The content of the invention
It is an object of the invention to provide a kind of Probabilistic Load computational methods cut down based on scene, to reduce
When enchancement factor ratio is higher in power network data fluctuations it is larger cause probabilistic load flow precise decreasing the problem of.
To reach above-mentioned purpose, present invention employs following technical scheme:
1.1) enchancement factor in power system is sampled, obtains initial scene library;
1.2) initial scene library is cut down, obtains typical scene;According to typical scene to complete in initial scene library
Portion's scene is classified, and obtains scene collection, and scene collection is calculated according to the scene classification and probability characteristics of initial scene library
Probability characteristics;
1.3) probabilistic load flow is carried out using the Probabilistic Load Flow algorithm based on Cumulants method respectively to each scene collection;
1.4) probability characteristics of scene collection is combined, the characteristics of tidal flow to each scene collection carries out probability superposition, obtains power network
Trend distribution character.
The step 1.1) specifically include following steps:The feature of enchancement factor is obtained according to the initial data of power system
Distribution, then carries out n times sampling using Monte Carlo Method to enchancement factor, often carries out single sample and just obtains an initial scene
And its probability characteristics.
The step 1.2) specifically include following steps:
1) initial scene is cut down with fast forword back-and-forth method, until the initial scene retained reaches the target of setting
Scene number n, typical scene is used as using the n of reservation initial scenes;
2) initial scene is classified according to the probability metrics between initial scene and typical scene, made just by classification
Beginning scene is converged according to typical scene, so that it is determined that the border between initial scene;
3) the probability q occurred for j-th of scene collectionjObtained by following formula:
Wherein, i ∈ j represent that i-th of scene will be divided into j-th of scene collection, p in initial scene libraryiFor initial scene
The probability that i-th of scene occurs in storehouse.
For certain initial scene oi, the initial scene and each typical scene s are calculated respectivelyjBetween probability metrics;Should
Initial scene is divided into a class with the minimum corresponding typical scene of probability metrics, i.e., i-th of scene will be divided in initial scene library
To scene collectionpic(oi,sj) it is scene oiAnd sjBetween probability metrics, c (oi,sj)
For scene oiAnd sjBetween Euclidean distance.
The step 1.3) specifically include following steps:
1) modal equation and branch equation in power system are subjected to Taylor expansion at typical scene, then pass through line
Property, obtain sensitivity matrix and transfer matrix;
2) all scenes are concentrated for scene, the L rank central moments of scene are calculated using the method for statistics, according to the L
Rank central moment calculates the L rank cumulant that power network injects variable, and L ranks cumulant, the sensitivity square of variable are injected according to power network
Battle array and transfer matrix calculate the L rank cumulant for obtaining electric network state variable;
3) utilize Gram-Charlier series expansions, by the Probability Characteristics of electric network state variable be expressed as normal state with
The series of machine variables L order derivative composition, series coefficients are determined according to the L ranks of electric network state variable standardization cumulant.
The value of the L is 5~7.
The standardization cumulant is calculated according to below equation to be obtained:
Wherein, giFor the i-th rank cumulant after standardization,For the i powers of electric network state variable standard deviation, χiFor electricity
I-th rank cumulant of netted state variable, i=1 ..., L.
The computational methods of the L ranks central moment are as follows:
Wherein, E is scene desired value, λiI-th of the scene concentrated for scene, m is the number of scene collection Scene, βlIt is
The l rank central moments of scene, l=1 ..., L.
The step 1.4) specifically include following steps:The Probabilistic Load Flow characteristic of scene collection is carried out into probability according to the following formula to fold
Plus, obtain the trend distribution character of power network:
Wherein, f and F are respectively the probability density function vector sum cumulative distribution function vector of electric network swim, fjAnd FjRespectively
For the probability density function vector sum cumulative distribution function vector of electric network swim under j-th of scene collection, qjFor correspondence scene collection
Probability, n is the number of scene collection.
Beneficial effects of the present invention are embodied in:
The present invention is cut down initial scene, the typical scene for classification is formed, by being divided initial scene
Class and probability reallocation, it is possible to reduce due to the Cumulants method calculation error that data fluctuations are big and cause, by each field
Jing Ji characteristics of tidal flow is overlapped, so as to obtain the overall trend distribution character of power network.The invention enables in face of it is complicated,
The solution of electric network swims in large scale, containing a variety of enchancement factors (regenerative resource, energy storage, controllable burden etc. at high proportion) is asked
Topic, computational accuracy is also ensured while solution efficiency is improved, calculated results can as the reference for Electric Power Network Planning according to
According to engineering actual person can deploy correlative study work accordingly.
Embodiment
The present invention is described in further details with reference to the accompanying drawings and examples.
1. the generation of scene, reduction and the formation of scene collection
1) scene is generated
Firstly the need of generation database.Optimization expectational model containing stochastic variable can be with following form it is convex random excellent
The model of expected value of change is represented:
As the random vector parameter ω in Expection optimal time of the policymaker according to actual conditions function of pursuit of the objective, formula (2)
Following (or observing) Discrete Stochastic number, these following random numbers clothes in the cards in the cards will be converted into
From random vector parameter ω probability distribution, Discrete Stochastic number here is " scene ".
In Practical Project problem, original probability estimates P by continuously or by many discrete scenes constituting, it is difficult to logical
The method for crossing parsing tries to achieve the optimal desired value of formula (2), it is necessary to which originally discrete or continuous probability measure P is further discrete
Change, i.e., with limited huge sample approximate representation probability measure P.Finite number amount is obtained according to probability-distribution function huge
Sample approximate representation probability measure P is exactly " scene generation ".Enchancement factor in each pair power system completes single sample, just gives birth to
Into a scene.Multiple sampling is carried out, number (or the sampling time in initial scene library, initial scene library comprising scene is built
Number) depending on the precision and the dimension of scene needed for research.
2) scene is cut down
The number comprising enchancement factor is more and higher to required precision in usual power system, and initial scene library is often wrapped
Containing thousands of scenes.The data error that enchancement factor data fluctuations are brought can influence the accuracy of follow-up result of calculation,
It is therefore desirable to carry out scene reduction to initial scene library.But so-called scene abatement feature is in this patent:1) using quick
Forward selection procedures selection typical scene (desired value number is general between 5~10), is entered using typical scene to initial scene library
Row subregion (is classified) to initial scene, forms scene collection;2) probability reallocation is carried out to the scene collection obtained after classification,
Obtain the corresponding probability of scene collection.Based on two above feature, the present invention obtains multiple scene collection and scene collection is corresponding
Probability, compared to the degree of fluctuation for equally reducing enchancement factor with initial scene library inside each scene collection, helps to carry
The computational accuracy of high Cumulants method.
The known various approximate models and structure for some particular problem for modelling shown in formula (2).Wherein ζ
The mathematical notation that structure probability is estimated is as follows:
The reality of formula (3) is meant that:" solve Stochastic Optimization Model ∫ in formula (2)ΩF (ω, x) Pd ω expectation optimal value "
With " solving Stochastic Optimization Model ∫ΩF (ω, x) Qd ω expectation optimal value " it is of equal value.Therefore the P scales in the formula (2) compared with
It is big and when be difficult to be fully described, probability measure Q solving models can be simplified using the approximate of P." how to obtain optimal letter
It is exactly described " scene reduction " to change scene collection Q ".It is inherently an optimization problem on Q that scene, which cuts down problem,.
The reduction speed of fast forword back-and-forth method, cuts down thought intuitively, therefore the scene of the present invention is cut down in algorithm
Using the method.In order to illustrate fast forword back-and-forth method, it is necessary to introduce optimal reduction problem:
Wherein, N is initial scene number, and n is the target scene number after cutting down, c (ωi,ωj) it is scene ωiAnd ωjBetween
Euclidean distance, piFor scene ωiThe probability of generation.Formula (4) illustrates that initial scene set { 1,2 ..., N } is divided into
Two parts, a part be cut in scene numbering set J, another part be remain scene numbering gather 1,
2 ..., N } J, in the optimization problem, object function is minimum DJValue, optimized variable is J.
The problem of in normal conditions following formula (4), does not have effective algorithm, but is directed to the scene number #J=N-1 being cut in
This case, the solution of the problem becomes relatively easy.Fast forword back-and-forth method is to be based on this case, now equivalent to choosing
A point u is selected, it is not cut in, formula (4) is converted into:
Wherein, u is the point that need not be deleted, and in addition all points are all deleted, and so just have selected one least
The scene that can be cut in, then continues similarly with selection and the probability redistribution problem for considering other scenes again.Fast forword
The step of back-and-forth method, is as follows:
1. initial calculation
OrderFirst is selected
The point that need not be deletedJ1={ 1,2 ..., N } { u1}。
2. cycle calculations
It is determined that i-th of (i>1) the point u that need not be deleted being selectediDuring,Computational methods it is different
In initial calculation, it is necessary to use the value in the i-th -1 time calculating, i.e.,For i-th -1
The individual point that need not be deleted chosen,The point being selected for i-th
Ji=Ji-1\{ui}。
3. probability is reallocated
Probability reallocation is substantially exactly that N number of scene in initial scene library is classified.If oi(i=1 ..., N) be
Any scene in initial scene library, piThe probability that (i=1 ..., N) occurs for any scene in initial scene library, sj(j=
1 ..., it is n) any scene (typical scene) in reservation scene library.The present invention is according between initial scene and typical scene
Probability metrics is classified to initial scene, i.e., for any one initial scene, and the initial scene and each allusion quotation are calculated respectively
Probability metrics between type scene;The initial scene is divided into one with obtaining typical scene corresponding during probability metrics minimum value
I-th of scene will be divided into scene collection in class, i.e., initial scene libraryI ∈ j are denoted as,
pic(oi,sj) it is scene oiAnd sjBetween probability metrics.Any scene collection setjThe probability q of appearancejObtained by following formula:
3) formation of scene collection
Present invention definition converges to typical scene sjAll initial scenes (including scene sjItself) constitute scene collection
setj, setjIn the initial scene number that includes be numj, then have:
setj={ oi|i∈j} (7)
The classification to initial scene is so completed, that is, forms scene collection.
2. the probabilistic load flow cut down based on scene
Cumulants method is very sensitive to the fluctuation of data, therefore the data that a high proportion of enchancement factor is brought in power network
Larger fluctuation and randomness can cause error calculated to increase.Because carrying out Taylor expansion at benchmark operating point
The high-order term of 2 times and the above is have ignored during linearisation, if enchancement factor fluctuation is larger, then the error of linearisation can also increase therewith
Greatly.The present invention has carried out division operation to initial scene library, and the border between scene sample is determined using scene classification method,
The fluctuation of enchancement factor is limited in the subregion where it, i.e., inside scene collection.This equally reduces the fluctuation of enchancement factor
Degree, so as to improve the computational accuracy of Cumulants method.The Cumulants method that the present invention is used is illustrated below.
1) scene collection (node injection variable) cumulant is calculated
The number of scenes included in initial scene library is N, and scene integrates number as n, and the probability after fast forword selection divides again
The classifying and numbering to N number of scene in initial scene library is completed with process, multiple scene collection are formed.Institute is concentrated for research scene
Some scene samples, each rank central moment for obtaining scene is calculated using the method for statistics:
Wherein, E is scene desired value, λiI-th of the scene concentrated for research scene, m concentrates scene for research scene
Number, βlIt is the l rank central moments of scene.
Relation between central moment and cumulant is shown below:
Wherein, γlFor l rank cumulant.The present invention takes preceding 7 rank cumulant, and its precision, which is met, to be required.According to formula
(9)-(11), obtain the preceding 7 rank cumulant of research scene collection, for subsequently calculating.
2) Cumulants method
In order to which application of the Cumulants method in probabilistic load flow is better described, power system linearisation is firstly introduced into
Thought.The modal equation and branch equation of power system are represented with the form of matrix, and in benchmark operating point (i.e. in scene collection
Typical scene) place carries out Taylor series expansion to it, ignores the high-order term of 2 times and the above:
In view of being met at benchmark operating point:
Just lienarized equation of the benchmark operating point to random perturbation Δ W can be obtained:
In formula:Subscript 0 represents benchmark operating point, and W is that node injects variable, and X is node state variable, and Z is membership
Variable.S0For sensitivity matrix, T0For transfer matrix.
Cumulant has two critical natures:
1. additive property:Each rank cumulant of independent random variable sum be equal to each stochastic variable each rank cumulant it
With.
2. it is homogeneity:The r ranks half that r (r >=1, r is integer) rank cumulant of a times of stochastic variable is equal to the variable are constant
The a of amountrTimes.
Utilize the node injection rate shown in this two critical natures and formula (14) and node state amount and line status
Relation between amount, just can be obtained according to the cumulant of node injection rate node state amount and line status amount partly not
Variable.
The random change is determined by each rank cumulant of stochastic variable (state variable such as node voltage, line power)
The method of amount probability characteristics has many kinds.Gram-Charlier series expansions have stronger representativeness, its series expansion mode
It is as follows:
Wherein, N (t) is normpdf, and t is the stochastic variable after standardization, χiFor stochastic variable
I-th rank cumulant, giFor the i-th rank cumulant after standardization, Hi(t) it is the i-th rank Hermite multinomials, EX is random change
Measure X desired value, δXFor stochastic variable X standard deviation,For the i powers of stochastic variable X standard deviations.
3) scene reduction and the combination of Cumulants method
Cut down by scene and probability reallocation link, obtain multiple scene collection setjAnd its corresponding probability qj;Each
Cumulant inside scene collection can also be obtained by the method for statistics.Each scene collection, which possesses, uses Cumulants method
Carry out all key elements of probabilistic load flow.The Cumulants method key step cut down based on scene is as shown in Figure 3:
1. scene reduction is carried out, the probability that typical scene, corresponding scene collection and scene collection occur is obtained, using statistics
Method calculates each rank cumulant obtained under scene collection.
2. it is directed to scene collection setj, because it possesses whole key elements needed for Cumulants method, using cumulant
Method carries out probabilistic load flow, obtains the Probability Characteristics of the scene collected state amount, the probability density letter comprising quantity of state
Number fj(x) with cumulative distribution function Fj(x)。
3. the probability density function and probability-distribution function of certain quantity of state (i.e. state variable) under each scene collection are calculated, is led to
Cross the Probability Characteristics (Fig. 4) that probability principle of stacking obtains the quantity of state:
All quantity of states (including active power, reactive power on node voltage amplitude, phase angle and branch road) are pressed
Illuminated (18) is calculated, i.e., the characteristics of tidal flow to each scene collection is overlapped, then obtains the overall trend distribution character of power network.
Simulation example
1. generate database
Initial data acquired in arranging, such as main distribution network structure, enchancement factor characteristic, system initial parameter, and
Sliding-model control is carried out to data according to its probability-distribution function, according to the distribution characteristics of enchancement factor in power system to wind
These enchancement factors such as machine, photovoltaic, energy storage device, controllable burden are sampled, and are often completed single sample and are just obtained a scene.
Multiple sampling is carried out, great amount of samples is obtained, initial scene library is generated.So that number of scenes is 500 initial scene library as an example, initially
Scene distribution is as shown in figure 1, comprising 500 sample points in Fig. 1, represent 500 initial scenes, for convenience on plan
Shown and illustrated, each scene is only comprising 2 factors;The span of each factor is that the integer between 0-100 (is included
0 and 100);The probability that each scene occurs is identical, is all 0.2%, and probability sum is 1.
2. the initial scene library of pair generation carries out scene reduction
A large amount of original scenes are cut down with quick former generation method.Eliminated using quick former generation and field is carried out to initial scene library
Scape is cut down, and specific process of cutting down is referring to formula (2)-(6), and target scene number is 10, scene distribution such as Fig. 2 institutes after reduction
Show, it can be seen that typical scene is multiple cluster centres of initial scene library, in initial scene library, the field around typical scene
Scape distribution is more intensive.
3. form scene collection
Initial scene is classified according to formula (7), (8), scene collection is formed.
4. probabilistic load flow
In view of application background be containing blower fan, photovoltaic, energy storage device, controllable burden these enchancement factors power network, this hair
Bright use Cumulants method carries out the calculating of Probabilistic Load Flow to power network.
1) cumulant of scene collection, the connection set up between scene collection and cumulant are calculated according to formula (9)-(11)
System.
2) modal equation and branch equation in power system are subjected to Taylor expansion at typical scene, ignore it is secondary and
High-order term above, obtains formula (12), the lienarized equation of (14) form, so as to calculate the cumulant of state variable.
3) Gram-Charlier series is utilized, the Probability Characteristics of state variable are expressed as normal state according to formula (15)
The series of stochastic variable all-order derivative composition, series coefficients are determined according to each rank of state variable standardization cumulant.
5. scene cuts down the combination with Cumulants method
By the Probabilistic Load Flow characteristic comprehensive analysis of the probability of scene collection and each scene collection, obtain being investigated according to formula (18)
The probability density characteristicses of electric network state amount.
Simulation result shows, using the method for the present invention, compared with conventional Cumulants method, realizes initial scene library
Division operation, can equally reduce the degree of fluctuation of enchancement factor, so as to improve Cumulants method computational accuracy.
In a word, the present invention proposes the concept of scene collection, after realizing that scene is cut down, completes the classification to initial scene,
So cause scene between border determination sharpening, help to be combined with cumulant calculating power system load flow, to larger
Initial scene library cut down and obtain scene collection on the basis of, carry out probability with Cumulants method inside each scene collection
The calculating of trend, while improving efficiency, can be effectively ensured computational accuracy.Method proposed in the present invention can be applied to
The Load flow calculation of large-scale complex power grid containing regenerative resource at high proportion, can effectively be solved due to power system enchancement factor ripple
Cumulants method calculation of tidal current precise decreasing problem caused by moving greatly, it is ensured that the height of the probabilistic load flow of power network is accurate
True property.