Joural of Commuicatios Vol., No. 2, December 2 A New Method for Traffic Predictio i Emergig Mobile Networks Yuia Jia, Beili Wa, Liag Liag, Qia Zhao, Yu Zhag, ad Liag Tag2 College of Commuicatio Egieerig, Chogqig Uiversity, Chogqig, Chia 2 Chia Mobile Group Chogqig Co., Ltd. Chogqig, Chia Email: {yuia, wabeili, liagliag, zhaoqia, yzhag}@cqu.edu.c; tagliag2@cq.chiamobile.com Abstract With the icreasig popularity of mobile devices ad applicatios, emergig mobile etwork traffic exhibits special characteristics i temporal scale e.g., there is a scale variace betwee the etwork traffic o weekdays ad o weekeds. Although most existig methods have bee applied to data traffic predictio, few of them take such characteristic ito cosideratio. I this paper, by usig real data i mobile etworks, we adopt the etropy theory to reveal that the duratio of time-series give for predictio does't always have a positive impact ad that the ucorrelated precedig time-series also deteriorates the predictio accuracy. I view of this, partitioig the etwork traffic predictio ito weekdays ad weekeds perspective, we propose a method to predict the data traffic. Fially, we evaluate the proposed method through predictig the data traffic for a future time accordig to the historical data traffic i a real mobile etwork. I compariso with the work based o ARMA (Auto Regressive Movig Average) method, our proposed method ca reduce the Mea Absolute Percetage Error (MAPE) by 3.7% ad 43.8% o weekdays ad weekeds predictio, respectively. patters i 3G etworks. Other methods, such as Auto Regressive Itegrated Movig Average (ARIMA) [], eural etworks method [6], Kalma filterig method [7] ad wavelet method [8] etc, have also bee adopted to predict future traffic i 3G etworks. However, the methods metioed above usually predict the etwork traffic based o the precedig time-series ad the predictio accuracy deteriorates quickly as the duratio of time-series icreases. I this respect, authors i [9]-[] exploited the etropy theory to aalyze the traffic predictability i mobile etworks ad cocluded the appropriate duratio of precedig time-series used for traffic predictio. Alog with the evolutio of mobile etworks, we fid that the correlatio betwee etwork traffic ad social aspects i emergig mobile etworks is more complex tha that i traditioal 2G or 3G etworks, e.g., there is a scale variace betwee the etwork traffic o weekdays ad etwork traffic o weekeds with respect to user behaviors aspect. Therefore it is importat to aalyze this tred ad exploit it to improve the performace of etwork traffic predictio. Cosiderig that the social behaviors of users have a great ifluece o the etwork traffic. Specifically, from social behavior aspect, users o weekeds seem to sped more time o mobile devices ad mobile applicatios tha weekdays. There is a scale variace betwee the data traffic o weekdays ad data traffic o weekeds. Therefore, i this paper, we take such characteristic ito cosideratio, ad adopt etropy theory to demostrate that the predictio performace will be iflueced by the combiatio of temporal duratio. Partitioig the traffic predictio ito weekdays perspective ad weekeds perspective, we propose a predictio method cocetratig o combiig the data from correlated days. Moreover, as the days o weekeds are limited, we cosider usig a factor to compesate the variace of the data traffic o weekdays to improve the predictio o weekeds. Numerical results show that our method ca idetify the optimal temporal combiatio as prior iformatio ad predict the obective data traffic accurately. Therefore, our work provides a essetial uderstadig o traffic predictio i future etworks. The rest of this paper is orgaized as follows: Sectio II aalyzes the etwork traffic predictability i real mobile etworks. Sectio III reviews the model traiig. Idex Terms Mobile etworks, data traffic predictio, etropy theory, time-series, real data I. INTRODUCTION With the rapid developmet of mobile etworks, from the secod geeratio (2G) to the third geeratio (3G) ad the fourth geeratio (4G), the smart devices ad mobile applicatios have bee icreasig rapidly. Toward the fifth geeratio (G) of mobile etworks, umerous devices will be oied i ad the demad of etwork traffic will costatly rise. Therefore, precise traffic predictio is expected to esure the ormal operatio of the etwork ad achieve high resource utilizatio [], [2]. Whe 3G starts, the popularity of services leads to dramatic chages i etwork traffic characteristics. Cosequetly, some methods of traffic predictio have bee proposed i this era [3]-[8]. Authors i [3] took the otio of self-similarity that took place i 3G etworks ito accout ad proposed a method to estimate the mai parameters of etwork traffic. I [4], the authors proposed a order-k Markov model to predict traffic Mauscript received Jue, 2; revised December 7, 2. This work is supported by the Natioal High-tech R&D Program of Chia (863 Program) uder grat No. 2AAA76. This work is also sposored i part by Hitachi, Ltd. Correspodig author email: yuia@cqu.edu.c. doi:.272/cm..2.947-94 2 Joural of Commuicatios 947
Joural of Commuicatios Vol., No. 2, December 2 I Sectio IV, we itroduce the data traffic predictio. Numerical results are preseted i Sectio V, ad Sectio VI gives the coclusios. II. THEORETICAL ANALYSIS ON DATA TRAFFIC PREDICTION Iformatio theory [2] has played a key part i illustratig the geerality of iformatio cotet. Etropy theory, as its basic idea, offers a precise defiitio of iformatio cotet ad gives a effective method to measure its ucertaity. I this sectio, we adopt the etropy approach to gauge the data traffic predictability accordig to certai prior iformatio from historical data. A. Data Collectio We collect the data traffic from mobile etworks i Chia Mobile Group Chogqig Co., Ltd., Chia. The data set covers a large area of about 8243 km2 ad it covers more tha oe millio mobile users. Fig. shows the architecture of data collectio. The followig list shows the elemets from the leaves to the root of the tree: odes, core equipmets, et stream, etwork maagemet system, optical splitter ad DPI (Deep Packet Ispectio) aalysis system. Each ode, referrig to a computer or other devices, coects to the core equipmet through commuicatio lik. The data traffic is stored i et stream. Network maagemet system is used for resource moitorig ad aalysis. Durig the trasmissio, a part of data traffic is copied for aalysis by optical splitter. DPI aalysis system works as a filterig to collect ad aalysis the statistical iformatio ad our data traffic is collected from this system. Fig.. The architecture of data collectio i Chia Mobile Group Chogqig Co., Ltd. B. Theoretical Aalysis o Predictio The collected dataset icludes all types of applicatios (HTTP, P2P, IM etc.) i both rural ad urba areas i Chogqig, Chia. Each hour records the volume of total applicatios. After obtaiig the dataset, a processig procedure is coducted to ormalize the data traffic with the followig equatio: actualdata ( t) mi Data Nor _ Data( t) max Data mi Data where actualdata(t) deotes the collected dataset, ad Nor_Data(t) deotes the ormalized dataset. maxdata () ad midata are the maximal ad miimal data traffic values of the dataset, respectively. Fig. 2. Data traffic for two weeks. Probability of each level Mo.Tue.Wed.Thur. Fri. Sat. Su. Mo.Tue.Wed.Thur. Fri. Sat. Su. Mo. Time of Two Weeks (day).6.4.2 Mo. Su. 2 3 4 Levels of quatitized traffic Fig. 3. The probability of each level o Moday ad Suday. To ease the followig aalysis, the data durig a certai period i is quatized ito levels, which represet five differet volume of data traffic, from the lowest to the highest oe. Thus, with the ormalized dataset, the correspodig data traffic probabilities ca be obtaied. For example, Fig. 2 shows the data traffic two weeks. I this figure, a clear diural patter of data traffic is illustrated. For all days, there is a scale variace betwee weekdays ad weekeds. However, there is little scale variace amog days i weekdays ad days i weekeds respectively. This kid of patter ca be explaied that user behaviors have relatioship to data traffic. Fig. 3 depicts the probability of each level o Moday ad Suday i oe week. The etropy is employed to measure the ucertaity of evets [2]. Let X be a discrete radom variable with possible values {x,..., x } ad the correspodig probability is defied as H( X ) p( x )log p( x ) (2) i i b i where b is the base of the logarithm. Uless otherwise specified, we commoly take the logarithm to base 2. Accordig to the theory of etropy, the data traffic probability heavily depeds o the data traffic characteristics. For example, the data traffic i each week which is show i Fig. 2 would have a scale variace betwee weekdays ad weekeds. From social behavior perspective, users i weekeds would sped more time o mobile devices ad mobile applicatios tha weekdays. Give that the data traffic demad is highly liked to user behaviors. That's to say, the variety of etropy values is also more or less related to user behaviors. Therefore, we use etropy theory to describe the ucertaity of data traffic i mobile etworks. I additio, the accuracy of 2 Joural of Commuicatios 948
Joural of Commuicatios Vol., No. 2, December 2 data traffic predictio relies o the adopted model, but also requires a certai quatity of prior iformatio to reduce ucertaity [2]. I other words, predictio performace will be improved with the icrease of the amout of prior iformatio. Here, we take the coditioal etropy of two radom variables ad ito X {x, x2, x } Y { y, y2, y } i, p( y ) (3) p( xi, y ) Takig the etwork traffic o Moday ad o Suday for example, we calculate the etropy values of differet coditios. As idicated i Table I, amog the dataset, both the coditioal etropies of Moday ad Suday etwork traffic decrease rapidly, eve though the radom etropy of Moday or Suday is comparatively larger. Moreover, for Moday, the coditioal etropy with etwork traffic o weekdays (e.g., the preset Moday, or the preset Moday ad Tuesday) decreases more rapidly tha that o weekeds (e.g., the preset Saturday). For Suday, the coditioal etropy with etwork traffic o weekeds (e.g., the preset Suday, or the preset Saturday ad Suday,) decreases more rapidly tha that o weekdays (e.g., the preset Moday). This fidig is also applied to other days. Therefore, we coclude that, the differet combiatios of time-series have differet impacts o traffic predictability. I particular, the precedig time-series o weekeds has little cotributio to the performace of weekdays predictio, ad vice versa. I the ext sectio, we preset a method to determie the time-series combiatio from both weekdays perspective ad weekeds perspective. m yweekdays (t ) wi fi (t ) where fi (t ) (t=,2,...,. i=,2,...,m) reflects the combiatio sequece i o iterval t. yweekdays (t ) (t=,2,, ) deotes the combiatio of fi (t ). The combiatio weight vector for m combiatio sequeces is w (w, w2, wm )T, ad it satisfies the followig coditios: Next Mo. Next Su. Calculatio coditioed o Etropy Noe 2.2672 et w () w (6) where et=(,,...,). The preset week Next week Mo. Tue. Wed. Thur. Fri. Sat. Su. Mo. Tue. Wed. Thur. Fri. Sat. Su. yweekdays(t) TABLE I: CORRELATIONS BETWEEN SEVEN DAYS AFTER COMPENSATION Day (4) i red ict H ( X Y ) p( xi, y ) log Weekdays Perspective As for the variatio of data traffic o weekdays, we use the combiatio of previous days as the prior iformatio to predict the data traffic of days i ext weekdays, as show i Fig. 4 (a). First, we use the temporal combiatio of correlated days i a week to describe the day (e.g., Moday) we wat to predict i the ext week. The temporal combiatio of correlatio ca be built up as equatio (4). A. H ( X Y ) is To p cosideratio. The coditioal etropy defied as obtaied to idicate the volume of etwork traffic i itervals for the give week. We use the data from the first week to traiig the predictio model, tha we adopt the traiig model to predict the traffic i ext week. I this sectio, we maily itroduce the model traiig of our method. Pweekdays(t) (a) The preset week Next week Mo. Tue. Wed. Thur. Fri. Sat. Su. Mo. Tue. Wed. Thur. Fri. Sat. Su. The preset Mo..24 The preset Sat..72 The preset Mo. ad Tue..38 The preset Mo. Tue. ad Wed.. The preset Mo. Tue. ad Sat..3 Noe.8962 (b) The preset Mo..223 The preset Su..76 The preset Mo. ad Su..729 Fig. 4. (a) The flow diagram of the predictio method o weekdays perspective (i.e., for the data traffic predictio o Moday). (b) The flow diagram of the predictio method o weekeds perspective (i.e., for the data traffic predictio o Suday). The preset Sat. ad Su..99 The preset Mo. Sat. ad Su..438 III. CF(t) yweekeds(t) Pweekeds(t) Weekeds Perspective Cosiderig two days o weekeds for a week, we propose a factor to compesate the previous data traffic o weekdays. After the recostructio of the data, we use the combiatio of recostructed data to forecast the day (e.g., Suday) i ext weekeds, as Fig. 4 (b) show. The model ca be built up by the followig steps. B. MODEL TRAINING As etwork traffic is geerally cotiuous, we treat the volume of etwork traffic i time iterval [t, t+δt'] as a etry i the method. Therefore, legth sequece is 2 Joural of Commuicatios edict To pr 949
Joural of Commuicatios Vol., No. 2, December 2 Step oe: we propose a factor to compesate the small scale variace betwee weekdays ad weekeds. The factor i each iterval ca be calculated by equatios (7) ad (8). f () t pt () (7) f () t where fs () t (t=,2,...,) is the sequece of the predicted day o weekeds, ad f () t (=,2,...,. t=,2,..., ) is the combiatio sequece o weekdays ( days). pt () is the ratio of f () t ad fs () t. CF() t is the compesate factor o iterval t. s CF( t) ( p() p(2)... p( )) (8) Step two: we use the compesate factor to modify the data traffic o weekdays. The ew sequece o weekdays is obtaied by equatio (9). f ( ) ( ) ( ) ew _ t CF t f t (9) where f () _ t idicates the ew combiatio sequece ew compesated by factor CF() t. The, we use the ew sequeces to update the prior data o weekdays. Step three: we use compesated sequeces to predict the data traffic o weekeds as described by equatio (). m y ( t) w f ( t) () weekeds ew _ where y () t (t=, 2,..., ) is the combiatio of data weekeds traffic from correlated days that has bee compesated. I order to get the best effect of combiatio, we propose the optimizatio model that is described below. By solvig the model, we could get the optimal combiatio weight vector. For weekdays: max R weekdays For weekeds: max R weekeds t ( f ( t) f ) ( y ( t) y ) k A weekdays A () ( f ( t) f ) ( y ( t) y ) k A weekdays A t t t T ew st.. w ( f ( t) f ) ( y ( t) y ) s B weekeds B s B weekeds B t t (2) (3) ( f ( t) f ) ( y ( t) y ) T ew st.. w (4) where fk () t ad fs () t (t=,2,...) are the correspodig data traffic of the predicted day i last weekdays ad weekeds, respectively. f A is the average value of fk () t ad f deotes the average value of f () t (t=,2,...). B y A ad y B are the average values of yweekdays () t ad yweekeds () t, respectively. R is adopted to weigh the correlatio betwee two variables. I this work, we use R to weigh the correlatio betwee variable f () t weekdays ad variable yweekdays () t ad Rweekeds s to weigh the correlatio betwee variable fs () t ad variable yweekeds () t. Whe R, it idicates that the two variables are positively related. Whe R, it idicates that the two variables are egatively related. Whe R, it idicates that there is o liear correlatio betwee the two variables. Whe R, it idicates that there is total liear correlatio betwee the two variables, that is, the fuctio relatioship. Whe R ad R, there is a liear correlatio betwee the two variables. Ad if R is closer to, the liear relatioship betwee the two variables is strog close; if R is closer to, the liear correlatio betwee the two variables is weak close. After determiig the optimal combiatio weight vector, we use the equatios () ad (6) to get the predicted value. () weekdays Pweekdays ( t) ( yweekdays ( t) fk ( t)) () 2 Pweekeds ( t) ( yweekeds ( t) fs ( t)) (6) 2 P t ad P () t deote the predicted data traffic weekeds value o weekdays ad weekeds, respectively. IV. MOBILE DATA TRAFFIC PREDICTION I this paper, we cosider the data traffic i oe week, ad deote the time poits of a day by [,, 2,..., 23]. So there are 24 27 time poits. As is explaied i Sectio II, the value of is 24 ad the value of m is 6. Here we aalyze the optimal combiatio of data traffic sequeces o weekdays ad weekeds, respectively. A. Predictio o Weekdays We assume that the data traffic o ext Moday is to be predicted. The combiatio of data traffic i this week except for Moday is used to predict it. Accordig to the predictio model metioed above, we should fid the optimal umber of combiatio ad proper weight vector. I order to get the optimal umber of combiatio, we cosider the possible combiatio from to 6, accordig to the degree of correlatio betwee other 6 days. First, we calculate the degree of correlatio betwee 7 days. As is depicted i Table II, the data traffic o Moday has a high degree of correlatio with that o weekdays. The, we take 6 trials to get the probable predictio of Moday i ext week. For each trial we take the data traffic from k 2 Joural of Commuicatios 9
Joural of Commuicatios Vol., No. 2, December 2 the days that are most correlated to Moday ito cosideratio. The compositio of combiatio for each trial is listed as follows: : {Tue.}; Trial #2: {Tue., Thur.}; Trial #3: {Tue., Wed., Thur.}; Trial #4: {Tue., Wed., Thur., Fri.}; Trial #: {Tue., Wed., Thur., Fri., Sat.}; : {Tue., Wed., Thur., Fri., Sat., Su.}. After the determiatio of the compositio, we take the data traffic o Moday as the actual data fk () t, ad the combiatio of data traffic from correlated days deotes as yweekdays () t. The we calculate the best combiatio weight factor for each trial through equatio () ad (2). Table III depicts the best weight factor for each trial. TABLE II: CORRELATIONS BETWEEN SEVEN DAYS Correlatios Mo. Tue. Wed. Thur. Fri. Sat. Su. Mo..997.996.997.992.96.93 Tue..997.998.999.992.962.939 Wed..996.998.999.993.93.924 Thur..997.999.999.992.9.929 Fri..992.992.993.992.969.934 Sat..96.962.93.9.969.988 Su..93.939.924.929.934.988 TABLE III: THE BEST WEIGHT FACTOR FOR EACH TRIAL ON MONDAY PREDICTION Factor value w w 2 w 3 w 4 w w 6 Null Null Null Null Null Trial #2.4998.2 Null Null Null Null Trial #3.3333.3332.333 Null Null Null Trial #4.2679..972.348 Null Null Trial #.289.233.3872.933. Null.282..647.3..8 origial Trial #2 Trial #3 Trial #4 Trial # : 6: 2: 8: 23: Fig.. Predictio of 6 trials o Moday. With the obtaied weight factor, we ca easily calculate the predicted data o Moday i the ext week by equatios (). Fig. shows the predictio of 6 trials. As it demostrates, the curves for 6 trials fit well with the origial oe. B. Predictio o Weekeds We first aalyze the ew degree of correlatio with the data traffic o weekdays which is updated by equatio (9). As is depicted i Table IV, the correlatio betwee Suday ad weekdays i a week has bee much improved. Accordig to the correlatio, we listed the followig 6 trials: : {Tue.}; Trial #2: {Tue., Wed.}; Trial #3: {Tue., Wed., Thur.}; Trial #4: {Mo., Tue., Wed., Thur.}; Trial #: {Mo., Tue., Wed., Thur., Fri.}; : {Mo., Tue., Wed., Thur., Fri., Sat.}. After that, we take the data traffic o Suday as the actual data fs () t, ad let the combiatio of data traffic from correlated days be y () t. The best combiatio weekeds weight factor for each trial ca be calculated by equatio (3) ad (4). The, we use equatio (6) to calculate the predicted data o Suday i the ext week with the obtaied weight factor listed i Table V. Fig. 6 shows the predictio of 6 trials ad it illustrates that the predicted values of 6 trials close to the origial data o Suday. TABLE IV: CORRELATIONS BETWEEN SEVEN DAYS AFTER COMPENSATION Correlatios Mo. Tue. Wed. Thur. Fri. Sat. Su. Mo..99.99.99.992.98.998 Tue..99.998.998.99.982.999 Wed..99.998.999.993.989.999 Thur..99.998.999.993.989.999 Fri..992.99.993.993.992.99 Sat..98.982.989.989.992.988 Su..998.999.999.999.99.988 TABLE V: THE BEST WEIGHT FACTOR FOR EACH TRIAL ON SUNDAY PREDICTION Factor value w w 2 w 3 w 4 w w 6 Null Null Null Null Null Trial #2.4999. Null Null Null Null Trial #3.3333.3333.3334 Null Null Null Trial #4.294.63.227.3236 Null Null Trial #.69.27.467.296.793 Null.24.23.884.228.27.236 origial Trial #2 Trial #3 Trial #4 Trial # : 6: 2: 8: 23: Fig. 6. Predictio of 6 trials o Suday V. NUMERICAL RESULTS I this sectio, we aalyze the performace of our predictio. To begi with, we use the predicted data mius the actual data, as equatio (7) shows: e( t) predicteddata ( t) actualdata ( t) (7) 2 Joural of Commuicatios 9
Joural of Commuicatios Vol., No. 2, December 2 Fig. 2, we ca clearly see that our method has better performace tha covetioal ARMA method. Fially, we calculate the MAPE of the methods metioed above o the predictio i a week. Table VI depicts, our method ca reduce the MAPE by 3.7% ad 43.8% o weekdays ad weekeds predictio compared with ARMA method. where e(t ) (t=, 2,..., 23) is the error of predictio. The, the curve of predictio error is obtaied. Fig. 7 depicts the curves of predictio error of 6 trials o Moday, ad Fig. 8 depicts the curves of predictio error of 6 trials o Suday. Fig. 7 ad Fig. 8 show the predictio errors of 6 trials are obviously differet..48.46.44. -. - : MAPE Predictio Error (Gbps).42.4 Trial #2 Trial #3 Trial #4 Trial #.38 6: 2: 8: 23:.2..2 MAPE Predictio Error (Gbps) Fig. 9. MAPE of 6 trials o Moday. Fig. 7. Predictio error of 6 trials o Moday. -. The optimal combiatio.36 Trial #2 Trial #3 Trial #4 Trial # Trials of Time-Series Combiatio Trial #2 Trial #3 Trial #4 Trial # - :.. The optimal combiatio. 6: 2: 8: 23: Fig.. MAPE of 6 trials o Suday. Fig. 8. Predictio error of 6 trials o Suday MAPEweekdays f k (t ) Pweekdays (t ), t, 2 t f k (t ) (8) MAPEweekeds f s (t ) Pweekeds (t ), t, 2 t f s (t ) (9) To evaluate the accuracy of each trial, we cosider the mea absolute percetage error (MAPE) as the performace idex. The idexes of weekdays ad weekeds are give by equatio (8) ad (9), respectively. Actual data i a week ARMA method Proposed method Mo. Tue. Wed. Thur. Fri. Sat. Time of Week (day) Su. Fig.. The compariso of our method ad ARMA method o the predictio of the traffic i a week. 4 As is explaied i Sectio IV, the value of is 24 ad there are 24 time poits i our predictio. We ca get the MAPE for 6 trials by equatio (8) or equatio (9). Fig. 9 ad Fig. depict the MAPE of each trial o Moday ad Suday predictio, respectively. I view of this, we coclude that the optimal combiatio as prior iformatio o weekdays predictio is the two days that most correlated to the predicted day i the preset weekdays. O weekeds predictio, the optimal combiatio is the updated days that most correlated to the predicted day i the preset weekeds. Based o that fidig, we ca coduct the predictio i a week. Cosiderig ARMA method have good performace i predictio up to date [3], we compare our method with ARMA method o the predictio i a week, as show i Fig.. Fig. 2 shows the compariso of our method ad ARMA method o Moday ad Suday predictio. For the large umber of vertical coordiates, the differeces betwee predicted data ad actual data have bee cofused. To better illustrate the performace of predictio, we amplify the local part of the figure. From 2 Joural of Commuicatios Trial #2 Trial #3 Trial #4 Trial # Trials of Time-Series Combiatio 2 Actual data o Moday ARMA method Proposed method 8 6 4 2 : 6: 2: 8: 23: 8: 23: (a) - : Actual data o Suday ARMA method Proposed method 6: 2: (b) Fig. 2. The compariso of our method ad ARMA method: (a) o Moday predictio, ad (b) o Suday predictio. 92
Joural of Commuicatios Vol., No. 2, December 2 TABLE VI: THE MAPE OF OUR METHOD AND ARMA METHOD MAPE Proposed method ARMA method Weekdays.38.842 Weekeds.473.86 VI. CONCLUSIONS This paper has preseted a ovel data traffic predictio method from social behavior aspect. The proposed method cosiders the characteristics of temporal variatio from data traffic i a week. To detect the distict characteristics, we have adopted etropy theory to aalyze the effect of differet temporal precedig duratio o data traffic predictio. The study shows that the precedig duratio i weekeds has little cotributio to the performace of weekdays predictio, ad the precedig duratio i weekdays leads to the same result o weekeds predictio. Therefore, our method has partitioed the data traffic predictio ito weekdays ad weekeds perspective ad focused o the combiatio of data from correlated days. I additio, we have coducted extesive simulatio experimets based o the actual data traffic i previous week to predict the data traffic i future week. Numerical results depict that our method ca idetify the optimal temporal combiatio as prior iformatio for predictio, ad provide accurate data traffic predictio afterwards. ACKNOWLEDGMENT This work is supported by the Natioal High-tech R&D Program of Chia (863 Program) uder grat No. 2AAA76. This work is also sposored i part by Hitachi, Ltd. REFERENCES [] P. Demestichas, A. Georgakopoulos, D. Karvouas, K. Tsagkaris, et al., G o the Horizo: Key Challeges for the Radio-Access Network, IEEE Vehicular Techology Magazie, vol. 8, o. 3, pp. 47-3, 23. [2] R. Sattirau ad H. D. Schotte, Reliability Modelig, Aalysis ad Predictio of Wireless Mobile Commuicatios, i Proc. IEEE VTC 24 Coferece, Seoul, Korea, May 8-2, 24. [3] A. Kredzel, Y. Koucheryavy, J. Haru, ad S. Lopati, Method for Estimatig Parameterers of 3G Data Traffic, i Proc. IEEE ICC, Paris, Frace, Jue 2-24, 24. [4] K. Zhag ad L. Cuthbert, Traffic Patter Predictio i Cellular Networks," i Proc. IEEE ICCS, Guagzhou, Chia, November 9-2, 28. [] Y. Shu, M. Yu, J. Liu, ad O. Yag, Wireless Traffic Modelig ad Predictio Usig Seasoal ARIMA Models, i Proc. IEEE ICC Coferece, Achorage, Alaska, May -,23. [6] L. Zhu, L. Qi, K. Xue, ad X. Zhag, A ovel BP eural etwork model for traffic predictio of ext geeratio etwork, i Proc. IEEE ICNC Coferece, Tiai, Chia, Augest 4-6, 29. [7] M. Achir, Y. M. Ghamri-Doudae, ad G. Puolle, Predictive resource allocatio i cellular etworks usig kalma filters, i Proc. IEEE ICC Coferece, Achorage, Alaska, May -, 23. [8] X. Wag, Y. Re, ad X. Sha, WDRLS: A wavelet-based olie predictor for etwork traffic, i Proc. IEEE GLOBECOM Coferece, Sa Fracisco, USA, December -, 23. [9] R. Li, Z. Zhao, X. Zhou, ad H. Zhag, Eergy savig scheme i radio access etworks via compressive sesig-based traffic load predictio, Trasactios o Emergig Telecommuicatios Techologies, vol. 2, o. 4, pp. 2-29, Apr. 24. [] X. Zhou, Z. Zhao, R. Li, Y. Zhou, ad H. Zhag, The predictability of celluler etworks traffic, i Proc. IEEE ISCIT, Gold Coast, Australia, October 4, 22. [] R. Li, Z. Zhao, X. Zhou, J. Palicot, ad H. Zhag, The predictio aalysis of cellular radio access etwork traffic: From etropy theory to etworkig practice, IEEE Comuuicatios Magazie, vol. 2, o. 6, Ju. 24. [2] T. M. Cover ad J. A. Thomas. (26). Elemets of Iformatio Theory, Wiley-Itersciece. [Olie]. Available: http://www.amazo.com/elemets-iformatio-theory- Telecommuicatios-Processig/dp/476296 [3] M. Wei ad K. Kim, Itrusio detectio scheme usig traffic predictio for wireless idustrial etworks, Joural of Commuicatios ad Networks, vol. 4, o. 3, pp. 3-38, 22. Yuia Jia received his B.S. degree from Nakai Uiversity, Chia, ad his M.E. ad Ph.D. degrees i Egieerig from Osaka Uiversity, Japa, i 999, 23 ad 26, respectively. From 26 to 22, he was with Cetral Research Laboratory, Hitachi, Ltd., where he egaged i research ad developmet o wireless etworks, ad also cotributed to LTE/LTE-Advaced stadardizatio i 3GPP. He is ow a professor at the College of Commuicatio Egieerig, Chogqig Uiversity, Chogqig, Chia. He is the author of more tha 6 published papers, ad 2 grated patets. His research iterests iclude radio access techologies, mobile etworks, ad IoT. Dr. Jia has wo several prizes from idustry ad academia icludig the IEEE Vehicular Techology Society Youg Researcher Ecouragemet Award, the IEICE Paper Award, the Yokosuka Research Park R&D Committee YRP Award, ad the Top Youg Ivetors of Hitachi. Moreover, he was a research fellowship award recipiet of Iteratioal Commuicatio Foudatio i 24, ad Telecommuicatios Advacemet Foudatio Japa i 2. Beili Wa received her B.E. degree i School of Iformatio Egieerig from Southwest Uiversity of Sciece ad Techology (SWUST), Chia, i 22. She is curretly workig toward her M.E. degree i Iformatio ad Commuicatio Egieerig i Chogqig Uiversity. Her research iterests iclude traffic model ad traffic predictio for mobile etworks, with emphasis o emergig mobile etworks. Liag Liag received her B.E. ad M.E. degrees from the Southwest Uiversity of Sciece ad Techology (SWUST), Chia, i 23 ad 26, respectively, ad the Ph.D. degree i commuicatio ad iformatio system from the Uiversity of Electroic Sciece ad Techology of Chia (UESTC) i 22. She is curretly a lecturer i College of Commuicatio Egieerig, Chogqig Uiversity, Chogqig, Chia. Her research iterests iclude wireless commuicatio ad optimizatio, gree radio, ad wireless sesor etworks. 2 Joural of Commuicatios 93
Joural of Commuicatios Vol., No. 2, December 2 Qia Zhao received her B.E degree i Commuicatio Egieerig from Qufu Normal Uiversity, Chia, i 23. She is curretly workig toward her M.E degree i Electroic ad Commuicatio Egieerig from Chogqig Uiversity. She maily egages i the Iteret of Thigs. Yu Zhag received her B.E. degree i Commuicatio Egieerig from Chogqig Uiversity, Chia, i 24. She is curretly workig toward her M.E. degree i Iformatio ad Commuicatio Egieerig i the same uiversity. Her research iterests iclude radio resource maagemet ad optimizatio for mobile etworks, with emphasis o quality-of-experiece provisio. Liag Tag received his M.E degree from Chogqig Uiversity, Chia, i 29. He is curretly workig toward his Ph.D. degree i Iformatio ad Commuicatio Egieerig i the same uiversity. Moreover, his is a director i data bearig etwork ceter, Chia Mobile Group Chogqig Co., Ltd., Chogqig, Chia. He maily egages i wireless commuicatio ad radio resource maagemet for mobile etwork. 2 Joural of Commuicatios 94