CN109325638A - A kind of SDN method for predicting based on RBF neural - Google Patents
A kind of SDN method for predicting based on RBF neural Download PDFInfo
- Publication number
- CN109325638A CN109325638A CN201811331322.9A CN201811331322A CN109325638A CN 109325638 A CN109325638 A CN 109325638A CN 201811331322 A CN201811331322 A CN 201811331322A CN 109325638 A CN109325638 A CN 109325638A
- Authority
- CN
- China
- Prior art keywords
- prediction
- data
- rbf neural
- network
- sdn
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/10—Interfaces, programming languages or software development kits, e.g. for simulating neural networks
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Business, Economics & Management (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Strategic Management (AREA)
- Human Resources & Organizations (AREA)
- Economics (AREA)
- Mathematical Physics (AREA)
- Game Theory and Decision Science (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Development Economics (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Entrepreneurship & Innovation (AREA)
- Marketing (AREA)
- Operations Research (AREA)
- Quality & Reliability (AREA)
- Tourism & Hospitality (AREA)
- General Business, Economics & Management (AREA)
- Data Exchanges In Wide-Area Networks (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
A kind of SDN method for predicting based on RBF neural of the disclosure of the invention, belongs to cordless communication network technical field.SDN volume forecasting algorithm proposed by the present invention based on RBF neural has excellent nonlinear characteristic, processing especially suitable for nonlinearity system, it is trained by research history data record, one has the ability for concluding total data by suitably trained neural network.Secondly, RBF neural has a flexibly and effectively mode of learning, for the more other neural networks of structure, simpler, pace of learning is faster.Therefore, RBF neural can carry out more accurate prediction to network flow complicated and changeable.The present invention emulates the algorithm proposed using POX and Mininet, and simulation result shows proposed algorithm energy Accurate Prediction SDN changes in flow rate trend, has preferable estimated performance and lower prediction error.
Description
Technical field
The invention belongs to cordless communication network technical fields, more particularly to a kind of performance evaluation suitable for network and net
Network planing method.
Background technique
Volume forecasting has great importance for the performance evaluation and the network planning of network.In traditional TCP/IP network
In, the distributed network architecture makes the flexibility of network and intelligence not high, causes volume forecasting algorithm good
Applied in industry.Software defined network (Software Defined Networking, SDN) is as a kind of novel network
Framework there is control plane to separate with the decoupling of data plane, the programmable interface that opens and the control of logical set Chinese style etc.
Feature significantly improves so that SDN flexibility and intelligence have compared with traditional network.Therefore, the proposition of SDN is pre- to flow
The application of method of determining and calculating provides a good platform.And it is fewer about the volume forecasting research of SDN at present, further, since
The dynamic change of SDN flow and random sudden, it is relatively difficult for carrying out Accurate Prediction.It is therefore proposed that accurate and effective
SDN method for predicting be very important.
The research of method for predicting is the variation with network size and network application and continually changing.Its prediction side
The development experience of method four-stage: the first stage, network application type is few, and network size is small, this stage is mainly based upon
The volume forecasting of conventional model is studied.Second stage only has short range dependence, can not retouch due to the limitation of conventional model
State the long range dependent of network flow.So proposing the method for predicting based on self similarity model.Phase III, network rule
The increasingly parameter of increase and self similarity model of mould calculates excessively complexity, and the estimated performance of self similarity model is caused to decline, therefore,
It proposes based on intelligent algorithm prediction technique.Fourth stage, with the further expansion and network application type of network size
Diversification, single model cannot portray the complete characteristic of flow, propose the method for predicting based on built-up pattern.
In traditional network, volume forecasting is an important research direction.But traditional network is distributed network,
It can not neatly control, the flexibility of network is poor, intelligence degree is low, this causes volume forecasting algorithm that cannot apply well
In industry.
Exactly in this context, it needs to find a kind of more flexible efficient network deployment mode, is got over adapting to network
Carry out more flexible and changeable demand, reduces network complexity, and the paces of transmission via net can be accelerated.Therefore, software defined network
The concept of network is suggested and gradually gets the nod, and core technology OpenFlow individually extracts control layer from the network equipment
As controller.SDN has control plane and data plane decoupling, open programmable interface and the control of centralization etc.
Feature significantly improves so that SDN flexibility and intelligence have compared with traditional network.Therefore, its proposition is calculated to volume forecasting
The application of method provides a good platform.But it is also fewer about the volume forecasting research of SDN at present.Benefit of the invention
Accurate flux prediction model is established using these data training patterns with the statistical information that sampling algorithm obtains flow counter,
To the variation tendency of accurately and effectively predicted flow rate.
With the rise of SDN, research emphasis is transferred to SDN from traditional network by researcher.Therefore, based on the stream of SDN
Amount forecasting research also receives the extensive concern of researcher.In traditional TCP/IP network, the distributed network architecture makes
Flexibility and the intelligence for obtaining network be not high, causes volume forecasting that can not apply well.In SDN, control layer is separated
It is made into controller out, realizes the centralized control to data Layer.Therefore, it the deployment of volume forecasting algorithm and applies in SDN
Good application is arrived.
Summary of the invention
For disadvantage of the existing technology, due to the dynamic change of SDN flow and random sudden, to carry out accurate pre-
Survey be it is relatively difficult, when for predicting the network flow of high changeability, it is difficult to portray large scale network stream some models
The complicated non-linear relation of amount, influences the prediction effect of model to a certain extent.The present invention proposes a kind of based on RBF nerve
The SDN method for predicting of network.
Technical solution of the present invention is a kind of SDN method for predicting based on RBF neural, this method comprises:
Step 1: to SDN flow measurement and sampling;
The information for obtaining counter storage in flow entry in OpenFlow interchanger, obtains data from the sample survey;The counter
The information of storage recite belong to this stream message had received it is how many, and transmitting-receiving packet count, transmitting-receiving byte number and
Search number;
The initialization of step 2:RBF neural network and data prediction;
Data prediction is that the data from the sample survey that step 1 obtains is normalized;Because RBF neural to [0,
1] data are most sensitive, can accelerate the convergence rate of training process after normalization and avoid calculating in training process excessive
Out;
Step 3:RBF neural network learning;
Present invention employs newrb function, the RBF neural of newrb creation is the process continuously attempted to, is being created
It is constantly increasing the quantity of hidden layer and the number of neuron during building, until the error for meeting setting;
Step 4: sliding window prediction;
Sliding window is set, is successively slided in the data from the sample survey that sampling obtains, the data from the sample survey in sliding window is to need
The data to be trained;The variation characteristic of flow after multiple moment can be preferably captured by sliding window;Sliding window it is pre-
It is constant that the main thought of survey is to maintain data length, and sliding window supplements new data, while rejecting legacy data, the mould being built such that
Type is more able to reflect the feature of flow.
Step 5: introducing the measurement standard that mean square error MSE terminates as training;
MSE is the desired value of the difference square of flow estimation value and true value, can evaluate the variation degree of data, and MSE value is got over
It is small, it is higher to illustrate that prediction model describes experimental data accuracy;Conversely, prediction accuracy is lower;
Step 6: anti-normalization processing;
It saves data and carries out anti-normalization processing, obtain the actual prediction value of network flow.
Further, the constant duration methods of sampling is used in the step 1 or waits data packet samplings method to SDN flow
It is sampled;
Detailed process is as follows for the constant duration methods of sampling:
The network flow to be measured is selected first, the network path that the network flow of measurement passes through then is determined, so that it is determined that net
The interchanger that network stream passes through;Sampling period is set, and controller periodically sends FlowStatisticsRequest to interchanger and disappears
Breath, interchanger receive FlowStatisticsRequest message, send FlowStatisticsReply message to controller, so
The average information transmission rate in time interval is obtained divided by the sampling interval using the byte number obtained afterwards;Store these statistics
Information;
Detailed process is as follows for the constant duration methods of sampling:
The network flow to be measured is selected first, the network path that the network flow of measurement passes through then is determined, so that it is determined that net
The interchanger that network stream passes through;When the number-of-packet of interchanger counter increases fixed number 100,500 or 1000, it will count
The statistical information of device issues controller, and starts timer;Until next statistical message is put forward, time, control are calculated
Device obtains the time interval of controller using number-of-packet divided by two statistical messages;Finally store these statistical informations.
Further, the RBF neural initialization of the step 2, detailed process is as follows:
We will initialize neural network parameter first;Maximum frequency of training is 500, learning rate 0.0005,
Factor of momentum is 0.005 center vector, exports weight, and output offset takes the random number of [0,1], and extension constant takes [0.1,2.5]
Random number;
Data prediction described in step 2, detailed process is as follows:
The data sampled from SDN network are that have burst, randomness, and amplitude of variation is bigger, in order to improve mind
Pace of learning through network before training pattern starts, will be normalized data.This is because RBF neural
It is most sensitive to the data of [0,1], the convergence rate of training process can be accelerated after normalization and training process is avoided to fall into a trap
The spilling of calculation.
X=xnormal(xmax-xmin)+xmin (2)
Wherein, xnormalIndicate the normalized value of network flow sequence, xmaxAnd xminRespectively indicate network flow sequence most
Big value and minimum value, x indicate network flow sequence after normalization.
Further, RBF neural described in the step 3 learns, and detailed process is as follows:
RBF neural is by input layer, hidden layer and output layer composition;The number of plies for increasing hidden layer can increase artificial mind
Processing capacity through network, while prediction error is reduced, improve precision of prediction;But while increasing the number of plies of hidden layer
The complexity of network can improve, therefore the learning time of network will increase, and restrain when training slow;And the raising of precision of prediction
It can also actually be obtained by increasing node in hidden layer, the effect of RBF neural study is more held than increasing the number of plies
Easily observation and adjustment;The RBF neural uses newrb function, and the RBF neural of newrb creation is one continuous
The process of trial is constantly increasing the quantity of hidden layer and the number of neuron during creation, until meeting the mistake of setting
Until difference;Input layer is realized from x → Ri(x) Nonlinear Mapping, output layer are realized from x → Ri(x) Linear Mapping;
Input layer is mapped to the radial basis function of hidden layer in the RBF neural are as follows:
Wherein, Ri(x) respective layer sensing node is indicated, x is n dimensional input vector;ciIt is the center of i-th of basic function, with x
Vector with same dimension;σiIt is the vector of i-th of perception, it determines the width of the Basis Function Center point;M is respective layer
The number of sensing node;||x-ci| | indicate x and ciThe distance between;
Hidden layer is mapped to the function of output layer in the RBF neural are as follows:
Wherein ykIndicate the output of RBF neural, r is output node number, wikIt is weight.
Further, sliding window described in step 4 predicts that detailed process is as follows:
The size that modeling window is arranged is Wm, prediction window size is Wp;First in modeling window WmInterior foundation is about SDN
The RBF neural of flow, then using neural network in prediction window WpInside predicted and obtained the flow at corresponding moment
Value;When to prediction window WpAfter interior all time point predictions, W is slidedpA time point;At this moment, in new modeling window WmWith
Prediction window WpIt repeats the above process;
Slip heavy loads principle is introduced for convenience, the size that window is modeled in Fig. 2 is W-1, and the size of prediction window is 1,
Sliding window slides 1 unit every time;Wherein k-th of prediction window isModeling window isWhen window slides into the right+1 window of kth, x is increased compared with k-th of windowk+w-1, and eliminate
xk, the element in modeling window and prediction window at this moment is respectivelyWithWhereinFor When window continues to slide to the right, new prediction result is successively obtained The process of window sliding has embodied input results and has exported the dynamic change of result, to realize flow
Sliding window prediction.
Further, detailed process is as follows for method described in step 5:
Introduce the measurement standard that mean square error MSE terminates as training:
Wherein xiIndicate actual flow, xi' indicating traffic prediction value, k indicates that data on flows is made of k number strong point;
MSE refers to the desired value of the difference square of flow estimation value and true value, and MSE is used to evaluate the variation degree of data, MSE
It is worth smaller, illustrates that prediction model describes experimental data with better accuracy, conversely, prediction accuracy is poorer.
Further step-up error boundary is 0.01, and when MSE is less than 0.01, RBF neural training terminates.Algorithm is pre-
It is as shown in Figure 3 to survey step.
Beneficial effects of the present invention:
SDN volume forecasting algorithm proposed by the present invention based on RBF neural has excellent nonlinear characteristic, especially
It suitable for the processing of nonlinearity system, is trained by research history data record, one by suitably training
Neural network has the ability for concluding total data.Secondly, RBF neural has a flexibly and effectively mode of learning, structure compared with
Simpler for other neural networks, pace of learning is faster.Therefore, RBF neural can be to network flow complicated and changeable
Amount carries out more accurate prediction.Volume forecasting neural network based is relatively effective method in predicting network flow, pre-
Survey field achieves huge success.The present invention emulates the algorithm proposed using POX and Mininet, simulation result
Show proposed algorithm energy Accurate Prediction SDN changes in flow rate trend, there is preferable estimated performance and lower prediction error.
Detailed description of the invention
Fig. 1 is the flow chart of the SDN volume forecasting algorithm of the invention based on RBF neural;
Fig. 2 is RBF neural framework of the present invention;
Fig. 3 is that sliding window of the invention predicts schematic diagram;
Fig. 4 is that the truthful data of inventive flow prediction and the predicted value of data from the sample survey compare (1min);
Fig. 5 is that the truthful data of inventive flow prediction and the predicted value of data from the sample survey compare (5min);
Fig. 6 is that the truthful data of inventive flow prediction and the predicted value of data from the sample survey compare (10min);
Fig. 7 is that the predicted value of truthful data of the present invention and the predicted value of data from the sample survey compare (1min);
Fig. 8 is that the predicted value of truthful data of the present invention and the predicted value of data from the sample survey compare (5min);
Fig. 9 is that the predicted value of truthful data of the present invention and the predicted value of data from the sample survey compare (10min);
Figure 10 is the relative error (1min, 5min, 10min) of inventive flow prediction;
Figure 11 is the cumulative distribution function of constant duration of the present invention sampling prediction error;
Figure 12 is truthful data of the present invention and data from the sample survey prediction comparison (100packets);
Figure 13 is truthful data of the present invention and data from the sample survey prediction comparison (500packets);
Figure 14 is truthful data of the present invention and data from the sample survey prediction comparison (1000packets);
Figure 15 is the predicted value of truthful data of the present invention compared with the predicted value of data from the sample survey (100packets);
Figure 16 is the predicted value of truthful data of the present invention compared with the predicted value of data from the sample survey (500packets);
Figure 17 is the predicted value of truthful data of the present invention compared with the predicted value of data from the sample survey (1000packets);
Figure 18 is cumulative distribution function of the data packets such as the present invention every sampling prediction error.
Specific embodiment
The following further describes the present invention in detail with reference to the accompanying drawings and specific embodiments:
A kind of SDN volume forecasting algorithm based on RBF neural, the specific steps are as follows:
Step 1:SDN flow measurement and sampling
Detailed process is as follows for constant duration sampling algorithm:
The network flow to be measured is selected first, the network path that the network flow of measurement passes through then is determined, so that it is determined that net
The interchanger that network stream passes through.Sampling period is set, and controller periodically sends FlowStatisticsRequest to interchanger and disappears
Breath, interchanger receive FlowStatisticsRequest message, send FlowStatisticsReply message to controller, so
The average information transmission rate in time interval is obtained divided by the sampling interval using the byte number obtained afterwards.Store these statistics
Information is used for simulation analysis.
Detailed process is as follows for constant duration sampling algorithm:
The network flow to be measured is selected first, the network path that the network flow of measurement passes through then is determined, so that it is determined that net
The interchanger that network stream passes through.When the number-of-packet of interchanger counter increases fixed number (100,500,1000), it will count
The statistical information of number device issues controller, and starts timer.Until next statistical message is put forward, time, control are calculated
Device processed obtains the time interval of controller using number-of-packet divided by two statistical messages.These statistical informations are finally stored, are used
In simulation analysis.
The initialization of step 2:RBF neural network and data prediction
We will initialize neural network parameter first.Maximum frequency of training is 500, learning rate 0.0005,
Factor of momentum is 0.005 center vector, exports weight, and output offset takes the random number of [0,1], and extension constant takes [0.1,2.5]
Random number.
The data sampled from SDN network are that have burst, randomness, and amplitude of variation is bigger, in order to improve mind
Pace of learning through network before training pattern starts, will pre-process data, i.e. normalized.This is because RBF
Neural network is most sensitive to the data of [0,1], can accelerate the convergence rate of training process after normalization and avoid training
The spilling calculated in the process.Then anti-normalization processing is carried out after the completion of prediction, obtains actual prediction value.
X=xnormal(xmax-xmin)+xmin (2)
Wherein, xnormalIndicate the normalized value of network flow sequence, xmaxAnd xminRespectively indicate network flow sequence most
Big value and minimum value.
Step 3:RBF neural network learning
RBF neural is by input layer, hidden layer and output layer composition.The number of plies for increasing hidden layer can increase artificial mind
Processing capacity through network, while prediction error is reduced, improve precision of prediction.But while increasing the number of plies of hidden layer
The complexity of network can improve, therefore the learning time of network will increase, and restrain when training slow.And the raising of precision of prediction
It can also actually be obtained by increasing node in hidden layer, the effect of RBF neural study is more held than increasing the number of plies
Easily observation and adjustment.Present invention employs newrb function, the RBF neural of newrb creation is the mistake continuously attempted to
Journey is constantly increasing the quantity of hidden layer and the number of neuron during creation, until the error for meeting setting.
The present invention using radial basis function be Gaussian function, it have slickness it is good, radial symmetric, expression-form simply and
The good advantage of analyticity.Gaussian function form are as follows:
Wherein x is n dimensional input vector;ciIt is the center of i-th of basic function, there is the vector of same dimension with x;σiIt is i-th
The vector of a perception, it determines the width of the Basis Function Center point;M is the number of sension unit;||x-ci| | indicate x and ci
The distance between.Input layer is realized from x → Ri(x) Nonlinear Mapping, output layer are realized from Ri(x)→ykLinear Mapping,
That is:
Wherein r is output node number, wikIt is weight.
Step 4: sliding window prediction
The size that modeling window is arranged is Wm, prediction window size is Wp.First in modeling window WmInterior foundation is about SDN
The RBF neural of flow, then using neural network in prediction window WpInside predicted and obtained the flow at corresponding moment
Value;When to prediction window WpAfter interior all time point predictions, W is slidedpA time point.At this moment, in new modeling window WmWith
Prediction window WpIt repeats the above process.
Slip heavy loads principle is introduced for convenience, the size that window is modeled in Fig. 2 is W-1, and the size of prediction window is 1,
Sliding window slides 1 unit every time.Wherein k-th of prediction window isModeling window isWhen window slides into the right+1 window of kth, x is increased compared with k-th of windowk+w-1, and eliminate
xk, the element in modeling window and prediction window at this moment is respectivelyWithWhereinFor When window continues to slide to the right, new prediction result is successively obtained The process of window sliding has embodied input results and has exported the dynamic change of result, to realize flow
Sliding window prediction.
Step 5: algorithm design
Introduce the measurement standard that mean square error (Mean Squared Error, MSE) terminates as training:
Wherein xiIndicate actual flow, xi' indicating traffic prediction value, k indicates that data on flows is made of k number strong point.
MSE refers to the desired value of the difference square of flow estimation value and true value.MSE is used to evaluate the variation degree of data, MSE
It is worth smaller, illustrates that prediction model describes experimental data with better accuracy.Conversely, prediction accuracy is poorer.Herein, if
Setting bouds on error is 0.01, and when MSE is less than 0.01, RBF neural training terminates.Algorithm prediction steps are as shown in Figure 1.
Step 6: anti-normalization processing
The data sampled from SDN network are that have burst, randomness, and amplitude of variation is bigger, in order to improve mind
Pace of learning through network before training pattern starts, will pre-process data, i.e. normalized.This is because RBF
Neural network is most sensitive to the data of [0,1], can accelerate the convergence rate of training process after normalization and avoid training
The spilling calculated in the process.Therefore anti-normalization processing is carried out after the completion of prediction, obtain actual prediction value.
Middle solid line is the predicted value of true value on Fig. 4, Fig. 5, Fig. 6, and dotted line is the predicted value of sample value, middle solid line under Fig. 6
For the predicted value of true value, dotted line is the predicted value of sample value.From Fig. 4,5 and 6 as can be seen that truthful data and data from the sample survey
Predicted value can accurately predict the variation tendency of truthful data and data from the sample survey, and prediction effect is also extraordinary.
Solid line is the predicted value of true value in Fig. 7, Fig. 8, Fig. 9, and dotted line is the predicted value of sample value.It can be with from Fig. 7,8 and 9
Find out, the predicted value of data from the sample survey can be good at the changes in flow rate trend of approaching to reality data.This shows in certain sampling
Under interval, the predicted value of data from the sample survey can use to replace the predicted value of truthful data, so as to reach with lesser net
Network measurement expense carrys out predicted flow rate variation tendency.
Figure 10 is the relative error of volume forecasting.Herein, we calculate separately the average value for obtaining relative error.Pass through
Calculate it is found that with the sampling interval increase, relative error can also increase with it, and the effect of volume forecasting can also decline.Pass through
The precision of prediction that the average relative error of RBF neural prediction can be seen that sampling interval smaller flow is higher.
Figure 11 is the cumulative distribution function of constant duration sampling prediction error.When Figure 11 shows that the sampling interval is 1min,
Precision of prediction is highest.When relative error is 0.4, which can carry out 98% truthful data (1min) accurate pre-
It surveys, while Accurate Prediction can be carried out to 98% data from the sample survey (1min);When relative error is 0.2, which can be to 80%
Truthful data (5min) carry out Accurate Prediction, while can to 85% data from the sample survey (5min) carry out Accurate Prediction;When opposite
Error be 0.4 when, the algorithm can to 70% truthful data (10min) carry out Accurate Prediction, while can to 65% sampling number
Accurate Prediction is carried out according to (10min).Moreover, two curves of 1min are whole all in the upper surface of 5min and 10min, therefore predict
Precision is an advantage over 5min and 10min.
Figure 12 is the prediction comparison diagram (sampling interval 100packets) of truthful data and data from the sample survey.Figure 13 is true number
According to the prediction comparison diagram (sampling interval 500packets) with data from the sample survey.Figure 14 is the prediction pair of truthful data and data from the sample survey
Than (sampling interval 1000packets).From Figure 12,13 and 14 as can be seen that the predicted value of truthful data and data from the sample survey can
The stability of enough variation tendencies for accurately predicting truthful data and data from the sample survey, estimated performance and prediction result all than it is equal whens
Between interval sampling have further raising.
Figure 15 is the predicted value of truthful data and the predicted value comparison diagram (sampling interval 1min) of data from the sample survey.Figure 16 is true
The predicted value of real data and the predicted value comparison diagram (sampling interval 5min) of data from the sample survey.Figure 17 be truthful data predicted value and
The predicted value comparison diagram (sampling interval 10min) of data from the sample survey.From Figure 15,16 and 17 as can be seen that the predicted value of data from the sample survey
It is capable of the changes in flow rate trend of approaching to reality data well.
Figure 18 is the cumulative distribution function of relative error.Figure 18 explanation, when relative error is 0.4, which can be right
98% truthful data (100packets) carries out Accurate Prediction, while can carry out to 97% data from the sample survey (100packets)
Accurate Prediction;When relative error is 0.4, which can carry out Accurate Prediction to 91% truthful data (500packets),
Accurate Prediction can be carried out to 91% data from the sample survey (500packets) simultaneously;When relative error is 0.4, which can be right
62% truthful data (1000packets) carries out Accurate Prediction, at the same can to 60% data from the sample survey (1000packets) into
Row Accurate Prediction.And the sampling interval is smaller, the volume forecasting precision of the algorithm is higher.
Claims (7)
1. a kind of SDN method for predicting based on RBF neural, this method comprises:
Step 1: to SDN flow measurement and sampling;
The information for obtaining counter storage in flow entry in OpenFlow interchanger, obtains data from the sample survey;The counter storage
Information recite and belong to the message of this stream and had received how many, and transmitting-receiving packet count, transmitting-receiving byte number and lookup
Number;
The initialization of step 2:RBF neural network and data prediction;
Data prediction is that the data from the sample survey that step 1 obtains is normalized;
Step 3:RBF neural network learning;
Present invention employs newrb function, the RBF neural of newrb creation is the process continuously attempted to, is being created
The quantity of hidden layer and the number of neuron are constantly increasing in journey, until the error for meeting setting;
Step 4: sliding window prediction;
Sliding window is set, is successively slided in the data from the sample survey that sampling obtains, the data from the sample survey in sliding window is to need to instruct
Experienced data;
Step 5: introducing the measurement standard that mean square error MSE terminates as training;
MSE is the desired value of the difference square of flow estimation value and true value, can evaluate the variation degree of data, and MSE value is smaller, says
It is higher that bright prediction model describes experimental data accuracy;Conversely, prediction accuracy is lower;
Step 6: anti-normalization processing;
It saves data and carries out anti-normalization processing, obtain the actual prediction value of network flow.
2. a kind of SDN method for predicting based on RBF neural as described in claim 1, it is characterised in that the step
The constant duration methods of sampling is used in rapid 1 or data packet samplings method is waited to be sampled SDN flow;
Detailed process is as follows for the constant duration methods of sampling:
The network flow to be measured is selected first, the network path that the network flow of measurement passes through then is determined, so that it is determined that network flow
The interchanger of process;Sampling period is set, and controller periodically sends FlowStatisticsRequest message to interchanger, hands over
It changes planes and receives FlowStatisticsRequest message, transmission FlowStatisticsReply message is then sharp to controller
With the byte number of acquisition divided by the sampling interval, the average information transmission rate in time interval is obtained;Store these statistical informations;
Detailed process is as follows for the constant duration methods of sampling:
The network flow to be measured is selected first, the network path that the network flow of measurement passes through then is determined, so that it is determined that network flow
The interchanger of process;When the number-of-packet of interchanger counter increases fixed number 100,500 or 1000, by counter
Statistical information issues controller, and starts timer;Until next statistical message is put forward, time, controller benefit are calculated
The time interval of controller is obtained divided by two statistical messages with number-of-packet;Finally store these statistical informations.
3. a kind of SDN method for predicting based on RBF neural as described in claim 1, it is characterised in that the step
Rapid 2 RBF neural initialization, detailed process is as follows:
We will initialize neural network parameter first;Maximum frequency of training is 500, learning rate 0.0005, momentum
The factor is 0.005 center vector, exports weight, and output offset takes the random number of [0,1], extension constant take [0.1,2.5] with
Machine number;
Data prediction described in step 2, detailed process is as follows:
X=xnormal(xmax-xmin)+xmin (2)
Wherein, xnormalIndicate the normalized value of network flow sequence, xmaxAnd xminRespectively indicate the maximum value of network flow sequence
And minimum value, x indicate network flow sequence after normalization.
4. a kind of SDN method for predicting based on RBF neural as described in claim 1, it is characterised in that the step
The study of RBF neural described in rapid 3, detailed process is as follows:
The RBF neural uses newrb function, and the RBF neural of newrb creation is the mistake continuously attempted to
Journey is constantly increasing the quantity of hidden layer and the number of neuron during creation, until the error for meeting setting;It is defeated
Enter layer to realize from x → Ri(x) Nonlinear Mapping, output layer are realized from x → Ri(x) Linear Mapping;
Input layer is mapped to the radial basis function of hidden layer in the RBF neural are as follows:
Wherein, Ri(x) respective layer sensing node is indicated, x is n dimensional input vector;ciIt is the center of i-th of basic function, has with x
The vector of same dimension;σiIt is the vector of i-th of perception, it determines the width of the Basis Function Center point;M is respective layer perception
The number of node;||x-ci| | indicate x and ciThe distance between;
Hidden layer is mapped to the function of output layer in the RBF neural are as follows:
Wherein ykIndicate the output of RBF neural, r is output node number, wikIt is weight.
5. a kind of SDN method for predicting based on RBF neural as described in claim 1, it is characterised in that step 4 institute
The sliding window prediction stated, detailed process is as follows:
The size that modeling window is arranged is Wm, prediction window size is Wp;First in modeling window WmInterior foundation is about SDN flow
RBF neural, then using neural network in prediction window WpInside predicted and obtained the flow value at corresponding moment;When
To prediction window WpAfter interior all time point predictions, W is slidedpA time point;At this moment, in new modeling window WmAnd prediction
Window WpIt repeats the above process.
6. a kind of SDN method for predicting based on RBF neural as described in claim 1, it is characterised in that step 5 institute
Detailed process is as follows for the method stated:
Introduce the measurement standard that mean square error MSE terminates as training:
Wherein xiIndicate actual flow, xi' indicating traffic prediction value, k indicates that data on flows is made of k number strong point;
MSE refers to the desired value of the difference square of flow estimation value and true value, and MSE is used to evaluate the variation degree of data, and MSE value is got over
It is small, illustrate that prediction model describes experimental data with better accuracy, conversely, prediction accuracy is poorer.
7. a kind of SDN method for predicting based on RBF neural as claimed in claim 4, it is characterised in that setting misses
Poor boundary is 0.01, and when MSE is less than 0.01, RBF neural training terminates.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811331322.9A CN109325638A (en) | 2018-11-09 | 2018-11-09 | A kind of SDN method for predicting based on RBF neural |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811331322.9A CN109325638A (en) | 2018-11-09 | 2018-11-09 | A kind of SDN method for predicting based on RBF neural |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109325638A true CN109325638A (en) | 2019-02-12 |
Family
ID=65261180
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811331322.9A Pending CN109325638A (en) | 2018-11-09 | 2018-11-09 | A kind of SDN method for predicting based on RBF neural |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109325638A (en) |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110263969A (en) * | 2019-05-07 | 2019-09-20 | 西北农林科技大学 | A kind of shelf life apple quality Dynamic Forecasting System and prediction technique |
CN110363223A (en) * | 2019-06-20 | 2019-10-22 | 华南理工大学 | Industrial flow data processing method, detection method, system, device and medium |
CN111369042A (en) * | 2020-02-27 | 2020-07-03 | 山东大学 | Wireless service flow prediction method based on weighted federal learning |
CN111371644A (en) * | 2020-02-28 | 2020-07-03 | 山东工商学院 | Multi-domain SDN network traffic situation prediction method and system based on GRU |
CN112105048A (en) * | 2020-07-27 | 2020-12-18 | 北京邮电大学 | Combined prediction method based on double-period Holt-Winters model and SARIMA model |
CN114282658A (en) * | 2022-03-02 | 2022-04-05 | 北京航空航天大学 | Method, device and medium for analyzing and predicting flow sequence |
CN115022195A (en) * | 2022-05-26 | 2022-09-06 | 电子科技大学 | Dynamic flow measurement method for IPv6 network |
CN115442246A (en) * | 2022-08-31 | 2022-12-06 | 武汉烽火技术服务有限公司 | Flow prediction method, device, equipment and storage medium of data plane network |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103268525A (en) * | 2013-06-04 | 2013-08-28 | 南京大学 | WD-RBF (wavelet denoising-radial basis function)-based analogue prediction method of hydrological time sequence |
US20130262368A1 (en) * | 2010-09-27 | 2013-10-03 | Carmel-Haifa University Economic Corporation Ltd | Detecting change points in data streams |
CN103840988A (en) * | 2014-03-17 | 2014-06-04 | 湖州师范学院 | Network traffic measurement method based on RBF neural network |
CN104599500A (en) * | 2015-01-14 | 2015-05-06 | 南京信息工程大学 | Grey entropy analysis and Bayes fusion improvement based traffic flow prediction method |
CN104935460A (en) * | 2015-05-05 | 2015-09-23 | 浙江大学 | Multiple controller energy saving optimization method based on data surface flow prediction |
CN105049279A (en) * | 2015-06-19 | 2015-11-11 | 国家电网公司 | Communication trend flexibility configuration method and system |
-
2018
- 2018-11-09 CN CN201811331322.9A patent/CN109325638A/en active Pending
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20130262368A1 (en) * | 2010-09-27 | 2013-10-03 | Carmel-Haifa University Economic Corporation Ltd | Detecting change points in data streams |
CN103268525A (en) * | 2013-06-04 | 2013-08-28 | 南京大学 | WD-RBF (wavelet denoising-radial basis function)-based analogue prediction method of hydrological time sequence |
CN103840988A (en) * | 2014-03-17 | 2014-06-04 | 湖州师范学院 | Network traffic measurement method based on RBF neural network |
CN104599500A (en) * | 2015-01-14 | 2015-05-06 | 南京信息工程大学 | Grey entropy analysis and Bayes fusion improvement based traffic flow prediction method |
CN104935460A (en) * | 2015-05-05 | 2015-09-23 | 浙江大学 | Multiple controller energy saving optimization method based on data surface flow prediction |
CN105049279A (en) * | 2015-06-19 | 2015-11-11 | 国家电网公司 | Communication trend flexibility configuration method and system |
Non-Patent Citations (3)
Title |
---|
DENGFENG WEI: "Network traffic prediction based on RBF neural network optimized by improved gravitation search algorithm", 《NEURAL COMPUTING AND APPLICATIONS》 * |
冯明发等: "粒子群优化RBF神经网络的短时交通流量预测", 《计算机仿真》 * |
鹿如强: "基于深度学习的SDN流量预测研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110263969A (en) * | 2019-05-07 | 2019-09-20 | 西北农林科技大学 | A kind of shelf life apple quality Dynamic Forecasting System and prediction technique |
CN110363223A (en) * | 2019-06-20 | 2019-10-22 | 华南理工大学 | Industrial flow data processing method, detection method, system, device and medium |
CN111369042A (en) * | 2020-02-27 | 2020-07-03 | 山东大学 | Wireless service flow prediction method based on weighted federal learning |
CN111369042B (en) * | 2020-02-27 | 2021-09-24 | 山东大学 | Wireless service flow prediction method based on weighted federal learning |
CN111371644A (en) * | 2020-02-28 | 2020-07-03 | 山东工商学院 | Multi-domain SDN network traffic situation prediction method and system based on GRU |
CN111371644B (en) * | 2020-02-28 | 2023-04-07 | 山东工商学院 | Multi-domain SDN network traffic situation prediction method and system based on GRU |
CN112105048B (en) * | 2020-07-27 | 2021-10-12 | 北京邮电大学 | Combined prediction method based on double-period Holt-Winters model and SARIMA model |
CN112105048A (en) * | 2020-07-27 | 2020-12-18 | 北京邮电大学 | Combined prediction method based on double-period Holt-Winters model and SARIMA model |
CN114282658A (en) * | 2022-03-02 | 2022-04-05 | 北京航空航天大学 | Method, device and medium for analyzing and predicting flow sequence |
CN115022195A (en) * | 2022-05-26 | 2022-09-06 | 电子科技大学 | Dynamic flow measurement method for IPv6 network |
CN115022195B (en) * | 2022-05-26 | 2023-10-10 | 电子科技大学 | Flow dynamic measurement method for IPv6 network |
CN115442246A (en) * | 2022-08-31 | 2022-12-06 | 武汉烽火技术服务有限公司 | Flow prediction method, device, equipment and storage medium of data plane network |
CN115442246B (en) * | 2022-08-31 | 2023-09-26 | 武汉烽火技术服务有限公司 | Traffic prediction method, device, equipment and storage medium of data plane network |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109325638A (en) | A kind of SDN method for predicting based on RBF neural | |
CN103840988B (en) | A kind of network flow programming method method based on RBF neural | |
CN110675623B (en) | Short-term traffic flow prediction method, system and device based on hybrid deep learning | |
CN107293115B (en) | Traffic flow prediction method for microscopic simulation | |
CN109242140A (en) | A kind of traffic flow forecasting method based on LSTM_Attention network | |
CN108921822A (en) | Image object method of counting based on convolutional neural networks | |
CN109828304B (en) | Deep learning-based method for predicting lithologic sequence model by using seismic data | |
CN106570597A (en) | Content popularity prediction method based on depth learning under SDN architecture | |
CN108399745A (en) | A kind of city road network trend prediction method at times based on unmanned plane | |
CN108090508A (en) | A kind of classification based training method, apparatus and storage medium | |
CN105471631B (en) | Network flow prediction method based on traffic trends | |
CN109889452A (en) | Network context flow generation method and system based on condition production confrontation network | |
CN109189762A (en) | A kind of industry internet of things data analysis method, system and relevant device | |
CN109508788A (en) | A kind of SDN method for predicting based on arma modeling | |
CN110365503A (en) | A kind of Index and its relevant device | |
CN109886356A (en) | A kind of target tracking method based on three branch's neural networks | |
CN110147745A (en) | A kind of key frame of video detection method and device | |
US20140058985A1 (en) | Method and Device for Synthesis of Network Traffic | |
CN103268519A (en) | Electric power system short-term load forecast method and device based on improved Lyapunov exponent | |
CN106953854B (en) | A kind of method for building up of the darknet flow identification model based on SVM machine learning | |
CN109472030A (en) | A kind of system replys the evaluation method and device of quality | |
CN107945534A (en) | A kind of special bus method for predicting based on GMDH neutral nets | |
CN110991776A (en) | Method and system for realizing water level prediction based on GRU network | |
CN110972074A (en) | Track reconstruction method for single agent of multi-mobile agent group after finding lost position | |
CN107453921A (en) | Smart city system artificial intelligence evaluation method based on nonlinear neural network |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190212 |
|
RJ01 | Rejection of invention patent application after publication |