CN106600059B - Intelligent power grid short-term load prediction method based on improved RBF neural network - Google Patents

Abstract

The invention discloses a smart grid short-term load prediction method based on an improved RBF neural network, relates to the technical field of smart grids, and is used for determining a basis function center and improving the smart grid load prediction precision. The prediction method comprises the following steps: initializing a network; s2, calculating the center c of the basis function_i(ii) a S3, according to the center c of the basis function_iCalculating the variance ζ_i(ii) a S4, according to the center c of the basis function_iAnd the variance ζ_iComputing the output R of the hidden layer_i(ii) a S5, output R from hidden layer_iCalculating the output of the output layer; s6, calculating a prediction error E according to the mean square error and the function; s7, updating the connection weight of the hidden layer neuron and the output layer neuron in the neural network; s8, judging the prediction error E, and if the prediction error E is within an expectation range, ending the iterative computation; otherwise, the process returns to step S4, and the prediction error E is calculated again by iteration. The method and the device are used for predicting the load of the power grid.

Description

Intelligent power grid short-term load prediction method based on improved RBF neural network

Technical Field

The invention relates to the technical field of smart grids, in particular to a smart grid short-term load prediction method based on an improved RBF neural network.

Background

The rapid development of the smart grid generates a large amount of power consumption data (also called sample data), and the analysis of the sample data is of great significance. The sample data is applied to short-term load prediction by using a prediction method, so that the load prediction precision is improved, and the method plays an important role in safe scheduling and economic operation of the power system. The Radial Basis Function (RBF) neural network is the most widely applied prediction method in load prediction, because it is a local approximation network, can approximate any continuous Function with any precision, has the unique and best approximation characteristic, has no local minimum problem, and has simple topological structure and fast learning rate. The RBF neural network prediction method mainly comprises three parameters which influence the prediction precision and are respectively a basis function center, a basis function radius and a connection weight of a network hidden layer and an output layer. Wherein the connection weight is usually obtained by a gradient descent method. The influence of the base function center and the base function radius on the prediction accuracy is very large, so the existing research mainly focuses on how to determine the base function center and the base function radius of the RBF neural network. In the prior art, the following method is mainly adopted to calculate the center and radius of the basis function:

the first method is to calculate the center of the basis function and the radius of the basis function by using a clustering method (for example, a K-means method and an FCM method); the second uses heuristics (e.g., such as genetic methods and particle swarm methods) to compute the basis function centers and the basis function radii. The heuristic method is high in complexity and long in prediction time under the large-scale load data of the intelligent power grid, so that the clustering method is more suitable for determining the RBF neural network basis function center and the basis function radius under the large-scale load prediction of the intelligent power grid.

In addition, the RBF neural network basis function center is mainly determined by adopting an FCM method, but in the intelligent power grid load prediction, the FCM method is large in load data scale, multiple in dimensionality and complex in method, and finally the intelligent power grid load prediction accuracy is low.

Disclosure of Invention

The invention aims to provide a smart grid short-term load prediction method based on an improved RBF neural network, which is used for determining a basis function center and improving the smart grid load prediction precision.

In order to achieve the purpose, the invention adopts the following technical scheme:

the invention provides a smart grid short-term load prediction method based on an improved RBF neural network, which comprises the following steps:

s1, initializing the network;

s2, calculating the center c of the basis function_i；

S3, according to the center c of the basis function_iCalculating the variance ζ_i；

S4, according to the center c of the basis function_iAnd the variance ζ_iComputing the output R of the hidden layer_i；

S5, output R from hidden layer_iCalculating the output of the output layer;

s6, calculating a prediction error E according to the mean square error and the function;

s7, updating the connection weight of the hidden layer neuron and the output layer neuron in the neural network;

s8, judging the prediction error E, and if the prediction error E is within an expectation range, ending the iterative computation; otherwise, the process returns to step S4, and the prediction error E is calculated again by iteration.

Step S1 includes: determining input layer neuron number N_ⅠNumber of neurons in hidden layer N_ⅡNumber of neurons in output layer N_ⅢAnd initializing a learning rate η and a basis function overlap coefficient η, wherein the number N of neurons in the output layer_ⅢNumber of hidden layer neurons N1_ⅡI.e. the number of the centers of the basis functions.

Step S2 includes:

s21, S21, input fuzzy index m, iteration stop threshold value, PCA cumulative contribution rate factor and number N of basis function centers_ⅡClustered attribute weight ω_nOriginal data X; (ii) a Wherein, the sample data X ═ { X ═ X₁,x₂,…,x_NN is the number of sample points, x_jFor each sample point, x_j＝{x_j1,x_j2,…,x_jsJ is 1,2, … N, s represents the number of attributes contained in each sample point, i.e. the dimension of each sample point, sample data X is divided into K classes in the PCA-WFCM method, and the clustering center V is { V ═ V }₁,v₂,…,v_kK is more than or equal to 2 and less than or equal to N.

S22, PCA attribute dimension reduction is carried out on the sample data XAccording to the following formula, the dimension L after dimension reduction of each sample point is calculated

All dimensions before dimension L are retained as cluster attributes, where S represents the original dimension of each sample point, λ_nRepresenting the eigenvalue of covariance matrix in PCA, representing PCA cumulative contribution factor; calculating the attribute weight of each cluster according to the following formula:

n is 1,2, …, L, and weight omega is weighted according to the property of the cluster_nObtaining a weight vector W ═ { omega ═ of the clustering attributes_nAnd reduced dimension sample data X_new；。

S23, initializing a membership matrix U: let u be equal to or less than 0_ij≤1,

U＝{u_ijIn the formula, u_ijAnd K represents the number of clustering centers V.

S24, according to the membership matrix U and the sample data X after dimensionality reduction_newCalculating a clustering center V:

V＝{v_iwhere m is the fuzzy index, x_jRepresenting the jth sample point.

S25, according to the sample data X after dimension reduction_newAnd iteratively calculating a membership matrix U:

U＝{u_ijin the formula, m is a fuzzy index and represents the fuzzy degree of the membership degree matrix U, the larger the value of m is, the higher the fuzzy degree of the membership degree matrix U is, let m be 2, K represent the number of clustering centers V, d_ijRepresents each sample point x_jTo the center of the cluster v_iThe weighted Euclidean distance of (c) is calculated by the following formula, L represents eachDimension, omega, of sample points after dimensionality reduction_nThe attribute weight of the cluster is represented.

S26, calculating an objective function J according to the membership matrix U and the clustering center V:

wherein m is a blur index, d_ijRepresents each sample point x_jTo the center of the cluster v_iK represents the number of clustering centers V, and N represents the number of sample points.

S27, judging the target function J: if J^(t)-J^(t-1)|<Then the cluster center V, i.e., the basis function center c, is output_i(ii) a Otherwise, the process returns to step S24 until the formula | J is satisfied^(t)-J^(t-1)|<Stopping iterative calculation and calculating a clustering center V; where, the iteration stop threshold is represented, and t represents the number of iterations.

Step S3 includes: calculating the radius ζ of hidden layer neurons according to the following formula_i：ζ_i＝λmin_j‖c_i-c_j‖，i,j＝1,2,…N_Ⅱ(ii) a In the formula, c_iRepresenting the center of the basis function, N_ⅡRepresenting the number of hidden layer neurons, η represents the basis function overlap coefficients.

Step S4 includes: according to the sample data X after dimensionality reduction_newCenter of basis function c_iAnd the variance ζ_iAnd obtaining the output of the hidden layer:

i＝1,2,…N_Ⅱin the formula, N_ⅡRepresenting the number of hidden layer neurons.

Step S5 includes: according to the sample data X after dimensionality reduction_newAnd obtaining the output of the output layer:

in the formula, N_ⅡRepresenting the number of hidden layer neurons; w represents the connection weight of the ith hidden layer neuron to the output layer neuron, R_iOutput representing hidden layer。

Step S6 includes: the prediction error E is calculated using the mean square error sum function: let a set of input vectors { x_jJ-1, 2 … O and corresponding output value y_jJ-1, 2, … O as training samples,

wherein O is the number of samples and the prediction error

Step S7 includes: the connection weight W is updated according to the following formula:

where η denotes the learning rate, E denotes the prediction error, and q denotes the number of updates.

The clustering weight Z is calculated as follows:

the covariance matrix C is calculated by mapping an S-dimensional space data into L-dimensional subspace, wherein L<<S, let X ═ { X ═ X_nIs zero mean data, i.e.

Wherein N is 1,2, …, N,

t is a rank conversion symbol; and (3) carrying out eigenvalue decomposition on the covariance matrix C:

data x of one S dimension_iProjecting in the direction of main component L D, i.e. Y-XQ_L(ii) a Let the characteristic root λ of the covariance matrix C₁≥λ₂≥…≥λ_SAnd is and

as the contribution rate of the l-th principal component,

the cumulative contribution rate of the first L principal components, and the second after dimensionality reductionClustering attribute weight of k attributes:

1,2 …, L, and a feature vector set Q ═ Q₁,q₂,…,q_S]Characteristic value "l ═ diag (λ)₁,λ₂,…,λ_S) The cumulative contribution rate is made to be greater than 95%.

The method comprises the steps that a PCA-WFCM is adopted to determine the RBF basis function center, load data are subjected to PCA dimension reduction processing to obtain less irrelevant prediction input, so that the overlap of the RBF basis functions is reduced, the basis function center is better determined, and the complexity of a prediction algorithm is reduced; in addition, weighted FCM basis function clustering is adopted, and the variance contribution rates of different attributes obtained after PCA processing are weighted for the attributes after dimensionality reduction, so that clustering accuracy is improved, a more accurate basis function center is obtained, and load prediction accuracy is improved.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments will be briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art that other drawings can be obtained based on these drawings without creative efforts.

FIG. 1 is a basic structure of an RBF neural network according to an embodiment of the present invention;

fig. 2 is a flowchart of a short-term load prediction method of a smart grid based on an improved RBF neural network in the embodiment.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are some, not all, embodiments of the present invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

The invention relates to a weighted FCM clustering algorithm based on PCA dimension reduction, which is called PCA-FCM in a combined mode to obtain a short-term load prediction result of a smart power grid.

In addition, the iterative calculation involved in the invention refers to successive approximation, a rough approximate value is taken firstly, and then the initial value is repeatedly corrected by using the same or a plurality of formulas until the preset precision requirement is met. For example, the iterations involved in the present invention establish a trained model of the prediction error E, involving the output of the hidden layer, the output of the output layer, updates to the weighting values, and so forth.

Implicit layer center c in the invention_iAlso known as the basis function center c_i(ii) a Radius ζ of hidden layer neurons_iAlso known as variance ζ_i。

Example one

The embodiment provides a smart grid short-term load prediction method based on an improved RBF neural network. The method adopts PCA-WFCM clustering algorithm to determine RBF basis function center c_iAnd determining the connection weight of the hidden layer and the output layer of the RBF neural network by adopting a gradient descent method. The method for predicting the short-term load of the smart grid based on the improved RBF neural network is described in detail as follows:

firstly, establishing sample data X ═ X₁,x₂,…,x_N}，x_jFor each sample point, x_j＝{x_j1,x_j2,…,x_js1,2, … N for a total of N sample points, x_jRepresenting that each sample point contains s attributes, dividing sample data X into K classes in cluster analysis, wherein K is more than or equal to 2 and less than or equal to N; cluster center V ═ V₁,v₂,…,v_kAnd K cluster centers.

The Radial Basis Function (RBF) neural network is a feedforward neural network based on a function approximation theory, and compared with a BP (Back propagation) neural network, the Radial Basis Function (RBF) neural network has the advantages of good function approximation characteristic, simple structure, high training speed and the like. As shown in fig. 1, the RBF neural network is a three-layer neural network composed of an input layer 1, a hidden layer 2, and an output layer 3, and its structure is shown in the following figure.

The core idea of the RBF neural network is to combine radial basis functionsThe hidden layer space is formed as the base of the hidden layer unit, and the input vector is transformed in the hidden layer, so that the linearly inseparable data in the low-dimensional space is linearly separable in the high-dimensional space. The transformation of the input layer to the hidden layer is a non-linear transformation, while the hidden layer to the output layer is a linear transformation. The hidden layer transformation function is a radial basis function which is a non-negative nonlinear function with a radially symmetrical local distribution center. The connection weight between the input layer and the hidden layer of the RBF neural network is 1, the hidden layer completes parameter adjustment of the activation function, and the output layer adjusts the connection weight. In the RBF neural network, there are 3 parameters to be solved: center of basis function c_iWidth of the hidden layer, connection weight of the hidden layer to the output layer. The gaussian function is a commonly used basis function in RBF neural networks, so the output of the hidden layer neurons is:

wherein, c_iIs the center of the hidden layer, i.e. the center of the ith Gaussian function, N_ⅡNumber of neurons in the cryptic layer, σ_iIs the radius of the RBF hidden layer neuron, which can be expressed as:

ζ_i＝τmin_j‖c_i-c_j‖，i,j＝1,2,…N_Ⅱ； (2)

wherein, c_iRepresenting the hidden layer center, N_ⅡRepresenting the number of hidden layer neurons, η represents the basis function overlap coefficients.

The output of the RBF neural network is a linear combination of the outputs of all hidden layer neurons, and can be expressed as:

where W is the connection weight of the jth hidden layer neuron to the output neuron, R_iRepresenting the output of the hidden layer.

In the RBF neural network nonlinear approximation process, after a training sample is given, the algorithm needs to solve the following two key questionsTitle: 1) determining the network structure, i.e. determining the basis function center c of the RBF neural network_i(ii) a 2) Adjusting the connection weight omega of the hidden layer and the output layer_n. The selection of these parameters will affect the prediction performance of the RBF neural network, so before prediction, we need to select the optimal ω_nAnd c_iThe prediction performance of the RBF neural network is improved.

Connection weight omega of hidden layer and output layer_nTraining is generally performed by a gradient descent method. A set of input vectors x_jJ-1, 2 … O and corresponding output value y_jJ ═ 1,2, … O } as training samples, where K is the number of samples. The sum-mean-square error function is then:

to minimize the error function, the connection weight W is:

center of basis function c_iAnd determining by adopting a PCA-WFCM clustering method. The PCA-WFCM clustering algorithm is described in detail below.

The PCA-WFCM clustering algorithm is based on the traditional FCM clustering algorithm, reduces algorithm complexity by performing attribute dimension reduction on sample data, and further improves clustering performance by performing weighted FCM clustering by using each attribute variance contribution rate after dimension reduction as an attribute weight. The algorithm comprises two steps, wherein PCA dimensionality reduction is firstly carried out, then weighted FCM clustering is carried out, and each step is respectively introduced below.

Principal Component Analysis (PCA) is a linear dimensionality reduction algorithm that can map an S-dimensional spatial data into L-dimensional subspace, where L < S. mathematical calculations require computation of eigenvectors of the covariance matrix of the original data_nN is 1,2, …, N is zero mean data, i.e.

The covariance matrix C is defined as:

decomposing by characteristic values to obtain:

wherein Q is [ Q ]₁,q₂,…,q_S]For the set of characteristic vectors, Λ ═ diag (λ)₁,λ₂,…,λ_S) For the eigenvalues, the top L eigenvectors U may be utilized_L＝[u₁,u₂,…,u_L]Data x of one S dimension_iProjection to L D principal component direction is Y-XQ_L. Let the characteristic root λ of the covariance matrix C₁≥λ₂≥…≥λ_S> 0, definition

As the contribution rate of the l-th principal component,

the accumulated contribution rate of the first L principal components is generally more than 95% to achieve the purpose of dimension reduction and expect small information loss.

The PCA-WFCM algorithm is based on the FCM algorithm, which is a fuzzy clustering algorithm, namely a soft partitioning method. Each sample point cannot be strictly classified into a certain class, but belongs to a certain class with a certain degree of membership. Let u_ijRepresenting the membership degree of the jth sample point belonging to the ith class, the membership degree matrix and the clustering center are respectively U ═ U_ijV ═ V } and V ═ V_i}. The PCA-WFCM algorithm takes the importance of the attributes into consideration by giving different weights to the attributes after dimension reduction on the basis of the FCM algorithm. The weight of the attribute adopts the variance contribution rate of each attribute after the PCA reduces the dimension of the original data attribute, and the attribute with larger contribution rate shows that the attribute has larger function in the data set. The weight of the kth attribute after dimensionality reduction:

l＝1,2…,L。

the goal of the clustering algorithm is to maximize intra-class similarity and minimize inter-class similarity, and the similarity is measured in Euclidean distance. The algorithm therefore determines the cluster center V and the blur matrix U by minimizing an objective function of

Wherein

Wherein d is_ijIs a sample x_jTo the center of the cluster v_iWeighted euclidean distance of

In the formula (8), m is more than or equal to 1 and is a fuzzy weighting index, the fuzzy degree of the membership degree matrix U is represented, the higher m is, the higher the classified fuzzy degree is, usually m is 2, L represents the dimensionality of each sample point after dimensionality reduction, and Z represents the clustering weight value.

By calculating the differential of (8) and (9), we can obtain u_ijAnd v_iIs calculated by the formula

Wherein x_jRepresenting the jth sample.

Example two

According to the idea of the first embodiment, the neural network prediction in this embodiment needs to determine the input and output of the neural network and the number of hidden layer nodes in advance. The network inputs are determined by a series of parameters that affect the predicted values. Because of intelligenceThe load curve of the power grid user has good periodic characteristics, so that the daily periodic characteristics and the weekly periodic characteristics can be considered for influencing the load value at a certain moment, namely, the load value at the same moment on the day before the predicted moment and the load value at the same moment in the week before the predicted moment are selected. Specific prediction of input layer neuron number N_ⅠThe load value of the previous moment of the prediction point L (t-1), the load value of the two moments before the prediction point L (t-2), the load value of the same prediction point L (t-48) on the previous day, the load value of the previous moment of the same prediction point L (t-49) on the previous day, the load value of the next moment after the same prediction point L (t-47) on the previous day, the load value of the same prediction point L (t-48L 07) on the previous week, the load value of the previous moment before the same prediction point L (t-48 × 7-1) on the previous week, the load value of the next moment after the same prediction point L (t-48 × 7+1) on the previous week and a day type parameter, namely whether the load value is the end of the week or not are included_ⅢThe load at a certain time is predicted as 1. Number of hidden layer neurons N_ⅡAnd determining according to the prediction error minimum. All network inputs are processed by maximum and minimum normalization.

As shown in fig. 2, the smart grid short-term load prediction method based on the improved RBF neural network includes:

(1) and (5) initializing the network. Determining the number N of network input layer neurons according to the system input and output sequence_ⅠNumber of neurons in hidden layer N_ⅡNumber of neurons in output layer N_ⅢAnd initializing the learning rate η and the basis function overlap coefficient η.

(2) Calculating the center c of the RBF basis function_i. Determining the center of the basis function by adopting PCA-WFCM clustering algorithm, wherein the specific process comprises the following steps:

s21, inputting fuzzy index m, iteration stop threshold value and principal component cumulative contribution rate factor, and clustering number, namely number N of centers of basis functions_ⅡConnection weight ω_nSample data X;

and S22, performing PCA attribute dimension reduction processing on the sample data. According to the formula

L are obtained, all dimensions before L are reserved as cluster attributes, and the cluster attributes are obtained according toFormula (II)

n is 1,2, …, L, initializing each cluster attribute, and obtaining a connection weight vector W of the cluster attributes_nAnd reduced dimension sample data X_new(ii) a Wherein, X_new＝{x₁,x₂,…,x_gG is the number of sample points, x_gFor each sample point, xg ═ x_g1,x_g2,…,x_gsAnd s represents the number of attributes contained in each sample point, i.e. the dimension of each sample point.

S23, initializing a membership matrix U according to the formula (9);

s24, calculating a clustering center V according to the formula (12);

s25, calculating a membership matrix U according to the formula (11);

s26, calculating an objective function J according to the formula (8);

s27, if J^(t)-J^(t-1)If is, then the cluster center V, i.e., the basis function center c, is obtained_i(ii) a Otherwise, return to step S24;

(3) solving the variance ζ according to equation (2)_i。

(4) The output of the hidden layer is computed. According to the sample data X after dimensionality reduction_newHidden layer center c_iAnd the variance ζ_iThe output of the hidden layer is calculated according to equation (1).

(5) The output of the output layer is calculated according to equation (3).

(6) The prediction error E is calculated according to equation (4).

(7) The cluster weight value is updated according to equation (5).

(8) And (4) judging whether the iteration of the algorithm is finished or not, and if not, returning to the step (4).

The method comprises the steps that a PCA-WFCM is adopted to determine the RBF basis function center, load data are subjected to PCA dimension reduction processing to obtain less irrelevant prediction input, so that the overlap of the RBF basis functions is reduced, the basis function center is better determined, and the complexity of a prediction algorithm is reduced; in addition, weighted FCM basis function clustering is adopted, and the variance contribution rates of different attributes obtained after PCA processing are weighted for the attributes after dimensionality reduction, so that clustering accuracy is improved, a more accurate basis function center is obtained, and load prediction accuracy is improved.

The above description is only for the specific embodiments of the present invention, but the scope of the present invention is not limited thereto, and any person skilled in the art can easily conceive of the changes or substitutions within the technical scope of the present invention, and all the changes or substitutions should be covered within the scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the appended claims.

Claims (7)

1. A short-term load prediction method of a smart grid based on an improved RBF neural network is characterized by comprising the following steps:

s1, initializing the network;

carrying out dimensionality reduction on sample data of the smart power grid; carrying out PCA row weighting FCM (fuzzy C means) basic function clustering on the sample data subjected to PCA dimensionality reduction, wherein when carrying out weighted FCM basic function clustering on the sample data of the smart grid subjected to PCA dimensionality reduction, weighting the attribute clusters subjected to dimensionality reduction by the variance contribution rates of different attributes obtained in the PCA dimensionality reduction to obtain a basic function center c_iThe method comprises the following steps:

s2, calculating the center c of the basis function_i；

S3, according to the center c of the basis function_iCalculating the variance ζ_i；

The clustering weight value is calculated as follows:

the covariance matrix C is calculated by mapping an S-dimensional space data into L-dimensional subspace, wherein L<<S, let X ═ { X ═ X_nIs zero mean data, i.e.

Wherein N is 1,2, …, N,

t is a rank conversion symbol;

and (3) carrying out eigenvalue decomposition on the covariance matrix C:

data x of one S dimension_iProjecting in the direction of main component L D, i.e. Y-XQ_L；

Let the characteristic root λ of the covariance matrix C₁≥λ₂≥…≥λ_SAnd is and

as the contribution rate of the l-th principal component,

cumulative contribution rate for the first L principal components;

the cluster attribute weight of the kth attribute after dimensionality reduction:

set of eigenvectors Q ═ Q₁,q₂,…,q_S]Characteristic value "l ═ diag (λ)₁,λ₂,…,λ_S) Making the cumulative contribution rate greater than 95%;

according to the obtained basis function center c_iAnd variance ζ_iPerforming RBF neural network prediction on the sample data after the weighted FCM basis function clustering to obtain output layer output, which is specifically as follows:

s4, according to the center c of the basis function_iAnd the variance ζ_iComputing the output R of the hidden layer_i；

Step S4 includes:

according to the sample data X after dimensionality reduction_newCenter of basis function c_iAnd the variance ζ_iAnd obtaining the output of the hidden layer:

in the formula, N_ⅡRepresenting the number of hidden layer neurons;

s5, according to implicationOutput of layer R_iCalculating the output of the output layer so as to obtain a short-term load prediction result of the intelligent power grid;

s6, calculating a prediction error E according to the mean square error and the function;

s7, updating the connection weight of the hidden layer neuron and the output layer neuron in the neural network;

s8, judging the prediction error E, and if the prediction error E is within an expectation range, ending the iterative computation; otherwise, the process returns to step S4, and the prediction error E is calculated again by iteration.

2. The smart grid short-term load prediction method based on the improved RBF neural network as claimed in claim 1, wherein step S1 comprises:

determining input layer neuron number N_ⅠNumber of neurons in hidden layer N_ⅡNumber of neurons in output layer N_ⅢAnd initializing a learning rate η and a basis function overlap factor tau, wherein the number of neurons N of the output layer_ⅢNumber of hidden layer neurons N1_ⅡI.e. the number of the centers of the basis functions.

3. The smart grid short-term load prediction method based on the improved RBF neural network as claimed in claim 1, wherein step S2 comprises:

s21, inputting fuzzy index m, iteration stop threshold value, PCA cumulative contribution rate factor and number N of basis function centers_ⅡClustered attribute weight ω_nOriginal data X;

wherein, the sample data X ═ { X ═ X₁，x₂，…，x_NN is the number of sample points, x_jFor each sample point, x_j＝{x_j1，x_j2，…，x_jsJ is 1,2, … N, s represents the number of attributes contained in each sample point, i.e. the dimension of each sample point, sample data X is divided into K classes in the PCA-WFCM method, and the clustering center V is { V ═ V }₁，v₂，…，v_kK is more than or equal to 2 and less than or equal to N;

s22, conducting PCA attribute dimension reduction processing on the sample data X, and calculating the dimension L of each sample point after dimension reduction according to the following formula:

retain all dimensions before dimension L as cluster attributes;

where S represents the original dimension of each sample point, λ_nRepresenting the variance contribution rate in the PCA and representing the PCA cumulative contribution rate factor as a characteristic value;

calculating the attribute weight of each cluster according to the following formula:

weighting omega according to clustered attributes_nObtaining a weight vector W ═ { omega ═ of the clustering attributes_nAnd reduced dimension sample data X_new；

S23, initializing a membership matrix U:

let u be equal to or less than 0_ij≤1，

U＝{u_ij}；

In the formula u_ijRepresenting the membership degree of the jth sample point belonging to the ith class, and K representing the number of clustering centers V;

s24, according to the membership matrix U and the sample data X after dimensionality reduction_newCalculating a clustering center V:

V＝{v_i}；

wherein m is a blur index, x_jRepresents the jth sample point;

s25, according to the sample data X after dimension reduction_newAnd iteratively calculating a membership matrix U:

U＝{u_ij}；

wherein m is a fuzzy index and represents the fuzzy degree of the membership matrix U, the higher the value of m is, the higher the fuzzy degree of the membership matrix U is, let m be 2, K represents the number of the clustering centers V, d_ijRepresents each sample point x_jTo the center of the cluster v_iThe weighted euclidean distance of (a) is calculated by the following formula:

l denotes the dimension, ω, of each sample point after dimensionality reduction_nRepresenting the attribute weight of the cluster;

s26, calculating an objective function J according to the membership matrix U and the clustering center V:

wherein m is a blur index, d_ijRepresents each sample point x_jTo the center of the cluster v_iK represents the number of clustering centers V, and N represents the number of sample points;

s27, judging the target function J: if J^(t)-J^(t-1)If is, then the cluster center V, i.e., the basis function center c, is output_i(ii) a Otherwise, the process returns to step S24 until the formula | J is satisfied^(t)-J^(t-1)If the value is less than the threshold value, stopping iterative calculation, and calculating a clustering center V, wherein the value represents an iteration stop threshold value, and t represents iteration times.

4. The smart grid short-term load prediction method based on the improved RBF neural network as claimed in claim 1, wherein step S3 comprises:

calculating the radius σ of the hidden layer neurons according to the following formula_i：

σ_i＝τmin_j||c_i-c_j||，i,j＝1,2,…N_Ⅱ；

In the formula, c_iRepresenting the center of the basis function, N_ⅡRepresenting the number of hidden layer neurons and tau representing the basis function overlap coefficient.

5. The smart grid short-term load prediction method based on the improved RBF neural network as claimed in claim 1, wherein step S5 comprises:

according to the sample data X after dimensionality reduction_newAnd obtaining the output of the output layer:

in the formula, N_ⅡRepresenting the number of hidden layer neurons; w represents the connection weight of the ith hidden layer neuron to the output layer neuron, R_iRepresenting the output of the hidden layer.

6. The smart grid short-term load prediction method based on the improved RBF neural network as claimed in claim 1, wherein step S6 comprises:

the prediction error E is calculated using the mean square error sum function:

let a set of input vectors { x_jJ-1, 2 … O and corresponding output value y_jJ-1, 2, … O as training samples,

wherein O is the number of samples and the prediction error

7. The smart grid short-term load prediction method based on the improved RBF neural network as claimed in claim 1, wherein step S7 comprises:

the connection weight W is updated according to the following formula:

where η denotes the learning rate, E denotes the prediction error, and q denotes the number of updates.

Images