Predition Method for Channel Quality Indiator in LEO mobile Satellite Communiations Yadan Zheng *, Mingke Dong *, Wei Zheng *, Ye Jin *, Jianjun Wu * * Institution of Advaned Communiations, Peking University, Beijing, P.R. China, 100871 zhengyadan@pku.edu.n, just@pku.edu.n Abstrat CQI(Channel Quality Indiator) is an essential indiator for AMC(Adaptive Modulation and Coding) tehnique in terrestrial mobile system. Due to the long delay, and fast movement of LEO satellite, CQI predition is neessary to ensure effetive AMC in LEO mobile satellite ommuniation system. The omplete proedure and problem enountered when doing AMC in satellite system are introdued and the diffiulties of predition are analyzed. In order to obtain meaningful and feasible CQI predition results, a omplete predition sheme is proposed. For different evaluation angles and different UE speeds, Hallen s long-range predition model and a modified smooth-arima (Autoregressive Integrated Moving Average) are hosen to be applied in this sheme. Simulation results show that the predition performane is very well with the proposed method, whih an surely guarantee AMC performane. Keywords AMC, CQI Predition, LEO Satellite, Long Range Predition I. INTRODUCTION LEO mobile satellite system is very promising due to shorter transmission delay and smaller pathloss ompared with GEO satellite system. It an obtain global overage with several satellites. But the signal transmission situation is quite different from terrestrial system. The ombination of LEO satellites and terrestrial mobile ommuniation is a quite big hallenge. AMC (Adaptive Modulation and Coding) [1],[2] is utilized in many terrestrial systems suh as LTE and HSPA. It is an important method to improve system's apaity and data transmission effiieny. CQI (Channel Quality Indiator) [3] is alulated by UEs aording to the downlink referene signals, and is sent bak to enodeb to be an indiator for AMC. But in LEO mobile satellite ommuniation system, satellite hannels propagation delay will ause trouble for AMC. The signals arrying CQI are all outdated when arriving at enodeb. So CQI predition is of great needs to guarantee AMC performane. Many existing predition models are feasible for short-range predition in terrestrial system, suh as ARIMA (Autoregressive Integrated Moving Average)[4], long-range predition proposed by Hallen[5], Kalman filter[6], et. The predition parameter inludes CQI[4], Doppler shift[6], hannel matrix. All these models do predition aording to the input data sequene s autoorrelation. Their predition ability are limited when doing long-range predition muh longer than the sequene autoorrelation time. Therefore, in LEO mobile satellite system, to guarantee AMC performane, the reeived CQI data s appliation method and the predition model s seletion are all neessary to be taken into onsideration. In this paper, in order to solve the enountered problem when doing AMC in satellite system, several existing predition models are ompared. A existing model and a proposed modified model are seleted to obtain meaningful and feasible CQI predition results. In Setion II, LEO mobile satellite ommuniation system and the proedure and problem of CQI transmission in AMC are introdued in details. Setion III gives the onrete analysis for satellite hannel quality, and existing predition models limitation. In Setion IV, the proposed predition sheme is introdued, inluding introduing a modified ARIMA model, and the model seletion for various transmission situation. Setion V shows the simulation result. A summary for this paper an be seen in Setion VI. II. LEO MOBILE SATELLITE COMMUNICATIONS SYSTEM A. Satellite Propagation Link Model In LEO mobile satellite ommuniation systems, there are several satellites serving the system, working alone or together. In this paper, we only onsider the situation that a UE is served by one LEO satellite without inter-satellite ooperation. In one LEO satellite system, double-hop is the main transmission mode. The double-hop mode omprises two signal paths: one is from the starting UE, forwarded by a satellite, and finally arrives at the destination ground station; the other one is from the ground station, forwarded by a satellite similarly, and finally arrives at the destination UE. The omplete proess an be seen in Fig. 1. For example, in Iridium system, as the orbit altitude is relatively low, the double-hop transmission delay is about 10ms. In LEO system, typial delay of this double-hop model is 20-80ms. The influene of delay is the interest in this paper. ISBN 978-89-968650-0-1 799 January 27 ~ 30, 2013 ICACT2013
B. Adaptive Modulation and Coding (AMC) Adaptive modulation and oding (AMC) is an advaned tehnique to enhane data rate and make full use of hannel apaity by applying different, adaptive modulation order shemes. It has been applied in many terrestrial systems. In terrestrial LTE system, UE reports CQI to enodeb and enodeb uses it to do AMC. UE alulates CQI based on the downlink referene signals SNR, and feedbak to enodeb via uplink hannel. After reeiving CQI, enodeb will alloate suitable downlink MCS based on the CQI and the resoure distribution onditions. Overall, AMC is a stritly lose loop proess. Figure 1. Double-hop model of satellite system C. AMC in LEO Satellite System In LEO mobile satellite ommuniation system, AMC is also needed to improve system performane. The main diffiulties for AMC are the fast variation of hannel environment and the long transmission delay. The CQI data reeived by ground station for seleting proper MCS are outdated, and an not reflet real hannel quality. In order to guarantee AMC effetiveness, predition is neessary for LEO mobile satellite system. As the predition of hannel quality of LEO satellite is not that muh disussed in previous art, this paper wants to offer a heuristi idea. onsidered small sale fading and large sale fading for hannel quality analysis. [10] used three-state model to desribe diret signal: LOS onditions, moderate shadowing onditions, deep shadowing onditions. The transformation priniple between these three states is a Markov hin. In eah state, the detailed shadow fading situation varies under log-normal distribution, and with different distribution parameters for eah state. The small sale fading is of Riian distribution. The signal s variation an referene to Fig. 2 [10]. It an be seen that hannel quality s variation inludes fast variations, slow variations, and very slow variations, whih are orresponding to small sale fading in Riian distribution, large sale fading in log-normal distribution, and large sale states in a Markov hin. B. Predition Model In LEO mobile satellite ommuniation system, ground station will reeive CQI data feed baked by UE. One CQI data an be reeived every TTI (1ms) when doing periodi feedbak. The reeived CQI sequene is employed to do predition. Aording to the hannel quality model above, CQI sequene omprises regularity and randomness. Predition an be viewed as a filter to give a future result by using the inner regularity of the sequene and avoiding the influene of randomness as muh as possible. Different predition models may use different properties of the sequene and may have different predition ability. In general, the short-range predition performane is muh better than long-range predition. The limitation of a predition model is more and more obvious as the predit-range getting longer, whih may ause great impat on the AMC deision. III. SATELLITE CHANNEL QUALITY MODEL AND PREDICTION MODEL A. Satellite Channel Quality Model As introdued before, in LEO satellite system, the key hannel parameters needed to be taken into onsideration inludes: Doppler frequeny shift aused by satellite, Doppler spread aused by UE movement, the double-hop transmission delay. In this paper, we assume that the large Doppler frequeny shift an be deteted aurately and orreted without deviation. Models introdued by Loo[7], Lutz[8], and Corazza[9] are all ommonly used satellite hannel models. In this paper, a statisti model introdued in [10] is employed, whih is a detailed satellite hannel model based on Loo[7]. The model Figure 2. Different rates of hange of the various reeived signal omponents Two predition models are onsidered in this paper. ARIMA [4] is a ommonly used predition model in terrestrial system. Hallen s long-range predition [5] is introdued for the speifi situation in this paper. 1) Autoregressive Integrated Moving Average (ARIMA) ARIMA is a widely used hannel quality predition model. In terrestrial mobile system has already applied this ISBN 978-89-968650-0-1 800 January 27 ~ 30, 2013 ICACT2013
model to predit CQI in 1TTI (1ms) range[4]. Here is definition of ARIMA (p,d,q) model: l 1 p+ d zt() l = E[ z ] = φ z ( l j) + φ z θ a (1) t t+ l j t j t+ l j j t+ l j j= 1 j= l j= l Assume that reeived CQI state sequene is z, z, z t t 1 t 2 In Eq.1, ˆ () t z l stands for the minimum mean square error predition for l steps ahead of t. Assume μ as z t sequene s average value, and at stands for the z = z μ. This model has been interferene impulse. It has t t utilized in terrestrial with good performane4, and has a relative low omputation omplexity, and an output result in time. Due to the heavy dependene on sequene autoorrelation, the predition performane dereases quikly when the predition range is beoming longer than autoorrelation time. In this paper, the needed predition range is 40ms, as the TTI=1ms, 40 step preditions are needed under this situation. Aording to Eq. (1), 40-step predition will bring about huge omputation quantity, also with muh longer proess delay. Meanwhile, when the predition range is long enough, the output of this model will be the sequene s average value, whih loses the value of predition[11]. 2) Long-range predition[5] Long-range predition is firstly derived by Hellan. This method is based on AR model, and uses down sampled input data sequene as the predition input. If the down sample frequeny fs satisfies Eq. (2), where fd is the maximum Doppler shift, the predition an be done for a relative long-range with little information loss. The predition range an be as long as several times of sequene autoorrelation time, up to around one wavelength [5]. fs 2 fd (2) This method has a relative same omputation omplexity with ARIMA. But after downsampling, the input data s interval beomes larger, for example, 5ms, then the predition step will redue to 8, whih is muh smaller than 40 steps for ARIMA model and the omputation omplexity dereased onsequently. C. Channel Autoorrelation Analysis Aording to the analysis above, these predition models parameters and predition ability rely on the input sequene autoorrelation time. Therefore, it is neessary to analysis the CQI data, or hannel quality autoorrelation time before predition. In LEO mobile satellite system, the maximum Doppler spread is as follows: v fdm = f osθ (2) q Where v is the speed of UE, is the speed of light, f is the frequeny of arrier, θ is the angle between UE movement diretion and the radio waves inident diretion. The signal orrelation time is: 0.423 0.423 0.423 t = f = v = dm vf os f os θ (3) θ If S-band is used in this system, assume it has f=2.4ghz, then: 0.423 t = 8v osθ, 0.423 L =. (5) 8osθ A few orrelation time and maximum Doppler shift are alulated in Table 1. TABLE 1. CORRELATION PARAMETERS θ ( o ) V(m/s) fdm(hz) t (ms) 10 40 80 3 23.64 17.89 10 78.77 5.37 30 236.44 1.79 3 18.38 23.01 10 61.30 6.90 30 183.83 2.30 3 4.17 101.50 10 13.89 30.45 30 41.67 15 Aording to Table 1, influened by different UE speed and evaluation angle, the signal s orrelation time and length are quite various. IV. CQI PREDICTION METHOD IN LEO MOBILE SATELLITE COMMUNICATIONS From the analysis above, it seems that when θ is large and v is small, there is possibility to obtain deades milliseonds orrelation time for predition. Common predition model, suh as ARIMA is able to do good predition. When θ is small and v is large, the orrelation time is too short for diretly predition. A proper predition method should be seleted arefully for the latter situation. In the following disussion, both predition performane and omputation omplexity are taken into onsideration. When the data autoorrelation time is longer than predition range, ARIMA an be applied to do a 40-step predition. But taking the omplexity into onsideration, sine ARIMA does predition step by step, the 40-step predition brings about so muh omplexity and proess delay to the system. If Hallen s long-range predition is applied in this situation, predition step an be minimized. For example, when evaluation angle equals 80 o, and UE speed equals 3m/s, the minimum sample ISBN 978-89-968650-0-1 801 January 27 ~ 30, 2013 ICACT2013
rate is about 8HZ. Assume that the sample rate is 100HZ, than predition an be done with the data interval equals 10ms, and a 4-step is enough to obtain good performane. When the data autoorrelation time is shorter than predition range, as introdued before, Hallen s long-range predition model an do predition in this situation. But also taking omplexity into onsideration, when evaluation angle equals 40 o, and UE speed equals 30m/s, the minimum sample rate is about 366HZ, whih resulting the downsampled data interval is less than 3ms. If sample rate fs equals 500HZ, than a 20-step predition has to be done, thus also bringing about high omplexity and delay. So when fs is large, it s also improper to use Hallen s long-range predition model. Atually, downsampling is some kind of losing data inner disipline. If we want to inrease the sample rate to lower omplexity, the inner disipline must be made use of to guarantee the predition performane. In this paper, we propose to average the data sequene in the sample interval, not just take the data on sample point. The average an take advantage of the non-sample point data s information and smoothing away the noise influene to some extent, then enhaning the predition performane. This new method is referred as smooth-arima model in the following simulation. The predition sheme is proposed as follows: 1. Model Seletion: 1) When evaluation angle θ is large and UE speed v is small: The data autoorrelation time is longer than predition range: Hallen s long-range predition is applied. 2) When evaluation angle θ is small and UE speed v is large: When data autoorrelation time is a little shorter than predition range: Hallen s long-range predition is applied. When data autoorrelation time is muh shorter than predition range: is applied. 2. Determine the predition model s parameters aording to the evaluation angle and UE speed. 3. Do predition using the seleted model. V. SIMULATION AND ANALYSIS The predition results are obtained with different parameters in Table 2, inluding different elevations, different UE speed. Sine the predition model has relationship with data autoorrelation time, the parameter UE speed is seleted to ensure the data autoorrelation time is shorter than, almost equal to, longer than the predition range. TABLE 2. SIMULATION PARAMETERS Elevation ( o ) UE Speed (m/s) Predition Range (ms) 40,80 3, 30 40 Fig. 3 gives the predition result for autoorrelation time longer than predition range and the sample rate is larger than minimum sample rate limitation. It shows that all the three predition models do predition very well. 2 1 Figure 3. 2 1 ARIMA Predition performane when θ=80 o, v=3m/s, sample rate = 100HZ ARIMA 8 9 10 11 12 13 14 15 Figure 4. Predition performane when θ=40 o, v=3m/s, sample rate = 100HZ Fig. 4 gives the predition result for autoorrelation time a little shorter than predition range and the sample rate is larger than minimum sample rate limitation. It shows that the performane of downsampling or smoothing every 10ms is better than 40-step ARIMA model. Fig. 5 gives the predition result for autoorrelation time muh shorter than predition range and the sample rate is larger than minimum sample rate limitation. It shows that the performane of downsampling and smoothing are almost the same. Fig. 6 gives the predition result for autoorrelation time muh shorter than predition range and the sample rate is smaller than minimum sample rate limitation. It shows that the performane of downsampling drop signifiantly, and smooth ARIMA are almost the same as large sample rate. From the simulation above, it is obvious that our proposed sheme an do predition on every speifi situation. ISBN 978-89-968650-0-1 802 January 27 ~ 30, 2013 ICACT2013
2 1 Figure 5. 2 1 Figure 6. Predition performane when θ=40 o, v=30m/s, sample rate = 500HZ Predition performane when θ=40 o, v=30m/s, sample rate = 100HZ VI. CONCLUSION The problem enountered when doing CQI predition for AMC in LEO mobile ommuniation system is disussed in this paper. In order to obtain feasible and rational predition result to guarantee AMC performane, by using different predition model and different parameters, a omplete predition sheme is proposed in this paper. By onsidering both the performane and omplexity, a new smooth-arima model is proposed in this paper for the situation that evaluation angle is small and UE speed is large. Hallen s long-range predition model and a new proposed smooth-arima model are seleted to do predition in different situation. Simulation results show that the predition performane is very well with the proposed method. A onlusion an be drawn that the proposed omplete CQI predition sheme an surely guarantee AMC performane. ACKNOWLEDGMENTS This work is supported by the National Siene Foundation of China (Grant No. NFSC #61071083) and the National High-Teh Researh and Development Program of China (863 Program), No. 2012AA01A506. Corresponding author: Jianjun Wu; Phone: +86-10-62752848; E-mail: just@pku.edu.n. REFERENCE [1] 3GPP TS 36.213, Evolved Universal Terrestrial Radio Aess (E-UTRA); Physial layer proedures, URL: www.3gpp.org. [2] Shen, J., Suo, S., Quan, H., 3GPP Long Term Evolution: Priniple and System Design, People Post Press, Beijing, 2008, Chaps5.4.3. [3] Sesia, S., Toufik, I., Baker, M., LTE-UMTS Long Term Evolution From Theory to Pratie, Chinese version, People Post Press, 2009, pp. 160-165. [4] Shang, Y., Chen, X., Channel Quality Indiator Predition and Compensation Method and System, Patent No. CN101958765A, Jan. 2011. [5] Duel-Hallen, A., Hu, S., Hallen, H. Long-range predition of fading signals, Signal Proessing Magazine, IEEE, Vol. 17, May 2000, pp.62-75. [6] Heidari1, A. A., Khandani, K., MAvoy, D., Adaptive modelling and long-range predition of mobile fading hannels, IET Commun., 2010, Vol. 4, Iss. 1, pp. 39 50 [7] Loo, C., A Statistial Model for a Land Mobile Satellite Link, IEEE Transatioks On Vehiular Tehnology, Vol. VT-34, No. 3, Aug. 1985, pp. 122-127. [8] Lutz, Cygan, D., Dippold, M., Dolainsky, F.; Papke, W., The Land Mobile Satellite Communiation Channel-Reording, Statistis, and Channel Model, IEEE Transations On Vehiular Tehnology, Vol. 40. No. 2, May 1991, pp. 375-386. [9] Corazza, G. E., Vatalaro, F., "A statistial model for land mobile satellite hannels and its appliation to nonstationary orbit systems," IEEE Transations on Vehiular Tehnology, Vol. 43, No. 3, Aug. 1994, pp. 738-742. [10] Fontan, F.P.; Vazquez-Castro, M.; Cabado, C.E.; Garia, J.P.; Kubista, E., Statistial Modeling of the LMS Channel IEEE Transations On Vehiular Tehnology, Vol. 50, No. 6, Nov. 2001 pp. 1549-1567. [11] Cheng, H., Tan, P. N., Gao, J., Sripps, J., Multi Step ahead Time Series Predition, In Pro. of the Paifi-Asia Conf on Knowledge Disovery and Data Mining, Vol. 3918, 2006, pp765-774. Yadan Zheng was born in Hebei provine, China. She reeived the bahelor degree in eletroni information siene and tehnology from Peking University, Beijing, China, in 2010. Sine 2010, she has been an postgraduate student in Institution of Advaned Communiations, Peking University. Her researh interests are in the area of satellite mobile ommuniations and hannel oding. Jianjun Wu reeived his B.S., M.S. and Ph.D. degree from Peking University, Beijing, P.R.China, in 1989, 1992 and 2006, respetively. Sine 1992, he has joined the Shool of Eletronis Engineering and Computer Siene, Peking University, and has been appointed as an assoiate professor sine 2002. His researh interests are in the areas of satellite ommuniations, wireless ommuniations, and ommuniations signal proessing. ISBN 978-89-968650-0-1 803 January 27 ~ 30, 2013 ICACT2013