CN113837465A

CN113837465A - Multi-stage campus power short-term load prediction method

Info

Publication number: CN113837465A
Application number: CN202111109720.8A
Authority: CN
Inventors: 刘江永; 朱江; 周东访; 易灵芝; 王仕通; 刘波
Original assignee: Xiangtan University
Current assignee: Xiangtan University
Priority date: 2021-09-18
Filing date: 2021-09-18
Publication date: 2021-12-24

Abstract

The invention discloses a multi-stage campus power short-term load prediction method. The model classifies typical campus scenes by using a modified clustering algorithm KPCA-SSA-KMEANS, and then decomposes a load signal into a plurality of IMFs by using a modified VMD method (IVMD); performing complexity analysis on the IMF by using the dispersion entropy to combine and recombine the IMF with similar complexity; then respectively establishing an IVMD-DE-LSTM submodel for each sequence, and predicting a load point sequence with a specified length by adopting a rolling strategy; and finally, accumulating the predicted values of all the models to complete the prediction of the future time series. Finally, the method is applied to the scene of campus short-term load prediction, compared with the prior art, the method has the advantages of improving the short-term load prediction precision, reducing the training time and the like, and the prediction error is minimum.

Description

Multi-stage campus power short-term load prediction method

Technical Field

The invention relates to the technical field of power load prediction, in particular to a multi-stage campus power short-term load prediction model building method.

Background

The campus is a comprehensive body with multiple functions of teaching, scientific research, office work, life and the like, and has two significant characteristics of high population density and high building energy consumption. The effective method for forecasting the short-term load of the smart campus has important social value and economic value for power companies and campus users. From the electric power company perspective, not only can be better understand user's power consumption custom under the wisdom campus scene, and then better management electric wire netting system, ensure supply and demand balance, supplementary management department produces and overhauls the allotment of planning and generating set, can improve the stability of electric wire netting moreover, improve the waste of economic benefits and primary energy. From the campus user perspective, can provide the guarantee for personnel study in the school, work, life, improve user's environmental protection consciousness, standardize power consumption action etc..

The combined prediction method organically combines the traditional model method and the machine learning method, and respectively absorbs the advantages of a single method to achieve the purpose of improving the power load prediction precision. The combined prediction method integrates various methods according to different influence weights of different prediction methods on prediction results, and is mainly divided into a weighted average combined prediction method and a fitting degree optimal combined prediction method. The combined prediction method can absorb the advantages of each prediction model and effectively improve the prediction precision.

Disclosure of Invention

The method is a multi-stage combined power short-term load forecasting method which combines a clustering algorithm, a data preprocessing technology, a group intelligent optimization algorithm and a machine learning algorithm. In the first stage of the model, KPCA is adopted to process original multidimensional data to a low-dimensional space and determine the optimal k value of a cluster set, and the global optimization capability of SSA is utilized to determine the initial cluster centroid. The improved KMEANS clustering algorithm divides campus electricity utilization scenes of colleges into a plurality of typical scenes. And in the second stage, an improved VMD technology (IVMD) with higher decomposition efficiency is adopted to reasonably determine a K value and preprocess data. And in the third stage, the clustered and preprocessed power load data of a plurality of typical scenes are recombined by using the distributed entropy and then input into an LSTM model, so that respective IVMD-DE-LSTM prediction models are established. And finally, accumulating the predicted values of all the models to complete the prediction of the future time series. The campus typical scene analysis steps of the improved clustering algorithm are as follows:

step 1, inputting power consumption data.

And 2, preprocessing original data.

And 3, establishing a campus scene power utilization data multidimensional feature set.

And 4, determining the cluster number of the campus electricity utilization scenes. And 3, mapping the multi-dimensional campus scene electricity consumption data feature set established in the step 3 to a visual low-dimensional feature space, and describing the feature relationship among the electricity consumption scenes in the low-dimensional feature space. And determining the cluster number of the typical electricity utilization scenes of the campus according to the characteristic relation distribution condition.

Step 5, determining an initial optimal clustering center C₁,C₂,C₃…C_k。

And 6, clustering analysis is carried out on the campus typical electricity utilization scene. Using the cluster number of the k campus typical electricity utilization scenes determined in the step 4 as the optimal cluster number k of the KPCA-SSA-KMEANS algorithm and the initial optimal cluster center C determined in the step 5₁,C₂,C₃…C_kAnd performing cluster analysis, and dividing and outputting a typical electricity scene cluster of the campus.

The steps for proposing the improved VMD technique (IVMD) are as follows:

step 1, firstly, determining a penalty factor alpha, setting a value range of a mode K, setting the value range of the mode K as [2,20], setting an initial value as 2, carrying out variation mode decomposition on an original signal, and determining the relative entropy of intrinsic mode components (IMF) obtained by each decomposition through calculation to obtain the optimal solution of the corresponding K when the relative entropy is the minimum value.

And 2, determining the optimal solution of the penalty factor alpha according to the obtained K value of the optimal solution. The range of α is set to [100,2000], step 50. Then, the value of alpha corresponding to the minimum relative entropy is found to be the optimal alpha value. The K, α obtained at this time are both optimal.

Drawings

In order to make the reader more clearly understand the embodiments of the present patent, the following brief description of the drawings in the detailed description of the patent is provided:

FIG. 1 is a multi-stage campus power model prediction flow diagram.

Detailed Description

A multi-stage campus power short-term load prediction method comprises the following steps, and a flow chart is shown in figure 1.

Step 1, preprocessing data. And (4) performing front and back mean completion on the missing values, and dividing a training set, a test set and a verification set.

And 2, carrying out cluster analysis on the power load data of each scene. And (4) clustering and analyzing the high-school power load data by using a KPCA-SSA-KMEANS algorithm to obtain k typical power utilization scenes.

And 3, decomposing the load signal. And performing primary IVMD decomposition on the training set, and performing 0-1 normalization processing on a residual sequence generated by subtracting all IMFs from the obtained IMF and the original load signal.

And step 4, combining sequences. And (4) carrying out complexity analysis on the IMF and the residual sequence obtained in the step (3) by using a Dispersion Entropy (DE), and merging and recombining the subsequences with similar complexity.

And 5, constructing an IVMD-DE-LSTM submodel corresponding to each campus typical scene. And setting the super-parameters of the sub-models by using the prior knowledge, and connecting the output indexes of the sub-models to obtain respective integrated models.

And 6, combining the predicted values of the typical scenes of the campuses in the step 5 and outputting the combined predicted values as total predicted values.

Claims

1. A multi-stage campus power load prediction method is characterized in that: the improved clustering algorithm KPCA-SSA-KMEANS algorithm is applied to campus power load data to assist in completing power load prediction tasks, and the improved clustering algorithm avoids the problems that the optimal cluster number of the traditional KMEANS clustering algorithm is difficult to determine and the cluster center selection is influenced and is unstable; the improved VMD method is combined with the dispersion entropy and the LSTM, the problem that the number K of modes needs to be determined in advance in the traditional VMD algorithm is solved through the improved VMD technology, and the decomposition efficiency of the traditional VMD is improved; the decomposed IMF is recombined by using the dispersion entropy, so that the calculated amount of the whole work is reduced; the method provided by the invention has different load characteristics under different campus scenes; different scenes have certain relevance, and how to reasonably use a clustering algorithm to mine and summarize typical power utilization scenes in campus application scenes has important significance on the accuracy and the practicability of subsequent load prediction tasks.

2. The improved clustering algorithm KPCA-SSA-KMEANS algorithm according to claim 1 is characterized in that:

step 1, inputting an original high-dimensional data set L, and calculating a covariance matrix A of the high-dimensional data set after centralization processing;

step 2, a spectrum decomposition matrix A is used for calculating characteristic values;

step 3 retains the first 3 larger eigenvalues λ in the result of step 2₁，λ₂，λ₃And its corresponding feature vector xi₁，ξ₂，ξ₃Constructing a new three-dimensional feature space according to the 3 feature vectors;

step 4, mapping the original high-dimensional data set L to a new feature space;

step 5, determining the optimal clustering number k value according to the clustering condition of the sample points in the new feature space constructed in the step 4;

step 6, initializing a population;

step 7, finding out the maximum individual through fitness calculation, and recording the fitness and the average fitness of the maximum individual;

step 8, updating the positions of the finder, the follower and the alerter;

step 9, calculating the fitness again and finding out the maximum individual, and replacing the individual in the step 7 with the maximum individual;

step 10, judging the iteration times: if yes, outputting an optimal initial clustering center; if not, returning to the step 8 for circulation;

and 11, clustering operation is carried out, and a result is output.

3. The improved VMD algorithm of claim 1, characterized by:

step 1, firstly, determining a penalty factor alpha, setting a value range of a mode K, setting the value range of the mode K as [2,20], setting an initial value as 2, carrying out variation mode decomposition on an original signal, and determining the relative entropy of intrinsic mode components (IMF) obtained by each decomposition through calculation to obtain the optimal solution of the corresponding K when the relative entropy is the minimum value;