CN101706778A - Method for predicting power load - Google Patents

Abstract

The invention relates to a method for predicting power load and provides the power load prediction method which can better solve the problem of low accuracy. The invention provides application thereof in load prediction based on CURE algorithm. Application of the CURE algorithm in the load prediction comprises the following steps: (1) extracting data sample from a historical database of load prediction; (2) adopting hierarchical algorithm to cluster each sub-region of the data sample; and (3) clustering all the data in the sample and inputting representative points only including clusters found when all sub-regions cluster independently.

Description

A kind of method of load forecast

Technical field:

The present invention relates to a kind of Forecasting Methodology of electric load.

Background technology:

Load forecast is the important component part of energy management system and distribution management system, is the foundation of Power System Planning and traffic control, also is the necessary substance of electricity market commercial operation; The start and stop of the inner genset of electrical network can be reasonably arranged in load prediction accurately, reduce unnecessary rotation idle capacity, rationally arrange the turnaround plan of unit, under the condition of ordinary production that guarantees society and life, effectively reduce cost of electricity-generating, improve the economic and social benefits.

Short-term load forecasting is the basis that electric power system dispatching and planning authorities arrange to purchase the electricity plan and the formulation method of operation; Because electric load also has the randomness of self, so short-term load forecasting is a very complicated problems except being subjected to non-linear factors such as temperature, weather conditions influence.

People mainly comprise three aspects to the research of short-term load forecasting method: the research of the research of traditional algorithm, the research of modern algorithm and predicted application, it is different that each studies the emphasis of aspect, but, exist the not high problem of precision of prediction all the time because it is various to influence the factor of load prediction results.

Summary of the invention:

The present invention is exactly at the problems referred to above, and a kind of method that can well solve the load forecast of the not high problem of accuracy is provided.

For achieving the above object, the present invention adopts following technical scheme, the invention provides based on the application of CURE algorithm in load prediction the step of CURE algorithm in load prediction: (1), to extracting data sample out in the historical data base in the load prediction; (2), for each subregion, utilize the level algorithm to carry out cluster; (3), the total data in the sample is carried out cluster, input include only that each subregion finds during cluster alone bunch representative points.

The formula of CURE algorithm is as follows:

$\mathrm{Min}\underset{j}{\overset{m}{\mathrm{\Σ}}}\underset{i-1}{\overset{n}{\mathrm{\Σ}}}{p}_{\mathrm{ij}}d(x_j,y_i)$ Formula 1

Wherein define X=(x ₁, x ₂... x _m) be the sample data space, m is the total sample number order, x _jIt is j

Sample vector, n are the numbers that needs classification, y _iBe the center vector of i class, d (x _j, y _i) be x _jWith y _iCertain distance, as Euclidean distance, manhatton distance etc.

Formula 2

The purpose of optimizing is to obtain y=(y ₁, y ₂... y _m) make that formula 1 is satisfied.

Beneficial effect of the present invention:

The present invention has mainly used the cluster analysis in the data mining to the research of load forecast; Framework a kind of power load forecasting module based on the CURE clustering algorithm, the short-term electric load data are effectively predicted; And by mass data storage, the support of data mining and decision information can overcome the data finiteness effectively, imperfection and the influence of influence factor complicacy to predicting the outcome, performance unique advantage, realization economic worth.

Description of drawings:

Fig. 1 is a CURE algorithm synoptic diagram.

Embodiment:

(1), to extracting data sample out in the historical data base in the load prediction.When the data sample is carried out cluster, can be divided into two kinds of methods: a kind of is that whole sample datas are carried out cluster, and the method will make in the main memory that capacity is not enough and the system that makes can not finish by single pass.We adopt is that whole sample datas are divided into p district, and the data in each district are carried out cluster, like this can be the sample data of each subregion main memory of all packing into.

(2), for each subregion, utilize the level algorithm to carry out cluster.The level algorithm is actually and produces nested bunch collection, and the mode difference according to producing bunch collection can be divided into different level algorithms, and what we here adopted is agglomerative algorithm.Bunch number is n/ (p*q) for the first time, and wherein q is a certain constant.

Owing to use SCADA system acquisition data in the electric system, in the measurement of data, record, conversion, transmission course, all may cause fault and cause load data disappearance or unusual.The generation of abnormal data is at random, thereby the distribution in database also has uncertainty, all kinds of abnormal datas or at a time occur separately, or in same day continuous period, intersected to mix and occur, or on continuous many days same periods numerous situations such as cross-distribution.To the processing of abnormal data, be the key of impact prediction result's order of accuarcy.Can utilize two kinds of different technology suppressing exception points.First kind of technology is bunch deletion that will increases slowly.When bunch number when being lower than a certain threshold value, will only contain one or two member's bunch deletion; Second kind of technology is the final stage in cluster, with very little bunch deletion.

(3), the total data in the sample is carried out cluster, in order to guarantee in internal memory, to handle, input include only that each subregion finds during cluster alone bunch representative points.Use c point to represent each bunch, the entire database on the disk is carried out cluster.Data item in the database be assigned to nearest representative points represent bunch in.The set of representative points must be enough little adapting to the size of main memory, compare so each in n point all has with ck representative points.

Claims (2)

1. the method for a load forecast provides based on the application of CURE algorithm in load prediction, the step of CURE algorithm in load prediction: (1), to extracting data sample out in the historical data base in the load prediction; (2), for each subregion, utilize the level algorithm to carry out cluster; (3), the total data in the sample is carried out cluster, input include only that each subregion finds during cluster alone bunch representative points.

2. the method for a kind of load forecast according to claim 1 is characterized in that the formula of CURE algorithm is as follows:

$\mathrm{Min}\underset{j}{\overset{m}{\mathrm{\Σ}}}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{p}_{\mathrm{ij}}d(x_j,y_i)$ Formula 1

Wherein define X=(x ₁, x ₂... x _m) be the sample data space, m is the total sample number order, x _jBe j sample vector, n is the number that needs classification, y _iBe the center vector of i class, d (x _j, y _i) be x _jWith y _iCertain distance, as Euclidean distance, manhatton distance etc.

Formula 2

Obtain y=(y ₁, y ₂... y _m) make that formula 1 is satisfied.

Images