AU2016325186A1 - Bus load forecasting method - Google Patents

Abstract

The present invention discloses a method of bus load forecasting, the method comprises the following steps: using the horizontal comparison method to correct the abnormal value in the historical load data, 5 and using the gray correlation projection method to determine the key influencing factors of the bus load; using an improved K-means clustering method to classify the load curve with similar characteristics into one class, obtaining several typical load patterns, constructing the random forest classification model, and establishing a mapping relationship between influencing factors and clustering results; for each class of load pattern, using multiple linear regression method to train several forecasting models; using the random 10 forest classification model to determine the class of the day under test, and selecting the matching regression model to realize the load forecasting. The present invention adopts the data mining method to analyze the change rule of the bus load and to set up forecasting model base to realize the model matching in combination with the class of day under test, which improves the accuracy and real-time performance of short-term bus load forecasting, and provides more accurate decision support for power grid planning 15 and real-time scheduling.

Description

The present invention discloses a method of bus load forecasting, the method comprises the following steps: using the horizontal comparison method to correct the abnormal value in the historical load data, and using the gray correlation projection method to determine the key influencing factors of the bus load; using an improved K-means clustering method to classify the load curve with similar characteristics into one class, obtaining several typical load patterns, constructing the random forest classification model, and establishing a mapping relationship between influencing factors and clustering results; for each class of load pattern, using multiple linear regression method to train several forecasting models; using the random forest classification model to determine the class of the day under test, and selecting the matching regression model to realize the load forecasting. The present invention adopts the data mining method to analyze the change rule of the bus load and to set up forecasting model base to realize the model matching in combination with the class of day under test, which improves the accuracy and real-time performance of short-term bus load forecasting, and provides more accurate decision support for power grid planning and real-time scheduling.

DESCRIPTION

Bus Load Forecasting Method

CROSSREFERENCE TO RELATED APPLICATIONS [0001] This disclosure claims the priority of Chinese patent application No. 2016108123390, filed on September 9, 2016, and the entire contents thereof are incorporated herein by reference.

Technical Field [0002] The present invention relates to the field of power system engineering, and more particularly to a method of bus load forecasting.

Background [0003] Bus load forecasting is an important part of power system planning and the basis of economic operation of power system. Its forecasting results can better realize decentralized load management, which directly affects the analysis results of subsequent power grid safety check, which is very important for dynamic state estimation, reactive power optimization, local control of power plant, and reduction of power generation cost. Bus load is used as the bus-supplied load of substation, due to its own characteristics, such as the small base, the not strong stability, the not obvious change trend, and the excessive high frequency fluctuating components, which make it difficult to improve the forecasting accuracy. However, short-term bus load forecasting generally refers to real-time forecasting. It requires not only high forecasting accuracy but also fast calculation speed. Due to the characteristics of bus load itself and the inherent flaws of traditional forecasting methods, the research on system load forecasting is less.

[0004] Bus load forecasting generally uses regression analysis, neural network and support vector regression and other methods. Different forecasting methods extract historical load information from different perspectives, and have different requirements on the size and complexity and the like of sample data. When dealing with bus load forecasting problems considering multiple influencing factors, the regression analysis method is simple to implement, but it is not appropriate to deal directly with the bus load with strong fluctuation. Neural network and support vector regression have strong ability of non-linear fitting, which require not so strong the smoothness of the sample as the regression analysis method. However, when the input data dimension is too large and the sample size is large, the training of the model is easy to fall into local optimum, and the complexity is relatively higher, which reduces the forecasting accuracy and computational efficiency. And with the rise of large data technology and the rapid development of sensing technology, has formed a large power data, the internal characteristics of bus load become more complex, which is not conducive to the establishment of forecasting models. Therefore, the existing bus load forecasting algorithm cannot achieve a satisfactory forecasting effect.

Summary [0005] In view of the above, it is an object of the present invention to provide a method of bus load forecasting which can analyze the change rule of the bus load and improve the forecasting accuracy of the bus load while satisfying the real-time requirement.

[0006] Based on the above objects, the present invention provides a method of bus load forecasting, comprising:

inputting key influencing factor values of the bus under test as characteristic vectors into a pre-constructed classification model to obtain the class of day under test; the pre-constructed classification model is a result of clustering analysis based on the key influencing factors and the historical load data of the bus under test; establishing a mapping relationship between the key influencing factors and bus load;

selecting a corresponding multiple regression forecasting model from the multiple regression forecasting models obtained from the pre-training according to the class of day under test; forecasting the bus load under test, wherein the multiple regression forecasting models obtained from pre-training are a result of clustering analysis based on the historical load data of the bus under test; selecting the corresponding historical load data and influencing factor values for different typical load patterns as a forecasting model obtained from sample data training.

[0007] Optionally, the classification model is constructed by:

using the horizontal comparison method to select abnormal data from the historical bus load data of the bus under test, and correcting the abnormal data according to a preset algorithm to obtain the corrected load data;

using the gray correlation projection method to calculate the correlation degree between the influencing factors and the historical load data in combination with the corrected load data and the load-related influencing factors;

selecting a preset number of influencing factors corresponding to the correlation degree from large to small as the key influencing factors;

clustering analyzing the historical load data of bus under test according to the preset K-means clustering algorithm to obtain different typical load patterns;

according to the key influencing factors and the result of clustering analysis, using the random forest algorithm to construct the classification model, and establishing the mapping relationship between the key influencing factors and various bus loads.

[0008] Optionally, the step of selecting the abnormal data from the historical bus load data of the bus under test by the horizontal comparison method comprises:

cleaning the historical load data of the bus under test by the horizontal comparison method to identify the abnormal data; wherein, the horizontal comparison method using the change rate of the adjacent point load as a criterion for determining whether or not abnormal data is present, the change rate of the load is calculated as follows:

where ^e change rate of the bus load under test at time t of d day, and L(d,t) j_{s t}jj_e hus load value under test at time t of d day;

whether the sampling point data is abnormal data depends on:

L(d.t)-L(d.t-1) _t=1,

L(d,t —1)

L(d.t)-IXd-l,_q)

L(d,q)

Ja(d,t)<a^_iax (d,t), the load at time t is normal value va(d,t)>a^_iax (d,t),theloadat time t is abnormal value and

a.

(d,t) _fL(d-i,t)-L(d — i,t —1)-, · _{3 n} ^max{-ι/d r n-^} ’ ^{1 = X}’²’ ·

L(d — i,t -1) _rL(d-i,t)-L(d-i-l,q). . . _ max--—---—1 = 1,2, ¹ L(d—i,q) ' ,n,t = 1,2,,...,q .,n,t = 0 where ^amax _maxj_mum value of change rate of load at time t of d day during the first n days, i represents the number of days from the current d day, ^{— 1} — 1) j_{s t}he sampling point data at time (t-1) of (d-i) day, (q+1) sampling points per day; using the last m days of data to correct the abnormal data;

comparing the load value at the current time with the load value at the previous time in chronological order, correcting the abnormal data once the abnormal data detected, and using the corrected data as the comparison value of load value at the next time, completing the load data 15 correction.

[0009] Optionally, the step of correcting the abnormal data by the preset algorithm comprises:

correcting the selected abnormal data according to the following algorithm, the specific algorithm formula is:

L(d,t) = ^Iftd - l,t) + - 2,t) + ... + 2_mL(d - m,t) e(0,l)

ΣΛ⁼¹

J⁼¹ λ

where ^m is the weight coefficient, which is used to represent the influencing degree of the load at time t of (d-m) day on the load at time t of d day; P is the smoothing coefficient, ^{— ΠΊ}Τ ) j_s the bus load value under test at time t of (d-m) day.

[0010] Optionally, the step of using the gray correlation projection method to calculate the correlation 5 degree between the influencing factors and the historical load data and selecting a preset number of influencing factors corresponding to the correlation degree from large to small as the key influencing factors comprises:

First, determining a reference sequence and a comparison sequence, where the reference sequence is the corrected load data: X° ^— ; the comparison sequence is the load-related influencing factors: Ν — 1 i > ^xi..J J — 1,2,...,n . obtaining the sequence matrix of the reference sequence and the comparison sequence:

Xo		<XO1 ·	• %h?
X1	=	; ·
x„		s%i ·	• %m^

where m represents that each sequence has m elements, and n is the number of comparison sequences; X_o is the load sequence or reference sequence corresponding to the corrected load data, X_; is the i-th comparison sequence;

normalizing the matrix X as follows:

^eo = max : 1.2....,::% ¹¹¹¹¹¹ i 1.2...., m%

-(1-0.1) + 0.1 where mm and max are the minimum and maximum values of the row elements of

X..

¹ respectively, and ^1J is the normalized value of the element Xy of the matrix X, the value of 20 limited to 0.1 to 1;

the normalized matrix X is as follows:

is

-Ό1

V^enl ^e0m e , nm J p £.1=12 where ⁰ and ¹ ’ ’ ’ sequence;

are normalized reference sequence and normalized comparison the correlation coefficient -°'^{v 7} of the k-th element of normalized comparison sequence e_; and normalized reference sequence e₀ is calculated as follows:

mirr min, I — e₀, I +z?max, max, I — e₀, I

4„,(k) =-k-^{J P}' ”, ⁰¹ , i = 1,2,...,n, j = 1,2,...,

1¾ -e_ok +/)iimx, max le_e -e_0J I min. min. I e. — e_n I where ¹ J υ ^UJ is the minimum difference between the two levels, max. max. I e.. — e„. I . , . .... . , , n . . ., ..

¹ j u xj is the maximum diiierence between two levels; _1S the identification coefficient;

the gray correlation coefficient judgment matrix F is calculated as follows:

£oo(D ·	·· ^oo(j) ·	·· 4, (rn)
^(1)	· £n(j) ·	·· £n (m)
^oi(l) ·	· ^oi(j) ·	·· 20l(m)
£o„(D ·	'· £o„(j) ·	·· ^on(m)

^oo(J) where

1	.. 1 ..	1
^(1) ·	· £n(j) ·	· ^01 (m)
^oi(l) ··		• 20l(m)
£o„(D ·	·· £o„(j) ·	·· ^on(m)

is the correlation coefficient between the normalized reference sequence e₀ and its own

A (j) j-th element, and the value of >θθ ^{v J 7} is 1;

using the entropy weight method to give weights to column vectors of matrix E, except for the first row elements, the formula is as follows:

E- =ln(m) _i=1

ΣΡϋ^1ηΡϋ ln(m) j

ΣΧ /ΣΧ)¹¹¹^ ^ZZ^eij)’ j =1,2,...m i=l —Ew_; =-j = l,2,...,m '-Σε j=l where Ej is the information entropy of the data of the j -th column, Py is the proportion of the value of the 15 element ey in the column elements, Wj is the weight of the j-th column data, and the weight vector

W = (w_1?..., Wj,..., w_m) the weighted gray correlation coefficient judgment matrix F is as follows:

where ^e(b

W1 w/oiQ)	... Wj · ··· w^01(j) ·	·· Wm ·· wX(m)
w/oiQ) '	··· w^Oi(j) ··	'· wjoi(m)
W/onQ) ·	·· w^On(j) ·	·· wJOn(m)
Wj^oi(j)	is the weighted correlation cot

calculating the gray correlation projection value, the formula is as follows:

Z^wA(jWj j=! _

I m I m i = l,2, where D_; is the correlation projection value between the comparison sequence X_; and the bus load under test, D_; is between 0 and 1;

selecting the M factors which are the top of the projection values from large to small as the key influencing factors.

[0011] Optionally, the step of clustering analyzing the historical load data of bus under test according to the preset K-means clustering algorithm comprises:

using the corrected load data as a sample set, setting the sample set as iN’···’^xi’···’Nn) , load

X=fx X X ) C°=ic°c⁰ C° C°1 sequence as' ΐΛο’-’ it’···’ in J, the initial centroid set as ’’ ²'’ j’’’ ^{k J}, selecting one sample randomly as the first centroid, calculating the Euclidean distance to the first centroid separately for each of the remaining samples:

<ϊ = Σ(\-<ΐ) t=0 where m is the number of load sequences in the sample set, n represents that each load sequence has (n+1) data points, k is the number of centroids; ^X'¹ represents the load value at time (t+1) in the i-th load _c° sequence, ¹¹ represents the value of the (t + l)-th element in the first centroid vector;

selecting the sample with the largest Euclidean distance as the second centroid, then calculating the Euclidean distance to the second centroid separately for each of the remaining samples, then selecting the sample with the largest Euclidean distance as the third centroid, and so on until k centroids are determined;

calculating the Euclidean distance of each sample to all centroids separately, selecting the centroid with the smallest Euclidean distance from the current sample as the class to which the current sample belongs:

n ² d(k,cj) = Σ<Χ-c‘_t) , j =1,2,...,k,l =0,1,...,h t=0 where 1 is the number of iterations, x_; is the i-th sample, and ^J is the centroid vector of samples of class j before the (l+l)-th iteration;

updating the centroid of each class after all samples classified; let the number of samples of class j be ^mj and the sample set be , recalculating the centroid vector of class j as follows:

^Cj ⁼ (^CjO’^Cjl’-’^Cjt’---’^Cjn) »

where calculation formula of (t+l)-th component is:

XjteSj ^Xit , j =1,2, ,k determining whether or not the preset termination condition is reached, continuing the iteration calculation if not reached, until the historical load of the bus under test is classified as a typical load pattern of class k.

[0012] Optionally, the step of determining whether or not the preset termination condition is reached comprises:

after the completion of each iteration, calculating the distance interval between the various centroids before and after updating according to the following formula:

Dis' =11 c* -c'f¹ i = 1 2 k ^J ’·>····> , Di_sk represents the centroid distance interval of class j after the completion of 1-th iteration;

selecting the maximum value ^max{Dis. ), p- ^g maximum value is less than the difference tolerance , the algorithm terminates, otherwise reclassifying the samples to continue the iteration updating.

[0013] Optionally, the step of using the random forest algorithm to construct the classification model comprises:

' I ’O · _ /—y sampling with replacement s training sample sets ¹ , ^{1 —} EA···,^s _of the same size randomly from the historical load data sets;

sampling N influencing factors randomly as the characteristic attribute of each training sample set; wherein N<M;

training s sample sets TS; separately to generate the corresponding decision tree Tree;; remaining the characteristic attribute of the decision tree unchanged throughout the forest growth process;

except for the leaf nodes, each decision tree using the Gini index of the CART algorithm as the basis for node splitting:

Gini(d) = 1-£β 'i,d i=l

for each node, selecting the attribute with the smallest Gini index as the splitting attribute a; the Gini index of current node d split by splitting attribute a is:

Gini(a,d) = p_LGini(d_L) + p_RGini(d_R) where d_L and d_R are the left and right child nodes of node d, and P_L and P_R are the proportion of the left and right child nodes in the parent node;

continuing to split from top to bottom according to the above rules, until all nodes are split or marked as leaf nodes, that is, the decision tree growth is complete, there are a total of s trees, none of the trees needs pruning operation, leaf nodes of each tree correspond to some clustering result of clustering analysis; combining the s decision trees in order to obtain the random forest classification model;

when a test sample input, using each decision tree to classify the test sample, obtaining s classification results and selecting the class with the highest proportion as the class of test sample.

[0014] Optionally, establishing the multiple regression model by the multiple linear regression method, and determining the regression parameters by the least square method.

[0015] Optionally, the regression equation is:

f 00 = b₀ +b^i + b₂x_i2 + - · +t»A_n

X γ γ is the input sequence of sample k ™ is the value of the n-th influencing factor, ’i is the true value of the bus load corresponding to the output;

the input matrix X, the output matrix Y, and the coefficient matrix B are as follows:

X =

' 1	X11	X12 ·'	• Xln
					_Yi ’		bo
1	X21	^2 ’	•
				, Y =	y2	, B =	bl
1	xmi	Xm2 ·	• Xmn_		_ Yn _		_b„_

using the least squares method to determine the estimated parameter value of the regression equation:

(X^TX)-¹X^TY b„ [0016] As can be seen from the above, the bus load forecasting method provided by the present invention determines the key influencing factors of the load data by the gray correlation projection method. By means of K-means clustering and random forest classification method in combination with regression analysis method, and based on the characteristics of the sequence to establish statistical forecasting model separately, to mine the inherent rule of bus load data, to decompose the changing complex historical load into several classes of typical load patterns, to targeted train forecasting model, thus to complete the forecasting by selecting matching models according to the class of day under test. The method of bus load forecasting not only can improve the accuracy of the bus load forecasting, but also meet the requirement of real-time forecasting, so that the forecasting result is more stable and reliable.

Brief Description of the Drawings [0017] Figure 1 is a flow chart of a method of bus load forecasting according to one embodiment of the present invention;

[0018] Figure 2 is a flow chart of a method of bus load forecasting according to another embodiment of the present invention;

[0019] Figure 3 is a flow chart of the gray correlation projection method according to one embodiment of the present invention;

[0020] Figure 4 is a flow chart of the preset K-means clustering algorithm according to one embodiment of the present invention;

[0021] Figure 5 is a graph showing a result of clustering of load data in a bus load forecasting method according to the present invention;

[0022] Figure 6 is a flow chart of the random forest algorithm according to one embodiment of the present invention;

[0023] Figure 7 is a thumbnail view of the decision tree generation in the random forest algorithm according to the present invention;

[0024] Figure 8 is a graph showing a result of bus load forecasting in the bus load forecasting method according to the present invention.

Detailed Description [0025] In order to make the objects, technical solutions and advantages of the present invention become more clear and easy to understand, the present invention will be described in further detail with reference to specific embodiments in conjunction with the accompanying drawings.

[0026] It should be noted that all the expressions first and second used in the embodiments of the present invention are for the purpose of distinguishing two different entities having the same name or different parameters, and the use of first and second is for convenience of description only, and should not be construed as limiting the embodiments of the present invention, which will not be further described in detail in this regard in the following embodiments.

[0027] In order to overcome the shortcomings of short-term bus load forecasting in the prior art, the present invention adopts the data mining method in large data technology to cluster for several bus load patterns, and establishes the statistics forecasting models based on the characteristic rules of various load sequences in combination with regression analysis method, so as to select the corresponding model according to the class of day under test to complete the forecast, that is, to achieve model matching. Compared with the traditional method establishing the forecasting model directly, this method can better mine the change rule of bus load and improve the accuracy of the bus load forecasting by the time-domain decomposition of the historical load sequence; and the forecasting model base can be directly obtained once the forecasting model training is completed. The load of bus at all times of one day could be forecasted directly, it is not necessary to retrain the model before every forecasting to satisfy the real-time requirement of short-term forecasting.

[0028] With reference to Figure 1, in one embodiment of the present invention, a method of bus load forecasting is provided, comprising:

[0029] In step 101, inputting key influencing factor values of the bus under test of day under test as characteristic vectors into a pre-constructed classification model to obtain the class of day under test; the pre-constructed classification model is a result of clustering analysis based on the key influencing factors and the historical load data of the bus under test; establishing a mapping relationship between the key influencing factors and bus load;

[0030] In step 102, selecting a corresponding multiple regression forecasting model from the multiple regression forecasting models obtained from the pre-training according to the class of day under test; forecasting the bus load under test, wherein the multiple regression forecasting models obtained from pre-training are a result of clustering analysis based on the historical load data of the bus under test; selecting the corresponding historical load data and influencing factor values for different typical load patterns as a forecasting model obtained from sample data training.

[0031] In this way, by classifying the model classes to which day under test belongs firstly and then forecasting them using the forecasting model, the bus load can be forecasted more accurately, and the efficiency of bus load forecasting can be further improved, especially for the short-term bus load io forecasting. That is, the method of bus load forecasting according to the present invention realizes the efficient and accurate forecasting of the bus load based on the model matching.

[0032] The pre-constructed classification model is one of the key steps of the present invention, and therefore, in an alternative embodiment of the present invention, there is also provided a construction method of a classification model, comprising:

using the horizontal comparison method to select abnormal data from the historical bus load data of the bus under test, and correcting the abnormal data according to a preset algorithm to obtain the corrected load data; using the gray correlation projection method to calculate the correlation degree between the influencing factors and the historical load data in combination with the corrected load data and the load-related influencing factors;

selecting a preset number of influencing factors corresponding to the correlation degree from large to small as the key influencing factors; clustering analyzing the historical load data of bus under test according to the preset K-means clustering algorithm to obtain different typical load patterns;

according to the key influencing factors and the result of clustering analysis, using the random forest algorithm to construct the classification model, and establishing the mapping relationship between the key influencing factors and various bus loads.

[0033] Of course, the algorithm selected in the above-described construction method is merely an alternative to the embodiment of the present invention, and other suitable algorithms can be used for the construction of the model. But also does not limit the order between the above steps, only need to make the logical relationship with the normal algorithm steps can be.

[0034] With reference to Figure 2, a flow chart of a method of bus load forecasting according to another embodiment of the present invention is provided. The method of bus load forecasting comprising:

[0035] In step 201, obtaining and storing historical load data of the bus under test and load-related influencing factor data; wherein the influencing factors including factors such as meteorological information, day class and the like capable of affecting bus load.

[0036] In step 202, using the horizontal comparison method to select abnormal data from the historical bus load data of the bus under test, and correcting the abnormal data according to a preset algorithm to obtain the corrected load data; that is, preprocessing the historical load data of the bus under test, using the horizontal comparison method to process the historical load data of the bus under test to identify the abnormal data in load sequence and to correct the identified abnormal data.

[0037] In step 203, using the gray correlation projection method to calculate the correlation degree between the influencing factors and the historical load data in combination with the corrected load data and the load-related influencing factors; by determining the key influencing factors of the bus load, making the forecasting model has reasonable input dimension.

[0038] In step 204, clustering analyzing the historical load data of bus under test according to the preset K-means clustering algorithm, distributing the initial cluster centroids evenly in the sample space, and classifying the historical load data with similar change characteristics into one class, to obtain several typical load patterns. Distributing the initial cluster centroids evenly in the sample space according to the preset K-means clustering algorithm improves the convergence speed and clustering accuracy of the algorithm.

[0039] In step 205, according to the key influencing factors and the result of clustering analysis, using the random forest algorithm to construct the classification model, and establishing the mapping relationship between the key influencing factors and various bus loads.

[0040] In step 206, based on the result of clustering analysis of the historical load data of the bus under test, selecting the corresponding historical load data and influencing factor values for different typical load patterns, training the corresponding historical load data and influencing factor values to obtain different multiple regression forecasting models.

[0041] In step 207, inputting key influencing factor values of the bus under test of day under test as characteristic vectors into the classification model to obtain the class of day under test, selecting the corresponding multiple regression forecasting model according to the class of day under test to forecast the bus load under test.

[0042] It can be seen from the above embodiment that the bus load forecasting method provided by the present invention corrects the abnormal value in the historical load sequence of bus under test and determines the key influence factors of the bus load by the gray correlation projection method; uses the improved K-means algorithm to clustering analyze the corrected historical data, and uses the random forest algorithm to construct the classification model to establish the mapping relationship between the clustering results and key influencing factors; trains several multiple regression models for each class of load pattern; inputs the influencing factors of day under test into the random forest classification model to obtain the class of day under test, to select the matching regression model to complete the forecasting. The method of bus load forecasting not only improves the accuracy of bus load forecasting, but also meets the requirement of real-time forecasting, so that the forecasting result is more stable and reliable.

[0043] In some alternative embodiments of the present invention, the step of selecting the abnormal data from the historical bus load data of the bus under test by the horizontal comparison method comprises: cleaning the historical load data of the bus under test by the horizontal comparison method to identify and correct the abnormal data; wherein, the horizontal comparison method using the change rate of the adjacent point load as a criterion for determining whether or not abnormal data is present, the change rate of the load is calculated as follows:

[ΐχά,Ο-Ι/ά,ΐ-Ι) [ L(d,t —1) ’ _| L(d,t)-L(d-l,q) _|

L(d,q) where gjj^gg _{rate o}f ^_us |_oaj _un(j_er t_est _at time t of d day, and L(d,t) j_{s t}he ^_us load value under test at time t of d day;

whether the sampling point data is abnormal data depends on:

J a (d ,t)<aJJ_iax (d,t), the load at time t is normal value va(d,t)>a^_iax (d,t),theloadat time t is abnormal value and n

max (d,t)

L(d — i, t) — L(d — i, t — 1) . _{1 o} max{-_:-}, ϊ = 1,2, max{

L(d-i,t-l)

L(d - i, t) - L(d - i -1, q) L(d -i, q) }, i = l,2, ,n,t = 1,2,,...,q .,n,t = 0 where ^amax ;_{s t}jj_{e max}i_mum value of change rate of load at time t of d day during the first n days, i represents the number of days from the current d day, Ifod — i,t — 1) j_{s sam}pp_ng point data at time (t-1) of (d-i) day, (q+1) sampling points per day;

correcting the identified abnormal data according to the following formula:

L(d,t) = Λ, L(d — 1, t) + d — 2,t) + ... + 2_mL(d — m,t) ^< λ. =β(ϊ-βγ~\ e(o,i)

ΣΤ-¹ +1 λ

where ^m is the weight coefficient, which is used to represent the influencing degree of the load at time t of (d-m) day on the load at time t of d day; P is the smoothing coefficient, ^— mO is the bus load value under test at time t of (d-m) day, that is, using the last m days of data as corrected data.

[0044] The whole process is carried out in chronological order, and only the load value at the previous time is compared. When the anomaly data is detected, it is corrected immediately and is used as the comparison value of the next time data to complete the correction of all historical load data.

[0045] In some alternative embodiments of the present invention, with reference to Figure 3, the step of using the gray correlation projection method to calculate the correlation degree between the influencing factors and the historical load data and selecting a preset number of influencing factors corresponding to the correlation degree from large to small as the key influencing factors comprises: using the gray correlation projection method to calculate the correlation degree between each influencing factor and the bus load. Classifying the calculated gray correlation projection values and selecting the M factors with high value as the key influencing factors. The specific calculation steps are as follows:

[0046] In step 301, first, determining a reference sequence and a comparison sequence, where the reference sequence is the corrected load data: ^— iNn, Νκ,···, Nm ) _; the comparison sequence is the influencing factors (such as meteorological data, day class): ^— iNi’Nmi ¹ — 1,2,...,n .

obtaining the sequence matrix of the reference sequence and the comparison sequence:

Xo		<XO1 ·	• *0?
	=	; ·
X„_		Λ. ·	• Xnmy

where m represents that each sequence has m elements, n is the number of comparison sequence. X_o represents the reference sequence, that is, the load sequence; X; is i-th comparison sequence, that is, influencing factor sequence.

[0047] In step 302, normalizing the matrix X as follows:

. — min ^eu = j= 1,2,....m Xj max.

,. — mm.

j-1,2,...,m Xj

-(1-0.1) + 0.1 mm. , ₇ x. x. max. . ₇ x.

where ^J 9 is the minimum value of the row elements of ^1J, and ^J « is the maximum value of the row elements of eX, the value of ^1J is limited to 0.1 to 1.

The normalized matrix X is as follows:

e_; ^1J is the normalized value of the element χ_Η of the matrix <_e e ^{A C}01 · · · ^C0m where ⁰ and sequence.

e_i9i = 1,2,...,n are normalized reference sequence and normalized comparison [0048] In step 303, calculating the correlation coefficient, the correlation coefficient -^{(li v} ' of the k-th 15 element of normalized comparison sequence e_; and normalized reference sequence e₀ is calculated as follows:

min, min, I — e₀, I +/?max_j max, I — e₀, I 4„,(k) =--³ , ^{,J P}, ³ , °³ , i = 1,2,...,n, j = 1,2,...,

1¾ -e_ok l+pmax.maxj le_e -e_0J I , min. min. I e-- e_n. I . , . . , , ,, where ^{J J 1} is the minimum difference between the two levels, max-max , I e-— e_n, I . , . ,.-. , , , n . , ., .

j u is the maximum difference between two levels; f is the identification coefficient, preferably, the value of P is 0.5.

The gray correlation coefficient judgment matrix F is calculated as follows:

F =

X(D	··· ZooO ·	·· Zoo (rn)>		( 1	• 1 ··	' 1 Ί
Z01(D	·· £n(j) ·	·· Zoi(m)		ZoiQ) ·	• Z01(j) ·	• Zoi(m)
^oi(l)		·· Zoi(m)	—	ZOi(D ·	• ZOl(j) ··	• Zoi(m)
Jond)	·· ZonG) ·	·· Zon(m) y		Jo„(D '	'· Zon(J) ·	·· ZOn(mb

Λ (j) where j_s the correlation coefficient between the normalized reference sequence e₀ and its own / (j) j-th element, and the value of >θθ^{ν J7} is 1.

[0049] In step 304, using the entropy weight method to give weights to column vectors of matrix E, except 5 for the first row elements, the formula is as follows:

E. =ln(m) ti

ΣΡίΛΡίι ln(m) _i=1

ΣΧ ^ZZ^eij)’ j =1,2,...m l —Ew_; =-j =l,2,...,m πι-ΣΕ, j=l where Ej is the information entropy of the data of the j -th column, Py is the proportion of the value of the element ey in the column elements, Wj is the weight of the j-th column data, and the weight vector

W =
[0050] In step 305,	a weighted gray correlation judgment matrix F is obtained from step 304:
	W1	·· Wj ·	·· Wm
	w/oi(l) ·	·· w^01(j) ·	·· wX(m)
F' =
	w/oiQ) ·	·· w^Oi(j) ·	• wjoi(m)
	,w/on(!) ·	·· w^On(j) ·	• wJOn(m)?
	wi4i(j)
where	J 77 Ul \ J z	is the weighted correlation coefficient between the sequence e; and the j-th element of

e₀[0051] In step 306, calculating the gray correlation projection value, the formula is as follows:

D.

m

Σ ^wj^oi( j) · j=!

ill! I 111

Σ» j=i j=i where D; is the correlation projection value between the comparison sequence X; and the bus load, D; is between 0 and 1, the closer the value is to 1, the greater correlation degree; the closer the value is to 0, the weaker the correlation degree. Compared with the gray correlation, gray correlation projection value can be more comprehensive to reflect the degree of similarity of the development trends of two objects.

[0052] In step 307, selecting M factors with the largest projection values as the key influencing factors, the values of M could be set on demand, and could classify all gray correlation projection values firstly before the selecting process, selecting the top of M factors from large to small as the key influencing factors, it is possible without classifying.

[0053] In some alternative embodiments of the present invention, the step of clustering analyzing the historical load data of bus under test according to the preset K-means clustering algorithm comprises:

using the corrected historical load data as the sample set, and using the improved K-means algorithm to cluster the load characteristics of bus. Specific calculation process is as follows:

[0054] In step 401, the selecting of the initial centroid. In the traditional algorithms, the initial clustering centroid is selected randomly, and the improved algorithm proposed by the present invention distributes the initial centroid evenly in the sample space with distance as the criterion, and effectively reduces the influence of the initial centroid on the clustering result. Setting the sample set as ^— iN’···’ N ’···’ ^xm , load sequence as ^X' iNo’···’ Nt’···’ Nn) , the initial centroid set as sample is selected randomly as the first centroid.

C^u={c[\c°...,c)

X’} . One [0055] In step 402, calculating the Euclidean distance to the first centroid separately for each of the remaining samples:

^ = Σ(^-υθ) t=0 where m is the number of load sequences in the sample set, n represents that each load sequence has (n+1) data points, k is the number of centroids; ^X'¹ represents the load value at time (t+1) in the i-th load _c° sequence, ¹¹ represents the value of the (t + l)-th element in the first centroid vector;

selecting the sample with the largest Euclidean distance as the second centroid, then calculating the Euclidean distance to the second centroid separately for each of the remaining samples, then selecting the sample with the largest Euclidean distance as the third centroid, and so on until k centroids are determined. Optionally, k is 6.

[0056] In step 403, determining whether k centroids have been selected, if does, means that the selecting process has been completed and can proceed to the next step. If the number of centroids has not yet reached k, then continue the selecting process until reached.

[0057] In step 404, let obtained k centroids mutually correspond to different classes of load patterns separately; in other words, establishing the corresponding relationship, and the number of iterations this moment is denoted as 0.

[0058] In step 405, sample assignment or sample determination. Calculating the Euclidean distance of each sample to all centroids separately, selecting the load class corresponding to the centroid with the smallest Euclidean distance from the sample, assigning the sample to the load class. The calculation formula is as follows:

^ά<Χφ=Σ<Χ-4) ’ J⁼¹’²’ k,l =0,1,..,h t=0 where 1 is the number of iterations, x_; is the i-th sample, and ^J is the centroid vector of samples of class j before the (l+l)-th iteration.

[0059] In step 406, centroid updating: Calculating centroid vectors of each class, updating the centroid of each class after all samples assigned; let the number of samples of class j be and the sample set be recalculating the centroid vector of class j as follows:

^Cj ⁼ (^CjO’^Cjl’-’^Cjt’---’^Cjn) »

where (t+l)-th component is calculated as follows:

j =1,2, ,k [0060] In step 407, determining whether or not the termination condition is reached. In the present invention, in addition to selecting and setting the maximum number of iterations as the termination condition, it is also determined whether or not the algorithm can be terminated by defining the difference tolerance & . After the completion of each iteration, calculating the distance interval between the various centroids before and after updating according to the following formula:

018,=110,-0^11, j=l,2,...,k ! , .... . , _c , . _c , ^{J J} j ·' , DiSj represents the centroid distance interval ot class j alter the completion of 1-th iteration; each assignment updated, there are k calculations. Selecting the maximum value^max{kfi^sj b J 1,2,..., k _va|_ue j_s |_ess than the predefined difference tolerance, the algorithm terminates, otherwise proceeds to step 405 to continue the iteration updating.

[0061] This process classifies the bus historical load into load pattern of class k, the change trend of bus load in each load pattern is more obvious and more regular, which lays the foundation for constructing model and improving forecasting accuracy.

[0062] In step 408, obtaining k clustering results.

[0063] In a further embodiment of the present invention, the step of using the random forest algorithm to construct the classification model further comprises:

sampling with replacement s training sample sets i, ^{1 —} T^z’···’^sof S size randomly from the historical load data set S;

sampling N influencing factors randomly as the characteristic attribute of each training sample set according to M key influencing factors; wherein N<M;

training s sample sets TS; separately to generate the corresponding decision tree Tree;. Remaining the characteristic attribute of the decision tree unchanged throughout the forest growth process. Except for the leaf nodes, each decision tree using the Gini index of the CART algorithm as the basis for node splitting:

Gini(d) = 1-^ p*_d i=l where d is the current node, ^’d ;_{s t}jj_e proportion of target class i, k is the number of target class. The closer the Gini index is to 0, the higher the purity of the classification, the better the effect.

[0064] For each node, considering all cases where each characteristic attribute to node splitting, selecting the attribute with the smallest Gini index as the splitting attribute a; the Gini index of current node d split by splitting attribute a is:

Gini(a,d) = p_LGini(d_L) + p_RGini(d_R) where d_L and d_R are the left and right child nodes of node d, and P_L and P_R are the proportion of the left and right child nodes in the parent node.

[0065] Continuing to split from top to bottom according to the above rules, until all nodes are split or marked as leaf nodes, that is, the decision tree growth is complete, there are a total of s trees, none of the trees needs pruning operation, leaf nodes of each tree correspond to some clustering result of clustering analysis. When a test sample is input, each decision tree is used to classify the test sample, s classification results are obtained and the class with the highest proportion is selected as the class of test sample.

[0066] Combining the s decision trees in order to obtain the random forest classification model, thus to establish the mapping relationship between the influencing factors and the clustering results, and complete the determination of the class of day under test in real time, laying the foundation for the forecasting model matching.

[0067] In some alternative embodiments of the present invention, the step of obtaining different multiple regression forecasting models from the training further comprises:

[0068] Considering the clustering result of bus load and the real-time of short-term load forecasting, the present invention uses the multiple linear regression models for forecasting. To use bus load data of each class and key influencing factors as training samples to construct several forecasting models. Establishing the multiple regression model by the multiple linear regression method, and determining the regression parameters by the least square method.

[0069] Specifically, the i-th training sample in each class of sample set is X = i = l,2,...,m is ’ Yi) , where

Xi , m is the number of samples in each class of load sample set;

x x y is the input sequence of sample i, ⁱⁿ is the value of the n-th influencing factor, is the true value of the bus load corresponding to the output.

The regression equation is:

f 00 = b₀ +b^i + b₂x_l2 + · · · +b_A where b₀ is the regression constant term, bp^b2,...,b_n -_{s re}g_ressj_on coefficient, is the regression forecasting value.

[0070] The input matrix X, the output matrix Y, and the coefficient matrix B are as follows:

X =

' 1	X11	X12 '	• Xln
					_Yi ’		bo
1	X21	^2 ’	•
				, Y =	y2	, B =	bl
1	xmi	Xm2 ·	• Xmn_		_ Yn _		_b„_

using the least squares method to determine the estimated parameter value of the regression equation:

bi (X^TX)-¹X^TY inputting the key influencing factors at each time of day under test into the random forest classification model to obtain the class of day under test, to select the corresponding multiple regression forecasting model to complete the load forecasting.

[0071] The method of bus load forecasting of the present invention has the following effects: through the clustering analysis of the historical bus load, the establishment of statistical model and the model matching of the day under test, the change rule of the bus load is effectively mined, and more accurate forecasting results are obtained; meanwhile it is not necessary to train the model before every forecasting to further satisfy the real-time requirement of bus load forecasting.

[0072] In another embodiment of the present invention, a 110 kV bus active load of a city power grid company is used as a forecasting example.

[0073] In the first stage, the historical load data preprocessing section. The load value 24 hours of llOkV bus of a substation is selected as the forecasting object and the load data of one year before the forecasting day is used as training sample. The data format is shown in Table 1:

[0074] Table 1 Bus load data table

Date	The bus load value at each sampling time (Unit: MW)
00:00	01:00	02:00	03:00 ..	.. 20:00	21:00 22:00	23:00
01/01	52.6	49.5	46.7	44.7 ..	.. 49.9	46.1	46.8	46.2
12/31	52.0	48.5	42.9	41.3 ..	.. 69.0	65.3	60.5	59.5

[0075] Each row of data in Table 1 represents a 24-dimensional load sample sequence. There may be several abnormal values in historical load data due to instability of the acquisition systems, namely, abnormal data, which will seriously affect the forecasting accuracy.

[0076] The horizontal comparison method is used to correct abnormal data. Sampling cycle set to 1 hour, a total of 24 data points per day, the change rate of the load at the adjacent time is sequentially calculated in chronological order:

a(d,t) ! L(d,t)-L(d,t~23) _|

L(d,t-23) ^<! L(d,t)-L(d-1,23)! , L(d,23) t = 1,2, , t = 0 ,23 [0077] The maximum value of the change rate of load within 7 days before the sampling point is 15 calculated as the criterion of whether or not the sampling time data is an abnormal value. The judgment formula is as follows:

[0078] If the load data is abnormal value, then the data at the same time of last 3 days is corrected through the correction formula, the correction formula is as follows:

₂₀ L(d,t) = d - l,t) + 2,L(d - 2,t) + 2,L(d - 3,t) _ θ . . , ’ A 0.26 _an(|^3 θ·24 , the correction effect is the best.

[0080] In the second stage, the determination stage of key influencing factors. With reference to Figure 3, a specific implementation process of the gray correlation projection method is shown. Based on the meteorological data of the city where the city power grid company is located, the influencing factors are chosen as follows: maximum temperature (T_max), minimum temperature (T_min), real-time temperature (RT), Average wind speed (AW), relative humidity (RH), average precipitation (AP), day type (DT), and season type (ST), specific parameters as shown in Table 2:

[0081] Table 2 Influencing factors data

Sampling time	Influencing factors
Tmax	Tmin	RT	AW	RH	AP	DT	ST
01/01 00:00	10	-2	2	2.6	51	1.8	1	3
01/01 01:00	10	-2	0	2.6	51	6.5	1	3
06/13 14:00	31	18	30	0.8	64	25.7	2	4
10/01 00:00	25	10	15	1.2	59	19.6	3	1
12/31 23:00	8	-6	-1	2.3	48	0.9	1	3

[0082] In Table 2, the column data corresponding to each influencing factor is the comparison sequence X_;, and the load data at each sampling time is the reference sequence X_o. The day type is assigned {working day, weekend, holiday} = {1, 2, 3}, the season type is assigned {summer, winter, spring, autumn} = {4, 3, 2,1}, other influencing factors data is actually measured value.

[0083] The comparison sequence and the reference sequence are normalized by the formula, and the corresponding values are limited to 0.1 to 1 to obtain the normalized sequence matrix E:

<_e e ^{A C}01 · · · ^C0m y^em ''nm J [0084] Where ^nm represents the normalized value of element x_nm in the sequence matrix X. The number of comparison sequence n=8, m=8760 is the number of elements per sequence, ie, 24 sampling points per day for 365 days.

[0085] The gray correlation coefficient between the comparison sequence and the reference sequence is obtained by the formula of the correlation coefficient, and the gray correlation coefficient judgment matrix F is obtained. The entropy weight method is used to give weights to column vectors of matrix E, except for the first row elements, and the weight vector (w_x,W₂,..., W₈₇₆₀) correlation judgment matrix F i_s obtained in combination with matrix F:

where is obtained. The gray

W1 ·	WJ ·	' ' W8760
W1 ^01 (D ·	·· wA(J) ·	- W8760^01 (876θ)
	·· w^Oi(j) ··	· w876/0i(8760)
W^OnQ) ··	• w^On(j) ·	·· w876(/0n(8760)

4i(j) i is the correlation coefficient between the comparison sequence e_; and the j-th element of

Wreference sequence e₀. ^J is the weight of column vector j, where coefficient between the sequence e_; and the j-th element of e₀.

is the weighted correlation [0086] The gray correlation projection value D_; between the reference sequence X_; and the comparison sequence X_o is calculated as follows:

8760

D; =

Σ ^wj^oi( J) · j=l

8760 /8760 , 1 = 1,2,...,8 |Σ>ΑΟ>}²4Σ>7 [0087] Through the above calculation, the gray correlation projection value of 8 influencing factors is obtained as shown in Table 3:

[0088] Table 3 the gray correlation projection values of influencing factors

Tmax	Tmin	RT	AW	RH	AP	DT	ST
0.87	0.64	0.72	0.28	0.66	0.32	0.55	0.43

[0089] Average wind speed and average precipitation are weakly correlation factors, and the day type and season type are moderate correlation factors.

[0090] Classifying and selecting the maximum temperature, real-time temperature, relative humidity, minimum temperature, and day type as the key influencing factors for the trend of bus load under test.

[0091] In the third stage, the historical load clustering stage of the bus under test. With reference to Figure 4, a specific implementation process of the improved K-means algorithm is shown.

[0092] The determination of initial clustering centroid: There are a total of 365 sample sequences in the _c° historical load sample set; one sample is selected randomly as the first initial centroid vector ¹ , the _c°

Euclidean distance d to ¹ for each of the remaining samples is separately calculated as follows:

t=0

Y C where ^il represents the load value at time (t+1) in the i-th sample sequence, ^ll represents the value of the (t + l)-th element in the initial centroid vector of class 1.

[0093] The sample with the largest d is selected as the second initial centroid vector ² , then the Euclidean _c° distance to ² is calculated separately for each of the remaining samples, then the sample with the _c° largest d as ³ is selected, and so on until k centroids are determined. In the present embodiment, k = 6, 10 the bus load sequence is classified into 6 typical load patterns.

[0094] After 6 initial centroids are determined, the iterative algorithm is started and the samples are assigned. The Euclidean distances of each sample to all centroids are calculated, and each sample will be assigned to the class represented by the nearest centroid:

d(_Xi,c‘) = E(N-c‘_t) , j =1,2,...,6,1=0,1,...,1000 t=0

V, · where 1 is the number of iterations, x_; is the i-th sample, and ^J is the centroid vector of samples of class j before the (l+l)-th iteration.

[0095] The updating of the centroid vector: After each sample assignment, the centroid of load sample of each class needs to be updated. If the number of samples of class j is ^mj , the sample set ts , and then c¹ c¹ the elements ^Jtin the centroid vector ^J are updated as:

Σ.

_

Jl

Hlj

Nt , j =1,2,

Determine whether the termination condition of algorithm is reached. In addition to the maximum number of iterations, define difference tolerance & to determine whether the algorithm terminates. After the updating of centroid, the distance before and after the updating of each class centroid vector is calculated:

Dis¹; =11 C¹; - C¹;’¹ j =1,2,...,6,1=1,2,...1000

Selecting the maximum value^maxDiSj } _va|_ue j_s |_ess difference tolerance^ ^— 0.01 the algorithm terminates, otherwise continuing the iteration updating.

[0096] The above clustering process classifies the bus historical load into 6 load patterns, and the clustering results are shown in Figure 5. Class 1 consists of 82 days, mainly on rest days, Class 2, Class 5 mainly consist of 112 and 89 working days; The difference between these two load values is mainly due to seasonal factors; Class 3 consists of 25 days, the load is much higher than other classes due to extreme weather; Class 6 consists of 31 days, the load curve fluctuates greatly.

[0097] In the fourth stage, the construction stage of random forest classification model. With reference to Figure 6, a specific implementation process of the random forest algorithm is shown.

' I ’O · _

Sampling with replacement 100 training sample sets ¹ , ^{1 —} CA---,^sof S size randomly from the historical sample set S. The data of training sample set ί is shown in Table 4:

[0098] Table 4 training sample set

time	Tmax	RT	RH	Tmin	DT	Clustering result
01/01 00:00	10	2	51	-2	3	Class 5
02/05 09:00	5	3	64	-5	1	Class 2

[0099] 3 influencing factors are sampled randomly as the characteristic attribute for each training sample set.

[0100] 100 sample sets TS; are trained separately to generate the corresponding decision tree Tree;. The characteristic attribute of the decision tree Tree; is remained unchanged throughout the forest growth process. Except for the leaf nodes, each decision tree uses the Gini index of the CART algorithm as the basis for node splitting:

Gini(d) =l-£p²d i=l where d is the current node, is the proportion of target class i. The closer the Gini index is to 0, the higher the purity of the classification, the better the effect.

[0101] For each node, consider all the cases of node splitting with 3 characteristic attributes, and select the attribute with the smallest Gini index as the splitting attribute a. The Gini index of current node d split by splitting attribute a is:

₂₅ Gini(a,d) = p_LGini(d_L) + p_RGini(d_R) where d_L and d_R are the left and right child nodes of node d, and P_L and P_R are the proportion of the left and right child nodes in the parent node.

[0102] Continuing to split from top to bottom according to the above rules, until all nodes are split or marked as leaf nodes, that is, the decision tree growth is complete, there are a total of 100 trees, none of the trees needs pruning operation, leaf nodes of each tree correspond to some clustering result of clustering analysis. When a test sample is input, each decision tree is used to classify the test sample, 100 classification results are obtained and the class with the highest proportion is selected as the class of test sample.

[0103] Combine the 100 decision trees in order to obtain the random forest classification model, so as to establish the mapping relationship between the influencing factors and the clustering results. Figure 7 is a thumbnail view of the decision tree generation in the random forest classification model.

[0104] In the fifth stage, the training stage of the forecasting model. According to the clustering results, the bus load data and the key influencing factors in each load pattern are integrated into the training sample set, as shown in Table 5:

[0105] Table 5 training sample set

Sampling time	Sample sequence	Key Influencing factors	Corresponding load
Tmax	RT	RH	Tmtn	DT
06/13 14:00	XI	31	30	64	18	2	54.2
10/07 19:00	Xm	26	20	58	9	3	63.6

[0106] Where m Is the number of samples in each class of sample set; sample sequence i = l,2,...,m_wh_ % _; »

X;

^m .Where is the influencing factor sequence of sample “u y_i; is the true value of the bus load corresponding to the output.

[0107] Each class of load pattern needs to construct the forecasting model based on its own training sample set, and finally obtains 6 regression equations. The multiple regression equation is given by:

f (X) = b₀ + + b₂?q₂ + · · · + b₅ x_l5 where b₀ is the regression constant term, bpb₂,...,b_n -_{s re}g_ressj_on coefficient, ^'^^X'^is the regression forecasting value. The regression constant terms and coefficients for each regression equation are calculated from equation (22). The results are shown in Table 6:

[0108] Table 6 regression coefficient

	bo	bi	b2	b3	b4	b5
Equation 1	4.9	0.64	0.51	0.36	0.32	0.23

Equation 2	-8.5	0.57	0.60	0.41	0.26	0.18
Equation 3	6.5	0.48	0.66	0.35	0.43	0.20
Equation 4	-3.8	0.68	0.52	0.37	0.35	0.29
Equation 5	3.2	0.51	0.52	0.36	0.30	0.27
Equation 6	1.8	0.71	0.55	0.34	0.39	0.22

[0109] In the sixth stage, the stage of model matching and forecasting: The load data of the bus on June 21, 2016 is selected as the test sample, and the weather data and day type information of the day are as follows:

Time	RT	RH	Time	RT	RH	Time	RT	RH	Tmax	Tmin	DT
00:00	23	52	08:00	24	53	16:00	30	37	30	18	1
01:00	22	56	09:00	26	48	17:00	30	40
02:00	21	59	10:00	28	44	18:00	29	43
03:00	20	65	11:00	29	35	19:00	28	46
04:00	20	64	12:00	30	36	20:00	26	51
05:00	19	68	13:00	30	37	21:00	25	55
06:00	18	70	14:00	30	38	22:00	24	57
07:00	21	62	15:00	30	37	23:00	23	58
0110] Then, the corresponding influencing :	'actors ol	24 hours of the c	ay are input into the random forest

classification model separately, and result that the day under test belongs to load pattern of class 2 is obtained, and the model 2 (regression equation 2) is selected to complete the load forecasting, the results are shown in Figure 8. The maximum relative error of the forecasting result obtained by the method of the present invention is 2.89%, the minimum relative error is 0.32%, and the average relative error is only 1.29%; while the maximum relative error of the traditional multiple regression forecasting method is

3.06%, the minimum relative error is 0.28%, and the average relative error is 1.38%. Therefore, the accuracy of the forecasting method adopted in the present invention is obviously improved, and has wide application prospect.

[0111] It should be understood by those skilled in the art that the foregoing discussion of any embodiment is intended to be illustrative and not intended to suggest that the scope of the present disclosure, including the claims, be limited to such examples; and in the context of the present invention, the technical features of the various embodiments may also be combined, the steps may be implemented in any order, and there are many other variations of the various aspects of the present invention as described above, which are not provided in the details for the sake of brevity.

[0112] In addition, to simplify the description and discussion, and in order not to obscure the present invention, well-known power/ground connections to integrated circuit (IC) chips and other components may or may not be shown in the provided drawings. In addition, devices may be shown in block diagram form in order to avoid obscuring the present invention, and this also takes into account the fact that the details regarding the embodiments of these block diagram devices are highly dependent on the platform on which the present invention will be implemented (that is, these details should be fully within the purview of those skilled in the art). In the event that specific details (e.g., circuits) have been set forth to describe exemplary embodiments of the present invention, it will be apparent to those skilled in the art that the present invention may be implemented without these specific details or with variations in these specific details. Accordingly, these descriptions are to be regarded in an illustrative rather than a restrictive sense.

[0113] Although the present invention has been described in connection with specific embodiments thereof, many alternatives, modifications and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may use the discussed embodiments.

[0114] It is intended that the embodiments of the present invention cover all such alternatives, modifications and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalents, improvements, and the like within the spirit and principles of the present invention are intended to be embraced within the scope of the present invention.

Claims (10)

1. What is claimed is:

   1. A method of bus load forecasting, comprising:

   inputting key influencing factor values of the bus under test as characteristic vectors into a pre-constructed classification model to obtain the class of day under test; the pre-constructed classification model is a result of clustering analysis based on the key influencing factors and the historical load data of the bus under test; establishing a mapping relationship between the key influencing factors and bus load;

   selecting a corresponding multiple regression forecasting model from the multiple regression forecasting models obtained from the pre-training according to the class of day under test; forecasting the bus load under test, wherein the multiple regression forecasting models obtained from pre-training are a result of clustering analysis based on the historical load data of the bus under test;

   selecting the corresponding historical load data and influencing factor values for different typical load patterns as a forecasting model obtained from sample data training.
2. 2. The method of bus load forecasting according to claim 1, wherein the classification model is constructed by:

   using the horizontal comparison method to select abnormal data from the historical bus load data of the bus under test, and correcting the abnormal data according to a preset algorithm to obtain the corrected load data;

   using the gray correlation projection method to calculate the correlation degree between the influencing factors and the historical load data in combination with the corrected load data and the load-related influencing factors;

   selecting a preset number of influencing factors corresponding to the correlation degree from large to small as the key influencing factors;

   clustering analyzing the historical load data of bus under test according to the preset K-means clustering algorithm to obtain different typical load patterns;

   according to the key influencing factors and the result of clustering analysis, using the random forest algorithm to construct the classification model, and establishing the mapping relationship between the key influencing factors and various bus loads.
3. 3. The method of bus load forecasting according to claim 2, wherein the step of selecting the abnormal data from the historical bus load data of the bus under test by the horizontal comparison method comprises:

   cleaning the historical load data of the bus under test by the horizontal comparison method to identify the abnormal data; wherein, the horizontal comparison method using the change rate of the adjacent point load as a criterion for determining whether or not abnormal data is present, the change rate of the load is calculated as follows:

   L(d.t)-L(d.t-1) _t=1,

   L(d,t —1) _|L(d.t)-L(d-l,_q)|

   L(d,q) where _tjj_{e c}h_ang_{e ra}te of the bus load under test at time t of d day, and L(d,t) -_s ^_us load value under test at time t of d day;

   whether the sampling point data is abnormal data depends on:

   (d,t),theloadat time tis normal value ,a(d,t)>a₁’j_iax (d,t),theloadat time tis abnormal value and

   OL (d,t) _{i t})-_Ud , _{= 1>2}

   L(d — i,t -1) _max ud-ro-ud-'-tq) ' L(d—i,q) ' ,n,t = 1,2,,...,q .,n,t = 0

   10 where ^amax μ _tjj_{e max}i_mum value of change rate of load at time t of d day during the first n days, i represents the number of days from the current d day, ^— ~ k> is the sampling point data at time (t-1) of (d-i) day, (q+1) sampling points per day; using the last m days of data to correct the abnormal data;

   comparing the load value at the current time with the load value at the previous time in chronological

   15 order, correcting the abnormal data once the abnormal data detected, and using the corrected data as the comparison value of load value at the next time, completing the load data correction.
4. 4. The method of bus load forecasting according to claim 2, wherein the step of correcting the abnormal data by the preset algorithm comprises:

   20 correcting the selected abnormal data according to the following algorithm, the specific algorithm formula is:

   L(d,t) = 2jL(d — l,t) + 2₂L(d — 2,t) + ... + 2_mL(d — m,t) ^< λ.=β(ΐ-βγ~\ e(o,i) m

   ΣΛ=ι j=i λ

   where ^m is the weight coefficient, which is used to represent the influencing degree of the load at time t of (d-m) day on the load at time t of d day; P is the smoothing coefficient, ^{— ΠΊ}Α ) j_s the bus load value under test at time t of (d-m) day.
5. 5 5. The method of bus load forecasting according to claim 2, wherein the step of using the gray correlation projection method to calculate the correlation degree between the influencing factors and the historical load data and selecting a preset number of influencing factors corresponding to the correlation degree from large to small as the key influencing factors comprises:

   first, determining a reference sequence and a comparison sequence, where the reference sequence is the

   10 corrected load data: X° ^— i-Nn’Nk’···’Arm) _; the comparison sequence is the load-related influencing factors: X ^— > N2’---’ Nm) ¹ — 1,2,...,n . obtaining the sequence matrix of the reference sequence and the comparison sequence:

   X_o ^{_ <X}O1 · • X, = ; · X„_ Λι ·

   where m represents that each sequence has m elements, and n is the number of comparison sequences; X_o 15 is the load sequence or reference sequence corresponding to the corrected load data, X; is the i-th comparison sequence;

   normalizing the matrix X as follows:

   ^ea = max i 1.2.....,,, ^Xii ¹¹¹¹¹¹ i 1.2....,mNj

   -(1-0.1) + 0.1 where mm and max are the minimum and maximum values of the row elements of

   X··

   20 ^1J respectively, and ^1J is the normalized value of the element Xy of the matrix X, the value of limited to 0.1 to 1;

   the normalized matrix X is as follows:

   is

   -Ό1

   V^enl ^e0m e , nm / p £1=12 where ⁰ and ' ’ ’ ’ sequence;

   are normalized reference sequence and normalized comparison the correlation coefficient of the k-th element of normalized comparison sequence e_; and normalized reference sequence e₀ is calculated as follows:

   mirr min, I — e₀, I +z?max, max, I — e₀, I
6. 6,(k) =--j .j oj e-' ^{J OJ} , i = 1,2,...,n, j = 1,2,...,

   1¾ -e_ok l+pmax,max, le„ -e_0J I , min. min. I e. - e_n. I . , . . , , ,, where ^{J J 1} is the minimum difference between the two levels, max. max. I e.. — e„. I . , . ,... , ¹ j u is the maximum dinerence between two levels; v _1S the identification coefficient;

   the gray correlation coefficient judgment matrix F is calculated as follows:

   F =

   £oo(D · ·· ^oo(j) · ·· +oo<^m) r i • 1 ·· ' ¹ Ί ^oi CD · ^oi(j) · ·· £oi (m) ^oid) · • ^oi(j) · • ^01 (m) CD — ^(1) · £o„(D · ·· £o„(j) · ·· ^on(^m) y Jo„(D · ·· £o„(j) ·

   / (i) where >θθ^{v J7} is the correlation coefficient between the normalized reference sequence e₀ and its own / (i) j-th element, and the value of -^{(K) J 7} is I;

   using the entropy weight method to give weights to column vectors of matrix E, except for the first row elements, the formula is as follows:

   E_: =--— ΣΡϋ^1ηΡίι =—Γ7—Σ<Αι /ΣΧ)¹¹¹^ ^ZZ^eiP’ j =l’2,...m ^J ln(m) y 1-E _Wj = ln(m) _i=1 ^J—, j =l,2,...,m ^m_E^Ei j=l where Ej is the information entropy of the data of the j -th column, Py is the proportion of the value of the element ey in the column elements, Wj is the weight of the j-th column data, and the weight vector W = (w₁,...,w_j,...,w_m) the weighted gray correlation coefficient judgment matrix is as follows:

   ^W1 w/oiQ) ... _Wj · ··· w^₀₁(j) · ·· ^Wm ·· wX(m) w/oib) ' ··· w^_Oi(j) ·· '· wj_oi(m) W/onQ) · ·· w^_On(j) · ·· wJ_On(m) Wj^oi(j) is the weighted correlation cot

   where ^e(b calculating the gray correlation projection value, the formula is as follows:

   D.

   E^Wi^( j) · ^Wii i=l

   E(^wA(j))²^£(^wi)² where D_; is the correlation projection value between the comparison sequence X_; and the bus load under test, D_; is between 0 and 1, the closer the value is to 1, the greater correlation degree; the closer the value is to 0, the weaker the correlation degree;

   selecting the M factors which are the top of the projection values from large to small as the key 10 influencing factors.

   6. The method of bus load forecasting according to claim 2, wherein the step of clustering analyzing the historical load data of bus under test according to the preset K-means clustering algorithm comprises:

   using the corrected load data as a sample set, setting the sample set as ^— ίΚ’···’ K’···’ ^xm} , load x=fx x x ) C°=ic°c° c° c°) sequence as ' ΐΛο’-’ it’···’ in J, the initial centroid set as ¹ ’’ ¹' G’ ^{k J}, selecting one

   15 sample randomly as the first centroid, calculating the Euclidean distance to the first centroid separately for each of the remaining samples:

   t=0 where m is the number of load sequences in the sample set, n represents that each load sequence has (n+1) data points, k is the number of centroids; ^X'¹ represents the load value at time (t+1) in the i-th load _c°

   20 sequence, ¹¹ represents the value of the (t + l)-th element in the first centroid vector;

   selecting the sample with the largest Euclidean distance as the second centroid, then calculating the Euclidean distance to the second centroid separately for each of the remaining samples, then selecting the sample with the largest Euclidean distance as the third centroid, and so on until k centroids are determined;

   calculating the Euclidean distance of each sample to all centroids separately, selecting the centroid with the smallest Euclidean distance from the current sample as the class to which the current sample belongs:

   n ² ά<Λ,φ = Σ<χ-φ , j =1,2,...,k,l =0,1,...,h t=0 where 1 is the number of iterations, x_; is the i-th sample, and ^J is the centroid vector of samples of class j before the (l+l)-th iteration;

   updating the centroid of each class after all samples classified; let the number of samples of class j be ^mj and the sample set be , recalculating the centroid vector of class j as follows:

   ^Cj ⁼ (^CjO’^Cjl’-’^Cjt’---’^Cjn) where calculation formula of (t+l)-th component is:

   Σ.

   sAt j =1,2, ,k determining whether or not the preset termination condition is reached, continuing the iteration calculation if not reached, until the historical load of the bus under test is classified as a typical load pattern of class k.
7. 7. The method of bus load forecasting according to claim 6, wherein the step of determining whether or not the preset termination condition is reached comprises:

   after the completion of each iteration, calculating the distance interval between the various centroids before and after updating according to the following formula:

   Dis¹. =11 c¹. — c¹.^-1 II, j=l,2,...,k ! , .... . , _£ , . _£ , ^{J J} j ·' , DiSj represents the centroid distance interval of class j after the completion of 1-th iteration;

   selecting the maximum value ^max{Dis. }, _maxj_{mum va}|_ue j_s |_ess than the difference tolerance , the algorithm terminates, otherwise reclassifying the samples to continue the iteration updating.
8. 8. The method of bus load forecasting according to claim 2, wherein the step of using the random forest algorithm to construct the classification model comprises:

   ' I 'O · _ -| sampling with replacement s training sample sets ¹ , ^{1 —} TA···,^s _of the same size randomly from the historical load data sets;

   sampling N influencing factors randomly as the characteristic attribute of each training sample set; wherein N<M;

   training s sample sets TS; separately to generate the corresponding decision tree Tree;; remaining the characteristic attribute of the decision tree unchanged throughout the forest growth process;

   except for the leaf nodes, each decision tree using the Gini index of the CART algorithm as the basis for node splitting:

   Gini(d) = 1-£r 'i,d i=l for each node, selecting the attribute with the smallest Gini index as the splitting attribute a; the Gini index of current node d split by splitting attribute a is:

   Gini(a,d) = p_LGini(d_L) + p_RGini(d_R) where d_L and d_R are the left and right child nodes of node d, and P_L and P_R are the proportion of the left and right child nodes in the parent node;

   continuing to split from top to bottom according to the above rules, until all nodes are split or marked as leaf nodes, that is, the decision tree growth is complete, there are a total of s trees, none of the trees needs pruning operation, leaf nodes of each tree correspond to some clustering result of clustering analysis; combining the s decision trees in order to obtain the random forest classification model;

   when a test sample input, using each decision tree to classify the test sample, obtaining s classification results and selecting the class with the highest proportion as the class of test sample.
9. 9. The method of bus load forecasting according to claim 1, comprising establishing the multiple regression model by the multiple linear regression method, and determining the regression parameters by the least square method.
10. 10. The method of bus load forecasting according to claim 9, wherein the regression equation is:

    f 00 = b₀ +b^i +b₂^₂ + · · · +b_A where b₀ is the regression constant term, bpb₂,...,b_n -_{s re}g_ressj_on coefficient, Un) i_s the regression forecasting value; the i-th training sample in each class of sample set is^¹ >^x,, Y,_wh_ere ^={^1,^2,...,xj i = l,2,...,m , m is the number of samples in each class of load sample set; i x y.

    ', ⁱⁿ is the value of the n-th influencing factor, ^J1 is the true value is the input sequence of sample ¹, ⁱⁿ ii of the bus load corresponding to the output;

    the input matrix X, the output matrix Y, and the coefficient matrix B are as follows:

    ^X11 ^X12 ·' • ^Xln ^2 ’ ' '· ^xmi ^Xm2 · ·· ^Xmn

    ^_Yi ’ ’V , Y = y₂ , B = ^bl _ Yn _ A_

    using the least squares method to determine the estimated parameter value of the regression equation:

    (X^TX)-¹X^TY

    DRAWINGS

    Input key influencing factor values oflhe bus under test as eharaetcristie vectors into a pre-eonstrueted classification model to obtain the class of day under test

    Select a corresponding multiple regression forecasting model front the multiple regression forecasting models obtained from the pre-training according to the class of the day under test and forecast the bus load under test

    101

    102

    FIG.l

    Start

    201

    Obtain and store historical load data of the bus under test and load-related influencing

    Use the horizontal comparison method to correct abnormal data in the historical load

    202

    Use the gray correlation projection method to determine key influencing factors

    203

    Clustering analyzing the historical load data of bus under test according to the preset

    204

    Use the random forest algorithm to construct the classification model, and establish

    205

    Select the corresponding historical load data and influencing factor values for different t\picul load patterns, and train them to obtain different multiple regression forecasting models

    206

    Input key influencing factor values of the bus of day under test into the classification ni«.»del to obtain the class of day uniler test, ;in<.l select the corresponding multiple..................

    207

    End

    FIG.2 ί

    Start

    301

    Use the corrected historical load data as a reference sequence, and use the influencing factors as a comparison sequence

    302

    Normalize the elements of the reference sequence and the comparison sequence to obtain ; the normalized sequence matrix E

    Calculate the correlation coefficient between the reference sequence and corresponding i element of each comparison sequence to obtain the gray correlation coefficient judgment ' matrix F

    - 303

    304

    Use the entropy weight method to give weights to column vectors of matrix E

    305

    Obtain a weighted gray correlation judgment matrix from the matrix F and weight vector [ calculation

    306

    Calculate the gray correlation projection value of each influencing factor on historical load

    Select preset number of factors with the largest projection values as the key influencing factors

    - 307

    End

    FIG.3

    401

    Start

    402

    - 403

    404

    Obtain 6 initial clustering centroids, each clustering centroid corresponds

    Determine whether 1 is greater than a preset number threshold or whether the maximum interval before and after each class of centroid updating is less than the difference tolerance

    406

    408

    End

    FIG.4

    FIG.5 sample set S

    Sample with replacement training sample set TS1

    Sample with replacement training sample set TS100

    Sample 3 influencing factors randomly as the characteristic attribute

    Sample 3 influencing factors randomly as the characteristic attribute calculate Gini index 'or current each node

    Mark all child nodes that belong to the same class as leaf nodes (except for the leaf nodes), and select the splitting^ ttrihnte

    Use the splitting attribute to binary split the current node into sub-nodes

    Treel generate decision tree Treel

    Mark all child nodes that belong to the same class as leaf nodes calculate Gini index for current each node (except for the leaf nodes), and select the spl itt ing^at tribute

    Use the splitting attribute to binary split the current node into sub-nodes random forest classification model

    FIG.6

    Treel

    TreelOO

    FIG.7

    FIG.8