CN105205547A - Bus load prediction algorithm based on similarity matching of multiple uncertain factors - Google Patents

Abstract

The invention discloses a bus load prediction algorithm based on similarity matching of multiple uncertain factors. A correlation factor mapping base is constructed firstly; similarity of all historical days to a day to be predicted is calculated by performing mode matching analysis on recent d days of historical sampling data, and similarity of all the historical days is ordered; and the top n days are selected to act as similar days, and virtual bus load is predicted, based on which all industry load of the prediction day is split. The system load prediction algorithm considering influence of meteorological factors, week types, change of operation modes and other correlation factors is adopted on the basis that the industry load characteristics are parsed by the factors allocated point by point, the key user bus load nodes influencing the whole network load are searched, the characteristic change law of various load components of a power grid is elaborately mastered, load prediction hierarchical management is enhanced, and thus scientific and refinement level of load prediction can be comprehensively enhanced.

Description

A kind of bus load prediction algorithm based on multiple uncertain factor similarity mode

Technical field

The present invention relates to a kind of prediction algorithm, specifically a kind of bus load prediction algorithm based on multiple uncertain factor similarity mode.

Background technology

Compared with bus load prediction is predicted with system loading, there is following characteristics: the radix of (1) bus load prediction is less, and institute's on-load may be single load kind, also may be the comprehensive of several load kinds; For single load type, although its part throttle characteristics is obvious, load variations may be comparatively large, not easily holds the load level of each period, as steel plant's impact load etc.; (2) system median generatrix substantial amounts, and the Changing Pattern of every bar bus differs from one another, and cannot accomplish to analyze one by one; (3) due to the impact of power supply area user behavior, bus load less stable, easily produces sudden change; (4) have part bus, be subject to the impact of small power supply generating, load transfer and scheduled overhaul, load rule is also not obvious; (5) data of historical accumulation are not enough accurate, often have abnormal data to occur; (6) large by related factors such as changes of operating modes, small power plant mounting, overhaul of the equipments, load transfers.Therefore, these distinctive features above add the difficulty of bus load prediction.

The Forecasting Methodology being applied to bus load prediction field at present can be classified as two large classes: based on the Forecasting Methodology of bus load Self-variation rule and the Forecasting Methodology based on system loading distribution.The feature of two kinds of Forecasting Methodologies is as follows: (1) bus load has its distinctive Changing Pattern, can apply mechanically the certain methods of system loading prediction when carrying out bus load prediction.But unstable unlike the Changing Pattern of bus load self with system loading, also easily produce sudden change, cause using for reference merely system loading Forecasting Methodology and can not obtain very high precision of prediction.

(2) utilize distribution factor method to distribute system loading, forecast reason is: the first step is predicted system loading, obtains the load value of following a certain moment the whole network; Carry out bus load prediction afterwards, the factor distributes to every bar bus system loading according to a certain percentage.

Traditional distribution factor method is the proportional distribution from macroscopic perspective, lack micro-analysis, and fully cannot excavate the self character of bus load, for maintenance, turn the effective treating method of bus load change shortage that the modes such as confession change generation, therefore precision of prediction is not high, cannot meet the needs of fine-grained management.

Summary of the invention

The object of the present invention is to provide a kind of bus load prediction algorithm based on multiple uncertain factor similarity mode, to solve the problem proposed in above-mentioned background technology.

For achieving the above object, the invention provides following technical scheme:

A kind of bus load prediction algorithm based on multiple uncertain factor similarity mode, first Correlative Factor Mapping storehouse is built, by carrying out pattern match analysis to nearest d days historical sample data, calculate the similarity of all history days and day to be predicted, and the similarity of all history days is sorted, forward n days of selected and sorted is as similar day, virtual bus load is predicted, split the every profession and trade load of prediction day on this basis, (1) determines the apportion model that virtual bus load is predicted; (2) rolling schedule model parameter is determined; (3) time of specifying and virtual bus load are predicted; (4) every profession and trade load is calculated;

The virtual bus load corresponding to prediction day is assigned on every profession and trade load, first a scale load value is defined to each load, its pointwise is added the scale load just forming upper level loading zone, calculate the scale load ratio of each load to upper level load afterwards, pro rata distribute predicted load afterwards.

p _DK＝K _DK·P _DF(k＝1,2,…,n)

$K_{DK}=\frac{p_{ok}}{\underset{j=1}{\overset{n}{\Σ}}{p}_{oj}},(k=1,2,...,n)$

Wherein:

P _dK-bus k predicted load;

K _dK-bus k burden apportionment coefficient (constant);

P _dF-upper level predicted load, if most higher level is exactly system loading predicted value;

P _oK-bus k load criterion value (usually getting the peak load in day or week);

K=1,2 ..., n-bus sequence number

As the present invention's further scheme: the first step build Correlative Factor Mapping database select the principal element affecting load variations comprise weather pattern, max. daily temperature, minimum temperature, week type, set up Index Mappings database, then Different factor value is mapped in an interval that can mutually compare, form Discrete Mapping pair by carrying out sampling to the mapping function of each correlative factor, between sample point, interpolation tries to achieve the value after mapping; Second step, determine the similarity of day characteristic quantity according to the factors quantization index after mapping, choose similar day, be provided with i, j two days, the factors quantization index of its each day is respectively x _ik, x _jk, k=1 ~ m, the wherein number of quantization factor of m for considering every day, x _ik, x _jkbe nonnegative number, it is as follows with jth sky calculating formula of similarity to define i-th day:

$r_{ij}=\underset{k=1}{\overset{m}{\&Sigma;}}(x_{ik}\&CenterDot;x_{jk})/\sqrt{\left(\underset{k=1}{\overset{m}{\&Sigma;}}{x}_{ik}^2\right)\&CenterDot;\left(\underset{k=1}{\overset{m}{\&Sigma;}}{x}_{jk}^2\right)},$ Here x _ikand x _jkbe by former correlative factor value linear mapping to [0,1] between value, similarity concept is used for the degree of closeness of correlative factor between description two days, r _ijmore close to 1, then the correlative factor of two days is more similar, and the distribution of load is also more close; 3rd step, to the data of recent n days that choose as forecast sample collection, asks for the similarity r of each day and day to be predicted in history _i0, i=1 ... n, adopts following formula to be normalized sort to similarity afterwards, select the d that similarity is larger, the d < sample data of n days is as similar day sample preliminary data.

Compared with prior art, the invention has the beneficial effects as follows: the present invention takes pointwise distribution factor to resolve on the basis of industry load character, consider meteorologic factor, week type, the related factor such as changes of operating modes system loading prediction algorithm, the emphasis user bus load node of the whole network load is affected by research, become more meticulous and hold the characteristic variations rule of the various load composition of electrical network, strengthen load prediction differentiated control, improve the scientific and level that becomes more meticulous of load prediction comprehensively.

Accompanying drawing explanation

Fig. 1 is the bus load prediction process flow diagram based on the bus load prediction algorithm of multiple uncertain factor similarity mode;

Fig. 2 is tree-shaped constant load model schematic diagram in the bus load prediction algorithm based on multiple uncertain factor similarity mode;

Fig. 3 is system loading accuracy rate and qualification rate tabular drawing in the bus load prediction algorithm based on multiple uncertain factor similarity mode;

Fig. 4 is industrial trade load accuracy rate tabular drawing in the bus load prediction algorithm based on multiple uncertain factor similarity mode;

Fig. 5 sets based on the mixing load model of the bus load prediction algorithm median generatrix load model of multiple uncertain factor similarity mode.

Embodiment

Below in conjunction with the accompanying drawing in the embodiment of the present invention, be clearly and completely described the technical scheme in the embodiment of the present invention, obviously, described embodiment is only the present invention's part embodiment, instead of whole embodiments.Based on the embodiment in the present invention, those of ordinary skill in the art, not making the every other embodiment obtained under creative work prerequisite, belong to the scope of protection of the invention.

Refer to Fig. 1 ~ 5, in the embodiment of the present invention, a kind of bus load prediction algorithm based on multiple uncertain factor similarity mode, first Correlative Factor Mapping storehouse is built, by carrying out pattern match analysis to nearest d days historical sample data, calculate the similarity of all history days and day to be predicted, and the similarity of all history days is sorted, forward n days of selected and sorted is as similar day, virtual bus load is predicted, on this basis the every profession and trade load of prediction day is split, (1) apportion model that virtual bus load is predicted is determined, (2) rolling schedule model parameter is determined, (3) time of specifying and virtual bus load are predicted, (4) every profession and trade load is calculated,

The virtual bus load corresponding to prediction day is assigned on every profession and trade load, first a scale load value is defined to each load, its pointwise is added the scale load just forming upper level loading zone, calculate the scale load ratio of each load to upper level load afterwards, pro rata distribute predicted load afterwards.

p _DK＝K _DK·P _DF(k＝1,2,…,n)

$K_{DK}=\frac{p_{ok}}{\underset{j=1}{\overset{n}{\&Sigma;}}{p}_{oj}},(k=1,2,...,n)$

Wherein:

P _dK-bus k predicted load;

K _dK-bus k burden apportionment coefficient (constant);

P _dF-upper level predicted load, if most higher level is exactly system loading predicted value;

P _oK-bus k load criterion value (usually getting the peak load in day or week);

K=1,2 ..., n-bus sequence number

The first step build Correlative Factor Mapping database select the principal element affecting load variations comprise weather pattern, max. daily temperature, minimum temperature, week type, set up Index Mappings database, then Different factor value is mapped in an interval that can mutually compare, form Discrete Mapping pair by carrying out sampling to the mapping function of each correlative factor, between sample point, interpolation tries to achieve the value after mapping; Second step, determine the similarity of day characteristic quantity according to the factors quantization index after mapping, choose similar day, be provided with i, j two days, the factors quantization index of its each day is respectively x _ik, x _jk, k=1 ~ m, the wherein number of quantization factor of m for considering every day, x _ik, x _jkbe nonnegative number, it is as follows with jth sky calculating formula of similarity to define i-th day:

$r_{ij}=\underset{k=1}{\overset{m}{\&Sigma;}}(x_{ik}\&CenterDot;x_{jk})/\sqrt{\left(\underset{k=1}{\overset{m}{\&Sigma;}}{x}_{ik}^2\right)\&CenterDot;\left(\underset{k=1}{\overset{m}{\&Sigma;}}{x}_{jk}^2\right)},$ Here x _ikand x _jkbe by former correlative factor value linear mapping to [0,1] between value, similarity concept is used for the degree of closeness of correlative factor between description two days, r _ijmore close to 1, then the correlative factor of two days is more similar, and the distribution of load is also more close; 3rd step, to the data of recent n days that choose as forecast sample collection, asks for the similarity r of each day and day to be predicted in history _i0, i=1 ... n, adopts following formula to be normalized sort to similarity afterwards, select the d that similarity is larger, the d < sample data of n days is as similar day sample preliminary data.

Bus load model: to large-scale power system, the inconsistency of load type and Region dividing will be considered and set up hierarchical tree shaped model from top to bottom, contain the hierarchical relationship meeting type, region, refer to Fig. 5, in this load hierarchical relationship, system loading is in top layer, load type divides and is put into the second layer, and load type comprises industry, business etc., and Region dividing is put into third layer, comprise area, transformer station etc., bus load is positioned at the 4th layer.Ground floor system loading _pDFto second layer type load P _tibetween partition factor be change in time; Second layer type load P _tito third layer region load P _vjbetween partition factor both can change in time, also can be constant; Third layer region load P _vjto the 4th layer of bus load P _dkbetween partition factor be generally constant.In order to improve precision of prediction, two type loads of this model transformer station can separately be predicted.In the electrical network of reality, load type on bus is that mixing exists, and the external factor affecting bus load is also different, as meteorologic factor, changes of operating modes, overhaul of the equipments, load transfer etc., so the load composition allocation proportion of static model is different.

Prediction performance assessment criteria: adopt point load data every day 96 (every day 00::15 ~ 24:00, every 15min mono-point), main performance assessment criteria is as follows:

(1) relative error of the per period j of single busbar load i

(2) day bus load predictablity rate (%)

$(1-\sqrt{\frac{1}{N}\underset{i=1}{\overset{N}{\&Sigma;}}{A}_i^2})\&times;100\%$

Wherein: for all bus statistical errors of period k, M to refer in region the bus load sum of examination, hop count when N refers to every daily forecast total.The accuracy rate of single bus can calculate by M=1.

(3) day bus load prediction qualification rate (%)

$\mathrm{\&eta;}=\sqrt{\frac{1}{N}\underset{j=1}{\overset{N}{\&Sigma;}}\left(\frac{{\mathrm{\&delta;}}_n}{M_j}\right)}.$

Prediction work principle

Gathered for industry total load by each segmented industry load, be defined as a virtual bus load.First this paper predicts virtual bus load value of a certain moment, is then assigned to every class industry, adopts distribution factor method to carry out the parsing of industry load.First Correlative Factor Mapping storehouse is built, by carrying out pattern match analysis to nearest d days historical sample data, calculate the similarity of all history days and day to be predicted, and the similarity of all history days is sorted, forward n days of selected and sorted is as similar day, virtual bus load is predicted, on this basis to prediction day every profession and trade load split, process flow diagram as shown in Figure 1: (1) determines the apportion model that virtual bus load is predicted; (2) rolling schedule model parameter is determined; (3) time of specifying and virtual bus load are predicted; (4) every profession and trade load is calculated.

The virtual bus load corresponding to prediction day is assigned on every profession and trade load, the general model adopting hierarchical tree structure.Virtual bus load corresponds to trunk, and industry load corresponds to the end of each branch.The simplest algorithm of this model be by upper level load in proportion coefficient be assigned to next stage load, scale-up factor is constant in each time period.This model is generally used in very little system with in the bus load prediction of the bottom.

Fig. 2 is the simplest tree-shaped constant load model, first a scale load value is defined to each load, its pointwise is added the scale load just forming upper level loading zone, calculates the scale load ratio of each load to upper level load afterwards, pro rata distribute predicted load afterwards.

p _DK＝K _DK·P _DF(k＝1,2,…,n)

$K_{DK}=\frac{p_{ok}}{\underset{j=1}{\overset{n}{\&Sigma;}}{p}_{oj}},(k=1,2,...,n)$

Wherein:

P _dK-bus k predicted load;

K _dK-bus k burden apportionment coefficient (constant);

P _dF-upper level predicted load, if most higher level is exactly system loading predicted value;

P _oK-bus k load criterion value (usually getting the peak load in day or week);

K=1,2 ..., n-bus sequence number.

Concrete steps are as follows:

(1) Correlative Factor Mapping database is built

Select the principal element affecting load variations to comprise weather pattern, max. daily temperature, minimum temperature, week type, set up Index Mappings database.Then Different factor value is mapped in an interval that can mutually compare, form Discrete Mapping pair by carrying out sampling to the mapping function of each correlative factor, between sample point, interpolation tries to achieve the value after mapping.

(2) Similarity Measure

Determine the similarity of day characteristic quantity according to the factors quantization index after mapping, choose similar day.Be provided with i, j two days, the factors quantization index of its each day is respectively x _ik, x _jk(k=1 ~ m), the wherein number of quantization factor of m for considering every day, x _ik, x _jkbe nonnegative number.It is as follows with jth sky calculating formula of similarity to define i-th day:

$r_{ij}=\underset{k=1}{\overset{m}{\&Sigma;}}(x_{ik}\&CenterDot;x_{jk})/\sqrt{\left(\underset{k=1}{\overset{m}{\&Sigma;}}{x}_{ik}^2\right)\&CenterDot;\left(\underset{k=1}{\overset{m}{\&Sigma;}}{x}_{jk}^2\right)}$

Here x _ikand x _jkbe by former correlative factor value linear mapping to [0,1] between value.Similarity concept is used for the degree of closeness of correlative factor between description two days, r _ijmore close to 1, then the correlative factor of two days is more similar, and the distribution of load is also more close.

(3) pattern-recongnition method

To the data of recent n days that choose as forecast sample collection, ask for the similarity r of each day and day to be predicted in history _i0(i=1 ... n), following formula is adopted to be normalized.

$r_{i0}^{\&prime;}=r_{i0}/\underset{i=1}{\overset{n}{\&Sigma;}}{r}_{i0}$

Afterwards similarity is sorted, select the sample data in d (the d < n) sky that similarity is larger as similar day sample preliminary data.Suppose that the load of each day is in history y _it(i=1 ..., n; T=1 ..., T), T is every daily load sampling number (generally equaling 96 points), then the system loading data of day to be predicted the weighted mean value of this d days each daily loads, shown in following formula.

${\hat{y}}_{0t}=\underset{i=1}{\overset{d}{\&Sigma;}}{r}_{i0}^{\&prime;}{y}_{it},(t=1,...,T)$

(4) its every profession and trade load distribution coefficient lambda is asked according to each point bus load data of each day in history _it, access the system loading predicted value of all T points of day to be predicted afterwards then can obtain all bus load values of day to be predicted., such as formula shown in.

Example is studied

With Guangxi province industrial trade load in September for example is studied.The historical sample adopted is daily load data and weather data in September 21 1 day ~ 2013 September in 2013, predicts the 7 class industrial trade loads on September 28,22 days ~ 2013 September in 2013, and it is 96 points that prediction every day is counted.Forecast reason carries out according to abovementioned steps, first carries out the prediction of total bus load, calculates 7 class industrial load data afterwards according to distribution factor method.Performance assessment criteria selects average day bus load predictablity rate and average day bus load prediction qualification rate.

In order to the applicability of verification model, the load prediction of getting on September 22nd, 2013 to 7 days on the 28th is verified, the system loading precision of prediction obtained is as shown in table 1 with prediction qualification rate, consensus forecast precision and the consensus forecast qualification rate of 7 days are 97.91% and 95.26% respectively, have exceeded the highest appraisal standards of national grid.

According to distribution factor method and system loading, the pointwise load of 7 days each industrial trades can be doped, Fig. 4 display be the accuracy rate of 7 days 7 class industrial trade loads, all up to more than 95%, convincingly demonstrated the feasibility of this algorithm.

Claims (2)

1. the bus load prediction algorithm based on multiple uncertain factor similarity mode, it is characterized in that, first Correlative Factor Mapping storehouse is built, by carrying out pattern match analysis to nearest d days historical sample data, calculate the similarity of all history days and day to be predicted, and the similarity of all history days is sorted, forward n days of selected and sorted is as similar day, virtual bus load is predicted, split the every profession and trade load of prediction day on this basis, (1) determines the apportion model that virtual bus load is predicted; (2) rolling schedule model parameter is determined; (3) time of specifying and virtual bus load are predicted; (4) every profession and trade load is calculated;

The virtual bus load corresponding to prediction day is assigned on every profession and trade load, first a scale load value is defined to each load, its pointwise is added the scale load just forming upper level loading zone, calculate the scale load ratio of each load to upper level load afterwards, pro rata distribute predicted load afterwards.

p _DK＝K _DK·P _DF(k＝1,2,...,n)

Wherein:

P _dK-bus k predicted load;

K _dK-bus k burden apportionment coefficient (constant);

P _dF-upper level predicted load, if most higher level is exactly system loading predicted value;

P _oK-bus k load criterion value (usually getting the peak load in day or week);

K=1,2 ..., n-bus sequence number.

2. the bus load prediction algorithm based on multiple uncertain factor similarity mode according to claim 1, it is characterized in that, the first step build Correlative Factor Mapping database select the principal element affecting load variations comprise weather pattern, max. daily temperature, minimum temperature, week type, set up Index Mappings database, then Different factor value is mapped in an interval that can mutually compare, form Discrete Mapping pair by carrying out sampling to the mapping function of each correlative factor, between sample point, interpolation tries to achieve the value after mapping; Second step, determine the similarity of day characteristic quantity according to the factors quantization index after mapping, choose similar day, be provided with i, j two days, the factors quantization index of its each day is respectively x _ik, x _jk, k=1 ~ m, the wherein number of quantization factor of m for considering every day, x _ik, x _jkbe nonnegative number, it is as follows with jth sky calculating formula of similarity to define i-th day:

here x _ikand x _jkbe by former correlative factor value linear mapping to [0,1] between value, similarity concept is used for the degree of closeness of correlative factor between description two days, r _ijmore close to 1, then the correlative factor of two days is more similar, and the distribution of load is also more close; 3rd step, to the data of recent n days that choose as forecast sample collection, asks for the similarity r of each day and day to be predicted in history _i0, i=1 ... n, adopts following formula to be normalized sort to similarity afterwards, select the d that similarity is larger, the d < sample data of n days is as similar day sample preliminary data.