is also the important parameter describing wind load.Mean wind speed on the different directions of same place is substantially uneven, and the yardstick that large structure is in different directions often has obvious difference, particularly bridge structure is very big in the yardstick along bridge axes direction and vertical axis direction and the difference such as rigidity, vibration performance, thus this usually to study the Joint Distribution of wind speed and direction to be necessary to introduce wind direction.But, for wind speed, the research for wind direction the most extremely lacks, and its difficulty is mostly derived from following two aspects: one, meteorological observatory wind direction record many employings orientation method and without concrete numerical value；Two, wind angle is periodicity variable.Wind direction is distributed, it is possible to use trigonometric function matching wind direction frequency rectangular histogram sets up the continuous probability density function of wind angle.But the method can not provide fixing probability Distribution Model.At present, the most commonly used wind direction pdf model is VonMises distribution and mixed model based on VonMises distribution.

In order to preferably study the feature of wind load, need to consider wind speed and direction the two factor simultaneously.And the Joint Distribution of wind speed and direction belongs to angle linear distribution, its modeling method can be roughly divided into two categories below: indirect method and direct method.Indirect method is the Joint Distribution first drawing its dependent variable, the most indirectly obtains the Joint Distribution of wind speed and direction.As assume down wind and horizontal wind direction and wind velocity be separate normal variate and both standard deviations equal, being multiplied by the probability density function of the two variable, then by down wind and beam wind, the relation derivation of the wind speed on the two direction and total wind speed and direction goes out Joint Distribution.Direct method is the Joint Distribution directly being constructed both by the marginal probability density function of wind speed and direction, straightforward procedure just assume that two variablees are separate, angle variables obey VonMises distribution, the Joint Distribution model of wind speed variable Normal Distribution.But, wind speed and direction is not separate, be simply multiplied can not be correct the joint distribution function drawing wind speed and direction, then wind speed and direction Joint Distribution is represented by marginal distribution function and the product of two variant correlation coefficient distribution functions of two variablees.

Summary of the invention

The present invention to overcome the deficiency of tradition wind speed and direction distribution modeling method, uses direct method to propose a kind of Characteristics of Wind Field statistical distribution modeling method based on genetic algorithm.

A kind of Characteristics of Wind Field statistical distribution pattern method for building up based on genetic algorithm of the present invention, is made up of following three parts:

One, wind speed and direction Joint Distribution model

The present invention uses direct method to build wind speed and direction Joint Distribution model, and wind speed and direction Joint Distribution probability density function can be expressed as f_V,Θ(v, θ)=2 π g (γ) f_V(v)f_Θ(θ).This function is made up of marginal probability density function and their correlation coefficient function of wind speed and direction, as follows:

(1) the probability density function f of wind speed_VV () obeys finite mixtures weibull distribution, its probability density function can be expressed as:

$f_V\left(v\right)=\underset{l=1}{\overset{N}{\Σ}}{w}_l\frac{{\mathrm{\α}}_l}{{\mathrm{\β}}_l}{\left(\frac{v}{{\mathrm{\β}}_l}\right)}^{{\mathrm{\α}}_l-1}\exp \[-{\left(\frac{v}{{\mathrm{\β}}_l}\right)}^{{\mathrm{\α}}_l}\]$

Wherein, N is component number, and l is the label of each component, and v is the wind speed of average wind, β_lFor range parameter, α_lFor formal parameter, w_lFor each component hybrid weight, and meet:

$\underset{l=1}{\overset{N}{\Σ}}{w}_l=1$ And w_l≥0

(2) the probability density function f of wind direction_Θ(θ) obeying VonMises distribution, its probability density function can be expressed as:

$f_{\mathrm{\Θ}}\left(\mathrm{\θ}\right)=\underset{j=1}{\overset{N}{\Σ}}\frac{{\mathrm{\ω}}_j}{2{\mathrm{\πI}}_0\left(k_j\right)}\·\exp \[k_{j}cos(\mathrm{\θ}-{\mathrm{\μ}}_j)\]$

$I_0\left(k_j\right)=\frac{1}{\sqrt{2\mathrm{\π}}}{\∫}_0^{2\mathrm{\π}}\exp \[k_{j}cos\mathrm{\θ}\]d\mathrm{\θ}$

Wherein, N is component number, and j is the label of each component, and θ is the wind direction that average wind wind speed is corresponding, k_jFor lumped parameter, μ_jFor mean wind direction, I₀(k_j) be 0 rank revise the primal Bessel function, ω_jFor each component hybrid weight, and meet:

$\underset{l=1}{\overset{N}{\Σ}}{\mathrm{\ω}}_l=1$ And ω_l≥0

(3) correlation function g (γ) is the probability density function of correlation coefficient γ, and wherein γ can be expressed as:

$\mathrm{\γ}=2\mathrm{\π}\[{\∫}_0^v{f}_V\left(v\right)dv-{\∫}_0^{\mathrm{\θ}}{f}_{\mathrm{\Θ}}\left(\mathrm{\θ}\right)d\mathrm{\θ})\]$

If γ≤0, then γ=2 π+γ

The probability density function g (γ) of correlation coefficient also obeys VonMises distribution, and its probability density function can be expressed as

$g\left(\mathrm{\γ}\right)=\underset{j=1}{\overset{N}{\Σ}}\frac{{\mathrm{\ω}}_j}{2{\mathrm{\πI}}_0\left(k_j\right)}\·\exp \[k_{j}cos(\mathrm{\γ}-{\mathrm{\μ}}_j)\]$

$I_0\left(k_j\right)=\frac{1}{\sqrt{2\mathrm{\π}}}{\∫}_0^{2\mathrm{\π}}\exp \[k_{j}cos\mathrm{\γ}\]d\mathrm{\γ}$

Wherein, N is component number, and j is the label of each component, and γ is the correlation coefficient of average wind wind speed and direction, k_jFor lumped parameter, μ_jFor the meansigma methods of correlation coefficient, I₀(k_j) be 0 rank revise the primal Bessel function, ω_jFor each component hybrid weight, and meet:

$\underset{l=1}{\overset{N}{\Σ}}{\mathrm{\ω}}_l=1$ And ω_l≥0

Two, parameter estimation algorithm

Owing to wind speed is stochastic variable, only use finite mixtures distribution function just can fully describe the multi-modal statistical property of wind speed.In order to estimate the parameter in limited mixed distribution function, propose multiple method in recent years, including regression analysis, graphing method, method of least square, moments estimation method, Maximum Likelihood Estimation Method etc..EM algorithm is built upon a kind of New Algorithm on Maximum Likelihood Estimation Method basis, and this algorithm is by introducing some hidden variables so that the Maximum-likelihood estimation of parameter becomes easy.EM algorithm is presently the most conventional method for parameter estimation, EM algorithm clear thinking, calculates simple, but due to the stringency on its mathematical theory basis, is only applicable to the parameter estimation of one-dimension random variable distribution function.And traditional optimized algorithm is gradient based on object function or higher derivative, utilize steepest gradient method to obtain, be easily caused optimization process and be absorbed in local optimum.

Along with bionic huge progress, development in recent years has been got up a kind of effective optimization tool genetic algorithm.Genetic algorithm is exactly to simulate a kind of algorithm of biological evolution and development process.When use genetic algorithm, first individuality to be optimized encoded and generate initial population, from population, selecting more excellent individuality by selection opertor, using the operations such as intersection, variation thus produce a new generation population and participate in the coding calculating of next round.By iterating, until meeting the condition of convergence, now in population, optimum individual is considered as i.e. globally optimal solution.Genetic algorithm has certain collimation, its search is more to many processes, and conventional search methods is usually the search procedure of point-to-point, therefore genetic algorithm can rapid optimizing in bigger design space, and carry out optimizing with population scale, some individualities are absorbed in local optimum and have no effect on whole evolution process, have stronger of overall importance.Due to this algorithm, to have search speed fast, it is to avoid being absorbed in the advantages such as local optimum, the present invention uses the parameter that genetic algorithm can be estimated in model the most accurately.

Three, data process

The Data Source of the present invention is bridge health monitoring system.Anemoclinograph one on bridge floor has two kinds: ultrasonic type anemoclinograph and mechanical type anemoclinograph.Can obtain will being better than in substantial amounts of initial data, quantity and precision the data of weather station from the collecting device of bridge health monitoring system.But the initial data collected in bridge health monitoring system is the real time data of wind speed and direction, need initial data is processed, obtain the wind speed and direction of average wind needed for Characteristics of Wind Field is analyzed.China specify average wind time away from being typically taken as 10 minutes, it is therefore desirable to the data of the wind speed and direction of 10 minutes are averaging, count the wind speed and direction data of average wind.

The problem that the invention solves the problems that the following aspects:

One, wind speed and direction data volume in traditional method is solved few, incomplete problem.The data of conventional process are to obtain from local weather station, and general obtain is day maximum average mark speed, and wind direction is also to there is no concrete numerical value by azimuth recording.The wind speed and direction data that this method uses are to be provided by bridge health monitoring system, not only can obtain real-time data, and wind direction data are angularly records, provide excellent basis for wind speed and direction Joint Distribution model parameter estimation.

Two, the problem that traditional method cannot obtain a fixing probability Distribution Model is solved.When wind speed and direction is studied by traditional method, it is common that respectively the wind speed in some direction is studied, then the Wind speed model of all directions is showed respectively, it is impossible to provide fixing probability Distribution Model.And the present invention is when modeling, consider wind speed and direction the two factor, it is possible to represent the joint distribution function of wind speed and direction with a fixing probability density function simultaneously.

Three, traditional method is solved it cannot be guaranteed that the solution drawn is the problem of globe optimum.Traditional statistical analysis method is a kind of hill-climbing algorithm, makes its likelihood function value constantly increase by not stopping iteration, and this makes algorithm become a kind of local search algorithm, is only able to find local optimum.Genetic algorithm carries out optimizing with population scale, and some individualities are absorbed in local optimum and have no effect on whole evolution process, is a kind of global optimization approach.And initial value is chosen more sensitive by the convergence rate of traditional statistical analysis method and convergence point.Generally initial value is stochastic generation in its span, will find optimal solution along fixing route after initial value determines.Therefore, initial value choose the position directly influencing optimal solution.And the genetic algorithm that the present invention uses is basic without seriality, the property led requirement or strict mathematical theory to object function, therefore there is the wider array of suitability.

A kind of Characteristics of Wind Field statistical distribution pattern method for building up based on genetic algorithm of the present invention, is embodied as flow process as follows:

A. original wind speed and direction data are processed；

A1. with 10 minutes for for the moment away from, obtain the meansigma methods of ten minutes interior wind speed and direction data；

A2. wind speed and direction data being stored in excel or matlab file, wind speed is deposited in a column, and wind direction is deposited in a column, and with there is same a line away from interior wind speed and direction data one_to_one corresponding for the moment；

B. the parameter in wind speed probability density function is estimated；

B1. the scope of wind speed is between 0m/s to 20m/s.Taking an interval is [0,20], and its lower limit is more slightly smaller than minimum data, and its upper limit is more slightly larger than maximum data, and this interval is divided into 200 minizone R_V, each interval be spaced apart 0.1.So sample falls at interval R_VFrequency be represented byWherein K_vFor falling into interval R_VSample size, K is sample total, and ξ is interval R_VArea,For wind speed probability density function, x_vFor interval R_VThe x coordinate of central point, β is range parameter, and α is formal parameter, w_lFor each component hybrid weight；

B2. population scale is taken as 200, and maximum evolutionary generation is 5000, carries out 10 tests altogether, takes the maximum Search Results of fitness value as final argument estimated value.Its fitness function can be expressed as

$FIT=\frac{1}{\underset{v=1}{\overset{V}{\Σ}}{\left(\frac{q_v-f\left(x_v\right|w,\mathrm{\α},\mathrm{\β})\mathrm{\ξ}}{q_v}\right)}^2}$

Wherein f (x_v| w, α, β) it is wind speed probability density function.Work as q_vCloser to f (x_v| w, α, β) ξ time, the value of fitness function is the biggest.The value making fitness function reaches maximum range parameter β, formal parameter α, and each component hybrid weight w is the optimal solution of parameter；

B3. the estimates of parameters utilizing gained calculates the AIC value under different component number and Δ_cTo determine best composition number, component number therein starts until AIC value and Δ from 1_cTill no longer notable change.

Red pond information criterion AIC is a model selection criteria based on likelihood function value, is expressed as

AIC=2M-2ln (L)

The number of unknown parameter during wherein M is distributed model, represents the discount increasing parameter, and ln (L) is very big log-likelihood function, for weighing the goodness of fit of model.

Distance measure Δ_cIt is to derive from goodness of fit hypothesis testing.Δ_cValue is represented by

${\mathrm{\Δ}}_c=\underset{v=1}{\overset{T}{\Σ}}\frac{{(q_v-h_v)}^2}{q_v},$

Wherein T is interval division number, q_vIt is distributed in interval R for matching_vInterior frequency, h_vInterval R is fallen into for sample_vInterior probability.Select the component number that AIC value is corresponding with Δ c minima, establish optimum model of fit；

C. the parameter in wind direction probability density function is estimated；

C1. wind direction angular range is between 0 to 2 π.Taking interval is [0,2 π], and this interval is divided into 200 minizone R_V, each interval be spaced apart 0.1 π.So sample falls at interval R_VFrequency be represented byWherein K_vFor falling into interval R_VSample size, K is sample total, and ξ is interval R_VArea,For wind direction probability density function, x_θFor interval R_VThe x coordinate of central point, k_jFor lumped parameter, μ_jFor mean wind direction, w_jFor each component hybrid weight；

C2. population scale is taken as 200, and maximum evolutionary generation is 5000, carries out 10 tests altogether, takes the maximum Search Results of fitness value as final argument estimated value.Its fitness function can be expressed as

$FIT=\frac{1}{\underset{v=1}{\overset{V}{\Σ}}{\left(\frac{q_v-f\left(x_{\mathrm{\θ}}\right|w,k,\mathrm{\μ})\mathrm{\ξ}}{q_v}\right)}^2}$

Wherein f (x_θ| w, k, μ) it is wind direction probability density function.Work as q_vCloser to f (x_θ| w, k, μ) ξ time, the value of fitness function is the biggest.The value making fitness function reaches maximum lumped parameter k, mean wind direction μ, and each component hybrid weight w is the optimal solution of parameter；

C3. the estimates of parameters utilizing gained calculates the AIC value under different component number and Δ_cTo determine best composition number, component number therein starts until AIC value and Δ from 1_cTill no longer notable change.

Red pond information criterion AIC is a model selection criteria based on likelihood function value, is expressed as

AIC=2M-2ln (L)

The number of unknown parameter during wherein M is distributed model, represents the discount increasing parameter, and ln (L) is very big log-likelihood function, for weighing the goodness of fit of model.

Distance measure Δ_cIt is to derive from goodness of fit hypothesis testing.Δ_cValue is represented by

${\mathrm{\Δ}}_c=\underset{v=1}{\overset{T}{\Σ}}\frac{{(q_v-h_v)}^2}{q_v}$

Wherein T is interval division number, q_vIt is distributed in interval R for matching_vInterior frequency, h_vInterval R is fallen into for sample_vInterior probability.Select AIC value and Δ_cThe component number that minima is corresponding, establishes optimum model of fit；

D. the parameter in correlation coefficient probability density function is estimated；

D1. drawn the probability density function of wind speed and direction by two steps of B and C, then usedEach correlation coefficient corresponding to wind speed and direction data can be calculated；

D2. carry out C1, the step for of C2, C3, correlation coefficient data are processed；

D3. optimum lumped parameter k, the average correlation coefficient μ in correlation coefficient probability density function, and each component hybrid weight w are obtained；

E. wind speed and direction joint distribution function；

E1. calculate wind speed, wind direction, the probability density function of correlation coefficient by genetic algorithm more than, i.e. can obtain wind speed and direction probability density function f_V,Θ(v, θ)=2 π g (γ) f_V(v)f_Θ(θ)。

Compared with the existing methods, this method has a following advantage:

1, data used in the present invention are to be provided by bridge health monitoring system, will be better than in traditional method handled data in all fields.Bridge health monitoring system recorded data is wind speed and direction data all the time, and wind speed and direction data are accurate to metre per second (m/s) and degree respectively, has good precision；

2, the anemoclinograph in bridge health monitoring system is arranged on bridge floor, and the data recorded can well show bridge on-site near-earth wind characteristic, can design the foundation basic with checking computations offer for Wind-resistance of Bridges；

3, in bridge health monitoring system, anemoclinograph have recorded substantial amounts of initial data, can obtain studying required appointment data by processing these initial datas.This, compared to weather station recorded data, more can meet the needs of analysis；

4, solve and traditional statistical analysis method does not fully takes into account the direction problem to air speed influence, because wind speed is different in a different direction, and the wind in different directions is also different on the impact of bridge structure, it is impossible to only consider this factor of wind speed.The present invention sufficiently take into account this key factor of wind direction；

5, solve traditional statistical analysis method and do not take into account the problem that is mutually related between wind speed and direction.The marginal probability density function of wind speed and direction is relevant, and the marginal probability density function of both is simply simply multiplied by traditional processing method, and the model drawn does not has enough degree of accuracy.The present invention is when setting up wind speed and direction joint density function, it is contemplated that the dependency between both marginal probability density functions；

6, solve traditional statistical analysis method to fail to propose a fixing modeling method to the problem analyzing the Joint Distribution of wind speed and direction.The present invention proposes a fixing wind speed and direction joint distribution function f_V,Θ(v,θ)；

7, the problem that traditional statistical analysis method is only able to find local optimum is solved.The genetic algorithm that the present invention uses carries out optimizing with population scale, and some individualities are absorbed in local optimum and have no effect on whole evolution process, is a kind of global optimization approach；

8, compared to traditional statistical analysis method, it is fast that genetic algorithm has search speed, it is to avoid is absorbed in local optimum, and the suitability is high, result precision advantages of higher.

Accompanying drawing explanation

Fig. 1 is the parameter estimation flow chart of the present invention；

Fig. 2 is the genetic algorithm flow chart of the present invention.

Detailed description of the invention

With reference to the accompanying drawings, technical scheme is further illustrated.

A kind of Characteristics of Wind Field statistical distribution pattern method for building up based on genetic algorithm of the present invention, comprises the steps:

A. original wind speed and direction data are processed；

A1. with 10 minutes for for the moment away from, obtain the meansigma methods of ten minutes interior wind speed and direction data；

A2. wind speed and direction data being stored in excel or matlab file, wind speed is deposited in a column, and wind direction is deposited in a column, and with there is same a line away from interior wind speed and direction data one_to_one corresponding for the moment；

B. the parameter in wind speed probability density function is estimated；

B1. wind speed is between 0m/s to 20m/s.Taking an interval is [0,20], and its lower limit is more slightly smaller than minimum data, and its upper limit is more slightly larger than maximum data, and this interval is divided into 200 minizone R_V, each interval be spaced apart 0.1.So sample falls at interval R_VFrequency be represented by

B2. population scale is taken as 200, and maximum evolutionary generation is 5000, carries out 10 tests altogether, takes the maximum Search Results of fitness value as final argument estimated value.Its fitness function can be expressed as

$FIT=\frac{1}{\underset{v=1}{\overset{V}{\Σ}}{\left(\frac{q_v-f\left(x_v\right|w,\mathrm{\α},\mathrm{\β})\mathrm{\ξ}}{q_v}\right)}^2}$

B3. the estimates of parameters utilizing gained calculates the AIC value under different component number and Δ_cTo determine best composition number, component number therein starts until AIC value and Δ from 1_cTill no longer notable change.

Red pond information criterion AIC is a model selection criteria based on likelihood function value, is represented by

AIC=2M-2ln (L)

Distance measure Δ_cIt is to derive from goodness of fit hypothesis testing, Δ_cValue is represented by

${\mathrm{\Δ}}_c=\underset{v=1}{\overset{T}{\Σ}}\frac{{(q_v-h_v)}^2}{q_v}$

Select AIC value and Δ_cThe component number that minima is corresponding, establishes optimum model of fit；

C. the parameter in wind direction probability density function is estimated；

C1. wind speed is between 0 to 2 π.Taking an interval is [0,2 π], and this interval is divided into 200 minizone R_V, each interval be spaced apart 0.1 π.So sample falls at interval R_VFrequency be represented by $q_v=\frac{K_v}{K}\≈f\left(x_{\mathrm{\θ}}\right|w,\mathrm{\α},\mathrm{\β})\mathrm{\ξ};$

C2. population scale is taken as 200, and maximum evolutionary generation is 5000, carries out 10 tests altogether, takes the maximum Search Results of fitness value as final argument estimated value.Its fitness function can be expressed as

$FIT=\frac{1}{\underset{v=1}{\overset{V}{\Σ}}{\left(\frac{q_v-f\left(x_{\mathrm{\θ}}\right|w,k,\mathrm{\μ})\mathrm{\ξ}}{q_v}\right)}^2}$

C3. the estimates of parameters utilizing gained calculates the AIC value under different component number and Δ_cTo determine best composition number, component number therein starts until AIC value and Δ from 1_cTill no longer notable change.

Red pond information criterion AIC is a model selection criteria based on likelihood function value, is represented by

AIC=2M-2ln (L)

Distance measure Δ_cIt is to derive from goodness of fit hypothesis testing.Δ_cValue is represented by

${\mathrm{\Δ}}_c=\underset{v=1}{\overset{T}{\Σ}}\frac{{(q_v-h_v)}^2}{q_v}$

Select AIC value and Δ_cThe component number that minima is corresponding, establishes optimum model of fit；

D. the parameter in correlation coefficient probability density function is estimated；

D1. drawn the probability density function of wind speed and direction by two steps of B and C, then usedEach correlation coefficient corresponding to wind speed and direction data can be calculated；

D2. carry out C1, the step for of C2, C3, correlation coefficient data are processed；

D3. optimum lumped parameter k, the average correlation coefficient μ in correlation coefficient probability density function, and each component hybrid weight w are obtained；

E. wind speed and direction joint distribution function；

E1. calculate wind speed, wind direction, the probability density function of correlation coefficient by genetic algorithm more than, i.e. can obtain wind speed and direction probability density function f_V,Θ(v, θ)=2 π g (γ) f_V(v)f_Θ(θ)。

Content described in this specification case study on implementation is only enumerating of the way of realization to inventive concept; protection scope of the present invention is not construed as being only limitted to the concrete form that case study on implementation is stated, protection scope of the present invention also and in those skilled in the art according to present inventive concept it is conceivable that equivalent technologies means.

Claims (1)

1. a Characteristics of Wind Field statistical distribution pattern method for building up based on genetic algorithm, is embodied as flow process as follows:

A. original wind speed and direction data are processed；

A1. with 10 minutes for for the moment away from, obtain the meansigma methods of ten minutes interior wind speed and direction data；

A2. wind speed and direction data being stored in excel or matlab file, wind speed is deposited in a column, and wind direction is deposited in a column, and with there is same a line away from interior wind speed and direction data one_to_one corresponding for the moment；

B. the parameter in wind speed probability density function is estimated；

B1. the scope of wind speed is between 0m/s to 20m/s.Taking an interval is [0,20], and its lower limit is more slightly smaller than minimum data, and its upper limit is more slightly larger than maximum data, and this interval is divided into 200 minizone R_V, each interval be spaced apart 0.1.So sample falls at interval R_VFrequency be represented byWherein K_vFor falling into interval R_VSample size, K is sample total, and ξ is interval R_VArea,For wind speed probability density function, x_vFor interval R_VThe x coordinate of central point, β is range parameter, and α is formal parameter, w_lFor each component hybrid weight；

B2. population scale is taken as 200, and maximum evolutionary generation is 5000, carries out 10 tests altogether, takes the maximum Search Results of fitness value as final argument estimated value.Its fitness function can be expressed as

$FIT=\frac{1}{\underset{v=1}{\overset{V}{\Σ}}{\left(\frac{q_v-f\left(x_v\right|w,\mathrm{\α},\mathrm{\β})\mathrm{\ξ}}{q_v}\right)}^2}$

Wherein f (x_v| w, α, β) it is wind speed probability density function.Work as q_vCloser to f (x_v| w, α, β) ξ time, the value of fitness function is the biggest.The value making fitness function reaches maximum range parameter β, formal parameter α, and each component hybrid weight w is the optimal solution of parameter；

B3. the estimates of parameters utilizing gained calculates the AIC value under different component number and Δ_cTo determine best composition number, component number therein starts until AIC value and Δ from 1_cTill no longer notable change.

Red pond information criterion AIC is a model selection criteria based on likelihood function value, is expressed as

AIC=2M-2ln (L)

The number of unknown parameter during wherein M is distributed model, represents the discount increasing parameter, and ln (L) is very big log-likelihood function, for weighing the goodness of fit of model.

Distance measure Δ_cIt is to derive from goodness of fit hypothesis testing.Δ_cValue is represented by

${\mathrm{\Δ}}_c=\underset{v=1}{\overset{T}{\Σ}}\frac{{(q_v-h_v)}^2}{q_v},$

Wherein T is interval division number, q_vIt is distributed in interval R for matching_vInterior frequency, h_vInterval R is fallen into for sample_vInterior probability.Select the component number that AIC value is corresponding with Δ c minima, establish optimum model of fit；

C. the parameter in wind direction probability density function is estimated；

C1. wind direction angular range is between 0 to 2 π.Taking interval is [0,2 π], and this interval is divided into 200 minizone R_V, each interval be spaced apart 0.1 π.So sample falls at interval R_VFrequency be represented byWherein K_vFor falling into interval R_VSample size, K is sample total, and ξ is interval R_VArea,For wind direction probability density function, x_θFor interval R_VThe x coordinate of central point, k_jFor lumped parameter, μ_jFor mean wind direction, w_jFor each component hybrid weight；

C2. population scale is taken as 200, and maximum evolutionary generation is 5000, carries out 10 tests altogether, takes the maximum Search Results of fitness value as final argument estimated value.Its fitness function can be expressed as

$FIT=\frac{1}{\underset{v=1}{\overset{V}{\Σ}}{\left(\frac{q_v-f\left(x_{\mathrm{\θ}}\right|w,k,\mathrm{\μ})\mathrm{\ξ}}{q_v}\right)}^2}$

Wherein f (x_θ| w, k, μ) it is wind direction probability density function.Work as q_vCloser to f (x_θ| w, k, μ) ξ time, the value of fitness function is the biggest.The value making fitness function reaches maximum lumped parameter k, mean wind direction μ, and each component hybrid weight w is the optimal solution of parameter；

C3. the estimates of parameters utilizing gained calculates the AIC value under different component number and Δ_cTo determine best composition number, component number therein starts until AIC value and Δ from 1_cTill no longer notable change.

Red pond information criterion AIC is a model selection criteria based on likelihood function value, is expressed as

AIC=2M-2ln (L)

The number of unknown parameter during wherein M is distributed model, represents the discount increasing parameter, and ln (L) is very big log-likelihood function, for weighing the goodness of fit of model.

Distance measure Δ_cIt is to derive from goodness of fit hypothesis testing.Δ_cValue is represented by

${\mathrm{\Δ}}_c=\underset{v=1}{\overset{T}{\Σ}}\frac{{(q_v-h_v)}^2}{q_v}$

Wherein T is interval division number, q_vIt is distributed in interval R for matching_vInterior frequency, h_vInterval R is fallen into for sample_vInterior probability.Select AIC value and Δ_cThe component number that minima is corresponding, establishes optimum model of fit；

D. the parameter in correlation coefficient probability density function is estimated；

D1. drawn the probability density function of wind speed and direction by two steps of B and C, then usedEach correlation coefficient corresponding to wind speed and direction data can be calculated；

D2. carrying out C1, correlation coefficient data are processed by C2, C3 these three step；

D3. optimum lumped parameter k, the average correlation coefficient μ in correlation coefficient probability density function, and each component hybrid weight w are obtained；

E. wind speed and direction joint distribution function；

E1. calculate wind speed, wind direction, the probability density function of correlation coefficient by genetic algorithm more than, i.e. can obtain wind speed and direction probability density function f_V,Θ(v, θ)=2 π g (γ) f_V(v)f_Θ(θ)。