CN103748994B - A kind of load on host computers multisequencing combination forecasting method based on WAVELET PACKET DECOMPOSITION - Google Patents

Abstract

The present invention relates to a kind of load on host computers multisequencing combination forecasting method based on WAVELET PACKET DECOMPOSITION, belong to Computer Applied Technology field.The present invention is from non-linear, non-stationary information processing angle, by constructing many sequences, and utilize the several data relation that each sequence is relevant to load, by wavelet packet-SVR model and the incompatible prediction load on host computers of AR model group, to improve host load prediction precision.The method is suitable for the host load prediction to emphasis period or emphasis moment point.

Description

A kind of load on host computers multisequencing combination forecasting method based on WAVELET PACKET DECOMPOSITION

Technical field

The present invention relates to a kind of load on host computers multisequencing combination forecasting method based on WAVELET PACKET DECOMPOSITION, belong to Computer Applied Technology field.

Background technology

Conventionally, in distributed network environment, load estimation is divided into network load estimation (network load prediction) and Host Based load estimation (host load prediction).Host load prediction is mainly used in realizing the dynamic load leveling in distributed/parallel environment at present, and the HIDS in information security field, realizes the load on host computers abnormality detection based on prediction.For dynamic load leveling, host load prediction precision is higher, the distribution accuracy between each computer node is higher to task for balance sysmte, task because of the possibility of distributing the unbalanced situation that causes sub-distribution again to occur just lower, thereby effectively improve the efficiency that completes of task.For load on host computers abnormality detection, host load prediction precision is higher, abnormality detection system carves whether occur that abnormal judgement is just more accurate and timely at a time for load on host computers, thereby can more early to safety manager, send abnormal alarm, make denial of service etc. cause the attack that occurs in of the abnormal network attack of load on host computers to be just effectively controlled in early days, security incident is eliminated in bud, avoided network attack to cause damage to the network user.Conventionally, the every reduction by 1 of the value of the mean square error of host load prediction, just can be by the time advance of finding load abnormal about 1 second, and this Timing Advance is extremely valuable for the attack of taking precautions against Cyberspace.Therefore the core that how improving host load prediction precision becomes in above two applications studies a question.

The number of the sequence of using during according to prediction, can be divided into existing host load prediction unique sequence prediction and multisequencing combined prediction.

About unique sequence, predict, Yao Shuping etc. are at document < < Server Load Prediction Based on Wavelet Packet and Support Vector Regression > > (Proceedings of the 2006International Conference on Computational Intelligence and Security, 2006, Vol (2): 1016-1019) in Wavelet Packet Technique has been introduced to host load prediction, the host load prediction method of proposition based on wavelet packet-SVR (Wavelet Packet-SVR), it is the unique sequence Forecasting Methodology that a kind of precision of prediction is higher.

To seasonal effect in time series is olation as shown in Figure 1, its low frequency signal and high-frequency signal to upper strata decomposes Wavelet Packet Technique simultaneously.Under one-dimensional case, what it produced is a complete binary tree.This decomposition of wavelet packet is more careful, has improved the time frequency resolution of signal decomposition, is more beneficial to prediction.

In Fig. 1, node (0,0) represents without the original series decomposing.The node (i, i) of other layer represents j the sequence branch obtaining after i layer decomposes.As (1,0), (1,1) represent respectively the Liang Ge sequence branch that original series obtains after ground floor WAVELET PACKET DECOMPOSITION.

Prediction based on WAVELET PACKET DECOMPOSITION has multiple combination mode, and not certain node with the bottom, can combine the node of different levels and predict, as long as the direct sum in the corresponding space of selected node just covers primary signal space, does not overlap each other again.As selected node ((3,0), (3,1), (3,2), (3,3), (3,4), (3,5), (3,6), (3,7)) decomposition result of expression to original series, also can select node ((3,0), (3,1), (2,1), (1,1)) decomposition result of expression to original series.

Host load prediction side's ratio juris based on wavelet packet-SVR as shown in Figure 2.

Wherein,

for original series, be the chronological arrangement of top n load value of prediction moment t, be expressed as $H\left(t\right)={\left\{h_k\right\}}_{k=1}^N.$

Consider one-step prediction, ask h _n+1predicted value, be designated as

First H (t) is carried out to three layers of WAVELET PACKET DECOMPOSITION.From decomposition tree, read node (3,0), (3,1), (3,2), (3,3), the WAVELET PACKET DECOMPOSITION coefficient of (1,1), form 5 signal components, on three layers of wavelet decomposition basis, second layer detail signal (node (2,1)) is further decomposed.Consider ground floor detail signal (node (1,1)) be the random load component in original load signal, be the incremental loading being caused by incident, general ratio shared in whole load is very little, so do not need it further to be decomposed in prediction again.5 signal components are expressed as follows:

For the length of Shi Ge branch remains unchanged, each branch is carried out to single reconstruct, 5 branches after reconstruct are consistent with the length of original series, that is:

${\tilde{H}}_{11}=\{{\tilde{h}}_{11,i},1\&le;i\&le;N\}$

The forecast model of respectively 5 branches being set up as shown in Figure 2 carries out one-step prediction, obtains 5 predicted values, synthesized

that is:

${\hat{h}}_{N+1}={\hat{\tilde{h}}}_{30,N+1}+{\hat{\tilde{h}}}_{31,N+1}+{\hat{\tilde{h}}}_{32,N+1}+{\hat{\tilde{h}}}_{33,N+1}+{\hat{\tilde{h}}}_{11,N+1}---\left(3\right)$

In document, Simulation results shows: the host load prediction method precision of prediction based on wavelet packet-SVR is the Forecasting Methodology in host load prediction and the synthetic technology research > > (the journal .2007 of Beijing Institute of Technology, 27 (1): 42-45) based on non-stationary series higher than the people's such as Yao Shuping document < <.

But if need to the load in emphasis period, emphasis moment be monitored, the unique sequence method that above document proposes is also weak, multisequencing combination forecasting method is more effective.

Multisequencing combined prediction, refers to and utilizes the data that gather, and the structure sequence relevant to prediction from different perspectives, improves precision by the predicted value that combines each sequence.

The formal definitions of multisequencing combined prediction is:

If a certain forecasting object F is y in the actual value of a certain period _t(t=1,2 ..., n), N the time series of different attribute latent structure based on forecasting object F, is designated as S ₁(t), S ₂(t) ..., S _n(t), based on each unique sequence, predict respectively, obtain N predicted value

the final predicted value after Combination Forecasting

for:

${\hat{y}}_t=\mathrm{\&xi;}({\hat{y}}_t\left(1\right),{\hat{y}}_t\left(2\right),\&CenterDot;\&CenterDot;\&CenterDot;,{\hat{y}}_t\left(N\right))---\left(4\right)$

Wherein, ξ is linearity or nonlinear function.

Yao Shuping etc. propose a kind of based on small echo, neural net and autoregressive multisequencing combination forecasting method in the research > > of the web traffic combination forecasting method based on small echo (the journal .2006 of China Mining University, 35 (4): 540-544) at document < <.In literary composition, first web traffic is configured to two correlated serieses: historical series and similar value sequence.Then the similar value sequence AR model with steady feature is predicted; To having embodied, web traffic is non-linear, the historical series of non-stationary property, after wavelet decomposition and single reconstruct, adopts respectively neural net and autoregressive model prediction for each branch feature, adopts small echo-AR-BP model to predict.Finally combine predicting the outcome of two sequences and obtain final predicted value.Theory analysis and experiment show: owing to having increased similar value information (being the sequence of continuous some days synchronization load on host computers formations), document institute construction method is higher than unique sequence method precision of prediction.

But the weak point of the method is, for historical series, what the each branch after wavelet decomposition adopted is BP neural net and AR prediction, and the Chaotic time series forecasting research > > (Chinese University of Science and Technology thesis for the doctorate .2003) of document < < based on SVMs and wavelet theory explicitly points out, BP neural net is compared and is had open defect with SVMs (SVM) method, as learn to be easily absorbed in local minimum point, easily cause over-fitting and affect generalization ability of network etc.Above defect becomes the main cause of impact prediction precision.When SVM method is used for predicting, be called support vector regression (SVR) method.

Patent of invention < < host load prediction method > > (national defence patent of invention combining based on multisequencing of the propositions such as Hu Chang shakes, application number 200910121385.6) in the more much higher combined sequence Forecasting Methodology of a kind of precision of prediction has been proposed, principle is as shown in Figure 3.First the method utilizes data on flows structure historical series and the similar value sequence obtained.Then, for historical series, adopt small echo-AR-SVR-MA model to predict, obtain the predicted value 1 in t moment; For similar value sequence, carry out AR prediction, obtain t moment predicted value 2.Finally, calculate the mean value of two sequences predicted value, obtain final predicted value.

The defect of this Forecasting Methodology is that what historical series was carried out is wavelet decomposition, and wavelet decomposition only continues to decompose to the low frequency signal on upper strata each time, high-frequency signal is no longer processed, therefore be not as careful as WAVELET PACKET DECOMPOSITION, affect the time frequency resolution of signal decomposition, be unfavorable for prediction.

Summary of the invention

The object of the invention is to propose a kind of load on host computers multisequencing combination forecasting method based on WAVELET PACKET DECOMPOSITION in order further to improve host load prediction precision.The present invention is from non-linear, non-stationary information processing angle, by constructing many sequences, and utilize the several data relation that each sequence is relevant to load, by wavelet packet-SVR model and the incompatible prediction load on host computers of AR model group, to improve host load prediction precision.The method is suitable for the host load prediction to emphasis period or emphasis moment point.

A kind of load on host computers multisequencing combination forecasting method general frame design cycle based on WAVELET PACKET DECOMPOSITION of the present invention as shown in Figure 4.Specific implementation step is as follows:

Step 1, structure historical series and similar value sequence

First utilize data on flows structure historical series and the similar value sequence obtained, the structure historical series of introducing in the web traffic combination forecasting method research > > of concrete building method employing document < < based on small echo and the method for similar value sequence are constructed.

Step 2, for historical series, adopt wavelet packet-SVR model to predict, obtain the predicted value in t moment, be called predicted value 1

On the basis of step 1 structure historical series and similar value sequence, historical series is carried out to wavelet packet-SVR prediction, obtain the predicted value in t moment.Wavelet packet-SVR model that the present invention uses is the load estimation model based on wavelet packet-SVR of introducing in document < < Server Load Prediction Based on Wavelet Packet and Support Vector Regression > >.

Step 3, for similar value sequence, carry out AR prediction, obtain t moment predicted value, be called predicted value 2

On the basis of step 1 structure historical series and similar value sequence, similar value sequence is carried out to AR model prediction, obtain the predicted value 2 in t moment.The AR model that the present invention adopts is the similar value sequence A R forecast model using in the web traffic combination forecasting method research > > of document < < based on small echo.

The mean value of step 4, calculating two sequences predicted value, obtains final predicted value

On the basis of step 2 and step 3, calculate the mean value of two sequences predicted value, obtain final predicted value.

Beneficial effect

The present invention is directed to the demand of emphasis period, emphasis moment host load prediction, from non-linear, non-stationary information processing angle, traditional combined prediction thought is generalized to the combination of multisequencing, by constructing many sequences, and wavelet packet-SVR model is introduced to combined prediction, effectively raise host load prediction precision.

Accompanying drawing explanation

Fig. 1 is the WAVELET PACKET DECOMPOSITION principle schematic of prior art;

Fig. 2 is the load estimation illustraton of model based on wavelet packet-SVR of prior art;

Fig. 3 is a kind of multisequencing combined prediction schematic diagram of prior art;

Fig. 4 is a kind of load on host computers multisequencing combination forecasting method general frame design flow diagram based on WAVELET PACKET DECOMPOSITION of the present invention.

Embodiment

According to technique scheme, below in conjunction with embodiment, the present invention is described in detail.

Flow to certain unit Web server is collected, and the signal period is 60s, has 1440 data every day, collects altogether 61 days 87840 datas on flows.Front 60 day data that utilization is collected are predicted the flow of the 61st day 8:00-9:00.

Prediction adopts single step roll mode.Step is:

Step 1, structure historical series and similar value sequence

First utilize data on flows structure historical series and the similar value sequence obtained, the structure historical series of introducing in the web traffic combination forecasting method research > > of concrete building method employing document < < based on small echo and the method for similar value sequence are constructed.

Concrete building method is as follows:

The the 1. step: original load seasonal effect in time series expression formula as shown in Equation 5:

X＝x(t)＝{x(1)，x(2)，x(3)，…，x(k)，…}?(5)

The the 2. step: according to load time sequence expression formula, structure historical series and the similar value sequence relevant to load on host computers respectively.

If the two sequences relevant to moment t to be predicted is H (t), S (t):

H(t)＝{x(t-i)，N≥i≥1}?(6)

S(t)＝{x(t-j×T)，M≥j≥1}?(7)

The length that wherein T is one-period;

H (t) represents the chronological arrangement of top n load data in t moment, is called historical series;

S (t) represents the observation station load sequence corresponding with t in the adjacent M cycle before in t moment, is called similar value sequence.

The the 3. step: each element in two sequences is renumberd respectively, and H (t) and S (t) can be expressed as $H\left(t\right)={\left\{h_k\right\}}_{k=1}^N$ With $S\left(t\right)={\left\{s_j\right\}}_{j=1}^M.$

Step 2, for historical series, adopt wavelet packet-SVR model to predict, obtain the predicted value 1 in t moment

On the basis of step 1 structure historical series and similar value sequence, historical series is carried out to wavelet packet-SVR prediction, obtain the predicted value 1 in t moment.Wavelet packet-SVR model that the present invention uses is the load estimation model based on wavelet packet-SVR of introducing in document < < Server Load Prediction Based on Wavelet Packet and Support Vector Regression > >.

Specific implementation step is as follows:

The the 1. step: historical series is carried out to WAVELET PACKET DECOMPOSITION and reconstruct.

First select a female small echo, and determine and decompose number of plies L.The selected number of plies L=3 that decomposes of the present embodiment.

Original series is carried out to WAVELET PACKET DECOMPOSITION:

Decomposable process is as shown in table 1.Wherein represent total frequency band space that original series occupies, represent the n sub spaces on j yardstick.

The wavelet packet in table 1 original series space is divided

As shown in Table 1:

$U_{j-1}^n=U_j^{2n}\&CirclePlus;U_j^{2n+1}---\left(8\right)$

The sequence d of corresponding each subspace ^{j, n}represent, utilize wavelet decomposition algorithm:

$d_l^{j,2n}=\underset{k}{\mathrm{\&Sigma;}}{h}_{k-2l}{d}_k^{j-1,n}---\left(9\right)$

$d_i^{j,2n+1}=\underset{k}{\mathrm{\&Sigma;}}{g}_{k-2l}{d}_k^{j-1,n}---\left(10\right)$

Wavelet packet space layer-by-layer is segmented.Wherein h _n, g _nrespectively low pass, high pass resolution filter coefficient.

Decompose while proceeding to L layer, obtain 2 ^lindividual signal component.Utilize wavelet package reconstruction algorithm:

$d_l^{j-1,n}=\underset{k}{\mathrm{\&Sigma;}}[{\tilde{h}}_{l-2k}{d}_k^{j,2n}+{\tilde{g}}_{l-2k}{d}_k^{j,2n+1}]---\left(11\right)$

To each signal component reconstruct, obtain final signal component.Wherein

respectively low pass, high pass reconfigurable filter coefficient.

The the 2. step: the SVR prediction of each branch signal

With

represent any one branch signal, by the value d in the n before the t moment value prediction t moment _tproblem can be expressed as the problem of finding following corresponding relation F:

${\hat{d}}_t=F(d_{t-n},d_{t-n+1},\&CenterDot;\&CenterDot;\&CenterDot;,d_{t-1})---\left(12\right)$

Concerning the training of regression model, by N training sample, just can build N-n training sample pair, i.e. input (d ₁, d ₂..., d _n), the corresponding d that is output as _n+1; Input (d ₂, d ₃..., d _n+1), the corresponding d that is output as _n+2; The rest may be inferred.

Specific algorithm step is:

A. data are prepared.The data that collect are divided into two parts, and a part, as training data, is designated as

another part, as test data, is designated as

first training data is organized as to learning sample to (x take n as step-length _i, y _i), as shown in table 2.

Table 2 SVR forecast model learning sample

B. choose suitable supporting vector machine model.

C. determine and return step-length n.

D. according to definite support vector topological structure, utilize training dataset

training pattern.

E. utilize test data set

in data carry out Single-step Prediction.

F. according to predicting the outcome, carry out model evaluation.Calculate precision of prediction, if precision meets predefined threshold value, finish algorithm, otherwise adjust parameter, go back to steps d.

The the 3. step: each branch prediction value synthetic

Synthetic method in the synthetic employing document < < Server Load Prediction Based on Wavelet Packet and Support Vector Regression > > of the each branch prediction value of the present invention, directly by corresponding the predicted value of each branch, be added i.e. formula (3)

Step 3, for similar value sequence, carry out AR prediction, obtain the predicted value 2 in t moment

On the basis of step 1 structure historical series and similar value sequence, similar value sequence is carried out to AR model prediction, obtain the predicted value 2 in t moment.The AR model that the present invention adopts is the similar value sequence A R forecast model using in the web traffic combination forecasting method research > > of document < < based on small echo.

Concrete operation step:

The the 1. step: the zero-mean of original series.

Random data utilization AR model is analyzed and must be met stationary time series hypothesis, original series can be carried out to zero-mean processing, establishes the mean value that μ is S (t), zero-mean { s _jobtain ordered series of numbers { y _t}:

y _t＝s _t-μ?(13)

{ y _tit is approximate stationary time series.

The the 2. step: determine model order p.

The mathematic(al) representation of AR model is:

Z _t＝φ ₁Z _t-1+φ ₂Z _t-2+…+φ _pZ _t-p+a _t?(14)

Wherein, p is the exponent number of AR model, φ _ifor the coefficient of model, be solve for parameter, a _tit is white noise.

Employing from low order to high-order one by one the mode of matching determine the rank of model, from p=1, start to find one by one and can make sequence prediction curve and the good exponent number of actual curve matching.Be specially:

Utilize Yule-Walker equation group:

$\mathrm{\&Phi;}=P_p^{-1}\mathrm{\&rho;}---\left(15\right)$

Wherein, Φ is coefficient vector, and ρ is auto-correlation vector, and Pp is autocorrelation matrix.

$\mathrm{\&Phi;}=\left[\begin{array}{c}{\mathrm{\&phi;}}_1\\ {\mathrm{\&phi;}}_2\\ .\\ .\\ .\\ {\mathrm{\&phi;}}_p\end{array}\right]$ $\mathrm{\&rho;}=\left[\begin{array}{c}{\mathrm{\&rho;}}_1\\ {\mathrm{\&rho;}}_2\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ {\mathrm{\&rho;}}_p\end{array}\right]$ $P_p=\left[\begin{array}{cccc}1& {\mathrm{\&rho;}}_1& \&CenterDot;\&CenterDot;\&CenterDot;& {\mathrm{\&rho;}}_{p-1}\\ {\mathrm{\&rho;}}_1& 1& & {\mathrm{\&rho;}}_{p-2}\\ \&CenterDot;& \&CenterDot;& & \&CenterDot;\\ \&CenterDot;& \&CenterDot;& & \&CenterDot;\\ \&CenterDot;& \&CenterDot;& & \&CenterDot;\\ {\mathrm{\&rho;}}_{p-1}& {\mathrm{\&rho;}}_{p-2}& \&CenterDot;\&CenterDot;\&CenterDot;& 1\end{array}\right]---\left(16\right)$

To φ _icarry out valuation, then obtain sequence { y _tapproximate prediction curve, and therefrom select a good model of matching.

The the 3. step: for selected exponent number p, solve φ by least square method _i(i=1,2,3 ..., p).

The the 4. step: utilize definite model to obtain predicted value

?

The mean value of step 4, calculating two sequences predicted value, obtains final predicted value

On the basis of step 2 and step 3, calculate the mean value of two sequences predicted value, obtain final predicted value.

In the present embodiment, two correlated series H (t) and S (t) are based on original load time series X=x (t)={ x (i), 1≤i≤87840} structure.Wherein, H (t) is 540 historical datas before moment t, and S (t) is the discharge record in nearest 60 days t moment, i.e. N=540 in formula (6), T=1440 in formula (7), M=60.

For effect of the present invention is described, utilize above-mentioned data, by following four kinds of methods, test respectively.

Method 1: the load on host computers multisequencing combination forecasting method based on WAVELET PACKET DECOMPOSITION that the present invention proposes;

Method 2: the Forecasting Methodology that patent < < host load prediction method > > who combines based on multisequencing proposes;

Method 3: wavelet packet-SVR model of mentioning in document < < Server Load Prediction Based on Wavelet Packet and Support Vector Regression > >, the method completes based on historical series;

Method 4: the host load prediction of document < < based on non-stationary series and the method for synthetic technology research > >, the method completes based on historical series.

Test result is as shown in table 2.

The mean square error of four kinds of Forecasting Methodologies of table 2

σ represents mean square error, and computing formula is:

$\mathrm{\&sigma;}={\mathrm{\&Sigma;}}_{t=1}^N{(x_t-{\hat{x}}_t)}^2/N---\left(17\right)$

Wherein, x _tflow actual value,

it is predicted value.

By test result, can obtain drawing a conclusion:

One, the precision of multisequencing combination forecasting method is higher than corresponding unique sequence Forecasting Methodology.In experiment, method 3 and method 4 are two kinds of unique sequence Forecasting Methodologies, and method 2 is the multisequencing combined predictions that carry out on the basis of method 4, and method 1 is the multisequencing combined prediction carrying out on the basis of method 3.Obviously, the precision of prediction of method 2 is higher than method 4; The precision of prediction of method 1 is higher than method 3.

Two,, in four kinds of methods, the method precision of prediction that the present invention proposes is the highest.

Claims (1)

1. the load on host computers multisequencing combination forecasting method based on WAVELET PACKET DECOMPOSITION, is characterized in that specific implementation step is:

Step 1, structure historical series and similar value sequence

First utilize data on flows structure historical series and the similar value sequence obtained, concrete building method is as follows:

The the 1. step: original load seasonal effect in time series expression formula as shown in Equation 5:

X＝x(t)＝{x(1)，x(2)，x(3)，…，x(k)，…}???(5)

The the 2. step: according to load time sequence expression formula, structure historical series and the similar value sequence relevant to load on host computers respectively;

If the two sequences relevant to moment t to be predicted is H (t), S (t):

H(t)＝{x(t-i)，N≥i≥1}???(6)

S(t)＝{x(t-j×T)，M≥j≥1}???(7)

The length that wherein T is one-period;

H (t) represents the chronological arrangement of top n load data in t moment, is called historical series;

S (t) represents the observation station load sequence corresponding with t in the adjacent M cycle before in t moment, is called similar value sequence;

The the 3. step: each element in two sequences is renumberd respectively, and H (t) and S (t) can be expressed as

with

Step 2, for historical series, adopt wavelet packet-SVR model to predict, obtain the predicted value 1 in t moment

On the basis of step 1 structure historical series and similar value sequence, historical series is carried out to wavelet packet-SVR prediction, obtain the predicted value 1 in t moment; Specific implementation step is as follows:

The the 1. step: historical series is carried out to WAVELET PACKET DECOMPOSITION and reconstruct;

First select a female small echo, and determine and decompose number of plies L; The selected number of plies L=3 that decomposes of the present embodiment;

Original series is carried out to WAVELET PACKET DECOMPOSITION:

Decomposable process is as shown in table 1; Wherein

represent total frequency band space that original series occupies,

represent the n sub spaces on j yardstick;

The wavelet packet in table 1 original series space is divided

As shown in Table 1:

The sequence d of corresponding each subspace ^{j, n}represent, utilize wavelet decomposition algorithm:

Wavelet packet space layer-by-layer is segmented; Wherein h _n, g _nrespectively low pass, high pass resolution filter coefficient;

Decompose while proceeding to L layer, obtain 2 ^lindividual signal component; Utilize wavelet package reconstruction algorithm:

To each signal component reconstruct, obtain final signal component; Wherein

respectively low pass, high pass reconfigurable filter coefficient;

The the 2. step: the SVR prediction of each branch signal

With

represent any one branch signal, by the value d in the n before the t moment value prediction t moment _tproblem can be expressed as the problem of finding following corresponding relation F:

Concerning the training of regression model, by N training sample, just can build N-n training sample pair, i.e. input (d ₁, d ₂..., d _n), the corresponding d that is output as _n+1; Input (d ₂, d ₃..., d _n+1), the corresponding d that is output as _n+2; The rest may be inferred;

Specific algorithm step is:

A. data are prepared: the data that collect are divided into two parts, and a part, as training data, is designated as

another part, as test data, is designated as

first training data is organized as to learning sample to (x take n as step-length _i, y _i), as shown in table 2;

Table 2 SVR forecast model learning sample

B. choose suitable supporting vector machine model;

C. determine and return step-length n;

D. according to definite support vector topological structure, utilize training dataset

training pattern;

E. utilize test data set

in data carry out Single-step Prediction;

F. according to predicting the outcome, carry out model evaluation; Calculate precision of prediction, if precision meets predefined threshold value, finish algorithm, otherwise adjust parameter, go back to steps d;

The the 3. step: each branch prediction value synthetic, is directly added i.e. formula (3) by corresponding the predicted value of each branch

Step 3, for similar value sequence, carry out AR prediction, obtain the predicted value 2 in t moment

On the basis of step 1 structure historical series and similar value sequence, similar value sequence is carried out to AR model prediction, obtain the predicted value 2 in t moment;

Concrete operation step:

The the 1. step: the zero-mean of original series;

Random data utilization AR model is analyzed and must be met stationary time series hypothesis, original series can be carried out to zero-mean processing, establishes the mean value that μ is S (t), zero-mean { s _jobtain ordered series of numbers { y _t}:

y _t＝s _t-μ???(13)

{ y _tit is approximate stationary time series;

The the 2. step: determine model order p;

The mathematic(al) representation of AR model is:

Z _t＝φ ₁Z _t-1+φ ₂Z _t-2+…+φ _pZ _t-p+a _t???(14)

Wherein, p is the exponent number of AR model, φ _ifor the coefficient of model, be solve for parameter, a _tit is white noise;

Employing from low order to high-order one by one the mode of matching determine the rank of model, from p=1, start to find one by one and can make sequence prediction curve and the good exponent number of actual curve matching; Be specially:

Utilize Yule-Walker equation group:

Wherein, Φ is coefficient vector, and ρ is auto-correlation vector, and Pp is autocorrelation matrix;

To φ _icarry out valuation, then obtain sequence { y _tapproximate prediction curve, and therefrom select a good model of matching;

The the 3. step: for selected exponent number p, solve φ by least square method _i(i=1,2,3 ..., p);

The the 4. step: utilize definite model to obtain predicted value

?

The mean value of step 4, calculating two sequences predicted value, obtains final predicted value

On the basis of step 2 and step 3, calculate the mean value of two sequences predicted value, obtain final predicted value.

Images