4847
Prediction of Network Traffic based on Improved Echo Network
Dr.P.Kalyani
1, P Krishna Kishore
21Kovur,Nellore (dist),Andhra PradeshIndia.
2Assistant ProfessorDept. of CSEChadalawada Ramanamma Engineering CollegeTirupati,Andhra Pradesh India.
1Kalyani.purini@gmail.com, 2krishna.boinapalli@gmail.com
Article History: Received: 11 January 2021; Revised: 12 February 2021; Accepted: 27 March 2021; Published
online: 10 May 2021
Abstract: The social occasion of network traffic has complex attributes like unconventionality, unsettling influence, good judgment, and non-linearity, which make it hard to expect network traffic. To address these marvelous pictures and improve presumption accuracy, this article proposes another framework for expecting network traffic subject to an improved resounding network. Regardless, to conform to its weakness and disorder, an network traffic rattle getting out calculation subject to close protection projection is proposed to diminish the blueprint disturbance of crude network traffic. Second, to guarantee advantageousness and non-linearity, an network traffic figure model relies upon an extra twofold circle resounding organization that perceives both an unsettling influence free network traffic course of action and an network traffic social affair of crude relationship as information. At last, the proposed framework is shown utilizing two authentic game-plans of network traffic information, and the age results show that the proposed strategy can give better execution in expecting network traffic than other comparable strategies.
Keywords: prediction, Network traffic, echo state network, Denoise, preserving local projection.
1. Introduction
Exact measure of the social event of network traffic is fundamental for the association and control of the network [1]. A get-together of network traffic with complex pictures like unusualness, uproar, advantageousness, and non-linearity is an exceptional kind of time plan that is hard to anticipate.
Reliable changes ahead of the pack of network clients and the intricacy of the network alliance climate make the game arrangement of network traffic wobbly and rough. Consequently, it is vital for pre-measure the arrangement of network traffic. Huang et al. [2] self-destruct the network traffic plan into different inside detached bits utilizing an improved generally exploratory mode decay technique, and from that point foresee each part in like manner utilizing a quantum neural network. Nie et al. [3] applies discrete wavelet change to eliminate the low and high recurrent parts from the network traffic gathering, and a brief timeframe later fabricated a critical conviction network model and a Gaussian model to anticipate these two bits, freely. Zhang et al. [4] duplicate the stage space of the network traffic strategy, and from there on gather a versatile vector descend into sin model supporting v for presumption. Meng et al. [5] Eliminate the ruckus in the network traffic movement utilizing quite far wavelet technique, by then apply the genuine vector machine lose the faith model for estimate.
Among different preprocessing strategies, disturbance reduction can enough decrease the unfavorable outcomes of flightiness and tumult in predicting network traffic. Numerous upheaval decline estimations have been proposed, including wavelet racket decline [6], wavelet noise decline with a sensitive breaking point [7] and neighborhood insurance of projection disturbance decline (LPP) [8]. Since LPP can maintain neighborhood tank limits, develop calculations and accelerate, it is by and large used in various spaces, for instance, strange association traffic area [9] and picture portrayal [10].
Since the gathering of association traffic is non-straight, regular direct assumption models, for instance, the autoregressive model [11], the autoregressive composed moving typical model [12] and the Poisson model [13] are difficult to satisfy the requirements of the association traffic. Assumption, while a phony neural association (ANN) model [14] with a strong non-straight gauge capacity offers better execution in expecting network traffic. Thusly, considering the ANN, the experts proposed a couple of models for expecting network traffic. Zhang et al. [15] propose an association traffic estimate model ward on an inherited computation and an extended based association of abilities to improve the accuracy of assumptions. Narejo and Pasero [16] examine significant conviction networks with three special developments and apply them in like way to expect network traffic. In any case, most ANN-based assessing models take unnecessarily long to iteratively smooth out the weight organizations, and they are likely not going to meet the steady essentials [17] ceaselessly network traffic assessing works out. In 2001, Jager [18] proposed another dreary neural association called the Echo State Network (ESN). Appeared differently in relation to ANN, ESN only necessities to set up the yield network [19], [20], which can improve the assessing capability. Regardless, in light of the subjective plan of the inventory, ESN saves a long
4848
exertion to make the archive, and the ability to manage the non-direct gathering is confined. Consequently, it is imperative to address the plan of the inventory to furthermore diminish the planning time and improve the viability of predicting network traffic, and the neural relationship in the vault ought to be built up to improve its ability to manage the progression of association traffic.To regulate complex network characters traffic assembling, another framework for expecting network traffic in this and network dependent upon the improved resounding status is proposed paper. In any case, to address the oddity and disorder of the network. Work traffic movement, network traffic commotion decrease assessment LPP-based racket getting out of crude affiliation traffic is advanced. Get-together. Second, keeping an eye on advantageousness and non-linearity. network traffic gathering, network traffic gauge network model with twofold circle reverberation express A position office (ESN-DLRS) is a work in progress, which it incorporates both the pound network traffic approach and the unpleasant network traffic. Work progression strategy as data. From this time forward, not just outstanding illustration of the general strength of network traffic with disturbance decay consistency, yet additionally solid relationship with the check the worth in the harsh strategy of network traffic can be discovered utilizing ESN-DLRS, which can improve surmise exactness. At long last, unsettling influence decay, input window size and suspicion execution dissected utilizing two legitimate networks dataset. The remainder of this record is made as follows: An area II gives a format of an improved ESN-based network traffic guess framework. District III portrays a LLP-based scratch-off assessment for network traffic. Bit IV proposes an network traffic presumption model ward on ESN-DLRS. Appearing and assessment are an examined in Area V. The report envelops with a regionVI.
Figure 1. General procedure for forecasting network traffic based on enhanced network echo.
2. Overview of network traffic prediction method based on improved esn
The overall strategy for the proposed technique is appeared in Figure 1. The presumption issue of this methodology can be depicted as expecting the worth of network traffic tr(th) as indicated by the harsh network traffic strategy Trk1(t) and the commotion dropped network movement. Traffic. Trden (t), where h is the gauge step. Trk1(t) can be passed on as:
Trk1 (t) = {tr (t − k1) , tr (t −k1 +1) , . . . , tr (t −1) , tr (t)} (1)
Where k1 is the information window size of the crude network traffic course of action, and tr(t) is the network traffic respect at time t. In the preprocessing stage, the denoised network traffic game-plan Trden (t) is acquired by the network traffic denoising assessment dependent upon LPP. Trk2 (t) can be bestowed as:
Trden (t) = ntrden (t − k2) , trden (t − k2 + 1) ,. . . , trden (t − 1) , trden (t) (2)
Where k2 is the data window size of the denoised network traffic movement, and trden(t) is the denoised network traffic respect at time t.
In the figure stage, the crude affiliation traffic gathering Trk1 (t) and the denoised network traffic blueprint Trden (t) are utilized in the interim as the responsibility of the affiliation traffic presumption model ward on ESN-DLRS. Along these lines, in Fig. 1, the information vector of information layer is u(t)
T2 vector assessment. The yield vector of yield layer can be settled as y(t) (tr(th)) T , where L1 is the yield vector assessment.
4849
1. Collect and standardize the network traffic movement information.2. Make preparing tests (utrain (t) , ytrain (t+h)), t= 1,2,… T, where utrain (t) = (Trk1 (t), Trk2 den (t))T. Trk2 den (t) is acquired by the network traffic denoising calculation dependent upon LPP notice in piece III.
3. Develop the network traffic figure model ward on ESN-DLRS alluded to in Piece IV. 4. Train the network traffic gauge model by utilizing tests worked in A condition of congruity 2.
5. Denoise the actually aggregated network traffic movement. Develop the information vector upred (t1) = (Trk1 (t1) , Trk2 shelter (t1)) and put it into the prepared guess model to secure the normal worth tr (t1 + h).
Detailed methodology for assessing network traffic depicts a framework reliant upon improved association resonation state. Like computation 1. Taking both Trk1 .t/and Trden k2.t/as territory cannot simply make the figure model evaluation the development of the instance of the overall stable all through movement of network traffic with commotion scratch-off, yet what's more change the typical worth using the relationship between's the fundamental connection traffic amassing and expected worth. Accordingly, expect network traffic the precision can be improved satisfactorily.
3. Network traffic denoising algorithm based on local preserving projection
To manage the passing quality and commotion of network traffic development, an association traffic denoising computation subject to LPP is depicted in this segment. Specifically, the stage space of connection traffic approach is reproduced and the close by neighborhood for each stage point is gotten settled the imitated stage space. In addition, a short period of time later, the imitated stage space framework is separated into the alliance traffic signal subspace and the disturbance subspace in each close by territory. Starting their forward, the straight hyper plane is made by the association traffic signal subspace. Finally, the unruly association traffic is overseen by even projection on the direct hyper plane, so the denoised network traffic development can be gotten. The unequivocal methodology of alliance traffic denoising evaluation subject to LPP is portrayed as Algorithm 2.
The network traffic denoising computation subject to LPP uses the close by area to keep up the local complex features of association traffic data, and utilizations the reasonable projection to lessen the racket in the connection traffic data, which can reasonably decrease the contradicting delayed consequence of transient quality and bedlam of network traffic gathering.
4. Network traffic prediction model based on echo state network with double loop reservoir structure
To deal with the sensibility and nonlinearity of network traffic assembling, an network traffic figure model ward on ESN-DLRS is depicted in this part. The model knows about veneration with its new development, computation and preparing.
A. CONSTRUCTION OF ESN-DLRS
ESN-DLRS contain three portions as shown in Fig. 1: the information layer, the twofold circles vault and the yield layer. The imperative stock is utilized as the data preparing layer which makes ESN having unprecedented memory limit and stunning nonlinear check limit. Considering nearby info circle vault (ALR) [23], a Twofold Circle Supply Improvement (DLRS) is organized by adding forward networks and examination connection between neurons with a similar neuronal reach. Separated and ESN, DLRS is fixed and connection between neurons in DLRS are maintained.
The point by point improvement strategy of ESN-DLRS is portrayed as Algorithm 2.
Algorithm 2 Construction Procedure of ESN-DLRS
1. Set the measure of neurons in the storage facility as N . Set the neuronal reach as d , which fulfills mod (N, d) 0. Set the fixed weight as r, r (0, 1).
2. Create the guideline circle by accomplice all of inside neurons methodical insider astute development. To be unequivocal, for every neuron I, I = 1, 2, . . . , N − 1, set the pieces of the storage facility network cross matrix W : wi,i+1 r, wi+1,i r. Precisely when I N, set wi, 1 r, w1,i r. Unequivocally wi,j is the relationship of the neuron I to the neuron j.
3. Gather the resulting circle. Recognize the essential neuron as a beginning stage, and accomplice it with the other neuron which is d neurons disengaged. Beginning now and for a significant length of time, accomplice this neuron and the going with one which is d neurons confined similarly. To be unequivocal, for the neuron I, I = 1, 1 + d, 1 + 2d, . . . , 1 + (N/d − 2)•d , set the pieces of the store alliance framework: wi,i+d r, wi+d,i r. Right when I 1 (N/d1) d , set wi,1 r, w1,i r. The headway of DLRS is finished.
It ought to be seen that the various circles supply setup can be made by utilizing essentially indistinguishable techniques. Regardless, it has been appeared in our past work [24] that DLRS has better execution. According to one perspective, ESN-DLRS have the fixed store structure and the fixed section loads, which adequately shorten the arranging time to fulfill the ceaseless fundamental of network traffic figure. Obviously, the neuronal
4850
relationships in ESN-DLRS are propped to improve its capacity to deal with the nonlinear network traffic gathering.
B. CALCULATION OF ESN-DLRS
At time t, the information vector of the data layer is u (t) T Trk1 (t) , Tr(t), the yield vector of the yield layer is y (t)(tr (th))T , and the internal state vector of twofold circle supply is x (t) (x1 (t) , x2 (t) , . . . , xN (t))T . The initiation conditions of inner neurons are for the most part calculated as:
x (t) = f in Winu (t) + Wx (t − 1) + Wback y (t − 1)
(3)
Where f in is the commencement furthest reaches of the stock neurons (regularly the distorted redirection work), Win is the data lattice and Wback is the output feedback matrix.
The outcomes are calculated as:
y (t) = f out Wout x (t) (4)
Where Wout is the outcome matrix and f out is the readout furthest reaches of neurons (typically the character work).
The joint responsibility of the denoised network traffic movement and the harsh network traffic gathering, also as the remarkable DLRS, improve ESN-DLRS's capacity to deal with the nonlinear network traffic game-plan.
Algorithm 3 Training Procedure of ESN-DLRS
1. Initialize ESN-DLRS parameters, including the reservoir size N and the spectral radius λ, 0 < λ < 1. 2. Bring samples (utrain (t) , ytrain (t h)), t 1, 2, . . . , T into ESN-DLRS for training.
3. Collect the internal state vectors x (t) from time point t0 to the time point T , obtain the state matrix
X = [x (t0) , x (t0 + 1) , . . . , x (T )].
4. Collect the expected output vectors, and obtain the expected output matrix:
Y = [y (t0 + h) , y (t0 + h + 1) , . . . y (T + h)].
5. Calculate the output connection matrix Wout according to linear regression by W
out Y X +, where X + is the pseudo-inverse of X.
C.TRAINING OF ESN-DLRS
The arranging arrangement of ESN-DLRS is portrayed as Calculation 3.
ESN-DLRS produce Win and W in a fixed development, and keep them unaltered during the entire arranging measure. Just Wout should be set up by the given models. Unquestionably ESN-DLRS can diminish the arranging time to fulfill the consistent need of network traffic presumption.
5. Simulations and analyses
To show the check execution, utilizing Matlab as the increase contraption, the proposed method is imitated and explored utilizing two authentic network traffic datasets. Dataset An includes the measure of information packs in each second procured from Beijing University of Posts and Telecommunications Backbone Nodes of China Education Network (BUPTBN) [25] a few days. Dataset B remembers the measure of pieces for each a little habit from United Kingdom Academic Network Backbone Nodes (UKANBN) [26] a few days. Fig. 2 shows 6000 information points of Dataset an and 4000 information points of Dataset B autonomously.
A. SIMULATION SETTINGS
For the network traffic denoising appraisal subject to LPP, the denoising area range ε is set to 0.6. For the network traffic assumption model ward on ESN-DLRS, the vile reach λ of the hold is set to 0.8 and the storage space size N is set to 80. The vault starting work gets the overstated deviation work, and the yield sanctioning work recognizes the character work. The yield examination system isn't used in the model. The neuronal stretch d of DLRS is set to 8 and the stock association W of DLRS is passed on as depicted in Section IV (A). The data window size of the unrefined connection traffic gathering is set as k1 > 1, 2, . . . , 10 , and the information window size of denoised network traffic gathering is set as k2 > 0, 1, . . . , 5 . For Dataset an and Dataset B, the standard piece of worker ranches are used for planning and the rest half are for attempting. The Normalized Mean Square Error (NMSE) is used to evaluate the figure exactness.
4851
FIGURE 2. Network traffic datasets. a) Dataset A. b) Dataset B. FIGURE 3. Noisy 2-D phase diagram. a) Dataset A.b) Dataset B
B. DENOISING PERFORMANCE ANALYSIS
To show the appropriateness of network traffic denoising assessment subject to LPP, the unpleasant network traffic blueprint and the denoised network traffic gathering are bankrupt some place close to utilizing the 2-D stage layout technique [27] and the maximal Lyapunov type methodology [28].
The rambunctious 2-D stage graphs can be gotten by normalizing the unpleasant network traffic gathering, as demonstrated in Fig. 2. The stage outline of Dataset An is astoundingly dumbfounded. The stage chart of Dataset B is more arranged than that of Dataset A, in any case by and large wavers suddenly. The Root Mean Square Error (RMSE) of stage focuses is 0.0562 for Dataset an and 0.0118 for Dataset B. It will overall be seen that the harsh
4852
network traffic movement has the variability character. The 2-D stage graph of the denoised network traffic strategy can be gotten by utilizing the network traffic denoising assessment dependent upon LPP, as shown in Fig. 3. The RMSE of stage focuses is lessened to 0.0187 for Dataset An and 0.0058 for Dataset B. It will overall be seen that the proposed assessment can agreeably lessen the inconstancy of network traffic assembling and make the arrangement smoother.The relationship of the maximal Lyapunov model is appeared in Fig. 5. Going before denoising, the maximal Lyapunov representation of Dataset An is 0.029, and the maximal Lyapunov sort of Dataset B is 0.0156, which shows that the unpleasant network traffic strategy has the issue character. In the wake of denoising with the network traffic denoising calculation dependent upon LPP, the maximal Lyapunov representation of Dataset An is 0.0255, and the maximal Lyapunov sort of Dataset B is 0.0134, which shows that the proposed assessment can attainably decrease the problem of network traffic groupings.
C. INPUT WINDOW SIZE ANALYSIS
The control of the input window sizes of equally the raw network traffic progression k1 and denoised network traffic sequence k2 on the calculation precision with the network traffic prediction model mentioned in Section IV is analyzed as shown in Fig. 5.
.
FIGURE 3. Denoised 2-D phase diagram. a) Dataset A. b) Dataset B.
4853
From the beginning, as can be seen from Fig. 4, when k2> 0, explicitly, the information simply contains the unrefined network traffic system, the doubt botch is much more noticeable than that of k2 > 0 for either Dataset An or Dataset B. It shows that adding the denoised network traffic social event to information can help the action model to get settled with the moving delineation of everything considered stable of connection traffic blueprint and enough improve the assumption accuracy.Moreover, as can be seen from Fig. 5, for Dataset A, when k1 > 1 and k2 > 1, the value of NMSE is the base which is 0.98378. Meanwhile, for Dataset B, when k1 > 1 and k2 > 4, the value of NMSE is the base which is 0.0037442. It shows that the ideal information window size of denoised network traffic development will be shifting due to different datasets or explicit examining ranges.
Finally, as can be seen from Fig. 5, the ideal information window size of the brutal network traffic development is 1. Since the justification using input brutal connection traffic gathering is to address the typical worth. As such, it is adequate to pick unquestionably the latest unforgiving network traffic regard which has the closest relationship with the typical worth as data.
FIGURE 5. Input window analysis of the raw traffic sequence and the denoised traffic sequence. a) Dataset A. b) Dataset B.
FIGURE 6. Training time analysis.
D. PREDICTION PERFORMANCE ANALYSIS
To show the assumption execution, the evaluation of check execution with other relative systems is showed up in Fig. 6 and Fig. 7. Fig. 6 shows the planning season of the notion models subject to Back Propagation (BP) neural network [29], ESN and ESN-DLRS in a relative selecting environment. It might be seen from Fig. 8 that the masterminding season of ESN-DLRS is clearly not everything thought about BP and incidentally not actually ESN for different datasets. ESN and ESN-DLRS require orchestrating time emphatically not actually BP, considering the way that ESN and ESN-DLRS embrace the speedy slide into wrongdoing structure instead of the incline drop framework and fundamentally need to set up the output structure. ESN-DLRS costs less planning time than ESN, which shows that the fixed stock development and the fixed part loads can what's more diminishing organizing time to satisfy the predictable need of network traffic assumption.
4854
Fig. 7 shows the NMSEs of the measure models subject to BP, ESN, Simple Circle Reservoir (SCR) [25], ESN-DLRS with essentially the undesirable association traffic gathering as data, in like way shows the notion execution of the check models subject to ESN, ESN-DLRS with both the rough network traffic development and the denoised network traffic plan which is gotten by network traffic denoising computation reliant upon LPP as information (ESN-LPP and ESN-DLRS-LPP), naturally. As demonstrated by the information window size evaluation in Section V (C), for ESN-LPP and ESN-DLRS-LPP, set k1> 1, k2 > 1 for Dataset A, and set k1 > 1, k2> 4 for Dataset B. It might be seen from Fig. 8 that the NMSE of ESN-DLRS-LPP is more unnoticeable than that of various frameworks for different datasets and explicit figure steps. By changing the strange hold plan to the fundamental circle narrative construction, SCR is fixed up subject to ESN, along these lines the NMSE of SCR looks like ESN. In particular, with simply the rough network traffic blueprint as information, ESN and SCR have upheld check execution over BP, and ESN-DLRS performs better wandered from ESN and SCR. This is in light of the fact that ESN and SCR have more grounded nonlinear speculation limit than BP, while ESN-DLRS support neuronal receptiveness to improve this cutoff. With a close to figure model, ESN-LPP performs better stood apart from ESN, which shows that taking both the rough association traffic plan and the denoised network traffic development as data can feasibly improve assumption precision. With a close to data, ESN-DLRS-LPP has upheld figure execution over ESN-LPP, which shows that adding the denoised network traffic get-together to enter and enduring DLRS in the interim can intertwine the potential gains of both and develop assumption exactness.FIGURE 7. NMSE analysis. a) Dataset A. b) Dataset B.
6. Conclusion
The network traffic order has the compound characters such as flexibility, disorder, suitability and nonlinearity, which brings lots of difficulties to network traffic prediction. In this article, a novel network traffic prediction technique based on enhanced echo state network is proposed. Initially, to knob the variability and disorder of network traffic sequence, a network traffic denoising algorithm based on LPP is planned to denoise the raw network traffic sequence. Secondly, to knob the suitability and nonlinearity of network traffic sequence, a network traffic prediction model based on ESN-DLRS is constructed, which takes both the denoised network traffic sequence and the raw network traffic sequence as input. Finally, denoising presentation, input window size and prediction presentation are analyzed using two definite network traffic datasets. Simulation outcomes reveal that the proposed method performs enhanced than other parallel methods. Our future work will focus on the adaptive gaining of the best input window sizes of both the raw network traffic sequence and denoised network traffic sequence.
4855
References
1. X. Cao, Y. Zhong, Y. Zhou, J. Wang, C. Zhu, and W. Zhang, ‘‘Interactive temporal recurrent convolution network for traffic prediction in data cen- ters,’’ IEEE Access, vol. 6, pp. 5276–5289, Feb. 2018.
2. W. Huang, J. Zhang, S. Liang, and H. Sun, ‘‘Backbone network traffic pre- diction based on modified EEMD and quantum neural network,’’ Wireless Pers. Commun., vol. 99, no. 4, pp. 1569–1588, Apr. 2018.
3. L. Nie, D. Jiang, S. Yu, and H. Song, ‘‘Network traffic prediction based on deep belief network in wireless mesh backbone networks,’’ in Proc. Wire- less Commun. Netw. Conf. (WCNC), San Francisco, CA, USA, Mar. 2017, pp. 1–5.
4. X. Zhang, Z. Wu, J. Zhang, and J. Chen, ‘‘An adaptive network traffic prediction approach for LDoS attacks detection,’’ Int. J. Commun. Syst., vol. 31, no. 7, p. e3505, 2017.
5. Q.-F. Meng, Y.-H. Chen, Z.-Q. Feng, F.-L. Wang, and S.-S. Chen, ‘‘Nonlin- ear prediction of small scale network traffic based on local relevance vector machine regression model,’’ Acta Phys. Sinica, vol. 62, no. 15, pp. 1–7, 2013.
6. W. Habib, T. Sarwar, A. M. Siddiqui, and I. Touqir, ‘‘Wavelet denoising of multiframe optical coherence tomography data using similarity measures,’’ IET Image Process., vol. 11, no. 1, pp. 64– 79, Jan. 2017.
7. Q. Zhang, H. Liou, X. Zhou, and Y. Wang, ‘‘Wavelet soft-threshold denois- ing method of power quality signal,’’ High Voltage Eng., vol. 32, no. 1, pp. 99–101, 2006.
8. X. He and P. Niyogi, ‘‘Locality preserving projections,’’ in Proc. Adv. Neural Inf. Process. Syst., vol. 16, no. 1, pp. 186–197, 2002.
9. X.-X. Wen, X.-R. Meng, Z.-Q. Ma, and Q.-Y. Kang, ‘‘Network traffic anomaly detection based on local projection denoise and FSVDD,’’ Appl. Res. Comput., vol. 30, no. 5, pp. 1523–1526, 2013. 10. Y. Zhai et al., ‘‘A modified locality-preserving projection approach for hyperspectral image
classification,’’ IEEE Geosci. Remote Sens. Lett., vol. 13, no. 8, pp. 1059–1063, Aug. 2016. 11. S. XU and B. Zeng, ‘‘Network traffic prediction model based on auto-regressive moving average,’’
J. Netw., vol. 9, no. 3, pp. 97–102, 2014.
12. H. Z. Moayedi and M. A. Masnadi-Shirazi, ‘‘Arima model for network traffic prediction and anomaly detection,’’ in Proc. Int. Symp. Inf. Technol., Kuala Lumpur, Malaysia, Aug. 2008, pp. 1–6. 13. T. Karagiannis, M. Molle, M. Faloutsos, and A. Broido, ‘‘A nonstation- ary Poisson view of Internet
traffic,’’ in Proc. INFOCOM, Hong Kong, Mar. 2004, pp. 1558–1569.
14. W. Du, S. Y. S. Leung, and C. K. Kwong, ‘‘Time series forecasting by neu- ral networks: A knee point-based multiobjective evolutionary algorithm approach,’’ Expert Syst. Appl., vol. 41, no. 18, pp. 8049–8061, 2014.
15. Y. M. Zhang, Y. B. Zhang, and L. Z. Lin, ‘‘Notice of retraction application of genetic algorithm and RBF neural network in network flow prediction,’’ in Proc. IEEE Int. Conf. Comput. Sci. Inf. Technol., Chengdu, China, Jul. 2010, pp. 298–301.
16. S. Narejo and E. Pasero, ‘‘An application of Internet traffic prediction with deep neural network,’’ in Proc. Workshop Neural Nets, Vietri sul Mare, Italy, 2018, pp. 139–149.
17. Z. Duan, Y. Yang, K. Zhang, Y. Ni, and S. Bajgain, ‘‘Improved deep hybrid networks for urban traffic flow prediction using trajectory data,’’ IEEE Access, vol. 6, pp. 31820–31827, 2018.
18. H. Jaeger, ‘‘The ‘echo state’ approach to analyzing and training recurrent neural networks-with an erratum note,’’ German Nat. Res. Center Inf. Technol., Sankt Augustin, Germany, Tech. Rep. 148, Jan. 2001.
19. H. Jaeger and H. Haas, ‘‘Harnessing nonlinearity: Predicting chaotic sys- tems and saving energy in wireless communication,’’ Science, vol. 304, no. 5667, pp. 78–80, Apr. 2004.
20. H. Jaeger, ‘‘Tutorial on training recurrent neural networks, covering BPRT, RTRL, EKF and the ‘echo state network’ approach,’’ German Nat. Res. Center Inf. Technol., Sankt Augustin, Germany, Tech. Rep. 159, Jan. 2002.
21. A. M. Fraser and H. L. Swinney, ‘‘Independent coordinates for strange attractors from mutual information,’’ Phys. Rev. A, Gen. Phys., vol. 33, no. 2, pp. 1134–1140, Feb. 1986.
22. L. Cao, ‘‘Practical method for determining the minimum embedding dimension of a scalar time series,’’ Phys. D, Nonlinear Phenomena, vol. 110, nos. 1–2, pp. 43–50, 1997.
23. X.-C. Sun, H.-Y. Cui, R.-P. Liu, J.-Y. Chen, and Y.-J. Liu, ‘‘Modeling deter- ministic echo state network with loop reservoir,’’ Frontiers Inf. Technol. Electron. Eng., vol. 13, no. 9, pp. 689–701, 2012.
4856
and its application in network traffic prediction,’’ in Proc. IEEE 22nd Int. Conf. Comput. Supported Cooperat. Work Design, Nanjing, China, May 2018, pp. 689–694.25. H. Cui, C. Feng, Y. Chai, R. P. Liu, and Y. Liu, ‘‘Effect of hybrid circle reservoir injected with wavelet-neurons on performance of echo state network,’’ Neural Netw., vol. 57, no. 9, pp. 141–151, 2014.
26. P. Cortez, M. Rio, M. Rocha, and P. Sousa, ‘‘Multi-scale Internet traffic forecasting using neural networks and time series methods,’’ Expert Syst., vol. 29, no. 2, pp. 143–155, 2012.
27. Z. Tang, T. Peng, and W. Wang, ‘‘A small scale network traffic prediction algorithm based on correlation analysis and local scale least squares sup- port vector machine,’’ Acta Phys. Sinica, vol. 63, no. 13, pp. 49–58, 2014.
28. Y. Zhang and W. Guan, ‘‘Prediction of multivariable chaotic time series based on maximal Lyapunov exponent,’’ Acta Phys. Sinica, vol. 58, no. 2, pp. 756–763, 2009.
29. [29] D. E. Rumelhart, G. E. Hinton, and R. J. Williams, ‘‘Learning repre- sentations by back-propagating errors,’’ Nature, vol. 323, pp. 533–536, Oct. 1986.
