CN109272058B - Integrated power load curve clustering method - Google Patents

Abstract

The invention discloses an integrated power load curve clustering method. Under the big background that the power grid company in China increases the innovation of the demand side and more effectively performs economic dispatching on the operation of the power system, the invention provides a fast and effective power load clustering algorithm. Firstly, carrying out coarse clustering on original power load data by using an SOM neural network to obtain a clustered class center; clustering the roughly clustered class centers by using a DBSCAN algorithm, and merging the class clusters corresponding to the similar class centers; and finally, removing the deviating elements in the cluster and putting the deviating elements in the most similar cluster to obtain an optimal clustering result.

Description

Integrated power load curve clustering method

Technical Field

The invention belongs to the technical field of big data, and particularly relates to an integrated power load curve clustering method.

Background

In the big data era, the power utilization rule and characteristics of users are obtained by clustering and analyzing historical power load data curves, and the clustering and analyzing method is an important content for economic dispatching and demand side management of a power system. The clustering effect of the traditional clustering algorithms such as the K-Mean algorithm, the SOM neural network and the FCM algorithm on the power load curve is not good. The K-Mean algorithm and the FCM algorithm both need to specify the clustering number before clustering, so that the algorithm is difficult to accurately express the electricity utilization rule of a user; the SOM neural network has a drawback of not necessarily converging when clustering data, although the number of clusters is not specified.

Disclosure of Invention

The invention aims to solve the defects of the prior art and provides a power load clustering algorithm integrating an SOM neural network and a DBSCAN algorithm. Firstly, acquiring original power load data, and carrying out coarse clustering on the original power load data by using an SOM neural network to obtain a clustering result C1; then clustering the class centers of the clustering result C1 by using a DBSCAN algorithm, and merging the class clusters corresponding to the class centers of the same class; and finally, providing a correction deviation element rule to adjust the clustering result, better clustering the power load curve data and obtaining the optimal clustering result.

The purpose of the invention can be achieved by adopting the following technical scheme:

an integrated power load curve clustering method comprises the following steps:

s1, initializing weight vector of SOM neural network neuron and winning neuron field radius N₀：

w_ij＝rand(0，1)，1≤i≤m，1≤j≤n (1)

Wherein, W_iIs the weight vector of the ith neuron of the output layer, W_i＝[w_i1，w_i2，...，w_in]；w_ijIs W_iThe value of the jth component of (a); n is the number of input layer nodes, and m is the number of output layer neurons; the initial winning neuron domain radius N is a larger value N₀；

S2 normalization of input vector X_kSum weight vector W for SOM output layer neurons_i：

Wherein, X_k＝[x_k1，x_k2，...，x_kn]The k-th input vector is obtained, and n is 9 and is the number of nodes of the input layer; w_iIs the weight vector of the ith neuron of the output layer, W_i＝[w_i1，w_i2，...，w_in]Wherein i is more than or equal to 1 and less than or equal to m; i X_kI is the input vector X_kEuclidean norm of; i W_iThe | | is the Euclidean norm of the weight vector of the ith neuron;

s3, calculating input vector X_kAnd weight vector W of output layer neurons_iEuclidean distance of (a), obtaining a winning neuron:

wherein d (X)_k，W_i) Representing the euclidean distance between the kth input vector and the weight vector of the ith neuron. So that d (X)_k，W_i) Get the bestThe small-valued neuron i is the input vector X_kThe winning neuron of (1);

s4, updating weight vectors of neurons in the topology field of the winning neurons:

W_t+1＝W_t+η(t，N_t)·(X_k-W_t) (5)

wherein W_tAs an input vector X_kη (t, N)_t) The formula of the learning rate function is as follows:

s5, updated learning rate η and field of winning neuron topology N:

the learning rate η is updated in the following manner:

η_t+1＝η_t-αt (7)

wherein α is a constant greater than 0;

the updating mode of the winning neuron topological field N is as follows:

N_t+1＝N_t-βt (8)

wherein β is a constant greater than 0;

s6, judging whether the SOM neural network converges:

if the learning rate η is less than η_minOr the maximum iteration number T is reached, the clustering result C is output₁And turning to S7, otherwise turning to S3, wherein η_minAnd T is a value preset by a user;

s7, calculating SOM neural network clustering result C₁Class center of (c)_i：

Wherein, X_i，X_jAre all clustering results C₁The element in the ith class cluster of (1); r is the number of elements contained in the ith cluster; d (X)_i，X_j) Is X_i，X_jThe Euclidean distance of; class center c_iThe group is represented by R;

s8, calculating the clustering result C in S7₁Class center of (c)_iSet of K-distances D of set R:

for data set R ═ { c ═ c₁，c₂，c₃，...，c_m}, calculating element c_iTo the subset s ═ { c of R₁，c₂，c₃，...，c_i-1，c_i+1，...，c_mThe distances of all elements in the distance list are sorted from small to large, and a sorted distance set D' is obtained₁，d₂，d₃，...,d_k，d_k+1，...，d_m}，d_kI.e. the K-distance, for each element c in the set R_iK-distances are all calculated to obtain K-distance set D ═ D for all points_k1，d_k2，...，d_km}；

S9 field radius E for initializing DBSCAN algorithm_psAnd minimum density MinPts:

the value of MinPts, specified by the user, is the K value of the K-distance in S8; radius of area E_psIf there are a plurality of points corresponding to the point of the maximum slope in the K-distance curve D calculated in S8, the average of these points is taken as the domain radius E_psA value of (d);

s10, clustering result C by using DBSCAN algorithm₁Class center of (c)_iClustering is carried out on the set R:

clustering R according to the determined parameters of the DBSCAN algorithm of S8 and S9 to obtain a clustering result C₂；

S11, clustering result C₁Performing parallel classification to obtain clustering result C₃；

S12, calculating a clustering result C₃C 'of'_i；

S13, calculating a clustering result C₃Average distance M of the ith cluster_i；

S14, calculating a clustering result C₃Offset element X 'of the ith cluster'_i；

S15, and dividing the deviation element in S14X′_iClassifying the data into similar classes to obtain a final clustering result C of the power load data₄。

Further, the step S11 is to cluster the result C₁Performing parallel classification to obtain clustering result C₃Further comprising:

clustering result C₂Class center c with same class cluster_i，c_jCorresponding clustering result C₁The data of the ith cluster and the jth cluster in the cluster are merged to obtain a clustering result C₃。

Further, the step S12 calculates a clustering result C₃C 'of'_iFurther comprising:

wherein, X_i，X_jAre all clustering results C₃The element in the ith class cluster of (1); r is the number of elements contained in the ith cluster; d (X)_i，X_j) Is X_i，X_jThe euclidean distance of (c).

Further, the step S13 calculates a clustering result C₃Average distance M of the ith cluster_iFurther comprising:

wherein, X_i，X_jAre all clustering results C₃The element in the ith class cluster of (1); r is the number of elements contained in the ith cluster; d (X)_i，X_j) Is X_i，X_jThe euclidean distance of (c).

Further, the step S14 calculates a clustering result C₃Offset element X 'of the ith cluster'_iFurther comprising:

if C₃Element X in the ith cluster_iTo class center c'_iIs greater than λ times the average distance, then X_iIs namely C₃One of the ith class clustersIs off element X'_iNamely:

wherein λ is a deviation factor, which is a constant greater than 1; the deviating elements are removed from the cluster.

Further, the step S15 is to deviate the element X 'in S14'_iClassifying into similar classes to obtain final clustering result C₄Further comprising:

calculating deviation element X 'removed from ith cluster'_iClass center c 'with other class clusters'_jIs Euclidean distance of X'_iMerge so that distance d (X'_i，c′_j) C 'taking a minimum value'_jIn the corresponding class cluster; obtaining a final clustering result C₄。

Further, before the step S1, the method further includes: power load data is acquired.

Preferably, α is 0.0002 and β is 0.0005.

Preferably, η_minThe value is 0.000002 and the maximum iteration number T is 2000.

Preferably, λ is 1.5.

Compared with the prior art, the integrated power load curve clustering method provided by the invention at least has the following beneficial effects or advantages: the calculation speed is high, and the clustering effect is good; the method can better meet the demand of power grid companies on power load cluster analysis.

Drawings

The present invention will be described in further detail with reference to the accompanying drawings;

FIG. 1 is a flow chart of an integrated power load curve clustering method of the present invention;

FIG. 2(a) is a graph of the clustering results of the first class load data according to the embodiment of the present invention;

fig. 2(b) is a diagram of the clustering result of the second type of load data according to the embodiment of the present invention.

Detailed Description

The embodiments of the present invention are described with reference to the accompanying drawings, and in order to prove the advantages of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below by verifying the power load data of a certain company.

The invention discloses an integrated power load curve clustering method, as shown in fig. 1, the specific implementation case implementation steps are as follows:

and S1, acquiring the power load data. Acquiring power load data from a certain factory, and selecting the power load data from 1/2018 to 8/15/2018 as experimental data. The format of the power load data is that the load is sampled every 15 minutes from 0.00 of the day to 23.45 of the day, and the power load data of 96 points in total form a power load curve of the day.

S2, initializing weight vector of SOM neural network neuron and winning neuron field radius N₀：

w_ij＝rand(0，1)，1≤i≤m，1≤j≤n (1)

Wherein, W_iIs the weight vector of the ith neuron of the output layer, W_i＝[w_i1，w_i2，...，w_in]；w_ijIs W_iThe value of the jth component of (a); n is 9 and m is 16, the number of input layer nodes is the number of output layer neurons; the initial winning neuron domain radius N is a larger value N₀＝8。

S3 normalization of input vector X_kSum weight vector W for SOM output layer neurons_i：

Wherein, X_k＝[x_k1，x_k2，...，x_kn]The k-th input vector is obtained, and n is 9 and is the number of nodes of the input layer; w_iIs the weight vector of the ith neuron of the output layer, W_i＝[w_i1，w_i2，...，w_in]Wherein i is more than or equal to 1 and less than or equal to m, and m is 16; i X_kI is the input vector X_kEuclidean norm of; i W_iAnd | | is the Euclidean norm of the weight vector of the ith neuron.

S4, calculating input vector X_kAnd weight vector W of output layer neurons_iEuclidean distance of (a), obtaining a winning neuron:

wherein d (X)_k，W_i) Representing the euclidean distance between the kth input vector and the weight vector of the ith neuron. So that d (X)_k，W_i) The neuron i which takes the minimum value is the input vector X_kThe winning neuron.

S5, updating weight vectors of neurons in the topology field of the winning neurons:

W_t+1＝W_t+η(t，N_t)·(X_k-W_t) (5)

wherein W_tAs an input vector X_kη (t, N)_t) The formula of the learning rate function is as follows:

(the initial learning rate η is 0.006 in the embodiment of the present invention)

S6, updated learning rate η and field of winning neuron topology N:

the learning rate η is updated in the following manner:

η_t+1＝η_t-αt (7)

wherein α is a constant greater than 0 (0.0002 for an embodiment of the present invention).

The updating mode of the winning neuron topological field N is as follows:

N_t+1＝N_t-βt (8)

wherein β is a constant greater than 0 (0.0005 for the present embodiment).

S7, judging whether the SOM neural network converges:

if the learning rate η is less than η_minOr the maximum iteration number T is reached, the clustering result C is output₁And S8, otherwise S4, wherein η_minAnd T is a value preset by the user (embodiment η of the present invention)_minThe value is 0.000002 and the maximum iteration number T is 2000).

S8, calculating SOM neural network clustering result C₁Class center of (c)_i：

Wherein, X_i，X_jAre all clustering results C₁The element in the ith class cluster of (1); r is the number of elements contained in the ith cluster; d (X)_i，X_j) Is X_i，X_jThe Euclidean distance of; class center c_iThe set is represented by R.

S9, calculating the clustering result C in S8₁Class center of (c)_iSet of K-distances D of set R:

for data set R ═ { c ═ c₁，c₂，c₃，...，c_m}, calculating element c_iTo the subset s ═ { c of R₁，c₂，c₃，...，c_i-1，c_i+1，...，c_mThe distances of all elements in the distance list are sorted from small to large, and a sorted distance set D' is obtained₁，d₂，d₃，...，d_k，d_k+1，...，d_m}，d_kI.e. the K-distance, for each element c in the set R_iK-distances are all calculated to obtain K-distance set D ═ D for all points_k1，d_k2，...，d_km}。

S10 initializing DBSCAN algorithmRadius of area E_psAnd minimum density MinPts:

the value of MinPts, specified by the user, is the K value of the K-distance in S9; radius of area E_psIf there are a plurality of points corresponding to the point of the maximum slope in the K-distance curve D calculated in S9, the average of these points is taken as the domain radius E_psValue of (1) (MinPts is 4 in the embodiment of the present invention, E)_psIs 1.25).

S11, clustering result C by using DBSCAN algorithm₁Class center of (c)_iClustering is carried out on the set R:

clustering R according to the determined parameters of the DBSCAN algorithm of S9 and S10 to obtain a clustering result C₂。

S12, clustering result C₁Performing merging: clustering result C₂Class center c with same class cluster_i，c_jCorresponding clustering result C₁The data of the ith cluster and the jth cluster in the cluster are merged to obtain a clustering result C₃。

S13, calculating a clustering result C₃C 'of'_i：

Wherein, X_i，X_jAre all clustering results C₃The element in the ith class cluster of (1); r is the number of elements contained in the ith cluster; d (X)_i，X_j) Is X_i，X_jThe euclidean distance of (c).

S14, calculating a clustering result C₃Average distance M of the ith cluster_i：

Wherein, X_i，X_jAre all clustering results C₃The element in the ith class cluster of (1); r is the number of elements contained in the ith cluster; d (X)_i，X_j) Is X_i，X_jOu's ofDistance.

S15, calculating a clustering result C₃Offset element X 'of the ith cluster'_i：

If C₃Element X in the ith cluster_iTo class center c'_iIs greater than λ times the average distance, then X_iIs namely C₃Wherein one of the ith cluster is offset from element X'_iNamely:

wherein λ is a deviation factor, which is a constant greater than 1; the deviating elements are removed from the cluster (λ is 1.5 in the present embodiment).

S16, separating the offset element X 'in S15'_iBelongs to class center c 'with minimum Euclidean distance'_jThe method comprises the following steps:

calculating deviation element X 'removed from ith cluster'_iClass center c 'with other class clusters'_jIs Euclidean distance of X'_iMerge so that distance d (X'_i，c′_j) C 'taking a minimum value'_jThe corresponding class cluster. Obtaining a final clustering result C₄. The clustering results of the first-type load data and the second-type load data are shown in fig. 2(a) -2 (b).

And S17, finishing the algorithm.

The above description is only for the preferred embodiments of the present invention, but the protection scope of the present invention is not limited thereto, and any person skilled in the art can substitute or change the technical solution of the present invention and the inventive concept thereof within the scope of the present invention.

Claims (9)

1. An integrated power load curve clustering method, characterized by comprising the steps of:

acquiring original power load data;

roughing raw power load data using SOM neural networkClustering to obtain a clustering result C₁Specifically, the method includes the following steps S1 to S6:

s1, initializing weight vector of SOM neural network neuron and winning neuron field radius N₀：

wij＝rand(0,1),1≤i≤m,1≤j≤n (1)

Wherein, W_iIs the weight vector of the ith neuron of the output layer, W_i＝[w_i1,w_i2,...,w_in]；w_ijIs W_iThe value of the jth component of (a); n is the number of input layer nodes, and m is the number of output layer neurons; the initial winning neuron domain radius N is a larger value N₀；

S2 normalization of input vector X_kSum weight vector W for SOM output layer neurons_i：

Wherein, X_k＝[x_k1,x_k2,...,x_kn]The k-th input vector is obtained, and n is 9 and is the number of nodes of the input layer; w_iIs the weight vector of the ith neuron of the output layer, W_i＝[w_i1,w_i2,...,w_in]Wherein i is more than or equal to 1 and less than or equal to m; i X_kI is the input vector X_kEuclidean norm of; i W_iThe | | is the Euclidean norm of the weight vector of the ith neuron;

s3, calculating input vector X_kAnd weight vector W of output layer neurons_iEuclidean distance of (a), obtaining a winning neuron:

wherein d (X)_k,W_i) Representing the kth input vector andeuclidean distances of weight vectors of i neurons; so that d (X)_k,W_i) The neuron i which takes the minimum value is the input vector X_kThe winning neuron of (1);

s4, updating weight vectors of neurons in the topology field of the winning neurons:

W_t+1＝W_t+η(t,N_t)·(X_k-W_t) (5)

wherein W_tAs an input vector X_kη (t, N)_t) The formula of the learning rate function is as follows:

s5, updated learning rate η and field of winning neuron topology N:

the learning rate η is updated in the following manner:

η_t+1＝η_t-αt (7)

wherein α is a constant greater than 0;

the updating mode of the winning neuron topological field N is as follows:

N_t+1＝N_t-βt (8)

wherein β is a constant greater than 0;

s6, judging whether the SOM neural network converges:

if learning rate η<η_minOr the maximum iteration number T is reached, the clustering result C is output₁And turning to S7, otherwise turning to S3, wherein η_minAnd T is a value preset by a user;

clustering result C using DBSCAN algorithm₁The cluster centers are clustered, the cluster corresponding to the similar cluster centers are merged to obtain a clustering result C₃Specifically, the method includes the following steps S7 to S11:

s7, calculating SOM neural network clustering result C₁Class center of (c)_i：

Wherein, X_i,X_jAre all clustering results C₁The element in the ith class cluster of (1); r is the number of elements contained in the ith cluster; d (X)_i,X_j) Is X_i,X_jThe Euclidean distance of; class center c_iThe group is represented by R;

s8, calculating the clustering result C in S7₁Class center of (c)_iSet of K-distances D of set R:

for data set R ═ { c ═ c₁,c₂,c₃,...,c_m}, calculating element c_iSubset S to R ═ { c₁,c₂,c₃,...,c_i-1,c_i+1,...,c_mThe distances of all elements in the distance list are sorted from small to large, and a sorted distance set D' is obtained₁,d₂,d₃,...,d_k,d_k+1,...,d_m}，d_kI.e. the K-distance, for each element c in the set R_iK-distances are all calculated to obtain K-distance set D ═ D for all points_k1,d_k2,...,d_km}；

S9, initializing the domain radius Eps and the minimum density MinPts of the DBSCAN algorithm:

the value of MinPts, specified by the user, is the K value of the K-distance in S8; radius of area E_psIf there are a plurality of points corresponding to the point of the maximum slope in the K-distance curve D calculated in S8, the average of these points is taken as the domain radius E_psA value of (d);

s10, clustering result C by using DBSCAN algorithm₁Class center of (c)_iClustering is carried out on the set R:

clustering R according to the determined parameters of the DBSCAN algorithm of S8 and S9 to obtain a clustering result C₂；

S11, clustering result C₁Performing parallel classification to obtain clustering result C₃；

Clustering result C according to modified deviation element rule₃Adjusting to obtain final clustering result C₄Specifically, the method includes the following steps S12 to S15:

s12, calculating a clustering result C₃C 'of'_i；

S13, calculating a clustering result C₃Average distance M of the ith cluster_i；

S14, calculating a clustering result C₃Offset element X 'of the ith cluster'_i；

S15, separating the offset element X 'in S14'_iClassifying the data into similar classes to obtain a final clustering result C of the power load data₄。

2. The method for clustering integrated power load curves according to claim 1, wherein the step S11 is performed on the clustering result C₁Performing parallel classification to obtain clustering result C₃Further comprising:

clustering result C₂Class center c with same class cluster_i，c_jCorresponding clustering result C₁The data of the ith cluster and the jth cluster in the cluster are merged to obtain a clustering result C₃。

3. The integrated power load curve clustering method according to claim 1, wherein the step S12 is to calculate a clustering result C₃C 'of'_iFurther comprising:

wherein, X_i,X_jAre all clustering results C₃The element in the ith class cluster of (1); r is the number of elements contained in the ith cluster; d (X)_i,X_j) Is X_i,X_jThe euclidean distance of (c).

4. The integrated power load curve clustering method according to claim 1, wherein the step S13 is to calculate a clustering result C₃Average distance M of the ith cluster_iFurther comprising:

wherein, X_i,X_jAre all clustering results C₃The element in the ith class cluster of (1); r is the number of elements contained in the ith cluster; d (X)_i,X_j) Is X_i,X_jThe euclidean distance of (c).

5. The integrated power load curve clustering method according to claim 1, wherein the step S14 is to calculate a clustering result C₃Offset element X 'of the ith cluster'_iFurther comprising:

if C₃Element X in the ith cluster_iTo class center c'_iIs greater than λ times the average distance, then X_iIs namely C₃Wherein one of the ith cluster is offset from element X'_iNamely:

wherein λ is a deviation factor, which is a constant greater than 1; the deviating elements are removed from the cluster.

6. The integrated power load curve clustering method according to claim 1, wherein the step S15 is to deviate the element X 'in S14'_iClassifying into similar classes to obtain final clustering result C₄Further comprising:

calculating deviation element X 'removed from ith cluster'_iClass center c 'with other class clusters'_jIs Euclidean distance of X'_iMerge so that distance d (X'_i,c′_j) C 'taking a minimum value'_jIn the corresponding class cluster; obtaining a final clustering result C₄。

7. The method of claim 1, wherein α is 0.0002 and β is 0.0005.

8. The integrated power load curve clustering method of claim 1, wherein η_minThe value is 0.000002 and the maximum iteration number T is 2000.

9. The integrated power load curve clustering method of claim 5, wherein λ is 1.5.

Images