Proceedings of the 2012 2nd International Conference on Computer and Information Application (ICCIA 2012)

CQI Prediction for Shadow Fading in LTE-compatible GEO Mobile Satellite

Communications System

Yadan Zheng, Mingke Dong, Ye Jin, Jianjun Wu†

Institution of Advanced Communications, EECS, Peking University
Beijing, 100871, China
E-mail: [email protected], [email protected]

Abstract—CQI(Channel Quality Indicator) is an essential method and the prediction model’s selection are all
indicator for AMC(Adaptive Modulation and Coding) necessary to take into consideration.
technique in LTE. Due to the long delay of GEO satellite In Section II, GEO mobile satellite communication
channel, CQI prediction is necessary to ensure effective AMC system and the procedure and problem of CQI transmission
in LTE-compatible GEO mobile satellite communications in AMC are introduced in details. Section III gives the
system. The complete procedure and problem encountered concrete analysis for satellite channel quality, and various
when doing AMC in satellite system are introduced and the prediction models’ comparison. In Section IV, the proposed
difficulties of prediction are analyzed. In order to obtain prediction scheme is introduced, including the CQI data
meaningful and feasible CQI prediction results, a complete
approximation and the model selection strategy. Section V
prediction scheme is proposed. By using approximate CQI
data which is smoothed and with only shadow fading influence,
shows the simulation result. A summary for this paper can
various prediction models are compared. Taking both be seen in Section VI.
computation complexity and prediction performance into
II. LTE-COMPATIBLE GEO MOBILE SATELLITE SYSTEM
account, two models, ARIMA (Autoregressive Integrated
Moving Average) and Kalman filter are chosen to be applied A. Satellite Propagation Link Model
in this scheme. Simulation results show that the prediction
In GEO mobile satellite communication systems, there
performance is very well with the proposed method, which can
surely guarantee AMC performance. exist two transmission modes, single-hop and double-hop
mode. Typical application in satellite systems is double-hop
Keywords—CQI; AMC; GEO; ARIMA; Kalman mode, which will be the basic transmission model for this
paper’s analysis. The double-hop model comprises two
I. INTRODUCTION signal paths: one is that signal transfers from the starting
GEO mobile satellite system relies on high space UE, is forwarded by a GEO satellite, and finally arrives at
satellites, thus making it very different from terrestrial the destination ground station; and the other one is that
system. It has the characteristics of wide downlink signal transfers from the ground station, is forwarded by a
broadcast coverage, not sensitive to ground situations, wide GEO satellite similarly, and finally arrives at the destination
band, good communication quality and also many other UE. The complete process can be seen in Fig. 1. In GEO
advantages. system, typical delay of double-hop mode is 540ms[7]. This
AMC (Adaptive Modulation and Coding)[1,2] is paper mainly focuses on the influence of the delay to AMC
utilized in many terrestrial systems, which is an essential technique.
method to improve system's capacity and data transmission
efficiency. CQI (Channel Quality Indicator)[3] is calculated
by UEs according to the downlink reference signals, and is
sent back to eNodeB for AMC. But in satellite system,
satellite channels’ long propagation delay will cause trouble
for AMC. So CQI prediction is of great needs to guarantee
AMC performance.
Many existing prediction models are feasible for Figure 1. Double-hop model of satellite system
short-range prediction in terrestrial system, such as ARIMA
(Autoregressive Integrated Moving Average)[4], long-range B. Adaptive Modulation and Coding (AMC)
prediction proposed by Hallen[5], Kalman filter[6], etc. The
prediction parameter includes CQI[4], Doppler shift[6], and Adaptive modulation and coding (AMC) is an advanced
channel matrix. All these models do prediction according to technique to improve data rate and make full use of channel
the input data sequence’s autocorrelation. Their prediction capacity by applying different, adaptive order modulation
ability are limited when doing long-range prediction much schemes. It has been applied in many terrestrial systems.
longer than the sequence’ autocorrelation time. Therefore, UE calculates CQI based on the downlink reference signals’
in mobile satellite communication system, to guarantee SNR, and feed it back to eNodeB via uplink channel. After
AMC performance, the received CQI data’s application receiving CQI, eNodeB will allocate suitable downlink

Published by Atlantis Press, Paris, France.

© the authors
0623
 Proceedings of the 2012 2nd International Conference on Computer and Information Application (ICCIA 2012)

MCS based on the CQI and the resource distribution needed in this situation. According to Eq. (1), 540-step
conditions. Overall, AMC is a strictly close loop process. prediction will bring about huge computation quantity, also
with much longer process delay. When the prediction range
C. AMC in GEO Satellite System is long enough, the output will be the sequence’s average
In GEO mobile system, AMC is also needed to improve value, which loses the value of prediction[12].
system performance. The mainly difference is the channel 2) Long-range prediction[5]: Long-range prediction
environment, including the transmission delay, Doppler is firstly derived by Hellan. This method is based on AR
frequency and so on. model, and uses down sampled input data sequence as the
To overcome the long delay problem, effective prediction input. If the down-sample frequency fs satisfies
prediction is a sensible method to provide meaningful CQI Eq. (2), where fd is the maximum Doppler shift, the
data, but there is not much discussion about the CQI prediction can be done for a relative long-range with little
prediction in satellite environment so far. The challenge is information loss. The prediction range can be as long as
that the channel variation situation and prediction model’s several times of sequence autocorrelation time, up to around
characteristics must be considered jointly. one wavelength[5].
fs ≥ 2 fd (2)
III. SATELLITE CHANNEL QUALITY MODEL AND
PREDICTION MODEL This method has a same computation complexity with
ARIMA. But after down-sampling, the input data’s interval
A. Satellite Channel Quality Model becomes larger, for example, when sample interval is 20ms,
Models introduced by Loo[8], Lutz[9], and Corazza[10] the prediction step will be reduced to 27, which is much
are all commonly used satellite channel models. In this smaller than 540 steps for ARIMA model and the
paper, a statistic model introduced in [11] is employed, computation complexity decreased consequently. But in this
which is a detailed satellite channel model based on Loo[8]. paper, assume S band is utilized. f c = 2.4GHZ , and
[11] used three-state model to describe direct signal: LOS λ = c / f c = 0.125m , it is also very short compared with
conditions, moderate shadowing conditions, deep 540ms which makes difficulties for this method as well.
shadowing conditions. The transformation principle 3) Kalman filter: Kalman filter is often used to do
between these three states is a Markov chin. In each state, channel filtering, but actually, it is also an autoregressive
the detailed shadow fading situation varies under prediction method[6]. Kalman filter is based on linear
log-normal distribution, and has different distribution algebra and hidden Markov model. Future data state and
parameters. The small scale fading is under Rician value will be obtained gradually step by step. The model
distribution. coefficients are calculated with the autocorrelation
B. Prediction Model coefficients.
Kalman filter traces the hidden property of the input
One CQI data received by eNodeB can be received
every TTI (1ms) when doing periodic feedback. Different sequence and performs very well for predicting small scale
prediction models may have different prediction ability. fading in terrestrial system in short time distance. But the
Here are some commonly used prediction models which are high computation complexity due to the matrix
utilized in terrestrial system. The analysis focuses on the multiplication and the divergence problem when prediction
range becomes longer are the limitation for Kalman filter.
prediction performance, and calculation complexity.
1) Autoregressive Integrated Moving Average
(ARIMA): ARIMA is a widely used channel quality IV. CQI PREDICTION SCHEME IN LTE-COMPATIBLE
prediction model. In terrestrial mobile system has already MOBILE SATELLITE COMMUNICATIONS SYSTEM
applied this model to predict CQI in 1TTI (1ms) range[4]. A. Data Approximation
Here is definition of ARIMA (p,d,q) model:
∧ l −1 ∧ p+d q According to above, most of these prediction models’
z t (l ) = Et [ zt +l ] = φ j z t (l − j ) +  φ j zt +l − j − θ j at +l − j (1) parameters and prediction abilities rely on input sequence
j =1 j =l j =l autocorrelation. Even with Hallen’s long-range method,
Assume that received CQI state sequence is only several times of autocorrelation’s prediction can be
zt , zt −1 , zt − 2  . zˆt (l ) is the minimum mean square error achieved, which is far away from the required range. So the
prediction for l steps ahead of t. Assume μ as feasible and meaningful CQI prediction in GEO system
cannot be achieved if using the models directly.
sequence’s average value, and at stands for the interference As the sequence’s autocorrelation property can not
impulse. It has zt = zt − μ . This model has a relative low satisfy the prediction models’ requirement, approximation
computation complexity, and can output result in time. can be introduced in this situation. According to [11],
Due to the heavy dependence on sequence satellite channel is influenced by both large scale and small
autocorrelation, model prediction performance decreases scale fading. Large scale fading, which is mainly caused by
quickly when the prediction range is longer than shadow, varies much slower than small scale, which is
autocorrelation time. In this paper, the required prediction caused by multipath and Doppler Effect. If smoothing small
range is 540ms, as the TTI=1ms, 540-step predictions are scale fading away, the left data sequence’s autocorrelation
time will increase significantly [13,14]. From the simulation

Published by Atlantis Press, Paris, France.

© the authors
0624
 Proceedings of the 2012 2nd International Conference on Computer and Information Application (ICCIA 2012)

result in [13], the channel quality with only large scale

1
fading autocorrelation is much longer then that with both 0.9
autocorrelation after smoothing
autocorrelation before smoothing

large and small scale fading, thus making it possible to meet 0.8

the prediction models’ requirement. What’s more, since the 0.7

data is smoothed, say averaged, for a period of time, the left 0.6

data interval becomes longer, then the prediction step 0.5

number reduces. The smooth method can also be various, 0.4

such as moving average, exponential smoothing [15], etc. In 0.3

this paper, moving average is applied to smooth away the 0.2

small scale fading influence. 0.1

B. Prediction Model Selection 0

0 0.5 1 1.5
Time(s)
2 2.5 3

According to the analysis above, we can see that Figure 2. Comparison of correlation time between slow
long-range prediction algorithm is basically an ARIMA variation and fast variation CQI data
model with less input data. While the approximate CQI data
is already down-sampled by smoothing, this model is B. Prediction Result with Approximate Data
worthless in this situation; Finally, ARIMA and Kalman The prediction results are obtained with different
filter are left to be considered. parameters in Table I, including different UE speed. Since
As said before, Kalman filter is a kind of autoregressive the prediction model has relationship with data
model as well, and the matrix’s coefficients can be autocorrelation time, the parameter UE speed is selected to
calculated using ARIMA model. Generally speaking, the ensure the data autocorrelation time is shorter than, almost
prediction performance of ARIMA and Kalman filter can equal to, longer than the prediction range. According to
be almost the same in some situations. [14], we smoothed the data for every 2.5 meters.
As to the multi-step prediction, ARIMA needs to
calculate data one step by one step, and when a new TABLE I. SIMULATION PARAMETERS
measured data arrives, the prediction has to be done again.
Elevation (o) UE Speed (m/s) Prediction Range (ms)
On the other hand, Kalman filter needs to do matrix 80 10,50 500
multiplication first and stores the result, and when a new
measured data arrives, the prediction needs to do just one 0.2
ARIMA(2,0,0)
matrix multiplication. So actually, the computation 0.18 ARIMA(15,0,1)
plain ARIMA
complexity of these two models must be judged by concrete 0.16 Long-range Prediction
mean
situations. When prediction step is small, the computation 0.14
Kalman(2,0,0)
Kalman(15,0,0)

complexity can be viewed almost the same, but Kalman 0.12

filter has a faster output as the coefficients are already

RMSE

0.1

stored. When prediction step is large, the matrix 0.08

multiplication complexity cannot be endured anymore and 0.06

the divergence cannot be neglected. 0.04
A prediction scheme is proposed as follows:
0.02
1) Approximation: Calculate the approximate CQI
0
data using moving average or other smooth method. 5 6 7 8 9 10
SNR(dB)
11 12 13 14 15

2) Prediction model selection: When UE travels

Figure 3. Prediction performance of different models
slowly (autocorrelation time much longer than prediction
range), select low order Kalman filter or ARIMA model. Fig. 3 gives the simulation comparison between low
When UE travels fast (autocorrelation time shorter than or order ARIMA model, high order ARIMA model, low order
same with prediction range), select high order ARIMA Kalman Filter, high order Kalman filter using approximate
model. data, and high order ARIMA model, long-range prediction,
3) Using approximate data to do prediction. mean of data, using original data. The result shows that the
RMSE of ARIMA model and long-range prediction with
V. SIMULATION AND ANALYSIS original data are the same with the RMSE of giving only
mean of data. This means that predicting with original data
A. Data Approximation Result is improper in this paper’s situation. In the following, we
The approximate data is compared with the original data use ARIMA model and Kalman filter to do predict.
in previous work [16]. The comparison shows that the
difference is quite small that we can smooth away the small
scale fading influence without much data loss. Meanwhile,
the sequence’s autocorrelation increases significantly in Fig.
2, which greatly benefits the prediction. The proposed
approximation method is feasible and meaningful to be
applied in AMC process in satellite mobile system.

Published by Atlantis Press, Paris, France.

© the authors
0625
 Proceedings of the 2012 2nd International Conference on Computer and Information Application (ICCIA 2012)

proposed complete CQI prediction scheme can surely

0.15

0.14
ARIMA(2,0,0) guarantee AMC performance.
ARIMA(15,0,1)
Kalman(2,0)

0.12
Kalman(15,0)
ACKNOWLEDGMENT
0.1 This work is supported by the National High-Tech
0.08
Research and Development Program of China (863
RMSE

Program), No. 2012AA01A506. Corresponding author:

0.06
Jianjun Wu; phone: +86-10-62752848; e-mail:
0.04
[email protected].
0.02
REFERENCE
0
5 6 7 8 9 10 11 12 13 14 15
[1] 3GPP TS 36.213, “Evolved Universal Terrestrial Radio Access
SNR(dB) (E-UTRA); Physical layer procedures”, URL: www.3GPP.org.
Figure 4. Prediction performance when UE travels at 10m/s [2] J. Shen, S. Suo, H. Quan, 3GPP Long Term Evolution: Principle and
System Design, P eople Post Press, Beijing, 2008, Chaps5.4.3.
Simulation is done for low-order ARIMA, high-order
[3] S. Sesia, I. Toufik, M. Baker, LTE-UMTS Long Term Evolution From
ARIMA, AR based low order Kalman filter, AR based high Theory to Practice, Chinese version, People Post Press, 2009, pp.
order Kalman filter for different parameters. 160-165.
From Fig. 4, when UE travels slow, both ARIMA and [4] Y. Shang, X. Chen, “Channel Quality Indicator Prediction and
Kalman filter of high order and low order perform almost Compensation Method and System”, Patent No. CN101958765A, Jan.
the same with each other. As the prediction step is small, 2011.
Kalman filter is enough to ensure the prediction [5] Duel-Hallen, A., Hu, S., Hallen, H. “Long-range prediction of fading
performance with low complexity. signals”, Signal Processing Magazine, IEEE, Vol. 17 , May 2000,
pp.62-75.
0.15 [6] A. A. Heidari1, K. Khandani, D. McAvoy, “Adaptive modelling and
0.14 ARIMA(2,0,0)
ARIMA(20,0,1)
long-range prediction of mobile fading channels”, IET Commun.,
Kalman(2,0) 2010, Vol. 4, Iss. 1, pp. 39–50
0.12 Kalman(20,0)
[7] K. Wang, Z. Gao, J. Du, J. Guo, Y. Cheng, J. Wu, “A Handoff
0.1 Method for single Hop Satellite Mobile Communication System”,
Sixth Annual Conference of the satellite communications new
business and new technology, Beijing, Mar. 2010, pp382-389.
RMSE

0.08

[8] C. Loo, “A Statistical Model for a Land Mobile Satellite Link”, IEEE
0.06
Transactioks On Vehicular Technology, Vol. VT-34, No. 3, Aug. 1985,
0.04
pp. 122-127.
[9] Lutz, D. Cygan, M. Dippold, F. Dolainsky, W. Papke, “The Land
0.02
Mobile Satellite Communication Channel-Recording, Statistics, and
Channel Model”, IEEE Transactions On Vehicular Technology, Vol.
0
5 6 7 8 9 10
SNR(dB)
11 12 13 14 15 40. No. 2, May 1991, pp. 375-386.
[10] G. E. Corazza, F. Vatalaro, "A statistical model for land mobile
Figure 5. Prediction performance when UE travels at 50m/s satellite channels and its application to nonstationary orbit systems,"
IEEE Transactions on Vehicular Technology, Vol. 43, No. 3, Aug.
When UE travels faster, as Fig. 5 shows, low order 1994, pp. 738-742.
model becomes worse than high order model, and AR based [11] F.P. Fontan, M. Vazquez-Castro; C.E. Cabado, J.P. Garcia, E. Kubista,
Kalman filter performs worse than ARIMA. “Statistical Modeling of the LMS Channel” IEEE Transactions On
Vehicular Technology, Vol. 50, No. 6, Nov. 2001 pp. 1549-1567.
VI. CONCLUSION [12] H. Cheng, P. N. Tan, J. Gao, J. Scripps, “Multi Step ahead Time
Series Prediction”, In Proc. of the Pacific-Asia Conf on Knowledge
The necessity and problem encountered of CQI Discovery and Data Mining, Vol. 3918, 2006, pp765-774.
prediction in AMC in LTE-compatible GEO mobile [13] F. Fontan, M.A. Vazquez-Castro, “S-band LMS propagation channel
satellite communication system are discussed in this paper. behaviour for different environments, degrees of shadowing and
In order to obtain feasible and rational prediction result to elevation angles”, IEEE Transactions On Broadcasting, Vol. 44, No.
guarantee AMC performance, a complete prediction scheme 1, Mar. 1998, pp.40-76.
is proposed in this paper by using approximate CQI data [14] P. Taaghol, R. Tafazolli, “Correlation model for shadow fading in
and proper prediction models, The approximation by land mobile satellite systems”, ELECTRONICS LETTERS, 17th, Vol.
33, No. 15, Jul. 1997, pp. 1287-1289.
smoothing small scale fading makes it possible to do
[15] Brown, R. G., Meyer, R. F., “The fundamental theorem of
prediction with existing prediction models. Comparisons of exponential smoothing,” Operations Research, Vol. 9, No. 5, Sep. -
realization complexity and performance between various Oct., 1961, pp. 673-687.
prediction models are provided as well. Simulation results [16] Y. Zheng, S. Ren, X. Xu, Y. Si, M. Dong, J. Wu, “Channel Quality
show that the prediction performance is very well with the Approximation Using Large Scale Model in GEO Mobile Satellite
proposed method. A conclusion can be drawn that the System”, The International Workshop on Electromagnetism and
Communication Engineering, Neimenggu, Mar. 2012

Published by Atlantis Press, Paris, France.

© the authors
0626