CN103049651A - Method and device used for power load aggregation - Google Patents

Method and device used for power load aggregation Download PDF

Info

Publication number
CN103049651A
CN103049651A CN2012105373292A CN201210537329A CN103049651A CN 103049651 A CN103049651 A CN 103049651A CN 2012105373292 A CN2012105373292 A CN 2012105373292A CN 201210537329 A CN201210537329 A CN 201210537329A CN 103049651 A CN103049651 A CN 103049651A
Authority
CN
China
Prior art keywords
class
sample
center
cluster
module
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN2012105373292A
Other languages
Chinese (zh)
Inventor
姚丽娟
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Aerospace Science and Industry Shenzhen Group Co Ltd
Original Assignee
Aerospace Science and Industry Shenzhen Group Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Aerospace Science and Industry Shenzhen Group Co Ltd filed Critical Aerospace Science and Industry Shenzhen Group Co Ltd
Priority to CN2012105373292A priority Critical patent/CN103049651A/en
Publication of CN103049651A publication Critical patent/CN103049651A/en
Pending legal-status Critical Current

Links

Images

Abstract

The invention discloses a method and device used for power load aggregation. The method and device used for power load aggregation comprise S1, acquiring sample data of n transformer substation comprehensive load static characteristics; S2, mapping the sample data to a Hilbert space through a gaussian kernel function and acquiring samples; S3, confirming an initial aggregation center by selecting K samples in the samples; S4, performing aggregation calculation on mapped samples in a core space by adopting a k-means algorithm, and assigning each sample to closest-type upper and lower approximations according to an upper and lower approximation method; S5, dynamically adjusting weights omega1, omegabnr according to a current iteration; S6, calculating a Jomega value according to an arithmetic convergence criteria, and judging whether |Jomega(t)- Jomega(t-1)|<= epsilon or t>=tmax, if yes, then generating a final collection and finishing, if not, then entering into S7; and S7, enabling t=t+1, reconfirming an aggregation center and transferring to S4. The method and device used for power load aggregation is simple, easy, fast and effective, aggregation results are reasonable, and has important significance on practicability of load modeling research.

Description

A kind of method and device for the electric load cluster
Technical field
The invention belongs to electric load clustering technique field, more specifically, relate to a kind of method for the electric load cluster and device.
Background technology
When carrying out Power System Analysis, transformer station's part throttle characteristics cluster be exactly with same electrical network different substation and in the different period transformer station's characteristic approach or similar poly-be a class, as take the heavy industry load as main transformer station, take the municipal administration life as main transformer station, in the transformer station month season load curve, day season load curve etc.And the characteristic of each class is described with same model, can set up the integrated load model of effectively describing transformer station's characteristic by cluster.With regard to load modeling practical, the classification of part throttle characteristics is of great significance with comprehensive tool, and it can arrive the work or physical labor intensity that alleviates the operation and maintenance engineering.The key that the classification of part throttle characteristics is practical with comprehensively being the realization load model, for setting up suitable transformer station's load model, most of algorithms are all introduced clustering method in the Load Characteristic Analysis.
At present, the method that is applied to the electric load cluster has several lower patterns: (1) traditional clustering algorithm and improvement algorithm: electrical network is carried out simulation analysis such as the load characteristics clustering algorithm based on improved k average (k-means), based on the electric load analysis of grid, based on electric load analysis of spectral clustering etc.These clustering algorithms all are that traditional clustering algorithm is applied in the electric system.(2) based on the clustering algorithm of intelligent optimization algorithm: such as the electric load cluster analysis based on population, based on electric load analysis of ant group etc.These algorithms are carried out combination and be applied to utilizing the global optimizing ability of intelligent optimization algorithm to improve clustering performance in the electric load analysis, also shortened the cluster time.(3) based on the combination of many algorithms: such as the electric load cluster analysis based on the ACO-PAM integration algorithm, based on power system load characteristic clustering method of Parallel Neural Network with Particle Swarm Optimization etc., these algorithms can be effectively in conjunction with the advantage of various algorithms, improve the accuracy of cluster, also enlarged simultaneously the scope of application of clustering method, cluster result is more readily understood, and commercial exploitation is worth wider.
Present power system load data all are complicated and diversified, such as the characteristics with abnormality, distributivity, polyphyly, higher-dimension, complicacy, non-linear and magnanimity.The applicability of above several Clusterings can not satisfy current demand for electric system far away, as not processing nonlinear data, can not process border object, can not process higher-dimension mixed type data etc.The result that these shortcomings cause present cluster to go out can not meet reality completely, and the cluster result availability is also relatively poor, can not make business decision for the supvr accurately and efficiently.
Summary of the invention
Defective for prior art, the object of the present invention is to provide a kind of method for the electric load cluster, be intended to solve clustering method of the prior art and can not process nonlinear data, can not process border object and can not process higher-dimension mixed type data and cause the cluster result poor availability, can not make accurately and efficiently the problem of business decision for the supvr.
For achieving the above object, the invention provides a kind of method for the electric load cluster, comprise the steps:
S1: obtain sample matrix
Figure BDA00002578237300021
This sample matrix comprises that n is capable, concrete sample of each behavior, from
Figure BDA00002578237300022
Arrive
Figure BDA00002578237300023
Each sample has d attribute, as each sample
Figure BDA00002578237300024
D row component, namely each element in Φ (x) matrix can be expressed as
Figure BDA00002578237300025
I=1 wherein, 2 ..., n, j=1,2 ..., d, sample All belong to the F space;
S2: determine initial cluster center C, C={C by in n the sample of sample matrix Φ (x), choosing k sample 1, C 2..., C k, wherein C represents the set of matrices at k class center, and k is clusters number, comprises that k is capable, and the concrete center of each each class of behavior is from C 1To C k, the center of each class also has d attribute accordingly, as every class center C pD row component, each element in the C matrix can be expressed as C Pj, p=1 wherein, 2 ..., k, j=1,2 ..., d,
Figure BDA00002578237300027
S3: make iterations t=1;
S4: described sample matrix Φ (x) is carried out cluster calculation, and various kinds is originally distributed to the upper and lower approximate collection of nearest class according to the method that upper and lower approximate collection is determined;
S5: according to the weight factor ω of current iteration number of times t to described lower approximate collection lWeight factor ω with described approximate collection BnrDynamically adjust;
S6: the target function value J that calculates the rough kernel cluster w, and judge | J w(t)-J w(t-1) |≤ε or t 〉=t MaxWhether set up, if then finish; If not, then enter step S7; ε is a parameter of determining according to actual conditions, t MaxBe the artificial iterations upper limit of determining;
S7: make t=t+1, redefine cluster centre C p, and change step S4 over to.
Further, described step S1 is specially: obtain the accurate sample of n transformer station's synthetic load static characteristics, consist of accurate sample matrix X={x 1, x 2..., x n, x i∈ R d, and obtain sample after by gaussian kernel function described accurate sample data being mapped to the Hilbert space
Figure BDA00002578237300031
Accurate sample x iAll belong to space R dIn; Described gaussian kernel function is: and F (x, y)=exp (β || x-y|| 2/ 2 σ 2), wherein σ is the width parameter of function, and β is that 1, x is the accurate sample before the conversion, and y is the sample after changing.
Further, in step S4, the method that upper and lower approximate collection is determined is specially:
Figure BDA00002578237300032
Figure BDA00002578237300033
If Then order
Figure BDA00002578237300035
Otherwise order
Figure BDA00002578237300036
γ=0.07 wherein;
Figure BDA00002578237300037
Represent i sample
Figure BDA00002578237300038
To the Euclidean distance at p class center,
Figure BDA00002578237300039
Represent i sample
Figure BDA000025782373000310
Minimum value in the Euclidean distance at q class center, p=1 wherein, 2 ... k, q=1,2 ... k, p ≠ q; Dist (C p, C q) represent the center of p class to the Euclidean distance at the center of q class,
Figure BDA000025782373000311
Be the upper approximate collection of q class,
Figure BDA000025782373000314
It is the lower approximate collection of p class.
Further, in step S5, adopt formula
Figure BDA000025782373000312
With
Figure BDA000025782373000313
Weight factor ω to lower approximate collection lWeight factor ω with described approximate collection BnrDynamically adjust.
Further, in step S6, adopt formula
Figure BDA00002578237300041
I=1,2 ..., n, the target function value J of calculating rough kernel cluster wWherein, w jBe the characteristic weighing coefficient, and
Figure BDA00002578237300042
The lower approximate object of concentrating of expression arrives the weighted sum of affiliated class centre distance,
Figure BDA00002578237300044
The approximate object of concentrating is to the weighted sum of affiliated class centre distance in the expression.
Further, in step S7, adopt formula
Figure BDA00002578237300045
Calculate p class cluster center C pWherein,
Figure BDA00002578237300046
Represent p class boundary set;
Figure BDA000025782373000411
Represent approximate collection under the p class,
Figure BDA00002578237300047
The upper approximate collection that represents the p class.
The present invention adopts the combination of k-means, kernel function and several algorithms of rough set, can carry out to the Power system load data of present complexity the cluster of efficiently and accurately, especially exist nonlinear data, border object data and higher-dimension blended data to process in the load data, also can process the noise data that exists in the load data simultaneously, reduce it to the impact of final cluster result, improve clustering precision.The present invention can also guarantee the situation of the realistic electric system of result's energy that cluster goes out, and intelligibility is stronger as a result.
The present invention also provides a kind of device for the electric load cluster, comprises the data acquisition module, data computation module, dynamic adjusting module, judge module and the circulation module that connect successively; Described data acquisition module is used for obtaining sample matrix
Figure BDA00002578237300048
This sample matrix comprises that n is capable, concrete sample of each behavior, from
Figure BDA00002578237300049
Arrive
Figure BDA000025782373000410
Each sample has d attribute, as each sample
Figure BDA00002578237300051
D row component, namely each element in Φ (x) matrix can be expressed as
Figure BDA00002578237300052
Described data computation module comprises initially birds of the same feather flock together center determination module and the computing module of birdsing of the same feather flock together that connects successively, and the described center determination module of initially birdsing of the same feather flock together is determined initial cluster center C, C={C by choose k sample in n the sample of sample matrix Φ (x) 1, C 2..., C k, wherein C represents the set of matrices at k class center, and k is clusters number, comprises that k is capable, and the concrete center of each each class of behavior is from C 1To C k, the center of each class also has d attribute accordingly, as every class center C pD row component, each element in the C matrix can be expressed as C PjThe described computing module of birdsing of the same feather flock together is used for described sample matrix Φ (x) is carried out cluster calculation, and various kinds is originally distributed to the upper and lower approximate collection of nearest class according to the method that upper and lower approximate collection is determined; Described dynamic adjusting module is used for according to the weight factor ω of current iteration number of times t to described lower approximate collection lWeight factor ω with described approximate collection BnrDynamically adjust; Described judge module is used for calculating the target function value J of rough kernel cluster w, and judge | J w(t)-J w(t-1) |≤ε or t 〉=t MaxWhether set up, if then generate final cluster and end; If not, then carry out cycle calculations by described circulation module; Described circulation module is used for redefining cluster centre when t=t+1 and is back to the described computing module and carry out cycle calculations of birdsing of the same feather flock together.
Further, described dynamic adjusting module adopts formula
Figure BDA00002578237300053
With
Figure BDA00002578237300054
Weight factor ω to lower approximate collection lWeight factor ω with described approximate collection BnrDynamically adjust.
Further, described judge module adopts formula
Figure BDA00002578237300055
Calculate the target function value J of rough kernel cluster wWherein, w jBe the characteristic weighing coefficient, and
Figure BDA00002578237300056
Figure BDA00002578237300057
The lower approximate object of concentrating of expression arrives the weighted sum of affiliated class centre distance,
Figure BDA00002578237300061
The approximate object of concentrating is to the weighted sum of affiliated class centre distance in the expression.
Further, described circulation module adopts formula
Figure BDA00002578237300062
Calculate p class cluster center C pWherein,
Figure BDA00002578237300063
Represent p class boundary set;
Figure BDA00002578237300065
Represent approximate collection under the p class,
Figure BDA00002578237300064
The upper approximate collection that represents the p class.
Device provided by the invention is studied the similarity feature of part throttle characteristics based on the clustering method of rough set and kernel function, can extract more objective, accurately the function quintessence's feature with type load, the algorithm simple and fast is effective, cluster result is reasonable, can be used as the fundamental basis based on the load characteristics recorder device installation position selection that measures load modeling, practical significant to load modeling research.
Description of drawings
Fig. 1 is the method realization flow figure that is used for the electric load cluster that the embodiment of the invention provides;
Fig. 2 is lower approximate, the upper approximate and Boundary Region schematic diagram of the set X that provides of the embodiment of the invention;
Fig. 3 is the modular structure block diagram of the device that is used for the electric load cluster that provides of the embodiment of the invention.
Embodiment
In order to make purpose of the present invention, technical scheme and advantage clearer, below in conjunction with drawings and Examples, the present invention is further elaborated.Should be appreciated that specific embodiment described herein only in order to explain the present invention, is not intended to limit the present invention.
Electric load changes the impact that is subject to factors, load such as heavy industry, the light industry load, extractive industry load, agricultural load, tertiary industry load, municipal administration life load etc., these loads also all can be subjected to such as weather temperature simultaneously, the impact in season etc., short-term electric load curvilinear motion form all larger difference can occur.For this electric load Variation Features, it is a class that the historical data of utilizing the rough kernel clustering algorithm will have similar load curve is gathered.
In embodiments of the present invention, have non-linear for Power system load data, the characteristic such as higher-dimension and border property, kernel function can be changed sample space, make data by the non-linear linear separability that becomes, kernel function is converted into the inner product operation of m-dimensional space the kernel function calculating of n dimension lower dimensional space simultaneously, thereby has solved cleverly " dimension disaster " problem of calculating in high-dimensional feature space, improve higher-dimension blended data clustering performance, also simplified sorting procedure; Rough set is processed border object by being similar to up and down the heavy dynamic adjustment of the centralization of state power, reduces border object to the impact of cluster result, simultaneously sample is belonged to inhomogeneous up and down approximate the collection, improves clustering precision.Consider various Algorithm Performances, several algorithms are carried out effective integration, propose a kind of kernel function clustering algorithm based on rough set.
Fig. 1 shows the method realization flow that is used for the electric load cluster that the embodiment of the invention provides, and specifically comprises:
S1: obtain sample matrix
Figure BDA00002578237300071
This sample matrix comprises that n is capable, concrete sample of each behavior, from
Figure BDA00002578237300072
Arrive
Figure BDA00002578237300073
Each sample has d attribute, as each sample
Figure BDA00002578237300074
D row component, namely each element in Φ (x) matrix can be expressed as
Figure BDA00002578237300075
I=1 wherein, 2 ..., n, j=1,2 ..., d, sample All belong to the F space;
S2: determine initial cluster center C, C={C by in n the sample of sample matrix Φ (x), choosing k sample 1, C 2..., C k, that is to say, the every delegation component in the initial cluster center C matrix is (from C 1To C k) all be from
Figure BDA00002578237300077
Arrive
Figure BDA00002578237300078
Select, wherein C represents the set of matrices at k class center, and k is clusters number, and C comprises that k is capable, and the concrete center of each each class of behavior is from C 1To C k, the center of each class also has d attribute accordingly, as every class center C pD row component, each element in the C matrix can be expressed as C Pj, p=1 wherein, 2 ..., k, j=1,2 ..., d,
Figure BDA000025782373000710
S3: make iterations t=1;
S4: described sample matrix Φ (x) is carried out cluster calculation, and various kinds is originally distributed to the upper and lower approximate collection of nearest class according to the method that upper and lower approximate collection is determined;
S5: according to the weight factor ω of current iteration number of times t to described lower approximate collection lWeight factor ω with described approximate collection BnrDynamically adjust;
S6: the target function value J that calculates the rough kernel cluster w, and judge | J w(t)-J w(t-1) |≤ε or t 〉=t MaxWhether set up, if then finish; If not, then enter step S7; ε is a parameter of determining according to actual conditions, t MaxBe the artificial iterations upper limit of determining;
S7: make t=t+1, redefine cluster centre C p, and change step S4 over to.
In embodiments of the present invention, the F space is vector space or linear space, and the Hilbert space is the space that has defined inner product, and the F space comprises the Hilbert space; R represents real number space, and d representation dimension, real number set are used R usually dExpression, i.e. R dThe real number space of expression d dimension.The Hilbert definition space is for having defined the space of inner product in linear space, the Hilbert space is a kind of special linear space.
In embodiments of the present invention, X is accurate sample matrix, and the set of its sample that is as the criterion comprises that n is capable, specifically the accurate sample x of each behavior 1To x n, each accurate sample has d attribute, and as d row component of each sample, each element in the X matrix can be expressed as x Ij, i=1-n, j=1-d); Accurate sample matrix X is specially: X = x 11 . . . x 1 d . . . x ij . . . x n 1 . . . x nd ; And Φ (x) is the sample matrix that X obtains later through mapping, and is similar with X, comprises that also n is capable, specifically each behavior sample Arrive Each sample also has d attribute accordingly, and as d row component of each sample, each element in Φ (x) matrix can be expressed as
Figure BDA00002578237300084
I=1-n, j=1-d, sample Φ (x) is specially:
Figure BDA00002578237300085
Wherein, the set of matrices at k class center is specially: C = C 11 . . . C 1 d . . . C pj . . . C k 1 . . . C kd ; The span of k
Figure BDA00002578237300087
Figure BDA00002578237300088
Be rule of thumb to obtain, also can adjust arbitrarily according to actual needs.
Wherein, step S1 is specially: obtain the accurate sample of n transformer station's synthetic load static characteristics, consist of accurate sample matrix X={x 1, x 2..., x n, x i∈ R d, and obtain sample after by gaussian kernel function described accurate sample data being mapped to the Hilbert space
Figure BDA00002578237300091
Accurate sample x iAll belong to space R dIn; Gaussian kernel function is: and F (x, y)=exp (β || x-y|| 2/ 2 σ 2), wherein σ is the width parameter of function, and β is that 1, x is the accurate sample before the conversion, and y is the sample after changing.
In embodiments of the present invention, as shown in Figure 2, upper approximate collection, lower approximate collection and boundary set are defined as follows: the U/R of equivalence classification family of given knowledge base K=(U, R) and U, and to each object subset
Figure BDA00002578237300092
And
Figure BDA00002578237300093
X closely can be defined as respectively about upper approximate and the lower of R of R:
Figure BDA00002578237300094
With
Figure BDA00002578237300095
On the basis of and lower approximate definition approximate at R, just the positive territory of definition set, the positive territory of R of bearing territory and Boundary Region: X are POS very easily R(X)=
Figure BDA000025782373000921
, namely its lower being similar to collects; The negative territory of the R of X is
Figure BDA00002578237300096
Namely its domain and upper approximate difference set; The Boundary Region of X is
Figure BDA00002578237300097
Namely the upper of X is similar to and lower approximate difference set.Wherein, be approximately equal to the union in its Boundary Region and positive territory on the R of X; To judge the set that element forms among the domain U that may belong to X according to knowledge R;
Figure BDA000025782373000919
To judge the set that element forms among the domain U that certainly belongs to X according to knowledge R.
In embodiments of the present invention, the method that upper and lower approximate collection is determined in step S4 is specially:
Figure BDA00002578237300099
Figure BDA000025782373000910
If Then order
Figure BDA000025782373000912
Otherwise order
Figure BDA000025782373000913
γ=0.07 wherein;
Figure BDA000025782373000914
Represent i sample To the Euclidean distance at p class center, Represent i sample
Figure BDA000025782373000917
Minimum value in the Euclidean distance at q class center, p=1 wherein, 2 ... k, q=1,2 ... k, p ≠ q; Dist (C p, C q) represent the center of p class to the Euclidean distance at the center of q class,
Figure BDA000025782373000918
Be the upper approximate collection of q class,
Figure BDA000025782373000920
It is the lower approximate collection of p class.
In embodiments of the present invention, along with the continuous increase of iterations, if so that the weights omega of lower approximate collection lProportion constantly increase the weights omega of upper approximate collection BnrCorresponding proportion constantly reduce, the efficient of algorithm and precision can increase to some extent.By analyzing, can determine: at initial phase, obtain at random any as initial cluster center, then according to the k-means algorithm sample is divided in the concrete a certain class.Before carrying out during several times Rough clustering, the upper approximate data volume of concentrating is larger, therefore gives approximate centralization of state power repeated factor ω like this BnrA relatively large number; Same, when algorithm ran to the later stage, most of samples had been summed up as the lower approximate collection of every class, at this moment ω lValue relatively improve, the nature efficient of algorithm is increased.Therefore, can be with ω l, ω BnrValue adjust dynamically according to following formula, particularly, in step S5, can adopt formula
Figure BDA00002578237300101
Weight factor ω to lower approximate collection lDynamically adjust; Adopt formula The weight factor ω of upper approximate collection BnrDynamically adjust.Wherein t represents the current iterations of algorithm, t MaxExpression algorithm maximum iteration time.
In embodiments of the present invention, in step S6, adopt formula
I=1,2 ..., n, the target function value J of calculating rough kernel cluster wWherein, w jBe the characteristic weighing coefficient, and
Figure BDA00002578237300104
The lower approximate object of concentrating of expression arrives the weighted sum of affiliated class centre distance, The approximate object of concentrating is to the weighted sum of affiliated class centre distance in the expression.
In embodiments of the present invention, the calculating of cluster centre is a very crucial problem in the nuclear space, in higher dimensional space, whole samples all is assigned to the upper of each cluster centre, lower approximate concentrating, what then reflect in the corresponding sample space is that sample in different classes of is to the percentage contribution of cluster.And according to the character of rough set as can be known the lower approximate object of concentrating necessarily be included in upper approximate concentrating, except the lower approximate object of concentrating, bunch also might comprise the object of the upper approximate collection that belongs to simultaneously other bunches.In step S7, adopt formula
Figure BDA00002578237300111
Calculate p class cluster center C pWherein,
Figure BDA00002578237300112
Represent p class boundary set;
Figure BDA000025782373001114
Represent approximate collection under the p class,
Figure BDA00002578237300113
The upper approximate collection that represents the p class; c kCluster centre, ω lThe lower approximate centralization of state power of representation class is heavy, ω BnrThe weight of the upper approximate collection of representation class.
The Clustering Algorithm of Kernel based on rough set that the embodiment of the invention provides adopts gaussian kernel function that sample is mapped to feature space, amplifies differences between samples; Combining rough set is effectively processed border object, and dynamically changes up and down approximate collection weights omega lAnd ω BnrProportion in iteration each time improves clustering precision; Adopt reliefF algorithm balance different attribute to the percentage contribution of cluster result, guarantee that further result that cluster goes out meets the actual conditions of electric load cluster.Simultaneously, the clustering precision that draws by this algorithm is higher, and the convergence of algorithm time is also very fast, and cluster result relatively meets the actual conditions of Power system load data.Also proved the superiority of this algorithm in the power system load cluster, load characteristics clustering based on this algorithm provides effective way for the transformer station that the cloth measuring point is not installed sets up utility model simultaneously, and cluster result and cluster centre that this algorithm of while draws provide important reference frame for further carrying out the practical work of load modeling.
In embodiments of the present invention, in Rough Set Clustering, according to the theory of the upper and lower approximate collection of rough kernel, can there be the fuzzy situation in bunch border in final cluster result, and namely bunch border can be fully not definite.The decision method of the upper and lower approximate collection scope of rough set is so:
Figure BDA00002578237300114
Figure BDA00002578237300115
If
Figure BDA00002578237300116
Then order
Figure BDA00002578237300117
Otherwise order γ=0.07 wherein;
Figure BDA00002578237300119
Represent i sample
Figure BDA000025782373001110
To the Euclidean distance at p class center,
Figure BDA000025782373001111
Represent i sample Minimum value in the Euclidean distance at q class center, p=1 wherein, 2 ... k, q=1,2 ... k, p ≠ q; Dist (C p, C q) represent the center of p class to the Euclidean distance at the center of q class, Be the upper approximate collection of q class, C pIt is the lower approximate collection of p class.
In embodiments of the present invention, and in the electric load actual conditions, huge such as the effect that some attribute such as temperature, date is brought into play in cluster process, and very little even can ignore such as the effect of the attributes such as air pressure, wind-force.Therefore, utilize degree also maximum for what make the strong attribute of effect, so adopt the reliefF method that sample attribute is weighted processing, then the objective function of rough kernel cluster is shown below:
Figure BDA00002578237300121
I=1,2 ..., n, the target function value J of calculating rough kernel cluster wWherein, w jBe the characteristic weighing coefficient, and
Figure BDA00002578237300122
ω lAnd ω BnrThe lower approximate collection at class center and the weight of upper approximate collection when expression is calculated,
Figure BDA00002578237300123
The lower approximate object of concentrating of expression arrives the weighted sum of affiliated class centre distance, The approximate object of concentrating is to the weighted sum of affiliated class centre distance in the expression.
The clustering method that the present invention is based on rough set and kernel function is studied the similarity feature of part throttle characteristics, can extract more objective, accurately the function quintessence's feature with type load, its clear thinking, the algorithm simple and fast is effective, cluster result is reasonable, can be used as the fundamental basis based on the load characteristics recorder device installation position selection that measures load modeling, practical significant to load modeling research.The method is not only for the classification take transformer station's load structure ratio as the essential characteristic amount, and is effectively comprehensive.And can be generalized to the screening of industry typical user, also can be applicable to classification of dynamic load characteristics that local measures with comprehensive.
Fig. 3 shows the modular structure of the device that is used for the electric load cluster that the embodiment of the invention provides, and for convenience of explanation, only shows the part relevant with the embodiment of the invention, and details are as follows:
The device that should be used for the electric load cluster comprises data acquisition module 1, data computation module 2, dynamic adjusting module 3, judge module 4 and the circulation module 5 that connects successively, and wherein, data acquisition module 1 is used for obtaining sample matrix
Figure BDA00002578237300131
This sample matrix comprises that n is capable, concrete sample of each behavior, from
Figure BDA00002578237300132
Arrive Each sample has d attribute, as each sample D row component, namely each element in Φ (x) matrix can be expressed as
Figure BDA00002578237300135
I=1 wherein, 2 ..., n, j=1,2 ..., d, sample
Figure BDA00002578237300136
All belong to the F space; Data computation module 2 comprises successively and to connect initially birds of the same feather flock together center determination module 21 and the computing module 22 of birdsing of the same feather flock together, and the center determination module 21 of initially birdsing of the same feather flock together is determined initial cluster center C, C={C by choose k sample in n the sample of Φ (x) 1, C 2..., C k, wherein C represents the set of matrices at k class center, and k is clusters number, comprises that k is capable, and the concrete center of each each class of behavior is from C 1To C k, the center of each class also has d attribute accordingly, as every class center C pD row component, each element in the C matrix can be expressed as C Pj, p=1 wherein, 2 ..., k, j=1,2 ..., d,
Figure BDA00002578237300137
The computing module 22 of birdsing of the same feather flock together is used for described sample Φ (x) is carried out cluster calculation, and various kinds is originally distributed to the upper and lower approximate collection of nearest class according to the method that upper and lower approximate collection is determined; Dynamically adjusting module 3 is used for according to the weight factor ω of current iteration number of times t to described lower approximate collection lWeight factor ω with described approximate collection BnrDynamically adjust; Judge module 4 is used for calculating the target function value J of rough kernel cluster w, and judge | J w(t)-J w(t-1) |≤ε or t 〉=t MaxWhether set up, if then generate final cluster and end; If not, then carry out cycle calculations by described circulation module; Circulation module 5 is used for redefining cluster centre and be back to when t=t+1 birdsing of the same feather flock together computing module 22 and carrying out cycle calculations.
In embodiments of the present invention, dynamically adjusting module 3 can adopt formula
Figure BDA00002578237300138
With
Figure BDA00002578237300139
Weight factor ω to lower approximate collection lWeight factor ω with described approximate collection BnrDynamically adjust.
In embodiments of the present invention, judge module 4 can adopt formula
Figure BDA000025782373001310
Calculate the target function value J of rough kernel cluster wWherein, w jBe the characteristic weighing coefficient, and
Figure BDA00002578237300141
The lower approximate object of concentrating of expression arrives the weighted sum of affiliated class centre distance,
Figure BDA00002578237300143
The approximate object of concentrating is to the weighted sum of affiliated class centre distance in the expression.
In embodiments of the present invention, circulation module 5 can adopt formula
Figure BDA00002578237300144
Calculate p class cluster center C pWherein,
Figure BDA00002578237300145
The boundary set that represents p class center;
Figure BDA00002578237300147
The lower approximate collection that represents p class center,
Figure BDA00002578237300146
The upper approximate collection that represents p class center.
Device provided by the invention is studied the similarity feature of part throttle characteristics based on the clustering method of rough set and kernel function, can extract more objective, accurately the function quintessence's feature with type load, the algorithm simple and fast is effective, cluster result is reasonable, can be used as the fundamental basis based on the load characteristics recorder device installation position selection that measures load modeling, practical significant to load modeling research.
Those skilled in the art will readily understand; the above only is preferred embodiment of the present invention; not in order to limiting the present invention, all any modifications of doing within the spirit and principles in the present invention, be equal to and replace and improvement etc., all should be included within protection scope of the present invention.

Claims (10)

1. a method that is used for the electric load cluster is characterized in that, comprises the steps:
S1: obtain sample matrix
Figure FDA00002578237200011
This sample matrix comprises that n is capable, concrete sample of each behavior, from
Figure FDA00002578237200012
Arrive
Figure FDA00002578237200013
Each sample has d attribute, as each sample
Figure FDA00002578237200014
D row component, namely each element in Φ (x) matrix can be expressed as
Figure FDA00002578237200015
I=1 wherein, 2 ..., n, j=1,2 ..., d, sample
Figure FDA00002578237200016
All belong to the F space;
S2: determine initial cluster center C, C={C by in n the sample of sample matrix Φ (x), choosing k sample 1, C 2..., C k, wherein C represents the set of matrices at k class center, and k is clusters number, comprises that k is capable, and the concrete center of each each class of behavior is from C 1To C k, the center of each class also has d attribute accordingly, as every class center C pD row component, each element in the C matrix can be expressed as C Pj, p=1 wherein, 2 ..., k, j=1,2 ..., d, k gets
Figure FDA00002578237200017
In arbitrary value;
S3: make iterations t=1;
S4: described sample matrix Φ (x) is carried out cluster calculation, and various kinds is originally distributed to the upper and lower approximate collection of nearest class according to the method that upper and lower approximate collection is determined;
S5: according to the weight factor ω of current iteration number of times t to described lower approximate collection lWeight factor ω with described approximate collection BnrDynamically adjust;
S6: the target function value J that calculates the rough kernel cluster w, and judge | J w(t)-J w(t-1) |≤ε or t 〉=t MaxWhether set up, if then finish; If not, then enter step S7; ε is a parameter of determining according to actual conditions, t MaxBe the artificial iterations upper limit of determining;
S7: make t=t+1, redefine cluster centre C p, and change step S4 over to.
2. the method for claim 1 is characterized in that, described step S1 is specially: obtain the accurate sample of n transformer station's synthetic load static characteristics, consist of accurate sample matrix X={x 1, x 2..., x n, x i∈ R d, and obtain sample after by gaussian kernel function described accurate sample data being mapped to the Hilbert space
Figure FDA00002578237200021
Accurate sample x iAll belong to space R dIn; Described gaussian kernel function is: and F (x, y)=exp (β || x-y|| 2/ 2 σ 2), wherein σ is the width parameter of function, and β is that 1, x is the accurate sample before the conversion, and y is the sample after changing.
3. the method for claim 1 is characterized in that, in step S4, the method that upper and lower approximate collection is determined is specially:
Figure FDA00002578237200022
Figure FDA00002578237200023
If
Figure FDA00002578237200024
Then order
Figure FDA00002578237200025
Otherwise order
Figure FDA00002578237200026
γ=0.07 wherein;
Figure FDA00002578237200027
Represent i sample
Figure FDA00002578237200028
To the Euclidean distance at p class center, Represent i sample
Figure FDA000025782372000210
Minimum value in the Euclidean distance at q class center, p=1 wherein, 2 ... k, q=1,2 ... k, p ≠ q; Dist (C p, C q) represent the center of p class to the Euclidean distance at the center of q class,
Figure FDA000025782372000211
Be the upper approximate collection of q class,
Figure FDA000025782372000218
It is the lower approximate collection of p class.
4. the method for claim 1 is characterized in that, in step S5, adopts formula
Figure FDA000025782372000212
With
Figure FDA000025782372000213
Weight factor ω to lower approximate collection lWeight factor ω with described approximate collection BnrDynamically adjust.
5. method as claimed in claim 4 is characterized in that, in step S6, adopts formula I=1,2 ..., n, the target function value J of calculating rough kernel cluster wWherein, w jBe the characteristic weighing coefficient, and
Figure FDA000025782372000216
The lower approximate object of concentrating of expression arrives the weighted sum of affiliated class centre distance,
Figure FDA000025782372000217
The approximate object of concentrating is to the weighted sum of affiliated class centre distance in the expression.
6. method as claimed in claim 4 is characterized in that, in step S7, adopts formula Calculate p class cluster center C pWherein,
Figure FDA00002578237200032
Represent p class boundary set; The lower approximate collection that represents the p class, The upper approximate collection that represents the p class.
7. a device that is used for the electric load cluster is characterized in that, comprises the data acquisition module, data computation module, dynamic adjusting module, judge module and the circulation module that connect successively,
Described data acquisition module is used for obtaining sample matrix
Figure FDA00002578237200034
This sample matrix comprises that n is capable, concrete sample of each behavior, from
Figure FDA00002578237200035
Arrive
Figure FDA00002578237200036
Each sample has d attribute, as each sample
Figure FDA00002578237200037
D row component, namely each element in Φ (x) matrix can be expressed as
Figure FDA00002578237200038
I=1,2 ..., n, j=1,2 ..., d;
Described data computation module comprises initially birds of the same feather flock together center determination module and the computing module of birdsing of the same feather flock together that connects successively, and the described center determination module of initially birdsing of the same feather flock together is determined initial cluster center C, C={C by choose k sample in n the sample of sample matrix Φ (x) 1, C 2..., C k, wherein C represents the set of matrices at k class center, and k is clusters number, comprises that k is capable, and the concrete center of each each class of behavior is from C 1To C k, k gets
Figure FDA00002578237200039
In arbitrary value, the center of each class also has d attribute accordingly, as every class center C pD row component, each element in the C matrix can be expressed as C Pj, p=1,2 ..., k; The described computing module of birdsing of the same feather flock together is used for described sample matrix Φ (x) is carried out cluster calculation, and various kinds is originally distributed to the upper and lower approximate collection of nearest class according to the method that upper and lower approximate collection is determined;
Described dynamic adjusting module is used for according to the weight factor ω of current iteration number of times t to described lower approximate collection lWeight factor ω with described approximate collection BnrDynamically adjust;
Described judge module is used for calculating the target function value J of rough kernel cluster w, and judge | J w(t)-J w(t-1) |≤ε or t 〉=t MaxWhether set up, if then generate final cluster and end; If not, then carry out cycle calculations by described circulation module;
Described circulation module is used for redefining cluster centre when t=t+1 and is back to the described computing module and carry out cycle calculations of birdsing of the same feather flock together.
8. device as claimed in claim 7 is characterized in that, described dynamic adjusting module adopts formula
Figure FDA00002578237200041
With Weight factor ω to lower approximate collection lWeight factor ω with described approximate collection BnrDynamically adjust.
9. device as claimed in claim 7 is characterized in that, described judge module adopts formula Calculate the target function value J of rough kernel cluster wWherein, w jBe the characteristic weighing coefficient, and
Figure FDA00002578237200045
The lower approximate object of concentrating of expression arrives the weighted sum of affiliated class centre distance,
Figure FDA00002578237200046
The approximate object of concentrating is to the weighted sum of affiliated class centre distance in the expression.
10. device as claimed in claim 7 is characterized in that, described circulation module adopts formula
Figure FDA00002578237200047
Calculate p class cluster center C pWherein,
Figure FDA00002578237200048
Represent p class boundary set;
Figure FDA000025782372000410
The lower approximate collection that represents the p class,
Figure FDA00002578237200049
The upper approximate collection that represents the p class.
CN2012105373292A 2012-12-13 2012-12-13 Method and device used for power load aggregation Pending CN103049651A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN2012105373292A CN103049651A (en) 2012-12-13 2012-12-13 Method and device used for power load aggregation

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN2012105373292A CN103049651A (en) 2012-12-13 2012-12-13 Method and device used for power load aggregation

Publications (1)

Publication Number Publication Date
CN103049651A true CN103049651A (en) 2013-04-17

Family

ID=48062285

Family Applications (1)

Application Number Title Priority Date Filing Date
CN2012105373292A Pending CN103049651A (en) 2012-12-13 2012-12-13 Method and device used for power load aggregation

Country Status (1)

Country Link
CN (1) CN103049651A (en)

Cited By (19)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103872678A (en) * 2014-03-06 2014-06-18 国家电网公司 Load model identification method based on transformer substation measurement
CN104063480A (en) * 2014-07-02 2014-09-24 国家电网公司 Load curve parallel clustering method based on big data of electric power
CN104615722A (en) * 2015-02-06 2015-05-13 浙江工业大学 Blended data clustering method based on density searching and rapid partitioning
CN105021888A (en) * 2015-07-06 2015-11-04 广州供电局有限公司 Harmonic wave data monitoring method based on data clustering
CN105719194A (en) * 2016-01-11 2016-06-29 河南许继仪表有限公司 Automatic optimization type power load classification method
CN106055883A (en) * 2016-05-25 2016-10-26 中国电力科学研究院 Transient stability assessment input characteristic validity analysis method based on improved Sammon mapping
CN106067034A (en) * 2016-05-27 2016-11-02 中国电力科学研究院 A kind of distribution network load curve clustering method based on higher dimensional matrix characteristic root
CN106257870A (en) * 2015-06-16 2016-12-28 福特全球技术公司 The compression algorithm transmitted for the vehicle bus message of preview data
CN106446967A (en) * 2016-10-28 2017-02-22 国网福建省电力有限公司 Novel power system load curve clustering method
US9829880B2 (en) 2014-11-20 2017-11-28 General Electric Company System and method for modelling load in an electrical power network
CN107423769A (en) * 2017-08-03 2017-12-01 四川大学 Electric load curve adaptive clustering scheme based on morphological feature
CN107480731A (en) * 2017-09-06 2017-12-15 西安西热电站信息技术有限公司 A kind of EARLY RECOGNITION method of thermal power plant's automobile assembly welding Iine fault signature
CN108306284A (en) * 2018-01-05 2018-07-20 东南大学 A kind of online load modeling method measured based on local intelligence
CN108805174A (en) * 2018-05-18 2018-11-13 广东惠禾科技发展有限公司 clustering method and device
CN110165657A (en) * 2018-08-30 2019-08-23 中国南方电网有限责任公司 Consider substation's load characteristics clustering analysis method of user's industry attribute
CN110796159A (en) * 2019-09-12 2020-02-14 国网浙江省电力有限公司杭州供电公司 Power data classification method and system based on k-means algorithm
CN111476316A (en) * 2020-04-28 2020-07-31 福建师范大学 Method and system for clustering mean value of power load characteristic data based on cloud computing
CN112258231A (en) * 2020-10-22 2021-01-22 国网新疆电力有限公司经济技术研究院 Improved K-means algorithm-based power distribution network transformer substation site selection and volume fixing method and device
CN113761700A (en) * 2020-06-05 2021-12-07 国家电网有限公司华东分部 Load modeling and online correction method and system based on dynamic clustering

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101414366A (en) * 2008-10-22 2009-04-22 西安交通大学 Method for forecasting electric power system short-term load based on method for improving uttermost learning machine
CN101556464A (en) * 2009-05-22 2009-10-14 天津大学 Auto recommending method of urban power load forecasting module based on associative rules

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101414366A (en) * 2008-10-22 2009-04-22 西安交通大学 Method for forecasting electric power system short-term load based on method for improving uttermost learning machine
CN101556464A (en) * 2009-05-22 2009-10-14 天津大学 Auto recommending method of urban power load forecasting module based on associative rules

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
P LINGRAS: "Interval set clustering of web user with rough k-means", 《JOURNAL OF INTELLIGENT INFORMATION SYSTEM》, vol. 23, no. 1, 31 December 2004 (2004-12-31) *
周涛: "具有自适应参数的粗糙k-means聚类算法", 《计算机工程与应用》, vol. 46, no. 26, 31 December 2010 (2010-12-31) *
姚丽娟: "基于粒子群和粗糙集的聚类算法研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》, no. 09, 15 September 2012 (2012-09-15) *
姚丽娟等: "基于粒子群的粗糙核聚类算法", 《计算机应用研究》, vol. 29, no. 8, 31 August 2012 (2012-08-31), pages 2854 - 2857 *

Cited By (26)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103872678B (en) * 2014-03-06 2016-02-10 国家电网公司 A kind of load model identification method measured based on transformer station
CN103872678A (en) * 2014-03-06 2014-06-18 国家电网公司 Load model identification method based on transformer substation measurement
CN104063480A (en) * 2014-07-02 2014-09-24 国家电网公司 Load curve parallel clustering method based on big data of electric power
US9829880B2 (en) 2014-11-20 2017-11-28 General Electric Company System and method for modelling load in an electrical power network
CN104615722A (en) * 2015-02-06 2015-05-13 浙江工业大学 Blended data clustering method based on density searching and rapid partitioning
CN104615722B (en) * 2015-02-06 2018-04-27 浙江工业大学 Blended data clustering method with quickly dividing is searched for based on density
CN106257870A (en) * 2015-06-16 2016-12-28 福特全球技术公司 The compression algorithm transmitted for the vehicle bus message of preview data
CN105021888A (en) * 2015-07-06 2015-11-04 广州供电局有限公司 Harmonic wave data monitoring method based on data clustering
CN105021888B (en) * 2015-07-06 2018-09-18 广州供电局有限公司 Harmonic data monitoring method based on data clusters
CN105719194A (en) * 2016-01-11 2016-06-29 河南许继仪表有限公司 Automatic optimization type power load classification method
CN106055883A (en) * 2016-05-25 2016-10-26 中国电力科学研究院 Transient stability assessment input characteristic validity analysis method based on improved Sammon mapping
CN106055883B (en) * 2016-05-25 2022-09-02 中国电力科学研究院 Transient stability evaluation input feature validity analysis method based on improved Sammon mapping
CN106067034B (en) * 2016-05-27 2022-03-18 中国电力科学研究院 Power distribution network load curve clustering method based on high-dimensional matrix characteristic root
CN106067034A (en) * 2016-05-27 2016-11-02 中国电力科学研究院 A kind of distribution network load curve clustering method based on higher dimensional matrix characteristic root
CN106446967A (en) * 2016-10-28 2017-02-22 国网福建省电力有限公司 Novel power system load curve clustering method
CN107423769B (en) * 2017-08-03 2019-10-01 四川大学 Electric load curve adaptive clustering scheme based on morphological feature
CN107423769A (en) * 2017-08-03 2017-12-01 四川大学 Electric load curve adaptive clustering scheme based on morphological feature
CN107480731A (en) * 2017-09-06 2017-12-15 西安西热电站信息技术有限公司 A kind of EARLY RECOGNITION method of thermal power plant's automobile assembly welding Iine fault signature
CN108306284A (en) * 2018-01-05 2018-07-20 东南大学 A kind of online load modeling method measured based on local intelligence
CN108805174A (en) * 2018-05-18 2018-11-13 广东惠禾科技发展有限公司 clustering method and device
CN110165657A (en) * 2018-08-30 2019-08-23 中国南方电网有限责任公司 Consider substation's load characteristics clustering analysis method of user's industry attribute
CN110796159A (en) * 2019-09-12 2020-02-14 国网浙江省电力有限公司杭州供电公司 Power data classification method and system based on k-means algorithm
CN111476316A (en) * 2020-04-28 2020-07-31 福建师范大学 Method and system for clustering mean value of power load characteristic data based on cloud computing
CN111476316B (en) * 2020-04-28 2023-04-14 福建师范大学 Method and system for clustering mean value of power load characteristic data based on cloud computing
CN113761700A (en) * 2020-06-05 2021-12-07 国家电网有限公司华东分部 Load modeling and online correction method and system based on dynamic clustering
CN112258231A (en) * 2020-10-22 2021-01-22 国网新疆电力有限公司经济技术研究院 Improved K-means algorithm-based power distribution network transformer substation site selection and volume fixing method and device

Similar Documents

Publication Publication Date Title
CN103049651A (en) Method and device used for power load aggregation
Ye et al. A data-driven bottom-up approach for spatial and temporal electric load forecasting
Wu et al. A data mining approach combining $ K $-Means clustering with bagging neural network for short-term wind power forecasting
WO2018045642A1 (en) A bus bar load forecasting method
CN106022509B (en) Consider the Spatial Load Forecasting For Distribution method of region and load character double differences
CN108304978A (en) A kind of mid-term Electric Power Load Forecast method based on data clusters theory
CN105701596A (en) Method for lean distribution network emergency maintenance and management system based on big data technology
CN112149873B (en) Low-voltage station line loss reasonable interval prediction method based on deep learning
CN107730031A (en) A kind of ultra-short term peak load forecasting method and its system
CN104881706A (en) Electrical power system short-term load forecasting method based on big data technology
CN105389636A (en) Low-voltage area KFCM-SVR reasonable line loss prediction method
CN111461921B (en) Load modeling typical user database updating method based on machine learning
CN109461025A (en) A kind of electric energy substitution potential customers&#39; prediction technique based on machine learning
CN103679263A (en) Thunder and lightning approach forecasting method based on particle swarm support vector machine
CN110503256A (en) Short-term load forecasting method and system based on big data technology
Zhang et al. Short-term power load forecasting using integrated methods based on long short-term memory
CN103942606A (en) Residential electricity consumption customer segmentation method based on fruit fly intelligent optimization algorithm
CN108345908A (en) Sorting technique, sorting device and the storage medium of electric network data
CN110380444A (en) Distributing wind-powered electricity generation orderly accesses the method for planning capacity of power grid under a kind of more scenes based on structure changes Copula
CN110738232A (en) grid voltage out-of-limit cause diagnosis method based on data mining technology
CN111415049A (en) Power failure sensitivity analysis method based on neural network and clustering
Saffari et al. Deep convolutional graph rough variational auto-encoder for short-term photovoltaic power forecasting
Fontanini et al. A data-driven BIRCH clustering method for extracting typical load profiles for big data
Mohammad et al. Short term load forecasting using deep neural networks
CN115099511A (en) Photovoltaic power probability estimation method and system based on optimized copula

Legal Events

Date Code Title Description
C06 Publication
PB01 Publication
C10 Entry into substantive examination
SE01 Entry into force of request for substantive examination
C12 Rejection of a patent application after its publication
RJ01 Rejection of invention patent application after publication

Application publication date: 20130417