CN103400204B - Solar energy power generating amount Forecasting Methodology based on SVM Markov combined method - Google Patents

Abstract

The invention discloses a kind of solar energy power generating amount Forecasting Methodology based on SVM Markov combined method, the method comprises the following steps: (1) selects intensity of solar radiation, daily maximum temperature, relative humidity, wind speed as early warning factor；(2) a certain amount of sample data is collected according to early warning factor；(3) tentatively set up SVM regressive prediction model, and utilize sample data to be trained, determine SVM model structure；(4) the SVM model structure utilizing step (3) to obtain carries out photovoltaic power generation quantity preliminary forecasting；(5) application Markov approach is modified predicting the outcome；(6) predicted the outcome.The present invention utilizes support vector machine (SVM) to carry out regression analysis, and revised predicting the outcome by Markov approach, method agrees with photovoltaic generation feature, and the two is had complementary advantages, thus obtain predicting the outcome the most accurately, it is achieved the reliable prediction to photovoltaic power generation quantity.

Description

Solar energy power generating amount Forecasting Methodology based on SVM-Markov combined method

Technical field

The present invention relates to Solar use research field, particularly to a kind of sun based on SVM-Markov combined method Can method for forecasting photovoltaic power generation quantity.

Background technology

Nowadays, traditional fossil fuel energy is the most exhausted, and the harm that environment is caused by its burning simultaneously also becomes increasingly conspicuous, Energy crisis that traditional fuel is brought and environmental problem have become as the ultimate challenge of facing mankind.For human society can Sustainable development, countries in the world have been invested new and renewable sources of energy sight one after another, have been greatly developed and place high hopes, it would be desirable to Adjustment is restructured the use of energy present situation, it is ensured that mankind's energy security.Compared with water energy, wind energy, geothermal energy, bioenergy etc., solar energy The focus that people pay attention to is become so that it highlights exclusive advantage.Abundant solar radiant energy is inexhaustible, nexhaustible, and light Volt TRT noiseless, pollution-free, cheap, scaleable, it is easy to human freedom, extensively utilize.According to statistics, solar energy is per second Clock arrives the energy on ground and is up to 800,000 kilowatts, if the solar energy of earth surface 0.1% is transferred to electric energy, and number turnover is 5%, then Every annual electricity generating capacity is just up to 5.6 × 10¹²Kilowatt hour, is equivalent to 40 times of world's total energy consumption.Therefore, photovoltaic generation enjoys favor also It is used widely.

But, affected by extraneous complicated uncertain factor, there is randomness, undulatory property, intermittence, the most true in photovoltaic generation The shortcoming such as qualitative, and photovoltaic output also presents nonlinear relation with factor of influence, it is past that this results in photovoltaic generation power Toward being extremely unstable, economy, safe and stable operation on electrical network cause serious impact and threat.

Realize the prediction of solar energy power generating amount, it will help dispatching of power netwoks department overall arrangement normal power supplies and photovoltaic The cooperation of generating, in time adjust operation plan, reasonable arrangement power system operating mode.On the one hand, can effectively weaken Photovoltaic accesses the adverse effect bringing electrical network, improves the safety and stability of Operation of Electric Systems；On the other hand, it is possible to decrease The spinning reserve capacity of power system and operating cost, to make full use of solar energy resources, it is thus achieved that bigger economic benefit and society Can benefit.

At present, photovoltaic power generation quantity prediction can be largely classified into statistical method and Artificial Neural Network two class.Statistics Method is by historical data is carried out statistical analysis, utilizes theory of probability to find out its inherent law and for predicting；And it is artificial Sample data as input, through machine training test study, is set up forecast model and will be carried out pre-to future by neural net method Survey.Both the above method is applied in terms of photovoltaic power generation quantity prediction, but its having some limitations property of method, such as Regular and the strongest data message, both Forecasting Methodologies can reach higher precision of prediction, but photovoltaic generation is deposited In the feature such as randomness, undulatory property, using both approaches, effect is the most very poor, it is impossible to meet functional need.

Therefore, find a kind of method that photovoltaic power generation quantity can be carried out reliable prediction and there is important practical value.

Summary of the invention

Present invention is primarily targeted at the shortcoming overcoming prior art with not enough, it is provided that a kind of based on SVM-Markov The solar energy power generating amount Forecasting Methodology of combined method, the method utilizes support vector machine (SVM) to carry out regression analysis, and Being revised predicting the outcome by Markov approach, method agrees with photovoltaic generation feature, and the two is had complementary advantages, Thus obtain predicting the outcome the most accurately, it is achieved the reliable prediction to photovoltaic power generation quantity.

The purpose of the present invention is realized by following technical scheme: photovoltaic based on SVM-Markov combined method Generated energy Forecasting Methodology, comprises the following steps:

(1) suitable early warning factor is selected；

(2) a certain amount of sample data is collected according to early warning factor；

(3) tentatively set up SVM regressive prediction model, and utilize sample data to be trained, determine SVM model structure；

(4) the SVM model structure utilizing step (3) to obtain carries out photovoltaic power generation quantity preliminary forecasting；

(5) application Markov approach is modified predicting the outcome；

(6) predicted the outcome.

Affect a lot of because have of solar energy power generating amount, such as intensity of solar radiation, ambient temperature, relative humidity, Wind speed and setting angle, photoelectric conversion rate etc..Wherein, for setting angle, photoelectric conversion rate etc., device can be relied on to adjust Photovoltaic generation output effect is made to reach optimal with technological progress etc., and intensity of solar radiation, ambient temperature, relative humidity, wind speed All that people institute is uncontrollable, this some affect the key of photovoltaic generation output the most exactly.The output of photovoltaic cell Power is mainly affected by ambient temperature and intensity of solar radiation.And relative humidity, wind speed are by impact radiation indirect action In the increase of photovoltaic power generation quantity, such as relative humidity, the steam in air can stop Effective radiation.The reality of photovoltaic generation The result that output these trend just interacts.Analyze based on above, the early warning factor selected in described step (1) For: intensity of solar radiation, daily maximum temperature, relative humidity, wind speed are as early warning factor.Above early warning factor all can be from every day Weather forecast obtains.

Further, for intensity of solar radiation, it is mainly affected by weather conditions, and ultraviolet is solar radiation A part, both are essentially identical with the situation of change of weather conditions, therefore can select ultraviolet index approximate sign too Sun radiant intensity.Ultraviolet index excursion typically represents by the numeral of 0～15, and therefore, solar radiant energy can also be by Ultraviolet index is digitized characterizing, be entered as 1 the most successively, 2,3 ..., 14,15.

Preferably, described step (3) carries out pretreatment to the sample data gathered before setting up SVM regressive prediction model, bag Include the normalized etc. of the elimination of singular data, linear interpolation, data.To ensure the accuracy of institute's established model.

Support vector machine (SVM) method is built upon Statistical Learning Theory SLT (Statistic Learn Theory) On the basis of VC (Vapnik-Chervonenkis) dimension theory and Structural risk minization principle, according to finite sample information at mould Optimal compromise is sought, to obtaining best Generalization Ability between complexity and the learning capacity of type.SVM returns based on minimizing Structure risk rather than traditional empirical risk minimization, the basic thought of method is by a nonlinear mapping, Data sample is mapped to high-dimensional feature space, and carries out linear regression in this space.It is implemented as employing support to The regression model of amount machine, for sample data, finds the dependence between the input of this system data and output so that it is as far as possible Predict the output of the unknown exactly, be described as follows with mathematical linguistics:

A given data sample set { (x₁,y₁),(x₂,y₂),…,(x_i,y_i),…,(x_n,y_n), wherein, input to Amount x_i∈Rⁿ, export data y_i∈R,i=1,2,…,n.At RⁿOne optimal objective function f (x) of upper searching, to use y=f X () infers arbitrary input y value corresponding to x, make expected risk minimize simultaneously.

The basic thought setting up photovoltaic power generation quantity regressive prediction model is data to be reflected by nonlinear mapping Φ (x) It is mapped to higher dimensional space, and spatially carries out linear regression at this.

Optimal objective function f is:

f(ω,b)=ωΦ(x)+b

Through the answer algorithm of excess convexity double optimization problem, find globally optimal solution, and then above formula be represented by:

$f\left(x\right)=\underset{i=1}{\overset{l}{\mathrm{\Σ}}}(a_i-a_i^{*})K(x,x_i)+b$

Wherein, (x y) is kernel function, a to K_i,It it is learning sample solution when minimizing expected risk；B is the knot optimized Really.A according to the character of SVM regression function, only minority_i,It is not zero, the vector i.e. referred to as support that these parameters are corresponding Vector machine (SVM).

Specific to the present invention, described step (3) particularly as follows:

(3-1) sample data step (2) collected is as input data set, and exporting data is photovoltaic power generation quantity；

(3-2) kernel function is selected；

(3-3) SVM regressive prediction model is trained by sample data, it is thus achieved that support vector accordingly, and determine therefrom that this SVM The structure of regressive prediction model.

Further, the kernel function used in described step (3-2) is RBF function, it may be assumed that

$K(x,x_i)=\exp (-\frac{{|x-x_i|}^2}{{\mathrm{\σ}}^2}).$

The undulatory property of solar energy power generating is relatively big, is the stochastic process of a non-stationary, for limited sample, if only Support vector machine is utilized to train, it was predicted that model often imperfect stability.In view of state probability transfer square in Prediction of Markov Battle array has the ability following the trail of variable random fluctuation, and it has markov property in addition, if organically combining with SVM regression model Come, excavate macroscopic view change and the microoscillations rule of generating data sequence, it is possible to be effectively improved the precision of prediction of model. Concrete, the step that step of the present invention (5) application Markov approach is modified predicting the outcome is as follows:

(5-1) state grade of sample data is divided: utilize the SVM model structure forecast sample data that step (3) obtains, And compare with actual power value, obtain Relative Error δ；According to relative error δ of each test sample predictive value, really Determine the mobility scale of δ, and by the upper lower threshold value [δ residing for δ₀,δ_n] as state demarcation codomain, and determine state demarcation standard, build Vertical Markov state set S:S₁[δ₀,δ₁], S₂[δ₁,δ₂], S₂[δ₂,δ₃] ..., S_i[δ_i-1,δ_i] ..., S_n[δ_n-1,δ_n]；

(5-2) state transition probability matrix is set up: for often organizing photovoltaic power generation quantity relative error data δ, if δ is ∈ [δ_i-1, δ_i], i.e. event S_iOccur, then event is in state S_i, state S_iState S is transferred to through k step_jProbability be:

$P_{\mathrm{ij}}^{\left(k\right)}=\frac{N_{\mathrm{ij}}^k}{N_i};$

In formula,For sample state from S_iTo S_jTransfer number, N_iFor state S_iThe total degree of transfer occurs；Then k walks shape State transition probability matrix is:

$P^{\left(k\right)}=\left(\begin{array}{cccc}{P}_{11}^{\left(k\right)}& P_{12}^{\left(k\right)}& \·\·\·& P_{1n}^{\left(k\right)}\\ P_{21}^{\left(k\right)}& P_{22}^{\left(k\right)}& \·\·\·& P_{2n}^{\left(k\right)}\\ \·\·\·& \·\·\·& \·\·\·& \·\·\·\\ P_{n1}^{\left(k\right)}& P_{n2}^{\left(k\right)}& \·\·\·& P_{\mathrm{nn}}^{\left(k\right)}\end{array}\right);$

(5-3) determine state shift result: set the X (k) state probability vector as the k moment, X (0) be known initial time The state probability vector carved；P^(k)For state transition probability matrix, then triadic relation meets:

X(k)=X(0)P^(k)；

Determined state probability vector X (k) in k moment by above formula, choose column vector state maximum in acquired results and make For next step steering state；

(5-4) the photovoltaic power generation quantity initial predicted value obtained by step (4) is revised: object SVM regression model to be predicted is pre- The interval S of correction variation of measured value y_i[δ_i-1,δ_i], ask for the median in intervalThen predict object future time instance Predictive value, i.e. y^*=y(1-s_i)。

The present invention compared with prior art, has the advantage that and beneficial effect:

1, the present invention uses Support vector regression forecast model, and utilizes Markov approach to be modified, and has SVM concurrently Model and the advantage of Markov approach, can utilize less sample data to model, and forecasts general trend, is suitable for again undulatory property Bigger random sequence forecast, agrees with the feature of solar energy power generating, can predict photovoltaic power generation quantity easy, accurately.This If invention is widely popularized, strong help, cooperation normal power supplies and photovoltaic will be provided for dispatching of power netwoks department Generating, the most in time arranges operation plan, it is ensured that safe and stable, the reliability service of whole power system.

2, the present invention chooses intensity of solar radiation, daily maximum temperature, relative humidity, wind speed as early warning factor, is all too Most important influence factor in sun energy photovoltaic power generation quantity, plays decisive role, and data easily obtain from weather forecast simultaneously, right In the intensity of solar radiation that inconvenience directly obtains, have employed uitraviolet intensity be digitized characterize, feasible reliably.

Accompanying drawing explanation

Fig. 1 is the schematic flow sheet of the inventive method.

Detailed description of the invention

Below in conjunction with embodiment and accompanying drawing, the present invention is described in further detail, but embodiments of the present invention do not limit In this.

Embodiment 1

As it is shown in figure 1, solar energy power generating amount Forecasting Methodology based on SVM-Markov combined method includes following step Rapid:

S1: select suitable early warning factor.

The present embodiment select intensity of solar radiation, daily maximum temperature, relative humidity, wind speed as early warning factor.Above Early warning factor all can obtain from weather forecast every day.Wherein intensity of solar radiation also can be digitized by ultraviolet index Characterize, be entered as 1 the most successively, 2,3 ..., 14,15.

S2: collect a certain amount of sample data according to early warning factor.

S3: tentatively set up SVM regressive prediction model, and utilize sample data to be trained, determine SVM model structure.

SVM Forecasting Methodology achieves the nonlinear mapping between data space and feature space, can be effectively by data Various nonlinear operations in space develop into corresponding linear operation in feature space, and then greatly increase non-linear place Reason ability, can preferably solve the practical problems such as small sample, dimension non-linear, high, local minimum point.Set up SVM regression forecasting Before model, the sample data gathered is carried out pretreatment, including the elimination of singular data, linear interpolation, the normalized of data Deng.

According to analysis above, early warning factor intensity of solar radiation, daily maximum temperature, relative humidity, wind speed will be selected to make For input data set, and exporting data is photovoltaic power generation quantity, and above input data can obtain from weather forecast.

Linear function conventional in SVM, RBF (RBF), Sigmoid function etc., through comparing, select RBF Function, as kernel function, trains SVM regressive prediction model by sample data, it is thus achieved that support vector accordingly, and determine therefrom that this The structure of regression model.RBF function, it may be assumed that

$K(x,x_i)=\exp (-\frac{{|x-x_i|}^2}{{\mathrm{\σ}}^2}).$

S4: utilize the SVM model structure obtained to carry out photovoltaic power generation quantity preliminary forecasting.

Obtain predictive value y.

S5: application Markov approach is modified predicting the outcome.

Markov Chain describes a kind of status switch, and each of which state value depends on the most limited state, therefore may be used With the state present according to some variable and tendency of changes thereof, it was predicted that its shape being likely to occur within following a certain specific period State.This Forecasting Methodology is suitable for describing the bigger problem of stochastic volatility, such as the solar photovoltaic generating amount prediction described in the present embodiment.

Step is as follows:

(5-1) state grade of sample data is divided: utilize the SVM model structure forecast sample data that step (3) obtains, And compare with actual power value, obtain Relative Error δ；According to relative error δ of each test sample predictive value, really Determine the mobility scale of δ, and by the upper lower threshold value [δ residing for δ₀,δ_n] as state demarcation codomain, and determine state demarcation standard, build Vertical Markov state set S:S₁[δ₀,δ₁], S₂[δ₁,δ₂], S₂[δ₂,δ₃] ..., S_i[δ_i-1,δ_i] ..., S_n[δ_n-1,δ_n]；

(5-2) state transition probability matrix is set up:

According to Markov theory, for often organizing photovoltaic power generation quantity relative error data δ, experiment each time there may be Various states occurs, if event S_iOccur, then event is in state S_i.State S_iState S is transferred to through k step_jProbability be:

$P_{\mathrm{ij}}^{\left(k\right)}=\frac{N_{\mathrm{ij}}^k}{N_i};$

In formula,For sample state from S_iTo S_jTransfer number, N_iFor state S_iThe total degree of transfer occurs.Then k step State-transition matrix:

$P^{\left(k\right)}=\left(\begin{array}{cccc}{P}_{11}^{\left(k\right)}& P_{12}^{\left(k\right)}& \·\·\·& P_{1n}^{\left(k\right)}\\ P_{21}^{\left(k\right)}& P_{22}^{\left(k\right)}& \·\·\·& P_{2n}^{\left(k\right)}\\ \·\·\·& \·\·\·& \·\·\·& \·\·\·\\ P_{n1}^{\left(k\right)}& P_{n2}^{\left(k\right)}& \·\·\·& P_{\mathrm{nn}}^{\left(k\right)}\end{array}\right)$

(5-3) determine that state shifts result: P^(k)Metastatic rule between each state of the system that reflects, utilization state probability Transfer matrix, it may be determined that variable state in which and most probable value maxP thereof in photovoltaic power generation quantity forecast error ordered series of numbers_ij K (), so that it is determined that next step of variable turns to.

Markov Chain forecast model is represented by:

X(k)=X(0)P^(k)

In formula, X (k) is the state probability vector in k moment；X (0) is the state probability vector of initial time；P^(k)For state Transition probability matrix.

X (0) is by the initial time Determines selected, if there being 4 kinds of states, selected original state is in state S₁, then X is remembered (0)=(1,0,0,0), and X (k) is by calculating column vector Determines maximum in acquired results, if calculating gained maximum column vector It is the 3rd row, then remembers X (k)=(0,0,1,0).

(5-4) the photovoltaic power generation quantity initial predicted value obtained by step (4) is revised: the value to be predicted determined according to (3) After transfering state, i.e. can get the interval S of correction variation of object SVM forecast of regression model value y to be predicted_i[δ_i-1,δ_i], utilize Interval medianRevise the predictive value of prediction object future time instance, i.e. y^*=y(1-s_i)。

SVM regression model through Markov correction can reach the highest essence in solar energy power generating amount is predicted Degree, this effectively assessment being photovoltaic generation power provides strong guarantee.

S6: obtain final predicting the outcome.

The present embodiment uses Support vector regression forecast model, and utilizes Markov approach to be modified, and has SVM concurrently Model and the advantage of Markov approach, if being widely popularized, will provide strong help for dispatching of power netwoks department.

Above-described embodiment is the present invention preferably embodiment, but embodiments of the present invention are not by above-described embodiment Limit, the change made under other any spirit without departing from the present invention and principle, modify, substitute, combine, simplify, All should be the substitute mode of equivalence, within being included in protection scope of the present invention.

Claims (5)

1. solar energy power generating amount Forecasting Methodology based on SVM-Markov combined method, it is characterised in that include following step Rapid:

(1) suitable early warning factor is selected；Early warning factor is: intensity of solar radiation, daily maximum temperature, relative humidity, wind speed；Choosing Sign intensity of solar radiation is approximated by ultraviolet index；

(2) a certain amount of sample data is collected according to early warning factor；

(3) tentatively set up SVM regressive prediction model, and utilize sample data to be trained, determine SVM model structure；

(4) the SVM model structure utilizing step (3) to obtain carries out photovoltaic power generation quantity preliminary forecasting；

(5) application Markov approach is modified predicting the outcome, and step is as follows:

(5-1) state grade of sample data is divided: utilize the SVM model structure forecast sample data that step (3) obtains, and with Actual power value compares, and obtains Relative Error δ；According to relative error δ of each test sample predictive value, determine δ Mobility scale, and by the upper lower threshold value [δ residing for δ₀,δ_n] as state demarcation codomain, and determine state demarcation standard, set up Markov state collection S:S₁[δ₀,δ₁], S₂[δ₁,δ₂], S₂[δ₂,δ₃] ..., S_i[δ_i-1,δ_i] ..., S_n[δ_n-1,δ_n]；

(5-2) state transition probability matrix is set up: for often organizing photovoltaic power generation quantity relative error data δ, if δ is ∈ [δ_i-1,δ_i], i.e. Event S_iOccur, then event is in state S_i, state S_iState S is transferred to through k step_jProbability be:

$P_{ij}^{\left(k\right)}=\frac{N_{ij}^k}{N_i};$

In formula,For sample state from S_iTo S_jTransfer number, N_iFor state S_iThe total degree of transfer occurs；Then k step state turns Shifting probability matrix is:

$P^{\left(k\right)}=\left(\begin{array}{cccc}{P}_{11}^{\left(k\right)}& P_{12}^{\left(k\right)}& ...& P_{1n}^{\left(k\right)}\\ P_{21}^{\left(k\right)}& P_{22}^{\left(k\right)}& ...& P_{2n}^{\left(k\right)}\\ ...& ...& ...& ...\\ P_{n1}^{\left(k\right)}& P_{n2}^{\left(k\right)}& ...& P_{nn}^{\left(k\right)}\end{array}\right);$

(5-3) determine that state shifts result: setting the X (k) state probability vector as the k moment, X (0) is known initial time State probability vector；P^(k)For state transition probability matrix, then triadic relation meets:

X (k)=X (0) P^(k)；

Determined state probability vector X (k) in k moment by above formula, choose in acquired results maximum column vector state as under One step steering state；

(5-4) the photovoltaic power generation quantity initial predicted value obtained by correction step (4): object SVM forecast of regression model value y to be predicted The correction interval S of variation_i[δ_i-1,δ_i], ask for the median in intervalThen predict the prediction of object future time instance Value, i.e. y^*=y (1-s_i)；

(6) predicted the outcome.

Solar energy power generating amount Forecasting Methodology based on SVM-Markov combined method the most according to claim 1, its Being characterised by, the ultraviolet index excursion numeral of 0～15 represents, solar radiant energy is also by ultraviolet index number Wordization characterizes, be entered as 1 the most successively, 2,3 ..., 14,15.

Solar energy power generating amount Forecasting Methodology based on SVM-Markov combined method the most according to claim 1, its Being characterised by, described step (3) carries out pretreatment, including unusual to the sample data gathered before setting up SVM regressive prediction model The elimination of data, linear interpolation, the normalized of data.

Solar energy power generating amount Forecasting Methodology based on SVM-Markov combined method the most according to claim 1, its Be characterised by, described step (3) particularly as follows:

(3-1) sample data step (2) collected is as input data set, and exporting data is photovoltaic power generation quantity；

(3-2) kernel function is selected；

(3-3) SVM regressive prediction model is trained by sample data, it is thus achieved that support vector accordingly, and determine therefrom that this SVM returns The structure of forecast model.

Solar energy power generating amount Forecasting Methodology based on SVM-Markov combined method the most according to claim 4, its Being characterised by, the kernel function used in described step (3-2) is RBF function, it may be assumed that

$K(x,x_i)=\exp (-\frac{|x-x_i{|}^2}{{\mathrm{\σ}}^2}).$