CN106600059A - Intelligent power grid short-term load predication method based on improved RBF neural network - Google Patents

Abstract

The invention discloses an intelligent power grid short-term load prediction method based on an improved RBF neural network, relates to the technical field of intelligent power grid, and is used for determining the basis function center and improving the load prediction precision of the intelligent power grid. The prediction method includes: S1, performing network initialization; S2, calculating the basis function center ci; S3, calculating the variance [zeta]i according to the basis function center ci; S4, calculating the output Ri of a hidden layer according to the basis function center ci and the variance [zeta]i; S5, calculating the output of an output layer according to the output Ri of the hidden layer; S6, calculating a prediction error E according to a mean squared error and the function; S7, updating connecting weights of neurons of the hidden layer and neurons of the output layer in the neural network; and S8, determining the prediction error E, if the prediction error E is expected, ending iterative calculation, and otherwise, returning to step S4, and re-performing iterative calculation on the prediction error E. The method is used for predicting the load of the power grid.

Description

Based on the intelligent grid short-term load forecasting method for improving RBF neural

Technical field

The present invention relates to intelligent power grid technology field, more particularly to it is a kind of based on the intelligent grid for improving RBF neural Short-term load forecasting method.

Background technology

The fast development of intelligent grid generates substantial amounts of electricity consumption data (also known as sample data), to these sample datas It is analyzed and is significant.Sample data is applied in short-term load forecasting using Forecasting Methodology, so as to increasing productivity Precision of prediction, this sacurity dispatching and economical operation to power system plays an important role.RBF (Radial Basis Function, hereinafter referred to as RBF) neutral net is to be applied to most commonly used a kind of Forecasting Methodology in load prediction, Because it is a kind of partial approximation network, arbitrary continuation function can be approached with arbitrary accuracy, with unique optimal approximation properties and Without local minimum problem, and topological structure is simple, learning rate is fast.Mainly there are three ginsengs in RBF neural Forecasting Methodology Number affects precision of prediction, the respectively connection weight of Basis Function Center, basic function radius and network hidden layer and output layer.Its Middle connection weight is tried to achieve frequently with gradient descent method.The impact of Basis Function Center and basic function radius to precision of prediction is very Greatly, thus it is existing research be mainly concentrated in determine RBF neural Basis Function Center and basic function radius on.It is existing Mainly adopt in technology and calculate with the following method Basis Function Center and basic function radius：

The first calculates Basis Function Center and base letter using clustering methodology (for example, K-means methods and FCM methods) Number radius；Basis Function Center and base are calculated using heuristic (for example, such as genetic method and particle swarm optimization) second Function radius.Wherein, heuristic complexity is high, and predicted time is longer under the extensive load data of intelligent grid, therefore Under the extensive load prediction of intelligent grid, clustering procedure is more suitable for determining RBF neural Basis Function Center and basic function half Footpath.

In addition, for RBF neural Basis Function Center, it is main to be determined using FCM methods, but in intelligent grid load In prediction, FCM method load data scales are big, dimension is more, and method is more complicated, finally cause intelligent grid load prediction accuracy It is relatively low.

The content of the invention

It is an object of the invention to provide a kind of based on the intelligent grid short-term load forecasting side for improving RBF neural Method, for determining Basis Function Center, lifts intelligent grid load prediction precision.

To reach above-mentioned purpose, the present invention is adopted the following technical scheme that：

The present invention provides a kind of based on the intelligent grid short-term load forecasting method for improving RBF neural, based on improvement The intelligent grid short-term load forecasting method of RBF neural includes：

S1, netinit；

S2, calculating Basis Function Center c_i；

S3, according to Basis Function Center c_i, calculate variance ζ_i；

S4, according to Basis Function Center c_iAnd variance ζ_i, calculate the output R of hidden layer_i；

S5, according to the output R of hidden layer_i, calculate the output of output layer；

S6, predicated error E is calculated according to mean square error and function；

S7, the connection weight to hidden layer neuron in neutral net and output layer neuron are updated；

S8, predicated error E is judged, if predicated error E, in expected, iterative calculation terminates；Conversely, returning step Rapid S4, iterates to calculate predicated error E again.

Step S1 includes：Determine input layer number N_Ⅰ, hidden layer neuron number N_Ⅱ, output layer neuron number N_Ⅲ, and initialization learning rate η and basic function overlap coefficient η；Wherein, output layer neuron number N_Ⅲ=1, hidden layer nerve First number N_ⅡAs Basis Function Center number.

Step S2 includes：

S21, S21, input Fuzzy Exponential m, iteration stopping threshold epsilon and PCA accumulation contribution rate factor deltas, Basis Function Center Number N_Ⅱ, the attribute weights ω of cluster_n, initial data X；；Wherein, sample data X={ x₁,x₂,…,x_N, N is sample point Number, x_jFor each sample point, x_j={ x_j1,x_j2,…,x_js, j=1,2 ... N, s represents the category that each sample point is included The number of property, the i.e. dimension of each sample point, K classes, cluster centre V are divided in PCA-WFCM methods by sample data X ={ v₁,v₂,…,v_k, K is had, the scope of K is 2≤K≤N.

S22, PCA attribute dimension-reduction treatment is carried out to sample data X：Each sample point dimensionality reduction is calculated according to equation below Dimension L afterwardsRetain all dimensions before dimension L as cluster attribute, in formula, S represents each sample point Original dimension, λ_nThe characteristic value of covariance matrix in PCA is represented, δ represents the PCA accumulation contribution rate factors；According to equation below Calculate the attribute weights of each cluster：N=1,2 ..., L；According to the attribute weights ω of cluster_nClustered Weight vectors W={ the ω of attribute_nAnd dimensionality reduction after sample data X_new；.

S23, initialization subordinated-degree matrix U：Make 0≤u_ij≤1,U={ u_ij, in formula, u_ijRepresent j-th Sample point belongs to the degree of membership of the i-th class, and K represents the number of cluster centre V.

S24, according to subordinated-degree matrix U and sample data X after dimensionality reduction_new, calculate cluster centre V：V={ v_i, in formula, m is Fuzzy Exponential, x_jRepresent j-th sample point.

S25, according to sample data X after dimensionality reduction_new, iterate to calculate subordinated-degree matrix U：U= {u_ij, in formula, m is Fuzzy Exponential, represents the fuzziness of subordinated-degree matrix U, and the value of m is bigger, then the fuzziness of subordinated-degree matrix U It is higher, make m=2, K represent the number of cluster centre V, d_ijRepresent each sample point x_jTo cluster centre v_iWeighting it is European away from From being calculated by equation below：, L represents the dimension after each sample point dimensionality reduction, ω_nRepresent the attribute weights of cluster.

S26, according to subordinated-degree matrix U and cluster centre V calculating target function J： In formula, m is Fuzzy Exponential, d_ijRepresent each sample point x_jTo cluster centre v_iWeighted euclidean distance, K represents cluster centre The number of V, N represents sample point number.

S27, object function J is judged：If | J^(t)-J^(t-1)|<ε, then export cluster centre V, that is, in basic function Heart c_i；Conversely, then return to step S24, until meeting formula | J^(t)-J^(t-1)|<ε, then stop iterative calculation, in calculating cluster Heart V；In formula, ε represents iteration stopping threshold value, and t represents iterations.

Step S3 includes：The radius ζ of hidden layer neuron is calculated according to equation below_i：ζ_i=λ min_j‖c_i-c_j‖, i, j= 1,2,…N_Ⅱ；In formula, c_iRepresent Basis Function Center, N_ⅡHidden layer neuron number is represented, η represents basic function overlap coefficient.

Step S4 includes：According to sample data X after dimensionality reduction_new, Basis Function Center c_iAnd variance ζ_i, obtain hidden layer Output:I=1,2 ... N_Ⅱ, in formula, N_ⅡRepresent hidden layer neuron number.

Step S5 includes：According to sample data X after dimensionality reduction_new, obtain the output of output layer：In formula, N_ⅡRepresent hidden layer neuron number；W represent i-th hidden layer neuron with The connection weight of output layer neuron, R_iRepresent the output of hidden layer.

Step S6 includes：Predicated error E is calculated using mean square error and function：Make one group of input vector { x_j, j=1, 2 ... O } and correspondence output valve { y_j, j=1,2 ... O } and as training sample,

In formula, O is sample number, predicated error

Step S7 includes：Connection weight W is updated according to equation below：

In formula, η represents learning rate, and E is predicated error, and q is update times.

The computational methods of cluster weighted value Z are as follows：

Calculate covariance matrix C：A S dimension space data are made to be mapped in L n-dimensional subspace ns, wherein, L<<S, it is assumed that X ={ x_nBe zero-mean data, i.e.,Wherein, n=1,2 ..., N,T is to turn order symbol；It is right Covariance matrix C carries out Eigenvalues Decomposition：By data x of a S dimension_iTo L Dimension principal component direction projection, i.e. Y=XQ_L；If the characteristic root λ of covariance matrix C₁≥λ₂≥…≥λ_S, andFor l The contribution rate of individual principal component,For the accumulation contribution rate of front L principal component；K-th attribute is poly- after dimensionality reduction Generic attribute weights：L=1,2 ..., L；Characteristic vector intersection Q=[q₁,q₂,…,q_S], characteristic value Л=diag (λ₁, λ₂,…,λ_S), make accumulation contribution rate be more than 95%.

RBF Basis Function Centers are determined using PCA-WFCM, load data first passes through PCA dimension-reduction treatment, and acquisition is less not Related prediction input, so as to reduce the overlap of RBF basic functions, preferably determines Basis Function Center and reduces prediction algorithm complexity Degree；In addition, being clustered using weighted FCM basic function, the variance contribution ratio of the different attribute obtained after PCA is processed belongs to after dimensionality reduction Property weighting, so as to lift the cluster degree of accuracy, more accurately Basis Function Center is obtained, so as to improve load prediction precision.

Description of the drawings

In order to be illustrated more clearly that the embodiment of the present invention or technical scheme of the prior art, embodiment will be described below Needed for the accompanying drawing to be used be briefly described, it should be apparent that, drawings in the following description be only the present invention some Embodiment, for those of ordinary skill in the art, on the premise of not paying creative work, can be with attached according to these Figure obtains other accompanying drawings.

Fig. 1 is RBF neural basic structure in the embodiment of the present invention；

Fig. 2 is the flow chart in the present embodiment based on the intelligent grid short-term load forecasting method for improving RBF neural.

Specific embodiment

Below in conjunction with the accompanying drawing in the embodiment of the present invention, the technical scheme in the embodiment of the present invention is carried out clear, complete Site preparation is described, it is clear that described embodiment is a part of embodiment of the invention, rather than the embodiment of whole.Based on this Embodiment in bright, the every other enforcement that those of ordinary skill in the art are obtained under the premise of creative work is not made Example, belongs to the scope of protection of the invention.

The present invention is, based on the weighted FCM clustering algorithm of PCA dimensionality reductions, to be collectively referred to as PCA-FCM, obtains intelligent grid short term Predict the outcome.

In addition, the iterative calculation being related in the present invention, it is Approach by inchmeal to refer to, and first takes a coarse approximation, Then with same or several formula, this initial value is corrected repeatedly, until it reaches till predetermined accuracy is required.For example, in the present invention The iteration being related to sets up the training pattern of predicated error E, is directed to the output of hidden layer, the output of output layer, weighted value Update etc..

Hidden layer center c in the present invention_iAlso or referred to as Basis Function Center c_i；The radius ζ of hidden layer neuron_iAlso or referred to as Variance ζ_i。

Embodiment one

The present embodiment proposes a kind of based on the intelligent grid short-term load forecasting method for improving RBF neural.The method RBF Basis Function Center c are determined using PCA-WFCM clustering algorithms_i, using gradient descent method determine RBF neural hidden layer with Output layer connection weight.The following detailed description is based on the intelligent grid short-term load forecasting method for improving RBF neural：

Initially set up sample data X={ x₁,x₂,…,x_N, x_jFor each sample point, x_j={ x_j1,x_j2,…,x_js, j =1,2 ... N, have N number of sample point, x_jRepresent that each sample point includes s attribute, by sample data X in cluster analysis K classes are divided into, the scope of K is 2≤K≤N；Cluster centre V={ v₁,v₂,…,v_k, common K cluster centre.

RBF (RBF) neutral net is a kind of feedforward neural network based on function approaches theory, with BP (Back Propagation) neutral net is compared, and it has the advantages that good function approximation characteristic, simple structure and training speed are fast.As schemed Shown in 1, RBF neural is the three-layer neural network being made up of input layer 1, hidden layer 2, output layer 3, its structure below figure It is shown.

The core concept of RBF neural is, as the base of implicit layer unit, to constitute implicit sheaf space using RBF, Line translation is entered to input vector in hidden layer so that linear inseparable data are linear in higher dimensional space in lower dimensional space Can divide.Input layer to the conversion of hidden layer is nonlinear transformation, and hidden layer to output layer be linear transformation.Hidden layer becomes exchange the letters Number is RBF, and it is the radially symmetrical non-negative nonlinear function of local distribution.RBF neural input layer and Connection weight between hidden layer is 1, and hidden layer completes the parameter adjustment to activation primitive, and output layer is adjusted to connection weight It is whole.In RBF neural, the parameter for solving is needed there are 3：Basis Function Center c_i, the width of hidden layer, hidden layer is to output The connection weight of layer.Gaussian function is the basic function commonly used in RBF neural, therefore hidden layer neuron is output as：

Wherein, c_iFor hidden layer center, i.e., the center of i-th Gaussian function, N_ⅡFor hidden layer neuron number, σ_iIt is RBF hidden Radius containing layer neuron, can be expressed as：

ζ_i=τ min_j‖c_i-c_j‖, i, j=1,2 ... N_Ⅱ； (2)

Wherein, c_iRepresent hidden layer center, N_ⅡHidden layer neuron number is represented, η represents basic function overlap coefficient.

RBF neural is output as the linear combination of all hidden layer neuron outputs, can be expressed as：

Wherein, W is the connection weight of j-th hidden layer neuron and output neuron, R_iRepresent the output of hidden layer.

During RBF neural None-linear approximation, after giving training examples, algorithm needs to solve following two keys to ask Topic：1) determine network structure, that is, determine the Basis Function Center c of RBF neural_i；2) connection of hidden layer and output layer is adjusted Weights omega_n.The selection of these parameters can affect the estimated performance of RBF neural, therefore, before giving a forecast, we will choose Optimum ω_nAnd c_iTo lift the estimated performance of RBF neural.

The connection weight ω of hidden layer and output layer_nIt is general to be trained using gradient descent method.By one group of input vector { x_j,j =1,2 ... O } and correspondence output valve { y_j, j=1,2 ... O } and used as training sample, wherein K is sample number.Then mean square error and letter Number is：

In order to minimize error function, connection weight W is：

Basis Function Center c_iDetermined using PCA-WFCM clustering methods.PCA-WFCM clustering algorithms are described in detail below.

PCA-WFCM clustering algorithms are, based on tradition FCM clustering algorithms, by doing sample data attribute dimensionality reduction, to reduce calculating Method complexity, while doing weighted FCM cluster using each the property variance contribution rate after dimensionality reduction as attribute weights, is further lifted Clustering performance.The algorithm includes two steps, first does PCA dimensionality reductions, then does weighted FCM cluster, and each step is introduced separately below Suddenly.

Principal component analysis (PCA) is linear dimension-reduction algorithm, and a S dimension space data can be mapped in L n-dimensional subspace ns, Wherein L ＜ ＜ S.Requirement in mathematical computations calculates the characteristic vector of the covariance matrix of initial data.Assume X={ x_n, n= 1,2 ..., N are zero-mean data, i.e.,Defining covariance matrix C is：

Do Eigenvalues Decomposition to obtain：

Wherein Q=[q₁,q₂,…,q_S] it is characterized vectorial set, Λ=diag (λ₁,λ₂,…,λ_S) it is characterized value.Can be with profit With front L characteristic vector U_L=[u₁,u₂,…,u_L] by data x of S dimension_iIt is Y=XQ to L dimension principal component direction projections_L.If The characteristic root λ of covariance matrix C₁≥λ₂≥…≥λ_S>=0, definitionFor the contribution rate of l-th principal component,For the accumulation contribution rate of front L principal component.In order to reach the purpose of dimensionality reduction, and wish that information loss is little, Accumulation contribution rate is typically taken more than 95%.

Based on FCM algorithms, it is a kind of fuzzy clustering algorithm to PCA-WFCM algorithms, i.e. fuzzy partitioning method.Each Sample point can not strictly be divided into a certain class, but belong to a certain class with certain degree of membership.Make u_ijRepresent j-th sample Point belongs to the degree of membership of the i-th class, then subordinated-degree matrix and cluster centre are respectively U={ u_ijAnd V={ v_i}.PCA-WFCM algorithms It is that the different weights of attribute imparting take into account Attribute Significance after to dimensionality reduction on the basis of FCM algorithms.The power of attribute Value adopts variance contribution ratios of the PCA to each attribute after initial data attribute dimensionality reduction, the bigger attribute of contribution rate to represent that it is being counted According to concentrating, the effect for playing is also bigger.The weights of k-th attribute after dimensionality reduction：L=1,2 ..., L.

The target of clustering algorithm be make in class that similitude is maximum, similitude is minimum between class, and similitude adopts Euclidean distance Tolerance.Therefore algorithm determines cluster centre V and fuzzy matrix U by minimizing object function, and the object function is

Wherein

Wherein d_ijIt is sample x_jTo cluster centre v_iWeighted euclidean distance

M >=1 is FUZZY WEIGHTED index in formula (8), represents the fuzziness of subordinated-degree matrix U, and m is bigger, the fuzziness of classification It is higher, generally take m=2；L represents the dimension after each sample point dimensionality reduction, and Z represents cluster weighted value.

By to (8), (9) differential calculation, we can obtain u_ijAnd v_iComputing formula be

Wherein x_jRepresent j-th sample.

Embodiment two

According to the thought in embodiment one, neural network prediction needs that neutral net input has been determined in advance in the present embodiment Output and node in hidden layer.Network inputs are determined by the series of parameters for affecting predicted value.Because smart power grid user is born Lotus curve has good cyclophysis, therefore the load value at a certain moment is affected it is contemplated that characteristic diurnal periodicity and cycle are special Property, i.e. select the synchronization load value and the last week synchronization load value of the previous day prediction time.Concrete prediction input Layer neuron number N_Ⅰ=9, including 8 load points：Two moment loads before future position previous moment load value L (t-1), future position Value L (t-2), the previous day same future position load value L (t-48), the previous day same future position previous moment load value L (t-49), The previous day same future position later moment in time load value L (t-47), the last week same future position load value L (t-48 × 7), the last week Same future position previous moment load value L (t-48 × 7-1), the last week same future position later moment in time load value L (t-48 × 7+ 1).Also include day type parameter, i.e., whether be weekend.Prediction output neuron number is N_Ⅲ=1, that is, predict a certain moment Load.Hidden layer neuron number N_ⅡAccording to the minimum determination of predicated error.All-network input is all normalized using minimax Process.

As shown in Fig. 2 being included based on the intelligent grid short-term load forecasting method for improving RBF neural：

(1) netinit.Network input layer neuron number N is determined according to system input and output sequence_Ⅰ, hidden layer god Jing units number N_Ⅱ, output layer neuron number N_Ⅲ, and initialization learning rate η and basic function overlap coefficient η.

(2) RBF Basis Function Center c are sought_i.Basis Function Center is determined using PCA-WFCM clustering algorithms, idiographic flow is：

S21, input Fuzzy Exponential m, iteration stopping threshold epsilon and principal component accumulation contribution rate factor delta, clusters number is base Function Center Number N_Ⅱ, connection weight ω_n, sample data X；

S22, PCA attribute dimension-reduction treatment is done to sample data.According to formulaL is tried to achieve, before retaining dimension L All dimensions as cluster attribute, and according to formulaN=1,2 ..., L；Each cluster attribute is initialized, is obtained The connection weight vector W={ ω of cluster attribute_nAnd dimensionality reduction after sample data X_new；Wherein, X_new={ x₁,x₂,…,x_g, G is number for sample point, x_gFor each sample point, xg={ x_g1,x_g2,…,x_gs, s represents the category that each sample point is included The number of property, the i.e. dimension of each sample point.

S23, according to formula (9) initialize subordinated-degree matrix U；

S24, according to formula (12) calculate cluster centre V；

S25, according to formula (11) calculate subordinated-degree matrix U；

S26, according to formula (8) calculating target function J；

If S27, | J^(t)-J^(t-1)| ＜ ε, then obtain cluster centre V, that is, Basis Function Center c_i；Conversely, returning step Rapid S24；

(3) variance ζ is solved according to formula (2)_i。

(4) output of hidden layer is calculated.According to sample data X after dimensionality reduction_new, hidden layer center c_iAnd variance ζ_i, root The output of hidden layer is calculated according to formula (1).

(5) output of output layer is calculated according to formula (3).

(6) predicated error E is calculated according to formula (4).

(7) carried out clustering the renewal of weighted value according to formula (5).

(8) whether evaluation algorithm iteration terminates, if not terminating, return to step (4).

RBF Basis Function Centers are determined using PCA-WFCM, load data first passes through PCA dimension-reduction treatment, and acquisition is less not Related prediction input, so as to reduce the overlap of RBF basic functions, preferably determines Basis Function Center and reduces prediction algorithm complexity Degree；In addition, being clustered using weighted FCM basic function, the variance contribution ratio of the different attribute obtained after PCA is processed belongs to after dimensionality reduction Property weighting, so as to lift the cluster degree of accuracy, more accurately Basis Function Center is obtained, so as to improve load prediction precision.

The above, the only specific embodiment of the present invention, but protection scope of the present invention is not limited thereto, any Those familiar with the art the invention discloses technical scope in, change or replacement can be readily occurred in, all should contain Cover within protection scope of the present invention.Therefore, protection scope of the present invention should be defined by the scope of the claims.

Claims (9)

1. it is a kind of based on the intelligent grid short-term load forecasting method for improving RBF neural, it is characterised in that to include：

S1, netinit；

S2, calculating Basis Function Center c_i；

S3, according to Basis Function Center c_i, calculate variance ζ_i；

S4, according to Basis Function Center c_iAnd variance ζ_i, calculate the output R of hidden layer_i；

S5, according to the output R of hidden layer_i, calculate the output of output layer；

S6, predicated error E is calculated according to mean square error and function；

S7, the connection weight to hidden layer neuron in neutral net and output layer neuron are updated；

S8, predicated error E is judged, if predicated error E, in expected, iterative calculation terminates；Conversely, return to step S4, iterates to calculate predicated error E again.

2. according to claim 1 based on the intelligent grid short-term load forecasting method for improving RBF neural, its feature It is that step S1 includes：

Determine input layer number N_Ⅰ, hidden layer neuron number N_Ⅱ, output layer neuron number N_Ⅲ, and original chemical Practise speed η and basic function overlap coefficient η；Wherein, output layer neuron number N_Ⅲ=1, hidden layer neuron number N_ⅡAs base Function Center Number.

3. according to claim 1 based on the intelligent grid short-term load forecasting method for improving RBF neural, its feature It is that step S2 includes：

S21, input Fuzzy Exponential m, iteration stopping threshold epsilon and PCA accumulation contribution rate factor deltas, Basis Function Center number N_Ⅱ, gather The attribute weights ω of class_n, initial data X；

Wherein, sample data X={ x₁,x₂,…,x_N, N for sample point number, x_jFor each sample point, x_j={ x_j1, x_j2,…,x_js, j=1,2 ... N, s represents the number of the attribute that each sample point is included, the i.e. dimension of each sample point, Sample data X is divided into into K classes, cluster centre V={ v in PCA-WFCM methods₁,v₂,…,v_k, K is had, the scope of K For 2≤K≤N；

S22, PCA attribute dimension-reduction treatment is carried out to sample data X：Calculated after each sample point dimensionality reduction according to equation below Dimension L：

Retain all dimensions before dimension L as cluster attribute；

In formula, S represents the original dimension of each sample point, λ_nValue is characterized, variance contribution ratio in PCA is represented, δ represents that PCA tires out The product contribution rate factor；

The attribute weights of each cluster are calculated according to equation below：

${\mathrm{\&omega;}}_n=\frac{{\mathrm{\&lambda;}}_n}{{\mathrm{\&Sigma;}}_{n=1}^L{\mathrm{\&lambda;}}_n},n=1,2,...,L;$

According to the attribute weights ω of cluster_nObtain clustering the weight vectors W={ ω of attribute_nAnd dimensionality reduction after sample data X_new；

S23, initialization subordinated-degree matrix U：

Make 0≤u_ij≤1,U={ u_ij}；

In formula, u_ijRepresent that j-th sample point belongs to the degree of membership of the i-th class, K represents the number of cluster centre V；

S24, according to subordinated-degree matrix U and sample data X after dimensionality reduction_new, calculate cluster centre V：

V={ v_i}；

In formula, m is Fuzzy Exponential, x_jRepresent j-th sample point.

S25, according to sample data X after dimensionality reduction_new, iterate to calculate subordinated-degree matrix U：

U={ u_ij}；

In formula, m is Fuzzy Exponential, represents the fuzziness of subordinated-degree matrix U, and the value of m is bigger, then the fuzziness of subordinated-degree matrix U It is higher, make m=2, K represent the number of cluster centre V, d_ijRepresent each sample point x_jTo cluster centre v_iWeighting it is European away from From being calculated by equation below：

L represents the dimension after each sample point dimensionality reduction, ω_nRepresent the attribute weights of cluster；

S26, according to subordinated-degree matrix U and cluster centre V calculating target function J：

In formula, m is Fuzzy Exponential, d_ijRepresent each sample point x_jTo cluster centre v_iWeighted euclidean distance, K represents the number of cluster centre V, and N represents sample point number；

S27, object function J is judged：If | J^(t)-J^(t-1)|<ε, then export cluster centre V, that is, Basis Function Center c_i；Conversely, then return to step S24, until meeting formula | J^(t)-J^(t-1)|<ε, then stop iterative calculation, calculates cluster centre V, in formula, ε represents iteration stopping threshold value, and t represents iterations.

4. according to claim 1 based on the intelligent grid short-term load forecasting method for improving RBF neural, its feature It is that step S3 includes：

The radius ζ of hidden layer neuron is calculated according to equation below_i：

ζ_i=η min_j‖c_i-c_j‖, i, j=1,2 ... N_Ⅱ；

In formula, c_iRepresent Basis Function Center, N_ⅡHidden layer neuron number is represented, η represents basic function overlap coefficient.

5. according to claim 1 based on the intelligent grid short-term load forecasting method for improving RBF neural, its feature It is that step S4 includes：

According to sample data X after dimensionality reduction_new, Basis Function Center c_iAnd variance ζ_i, obtain the output of hidden layer:

$R_i\left(x_{new}\right)=\exp (-\frac{1}{2{{\mathrm{\&sigma;}}_i}^2}||x_{new}-c_i|{|}^2),i=1,2,...N_{II};$

In formula, N_ⅡRepresent hidden layer neuron number.

6. according to claim 1 based on the intelligent grid short-term load forecasting method for improving RBF neural, its feature It is that step S5 includes：

According to sample data X after dimensionality reduction_new, obtain the output of output layer：

$f\left(x_{new}\right)={\mathrm{\&Sigma;}}_{i=1}^{N_{II}}{\mathrm{WR}}_i\left(x_{new}\right);$

In formula, N_ⅡRepresent hidden layer neuron number；W represents the connection weight of i-th hidden layer neuron and output layer neuron Weight, R_iRepresent the output of hidden layer.

7. according to claim 1 based on the intelligent grid short-term load forecasting method for improving RBF neural, its feature It is that step S6 includes：

Predicated error E is calculated using mean square error and function：

Make one group of input vector { x_j, j=1,2 ... O } and correspondence output valve { y_j, j=1,2 ... O } and as training sample, formula In, O is sample number, predicated error

8. according to claim 1 based on the intelligent grid short-term load forecasting method for improving RBF neural, its feature It is that step S7 includes：

Connection weight W is updated according to equation below：

$W(q+1)=W\left(q\right)-\mathrm{\&eta;}\frac{\&part;E}{\&part;{\mathrm{\&omega;}}_n}\left(q\right);$

In formula, η represents learning rate, and E is predicated error, and q is update times.

9. according to any one of claim 1～8 based on improve RBF neural intelligent grid short-term load forecasting side Method, it is characterised in that the computational methods of cluster weighted value are as follows：

Calculate covariance matrix C：A S dimension space data are made to be mapped in L n-dimensional subspace ns, wherein, L<<S, it is assumed that X={ x_n} For zero-mean data, i.e.,Wherein, n=1,2 ..., N,T is to turn order symbol；

Eigenvalues Decomposition is carried out to covariance matrix C：

By data x of a S dimension_iPrincipal component direction projection is tieed up to L, i.e. Y=XQ_L；

If the characteristic root λ of covariance matrix C₁≥λ₂≥…≥λ_S, andFor the contribution rate of l-th principal component,For the accumulation contribution rate of front L principal component；

The cluster attribute weights of k-th attribute after dimensionality reduction：

Characteristic vector intersection Q=[q₁,q₂,…,q_S], characteristic value Л=diag (λ₁,λ₂,…,λ_S), make accumulation contribution rate be more than 95%.