CN101599871A

CN101599871A - A kind of SFARIMA network flow prediction method

Info

Publication number: CN101599871A
Application number: CNA2009100272950A
Authority: CN
Inventors: 丁元彬; 张顺颐; 颜学智; 王攀
Original assignee: NANJING XINWANG VIDEOTECH CO Ltd
Current assignee: Jiangsu Xinwang Tec Technology Co.,Ltd.
Priority date: 2009-05-27
Filing date: 2009-05-27
Publication date: 2009-12-09

Abstract

The present invention proposes network flow prediction method, and designed the algorithm of prediction.This method is by constantly sliding the delay compensation effect to the estimated sequence time of carrying out, and method is as follows: step 1: from the real traffic sequence, extract the sample array, and with its called after FArray, and three variable M of initialization, N, the value of m; Step 2: calculate the self similarity index H of sample array FArray, method available period figure method, R/S method, wavelet analysis method etc.Step 3: with the exponent number of AIC criterion sample estimates array.

If

, determine then that the exponent number of model is (p, q); Step 4: calculating arma modeling parameter [φ, θ]=ARMA (pbest, qbest), method is as follows: (1) estimates the parameter of autoregression part; (2) estimate the moving average coefficient.Step 5: design factor vector π _j=θ ₁π _J-1+ θ ₂π _J-2+ Λ+θ _qπ _J-q+ φ _jJ＞0 is π wherein ₀=-1, φ when j＞p+d _j=0; Step 6: use

The prediction network traffics.

Description

A kind of SFARIMA network flow prediction method

Technical field

The present invention is a kind of Forecasting Methodology of network traffics, is mainly used in the latency issue that solves volume forecasting, belongs to the Computer Applied Technology field.

Background technology

The problem of volume forecasting can simplified summary be: one group of data on flows X before the known current time _N-i, (i=0,1,2 Λ), then the flow X in the following a certain moment _N+kCan be by X _N-iDraw.The k here is called step-length, is the single step prediction when k=1, k＞1 o'clock multi-step prediction.Studies show that 80% network traffics are all measurable, the accuracy rate of prediction is relevant with selected time scale, the best scale that different sequences are corresponding different.This has illustrated that network traffics have high predictability.

The principle of predicting network flow problem is that sequence has dependency structure or fixing relation.Usually the Forecasting Methodology that adopts has the method and the nonlinear Time Series Method of linear time series analysis.Linear Time Series Method as the FARIMA model, is to exist the linear correlation structure in the supposition sequence; Nonlinear Time Series Method as the chaos time sequence method, is the characteristic that the supposition sequence has chaos, predicts by phase space reconfiguration.Volume forecasting based on small echo is that sequence is introduced wavelet field, divides stationary sequence because wavelet analysis is suitable for handling, and no longer has long correlation through the flow sequence behind the wavelet analysis, and more level and smooth.Also have based on the Forecasting Methodology of small echo a variety of, as wavelet neural network prediction, small echo forecast of regression model.

The subject matter that existing method for predicting exists has: predicated error is big, can only realize short-term forecast, can not realize on-line prediction, and prediction has time delay.

Summary of the invention

Technical problem: the purpose of this invention is to provide a kind of network flow prediction method, be used for solving the delay problem of present flow rate prediction.The present invention can improve the accuracy rate of predicting network flow.

Technical scheme: the present invention proposes network flow prediction method, and designed the algorithm of prediction.This method is by constantly sliding the delay compensation effect to the estimated sequence time of carrying out, and method is as follows:

Step 1): initialization sample array FArray, initialization M, N, the value of m;

Step 2): the self similarity index H that calculates the sample array;

Step 3): with the exponent number of AIC criterion sample estimates array;

Step 4): calculate the arma modeling parameter

Step 5): design factor vector

π wherein ₀=-1, when j＞p+d

Step 6): use

\hat{X} (h) = Σ_{j = 1}^{\infty} π_{j}^{(h)} {\hat{X}}_{t + h - j}

The prediction network traffics.

Detailed process is as follows:

Step 1: from the real traffic sequence, extract the sample array, with its called after FArray, and three variable M of initialization, N, the value of m;

Step 2: calculate the self similarity index H of sample array FArray, method available period figure method, R/S method, wavelet analysis method etc.

Step 3: with the exponent number of AIC criterion sample estimates array.

AIC (n, m) = \ln \hat{s_{a}^{2}} + 2 (n + m + 1) / N - - - (1)

If

AIC (p, q) = \min_{0 \leq n, m \leq L} AIC (n, m),

The exponent number of then determining model be (p, q).Wherein

Be the s of corresponding ARMA sequence _a ²The maximum likelihood estimated value.That is to say that we attempt p from low to high, the value of q is got the pbest and the qbest that can allow formula (1) get minimum value.Because along with the increase of exponent number, AIC is linear growth computing time.So we get pmax=qmax=5 and carry out exponent number and estimate here;

Step 4: calculate the arma modeling parameter

Method is as follows:

1) estimates autoregression parameter partly.We obtain according to the relational matrix of autoregressive coefficient and auto-correlation function:

By the value of sample point and auto-correlation function, we can estimate auto-correlation coefficient.

2) order

Can obtain Y _tCovariance function be:

γ_{k} (Y) = E [Y_{t} Y_{t + k}]

Use X _tCovariance estimate

Replace γ _k:

{ Y _tBe similar to and regard MA (q) sequence as, promptly

Y_{t} &cong; a_{t} - θ_{1} a_{t - 1} - θ_{2} a_{t - 2} - θ_{3} a_{t - 3},

Parameter and covariance function can be write as following equation, and we choose one group of initial value and carry out iteration:

\{\begin{matrix} {\hat{σ}}_{a}^{2} = {\hat{γ}}_{0} {(1 + {\hat{θ}}_{1}^{2} + Λ + {\hat{θ}}_{q}^{2})}^{- 1} \\ {\hat{θ}}_{1}^{2} = - {\hat{γ}}_{1} / {\hat{σ}}_{a}^{2} + {\hat{θ}}_{1} {\hat{θ}}_{2} + Λ + {\hat{θ}}_{q - 1} {\hat{θ}}_{q} \\ {\hat{θ}}_{2} = - {\hat{γ}}_{2} / {\hat{σ}}_{a}^{2} + {\hat{θ}}_{1} {\hat{θ}}_{3} + Λ + {\hat{θ}}_{q - 2} {\hat{θ}}_{q} \\ Λ \\ {\hat{θ}}_{q - 1} = - {\hat{γ}}_{q - 1} / {\hat{σ}}_{a}^{2} + {\hat{θ}}_{1} {\hat{θ}}_{q} \\ {\hat{θ}}_{q} = - {\hat{γ}}_{q} / {\hat{σ}}_{a}^{2} \end{matrix}

Step 5: design factor vector

π wherein ₀=-1, when j＞p+d

Step 6: use

\hat{X} (h) = Σ_{j = 1}^{\infty} π_{j}^{(h)} {\hat{X}}_{t + h - j}

The prediction network traffics.

Description of drawings

Fig. 1 is the flow chart of SFARIMA Forecasting Methodology.

Embodiment

Collection network flow at first, sampling instruments such as available WinPcap extract the flow sequence.Experiment porch is Matlab7.0, will capture underflow amount data as input, carries out volume forecasting with this method, observes output, and compares with actual flow.By our actual motion and test, well predicted network traffics, eliminated the prediction time delay, verified the accuracy of the method.

Claims

1. SFARIMA network flow prediction method is characterized in that the step that the SFARIMA network flow prediction method is comprised:

Step 2): the self similarity index H that calculates the sample array;

Step 3): with the exponent number of AIC criterion sample estimates array;

Step 4): calculate the arma modeling parameter

Step 5): design factor vector

π wherein ₀=-1, when j＞p+d

Step 6): use

\hat{X} (h) = Σ_{j = 1}^{\infty} π_{j}^{(h)} {\hat{X}}_{i + h - j}

The prediction network traffics.