1686 IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, VOL. 25, NO.

9, SEPTEMBER 2014

Fading Channel Prediction Based on Combination

of Complex-Valued Neural Networks and
Chirp Z-Transform
Tianben Ding and Akira Hirose, Fellow, IEEE

Abstract— Channel prediction is an important process for
channel compensation in a fading environment. If a future
channel characteristic is predicted, adaptive techniques, such as
pre-equalization and transmission power control, are applicable
before transmission in order to avoid degradation of commu-
nications quality. Previously, we proposed channel prediction
methods employing the chirp z-transform (CZT) with a linear
extrapolation as well as a Lagrange extrapolation of frequency-
domain parameters. This paper presents a highly accurate
method for predicting time-varying channels by combining a
multilayer complex-valued neural network (CVNN) with the
CZT. We demonstrate that the channel prediction accuracy of
the proposed CVNN-based prediction is better than those of the
conventional prediction methods in a series of simulations and
experiments.
Index Terms— Channel prediction, chirp z-transform
(CZT), complex-valued neural networks (CVNNs), fading,
frequency domain, high-capacity spatial-domain multiple
access (HC-SDMA).
I. I NTRODUCTION
P ERFORMANCE of mobile communications often suffers
from various fading phenomena. To reduce the adverse
effect, there are some techniques such as diversity reception
and error-correction coding. Other countermeasures are pre-
equalization and transmission power control. Such adaptive
techniques require channel prediction for the channel changes
in time [1], [2]. The prediction also leads to high-capacity
spatial-domain multiple access (HC-SDMA) [3] and multiuser
multiple-input multiple-output [4] systems.
There exist several methods in channel prediction. We can
predict channel characteristic using, e.g., linear prediction
based on linear extrapolation [5], [6] and autoregressive (AR)
model focusing on serial autocorrelation in time [7]–[10].
In addition, we may also employ recurrent neural network for
nonlinear prediction [11], [12]. Besides, we can use super-
resolution methods with high-precision spectral estimation
such as ROOT-MUSIC [13] and ESPRIT [14] to extrapolate
channel characteristics. These methods can be categorized into
Manuscript received January 30, 2013; revised January 30, 2014; accepted
February 9, 2014. Date of publication March 4, 2014; date of current version
August 15, 2014.
The authors are with the Department of Electrical Engineering and
Information Systems, University of Tokyo, Tokyo 113-8656, Japan (e-mail:
ding_tei@eis.t.u-tokyo.ac.jp; ahirose@ee.t.u-tokyo.ac.jp).
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TNNLS.2014.2306420
time-domain (TD) and frequency-domain methods. However,
each of them has problems more or less in its calculation cost
or prediction precision or both. For instance, the accuracy of
linear or nonlinear prediction in TD is not high in general
because the channel observed in the TD often yields irregular
variation. On the other hand, frequency-domain methods, such
as ROOT-MUSIC and ESPRIT, require large calculation cost
for estimating the channel.
To realize a low-cost and high-precision prediction, we
previously proposed the use of chirp z-transform (CZT) [15]
in Jakes model [1] for high-accuracy extraction of frequency-
domain channel parameters [16]. We also demonstrated
that the performance is greatly improved by combining
the CZT with linear extrapolation [17], [18] or Lagrange
extrapolation [19]. However, in a series of practical exper-
iments, we found that these methods fail to predict future
channels in some situations.
In this paper, we propose new prediction methods based
on complex-valued neural networks (CVNNs) [20] to predict
the channel characteristics of respective paths represented
by channel parameters estimated by the CZT. The CVNNs,
dealing with a complex-valued signal as an entity, is a powerful
framework in the signal processing area [21], [22]. It also
brings a flexible processing to the prediction of channel
characteristics, which is also complex-valued. We assume a
multilayer (ML) and a real-time-recurrent-learning (RTRL)
CVNN. Simulation results demonstrate that the CZT-combined
ML-CVNN (CZT-ML-CVNN)-based prediction can predict
the channel with a very high accuracy, and keep this high
accuracy even in situations, where conventional predictions
fail. In particular, we find that the proposed ML-CVNN-based
prediction presents a higher performance in the case of low
frequency-resolution transform, which is a necessary condition
to restrain the calculation cost. Furthermore, a series of exper-
iments also demonstrate that the proposed ML-CVNN-based
prediction is more accurate than the conventional predictions
in actual fading channels.
This paper is organized as follows. Section II describes
the combination of frequency and time domains for channel
prediction as well as conventional prediction methods. In
Section III, we propose the prediction methods based on
CVNNs. Then, Sections IV and V present their performance in
simulations and experiments, respectively. Finally, Section VI
provides the conclusions.

2162-237X © 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

Authorized licensed use limited to: Rohde & Schwarz. Downloaded on February 18,2020 at 13:00:00 UTC from IEEE Xplore. Restrictions apply.
DING AND HIROSE: FADING CHANNEL PREDICTION BASED ON COMBINATION OF CVNNs AND CZT 1687

II. C HANNEL P REDICTION W ITH C OMBINATION OF

F REQUENCY-D OMAIN AND T IME -D OMAIN T ECHNIQUES
A. Channel Model and CZT
According to the Jakes model, a fading channel can be
modeled as the summation of sinusoids at a receiver, which are
the multipath rays caused by scattering and reflection. Each
sinusoid can be characterized by a set of path parameters such
as amplitude am , Doppler frequency f m , and phase shift φm .
The channel characteristic c(t) as a function of time t is the
summation of M complex signal paths and expressed as

M
c(t) = am e j (2π f m t +φm ) . (1)
m=1

The Doppler frequency f m is given by

fc
fm = vcos(ϕ) (2)
c
where v and c are the mobile user’s speed and the speed of
light, respectively, fc is the carrier frequency, and ϕ is the
incident radio wave angle with respect to the motion of the
mobile user.
We can utilize channel prediction methods to reduce the
influence of fading. There are two types of methods for
channel prediction, which predict the future channel charac-
teristic from the past channel information: 1) TD method,
which predicts c(t) in TD directly and 2) frequency-domain
method, which passes the past channel characteristic to the fre-
quency domain to separate multiple paths, and then extrapolate
them.
Previously, we proposed the combination of the time and
frequency-domain treatments. As shown in Fig. 1, we apply
the Fourier transform to the past channel information (t ≤ 0)
to observe a Doppler frequency spectrum. The peaks represent
the Doppler-shift paths. We estimate the path parameters of
each path by finding the peak amplitude âm (t) and its Doppler Fig. 1. Conceptual illustrations representing (a) CZT to convert the past
shift frequency fˆm (t) in the Doppler spectrum, and the phase channel information into the frequency domain for the extraction of the
frequency-domain channel parameters, namely, (b) amplitude am and Doppler
shift φ̂m (t) in the phase spectrum presented in Fig. 1(c), frequency f m , and (c) phase shift φm , respectively.
corresponding to the Doppler frequency peak. Note that the
accuracy of the path estimation depends on the frequency
resolution. fˆm (−n/2), φ̂m (−n/2), are different from the parameters at the
In general, the Doppler shift frequencies fˆm (t) are much prediction period (t > 0). Thus, we have to predict the chan-
smaller than the carrier frequency. Therefore, we need to nel characteristic c̃(t) based on these estimated information
observe precisely the frequency spectrum in the region close in TD.
to the zero frequency. A conventional nonzero-padding-FFT
requires a long observation period. Even a zero-padding-FFT
increases the calculation cost. To overcome these issues, we B. Linear and Lagrange-Based Predictions
use the CZT to obtain the Doppler spectrum. The CZT realizes Fig. 1(a) is also a schematic diagram showing our previous
low calculation cost as well as a smooth frequency-domain methods. As previous works, we proposed a linear predic-
interpolation, resulting in a high estimation accuracy of the tion method and a prediction method based on Lagrange
peak frequencies fˆm (t) [15], [16]. extrapolation that predict the parameters, ãm (t), f˜m (t), φ̃m (t)
There is, however, a problem even in the CZT to obtain (t ≥ 0), to predict the channel characteristic c̃(t). The linear
a future channel characteristic (t > 0) from the past path prediction extrapolates ãm (t), f˜m (t), φ̃m (t) (t ≥ 0) using the
parameters. As shown in Fig. 1(a), we apply a CZT window parameters at t = −n/2 and −n, namely, âm (−n), fˆm (−n),
to obtain the Doppler spectrum. Since the CZT window φ̂m (−n) and âm (−n/2), fˆm (−n/2), φ̂m (−n/2) [17], [18]. The
function, such as Hann window, has length of n, extracts the Lagrange-based prediction extrapolates ãm (t), f˜m (t), φ̃m (t)
channel parameters with emphasis on the central part. There- (t ≥ 0) based on Lagrange extrapolation using the parameters
fore, the latest estimated path parameters, namely, âm (−n/2), at t = −n/2, −n and −3n/2 [19]. Then, according with the

Authorized licensed use limited to: Rohde & Schwarz. Downloaded on February 18,2020 at 13:00:00 UTC from IEEE Xplore. Restrictions apply.
1688 IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, VOL. 25, NO. 9, SEPTEMBER 2014
Fig. 2. Schematic diagram of the proposed channel prediction method based on the CZT and CVNNs prediction of frequency-domain channel characteristics
cm of dominant paths.
Jakes model, we obtained channel characteristics of respective
paths as
˜
c̃m (t) = ãm e j (2π f m t +φ̃m ) . (3)
Summation of the multiple complex paths c̃m (t) gives the
channel characteristic c̃(t).
In the process of phase prediction, a long interval of window
center generates a large phase increase/decrease, which leads
to violation of the sampling theorem, and fails to predict φ̃m (t)
properly. To solve this problem, we set the windows with an
interval shorter than n/2 as shown in Fig. 1(a), and avoid the
increase/decrease of φm beyond ±π.
These predictions present high performances in most cases,
and the Lagrange-based prediction is more accurate than the
linear prediction in some numerical experiments. However, we
found that these predictions cannot predict future channels in
some situations. For example, the accuracy of these predictions
deteriorates when scatterers’ distance has particular values.
In these cases, the linear and Lagrange-based predictions fail
to improve the performance or, in several cases, degrade the
communications quality.
C. Prediction Method Based on AR Model
There is a general prediction method based on an AR model.
The AR-model-based method predicts respective characteristic
c̃m (t) using past channel characteristics ĉm (t) presented by the
CZT as

IAR
c̃m (t) = dk (t)ĉm (t − k) (4)
k In this paper, we propose two different CVNN-based pre-
where IAR is the number of past channel samples, and dk (t)
is the AR coefficient. The AR coefficient dk (t) is updated
according to the steepest descent method, which minimizes
the difference of predicted c̃m (t) and a teacher signal ĉm (t).
III. P ROPOSAL OF CVNN-BASED P REDICTION
A. Signals in CVNNs
In this paper, we propose channel prediction methods based
on combination of CVNNs and CZT. The CVNNs that deal
Authorized licensed use limited to: Rohde & Schwarz. Downloaded on February 18,2020 at 13:00:00 UTC from IEEE Xplore. Restrictions apply.
with information in the complex domain is suitable for treating
rotation in the complex plane. The channel characteristic has
not only amplitude information, but also phase information.
The dominant part of channel changes of respective paths
is rotation in the complex plane. Therefore, the CVNNs
can bring about both high flexibility and low generalization
error in the channel prediction. In addition, we can deal
with channel characteristics of respective paths c̃m (t) =
ãm e j θ̃m (where θ̃m ≡ 2π f˜m t + φ̃m ) as complex entities in the
CVNNs. In this way, the channel prediction becomes more
appropriate to predict time-varying channel than prediction of
separate channel parameters ãm , f˜m , and φ̃m independently.
Fig. 2 represents the prediction method presented in this
paper. In this method, we set the windows with a shorter
interval again to avoid the increase/decrease of φm beyond ±π
as we did in the linear and Lagrange-based predictions. We
estimate ĉm (t) with the CZT at I points: t = −n/2 − I +
1, . . . , −n/2, and feed them to a CVNN to predict
c̃m (−n/2 + 1). After that, we predict c̃m (−n/2 + 2) by
using ĉm (t) estimated at t = −n/2 − I + 2, . . . , −n/2 and
c̃m (−n/2 + 1). We iterate this process to obtain c̃m (t) (t ≥ 0).
Summation of the multiple paths c̃m (t) gives the characteristic
c̃(t) (t ≥ 0). In this stage, we also know the actual ĉ(0).
Then, we combine ĉ(0) ≡ â(0)e j θ̂(0) with c̃(0) ≡ ã(0)e j θ̃(0)
as â(0)ã(t)e j (θ̃(t )+θ̂(0)−θ̃(0)) /ã(0). This is the finally predicted
channel.
B. ML CVNNs
diction methods. The first one is a CZT-ML-CVNN prediction.
Fig. 3 shows the construction of the ML CVNN used in
this paper to predict c̃m (t) for each path. It has input termi-
nals, a hidden-neuron layer and an output-neuron layer. The
input terminals of the layered CVNN distribute input signals,
cm (t − 1), …, cm (t − IML ), to the hidden-layer neurons as
their inputs z0 . In the same way, the outputs of the hidden-
layer neurons z1 are passed to the output-layer neuron as its
inputs. The output of the output-layer neuron is the prediction
result c̃m (t).
DING AND HIROSE: FADING CHANNEL PREDICTION BASED ON COMBINATION OF CVNNs AND CZT
Fig. 3. Layered CVNN.
The connection weight wl j i to the i th input at the j th
neuron in the layer l is expressed by its amplitude |wl j i | and
phase θl j i . The internal state of the neuron u l j is obtained as
the summation of the inputs weighted by wl j i , particularly
 
ul j ≡ wl j i z (l−1)i = |wl j i ||z (l−1)i |e j (θl j i +θ(l−1)i ) . (5)
i i
The output z l j is given by adopting an amplitude-phase-type
activation function f ap to u l j as
z l j = f ap (u l j ) = tanh(|u l j |)e( j arg(ul j )) .
We update the connection weights in the ML-CVNN as
follows. The past channel characteristics for t ≤ −n/2
of respective paths ĉm (t) are estimated from the Doppler
spectrum observed with the CZT. The time t represents the
window center of the CZT. The ML-CVNN regards this ĉm (t)
as the teacher signal and the previous channel characteristics
ĉm (t − 1), . . . , ĉm (t − IML ) as the input signals. The output
of the ML-CVNN z 2 is used to update the weights. Then, we
predict the channel characteristics of respective paths c̃m (t)
(t ≥ 0) with the connection weights resulting at t = −n/2.
To update the weights in this paper, we applied the steepest
descent method which updates weights to minimize the dif-
ference El ≡ |zl − ẑl |2 /2, where zl are the outputs signal
and ẑl are the teacher signal in layer l. Thus, the weight
amplitude |wl j i | and the phase θl j i are renewed as [23], [24]
∂ El
|wl j i |(r + 1)=|wl j i |(r ) − K 1
∂(|wl j i |)

=|wl j i |(r ) − K 1 (1 − |z l j |2 )
×(|z l j | − |ẑ l j |cos(θl j − θ̂l j ))|z (l−1)i |cosθlrot
|z (l−1)i |
−|z l j ||ẑ l j |sin(θl j − θ̂l j )
|u l j |
1 ∂ El
θl j i (r + 1)= θl j i (r ) − K 2
|wl j i | ∂θl j i
 of sight. In urban area, the number of dominant multiple
= θl j i (r ) − K 2 (1 − |z l j |2 )
×(|z l j | − |ẑ l j |cos(θl j − θ̂l j ))|z (l−1)i |sinθlrot
|z (l−1)i |
+|z l j ||ẑ l j |sin(θl j − θ̂l j )
|u l j |
Authorized licensed use limited to: Rohde & Schwarz. Downloaded on February 18,2020 at 13:00:00 UTC from IEEE Xplore. Restrictions apply.
1689
j i ≡ θl j −θ(l−1)i −θl j i and r is the number of learning
where θlrot
iteration, whereas K 1 and K 2 are learning constants. The
teacher signals ẑ l j are for the outputs z l j in layer l. They are
backpropagated through the layered network to yield ẑ (l−1)i
as
∗
ẑ (l−1)i = f ap (ẑ l∗j wl j i )
= tanh(|ẑ l∗j wl j i |)e( j (−θ̂l j )+θl j i ) . (9)
In other words, the teacher signals in the hidden layer ẑ1
are the signals obtained through the backpropagation of the
teacher signal ẑ 2 = ĉm (t) in (9).
In a general neural network, connection weights are updated
until theirs convergence. However, such update iterations
require a large learning cost that hinders a real-time appli-
cation. To solve this problem, we utilize the weak regularity
of the respective channel characteristics. That is, we stop
the update of the weights at a certain number of iteration
regardless of the convergence for ĉm (t − 1). Then, the weights
are used as the initial values in the weights update for ĉm (t).
By repeating this learning continuously, we can reduce the
learning cost and follow the respective characteristics, which
vary continuously.
C. Recurrent CVNNs Based on RTRL
Recurrent neural networks (RNNs) is widely used in time-
(6) domain prediction. A RNN learns channel characteristic with
recurrent output signals. The RTRL [25] requires only a single
weight update in unit time. We combine the CZT with a
recurrent CVNN updated by the RTRL to predict the respective
characteristics (CZT-RTRL-CVNN-based prediction).
IV. N UMERICAL E XPERIMENTS
A. Simulation Setup
The performance of the proposed CVNN-based predic-
tions is discussed in this section. We evaluate the accu-
racy of proposed methods in numerical experiments. We
apply the conventional methods for comparison, e.g., two
TD methods, namely, the TD-Linear and TD-RTRL-CVNN-
based prediction and the methods combined with CZT to
predict respective-path characteristics, namely, CZT-Linear,
CZT-Lagrange-based, and CZT-AR-model-based prediction as
well as nonprediction, which regards estimated âm (−n/2),
fˆm (−n/2), φ̂m (−n/2) as ãm (t), f˜m (t), φ̃m (t) (t ≥ 0) without
prediction (CZT only).
The geometrical setup is shown in Fig. 4 and the simulation
parameters are shown in Table I. There are two scatterers
 ji separate by x. The mobile user moves in the direction of
sinθlrot
ji (7) the arrow at v = 14 m/s. This is a hard situation for channel
prediction because of the fading channel varying very fast.
The base station (BS) and the mobile user have the line
paths are generally about three paths. The carrier frequency
is f c = 2 GHz.
To compare the effectiveness of the channel prediction
 ji
methods in actual application, we assume an orthogonal
cosθlrot
ji (8)
frequency-division multiplexing (OFDM) system. The OFDM
1690 IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, VOL. 25, NO. 9, SEPTEMBER 2014
Fig. 4. Geometrical setup used in the simulation. There are two scatterers,
a BS and a mobile user in the simulation field. The BS and the mobile user
have the line of sight.
TABLE I
S IMULATION PARAMETERS
Fig. 5. Block diagram of applied OFDM system.
systems can compensate frequency selective fading by itself,
but it requires some compensation for time-varying fading. The
block diagram of the OFDM is shown in Fig. 5. Table II lists
the system parameters following the standard of HC-SDMA as
well as the prediction parameters employed for the CZT-AR-
model-based, the CZT-ML-CVNN-based, and the both RTRL-
CVNN-based predictions. In this simulation, we assume a
QPSK as the modulation method and a time division duplex
(TDD), as shown in Fig. 6, which use a single frequency
band for uplink and downlink in a time division manner. The
number of OFDM subcarriers is 52. We predict the change
of the channel characteristic for a unit TDD frame (5 ms)
from t = 0. The center of windows are set with an interval of
1 frame. The prediction parameters were chosen in simulation
in such a way that the prediction error is minimized.
Authorized licensed use limited to: Rohde & Schwarz. Downloaded on February 18,2020 at 13:00:00 UTC from IEEE Xplore. Restrictions apply.
TABLE II
OFDM AND C HANNEL -P REDICTION PARAMETERS
Fig. 6. General TDD frame format.
B. Prediction Accuracy in Low-Resolution Conditions
Table III shows the calculation costs of conventional and
proposed methods in which S is the symbol number in a CZT
window as listed in Table II. The CZT-combined methods need
O(S log S) more calculation cost for CZT than TD methods.
The second terms for the CZT-combined methods express
prediction costs in the respective methods. The proposed
CZT-ML-CVNN-based prediction needs largest calculation
cost for the prediction. However, the calculation cost of
the second term is sufficiently small than that of the first
term even in the CZT-ML-CVNN-based prediction because
always S log S  M·RML ·JML ·IML for any practical window
length n. The inequality shows that the window lengths in
the CZT is the parameter most influential to the calculation
cost in the CZT-combined methods, including our proposed
methods. A short window length is required for reducing the
calculation cost, but it results in a low frequency resolution.
In such a tradeoff situation, we have to predict channel
characteristics c̃m (t) with low-frequency-resolution estima-
tions ĉm (t). Among various cases, we present a set of pre-
diction results when the window length is short. The length
of Hann window is 8 TDD frame. Fig. 7 shows a result
of the scatterers’ distance is x = 34 m. Fig. 7(a) shows
the predicted relative power in the prediction frame (5 ms),
whereas Fig. 7(b) is the predicted phase. The results of the
CZT-ML-CVNN-based prediction are near to the actual signal
while those of other methods case deviate from. Fig. 8 shows
the bit-error rate (BER) versus the signal-to-noise ratio in
the OFDM system with channel compensations using the
DING AND HIROSE: FADING CHANNEL PREDICTION BASED ON COMBINATION OF CVNNs AND CZT
TABLE III
C ALCULATION C OSTS OF T ESTED M ETHODS
Fig. 7. Predicted (a) relative power and (b) phase curves of the channel as
functions of time in the prediction period (t > 0).
Fig. 8. BER curves obtained for TD-Linear prediction, TD-RTRL-CVNN-
based prediction, CZT only, CZT-Linear prediction, CZT-Lagrange-based
prediction, CZT-AR-model-based prediction, CZT-RTRL-CVNN-based pre-
diction, CZT-ML-CVNN-based prediction, and actual channel characteristics
when the geometry causes severe change in the Doppler effects.
prediction results. The BER with the compensation based on
the CZT-ML-CVNN is worse only by 2 to 4 dB than that based
on the actual channel characteristic, whereas those based on
Authorized licensed use limited to: Rohde & Schwarz. Downloaded on February 18,2020 at 13:00:00 UTC from IEEE Xplore. Restrictions apply.
1691
the conventional methods degrade much more. In particular,
the performance of the CZT-AR-model, CZT-RTRL-CVNN,
and TD-Linear are worse by 7–10 dB. The other methods
completely fail to compensate the fading. The results show
that the prediction based on the CZT-ML-CVNN has a higher
performance even in the situation in which other predictions
fail to predict the fading channel.
In this case, the RTRL-CVNN shows medium BER perfor-
mance. Basically, a recurrent structure realizes simple influ-
ence of input signals on the output with first-order decay. It is
sometimes difficult for the RTRL-CVNN to predict a channel
in rapidly changing environment. However, the CZT-RTRL-
CVNN shows a result better than that of the TD-RTRL-CVNN.
C. Performance Dependence on Window Length
In this section, we analyze a relationship of the window
length of CZT and the prediction performances of each
method. In phase shift keying, the prediction result of phase
values is more influential than that of relative power to
the communications quality. We present the results of phase
prediction in respective prediction methods here. Fig. 9 shows
prediction-phase error (PPE) in the predicted phase accu-
mulated in a unit frame when the Hann window length n
and the scatterers’ distance x are changed for n = 6–50
and x = 2–50 m. Fig. 9(a) shows the error averaged
for the all distance as function of n. Fig. 9(b) presents the
maximum error observed when the distance is changed. In
addition, in Table IV, we list the averaged and max error
at relatively short window, n = 8, with the error in the TD
methods for comparison. In Fig. 9 and Table IV, a smaller
phase error directly means a higher performance. In quaternary
phase shift keying (QPSK), the prediction failure is caused
by the PPE beyond π/8 at a symbol. In this paper, a unit
TDD frame consists of s = 2500 symbol according to the
HC-SDMA standard. Thus, the prediction fails if the phase
error shown in Fig. 9 go over ∼1000 rad·symbol. This
is a rough estimation. In an actual system, there exists
noise. In an observation of the phase error values and
the BER, we found that 500-rad·symbol error corresponds
to 2–4 dB degradation in the BER, 1000 rad·symbol to
10–15 dB, and 1500 rad·symbol results in complete failure
of communications.
In many applications, the length should be sufficiently short
for the reduction of the calculation cost. In the region of n ≤ 15
in Fig. 9(a), we find that the error of the CZT-ML-CVNN is
the lowest among the tested methods. At the same time, we
find that the error curves rise up in most cases.
The peak phase error curves in Fig. 9(b) are somewhat
similar to the corresponding curves in Fig. 9(a). However,
the CZT-ML-CVNN-based prediction result shows the lowest
error over a wider window-length region. Simultaneously, the
peak error value in Fig. 9(b) is almost the same amount as the
mean error value in Fig. 9(a). These facts show that there exists
no serious degradation and that the stability of the CZT-ML-
CVNN-based prediction is very high. Again, the performance
of the CZT-ML-CVNN-based method is higher than those of
other methods.
1692 IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, VOL. 25, NO. 9, SEPTEMBER 2014
Fig. 9. (a) Mean and (b) peak phase errors for various prediction methods,
namely, CZT only, CZT-Linear prediction, CZT-Lagrange-based prediction,
CZT-AR-model-based prediction, CZT-RTRL-CVNN-based prediction, and
CZT-ML-CVNN-based prediction.
TABLE IV
M EAN AND P EAK PPE OF THE CZT-C OMBINED M ETHODS AT n = 8 AND
THE T IME -D OMAIN M ETHODS
D. Influence of Learning Iteration Number and Hidden
Neuron Number to Input Terminal Number Ratio
In our proposed method, iteration number of ML-CVNN-
weight update is a parameter influential to prediction accuracy
and cost. The influence to the prediction accuracy is shown
in Fig. 10. The window length n is 8, 20, 30, 40 and the
scatterers’ distance x is changed for x = 2–50 m. Each
situation is tested 25 times, and then the accumulated phase
Authorized licensed use limited to: Rohde & Schwarz. Downloaded on February 18,2020 at 13:00:00 UTC from IEEE Xplore. Restrictions apply.
Fig. 10. PPEs of CZT-ML-CVNN-based prediction for various n versus
iteration number of weight updates.
Fig. 11. PPE of CZT-ML-CVNN-based prediction versus ratio of hidden
neuron number to input terminal number.
error is averaged and plotted versus the iteration number.
The PPE decreases with the increase of iteration number,
and is less than or ∼500 rad·symbol when the iteration
number is over 3. Thus, the proposed method succeeds to
realize high prediction accuracy in a relatively small iteration
number.
To stop update at specific number is useful to reducing
learning cost. Learning cost affects prediction performance in
real-time applications. The learning cost of the RTRL is small
since the weights are updated only once for a single set of input
signals. On the other hand, the CZT-ML-CVNN phase error
for the case without iteration (RML = 1) takes a larger error
than the CZT-RTRL-CVNN phase error (709.5 rad·symbol).
However, the CZT-ML-CVNN-based prediction achieves a
phase error less than 500 rad·symbol by increasing the iteration
number only to 7 even in low frequency resolution (n = 8).
These facts show that the proposed CZT-ML-CVNN-based
method has an ability to predict the channel with a higher
accuracy and a small learning cost.
The ML-CVNN input terminal number IML and the hid-
den neural number JML are also influential parameters. The
influence to the prediction accuracy is shown in Fig. 11. We
evaluate the influence of the two parameters relatively by
fixing the ML-CVNN input terminal number IML and chang-
ing the ML-CVNN hidden neural number JML . The window
DING AND HIROSE: FADING CHANNEL PREDICTION BASED ON COMBINATION OF CVNNs AND CZT
Base Station (BS)
50m Mobile User (MU)
Fig. 12. Locations of the mobile user and the BS, including the movement
direction (Google Maps, modified).
TABLE V
E XPERIMENT PARAMETERS
Fig. 13. Block diagram of the system used in the actual experiment.
length n is 8 and the scatterers’s distance x is changed
for x = 2–50 m. Each situation for a certain set of x,
IML , and JML is tested 25 times, and then the accumulated
phase error is averaged and plotted versus the ratio of the
hidden neuron number to the input terminal number JML /IML .
The iteration number of ML-CVNN-weight update is 10 for
a single set of inputs. The PPE decreases with the increase
of the ratio, and is minimum at around JML /IML = 1. But,
this relationship between IML and JML has less influence
than other parameters because the fluctuation in the range of
460–490 is relatively small.
V. E XPERIMENTS IN ACTUAL E NVIRONMENT
A. Experiment Setup
In this section, we present the experimental results of
the proposed predictions in actual fading environment.
We conducted experiments for fading channels in actual com-
munications environment shown in Fig. 12. There are a mobile
user (MU) as a transmitter and a BS as a receiver in the
experiment site. Table V lists the experiment parameters. The
distance between the MU and the BS is ∼300 m, and the MU
moves in the direction of the arrow shown in Fig. 12 with a
Authorized licensed use limited to: Rohde & Schwarz. Downloaded on February 18,2020 at 13:00:00 UTC from IEEE Xplore. Restrictions apply.
1693
Fig. 14. Actual (a) relative power and (b) phase without modulation received
at the BS in Fig. 12.
velocity of 12.5 m/s. The block diagram of the transmitter
and the receiver is shown in Fig. 13. The MS transmits
1.297-GHz nonmodulation wave from a monopole antenna,
whereas the BS receives the wave at another monopole
antenna. The received channel signal is mixed with 1.287-GHz
local oscillator wave after amplified with a low-noise amplifier,
and then extracted as an intermediate frequency signal of
10 MHz. After passing it to an amplifier and a band-pass filter
with 2-MHz bandwidth, we sample the channel information by
60 MHz. For the compatibility with the HC-SDMA standard,
we thin down the time-sequential data at 500 kHz (downsam-
pling). Fig. 14 presents an example of the time-sequential data
showing relative power and phase observed at the BS. In this
section, We assume the conventional methods for comparisons
that have been described in Section IV.
B. Prediction Accuracy
We evaluate the performance of the prediction methods for
the measured channel characteristic by assuming the same
OFDM communications with the QPSK modulation as that
has been described in Section IV. We predict the characteristic
in a unit TDD frame (5 ms). The window length n is 8
again. Among various cases, we present a set of prediction
result in which the conventional prediction methods fail.
Fig. 15 are the results. The CZT-ML-CVNN predicts the
channel characteristic with higher accuracy, whereas the other
methods deviate from the actual channel. This fact shows the
CZT-ML-CVNN-based prediction has higher performance also
in the actual fading environment. The BER shown in Fig. 16
also indicates that the CZT-ML-CVNN-based prediction has a
higher performance than the conventional methods.
1694 IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, VOL. 25, NO. 9, SEPTEMBER 2014

Fig. 17. Number of communications failure (cumulative phase error

>1500 rad·symbol) in actual fading environment.

Fig. 15. Predicted (a) relative power and (b) phase curves of the channel as
functions of time in the prediction period. VI. C ONCLUSION
We proposed a channel prediction method based on the
CVNNs. The combination of the ML-CVNN and CZT brings
a highly accurate channel prediction even in a low frequency
resolution condition. The channel characteristic has not only
amplitude information, but also phase information. Since the
dominant part of channel changes of respective paths is
rotation in the complex plane, the CVNNs can bring about
both high flexibility and low generalization error in channel
prediction. We deal with respective characteristics as complex
entities. A series of simulations and experiments demon-
strated the communications with the CZT-ML-CVNN-based
prediction could realize better performance than that with the
TD-linear and TD-RTRL-CVNN-based prediction as well as
the CZT-combined linear, Lagrange-based, AR-model-based,
and RTRL-CVNN-based predictions.

for their insightful suggestions.
Fig. 16. BER curves obtained for TD-linear prediction, TD-RTRL-CVNN-
based prediction, CZT only, CZT-Linear prediction, CZT-Lagrange-based
prediction, CZT-AR-model-based prediction, CZT-RTRL-CVNN-based pre-
diction, CZT-ML-CVNN-based prediction, and actual channel characteristics.
C. Communications-Failure Rate
We conducted channel prediction for 70 trials by shifting the
prediction period with 1-ms step in the rapid fading interval of
Fig. 14. The phase error in the predicted phase is accumulated
in a unit frame. The Hann window length n is 8 again. Fig. 17
represents the number of prediction with phase error being
>1500 rad·symbol, which means a failure of communica-
tions in the OFDM system. It is found that the number
of communications failure in the CZT-ML-CVNN is smaller
than those in other methods. The CZT-ML-CVNN-based pre-
diction is highly effective also in actual fading-environment
experiment.
ACKNOWLEDGMENT
The authors would like to thank the anonymous reviewers
R EFERENCES
[1] W. C. Jakes, Microwave Mobile Communications, 2nd ed. New York,
NY, USA: Wiley, 1994.
[2] A. Duel-Hallen, “Fading channel prediction for mobile radio adaptive
transmission systems,” Proc. IEEE, vol. 95, no. 12, pp. 2299–2313,
Dec. 2007.
[3] High Capacity Spatial Division Multiple Access HC-SDMA, ANSI-ATIS
Standard 0700004-2005, Sep. 2005.
[4] A. J. Paulraj, D. A. Nabar, and H. Bolcskei, “An overview of MIMO
communications—A key to gigabit wireless,” Proc. IEEE, vol. 92, no. 2,
pp. 198–218, Feb. 2004.
[5] F. Maehara, F. Sasamori, and F. Takahata, “Linear predictive maximal
ratio combining transmitter diversity for OFDM-TDMA/TDD systems,”
IEICE Trans. Communicat., vol. E86-B, no. 1, pp. 221–229, 2003.
[6] H. Bui, Y. Ogawa, T. Nishimura, and T. Ohgane, “Performance eval-
uation of a multi-user MIMO system with prediction of time-varying
indoor channels,” IEEE Trans. Antennas Propag., vol. 61, no. 1,
pp. 371–379, Jan. 2013.
[7] T. Eyceoz, A. Duel-Hallen, and H. Hallen, “Prediction of fast fading
parameters by resolving the interference pattern,” in Proc. Asilomar
Conf. Signals, Syst., Comput., vol. 1. 1997, pp. 167–171.

Authorized licensed use limited to: Rohde & Schwarz. Downloaded on February 18,2020 at 13:00:00 UTC from IEEE Xplore. Restrictions apply.
DING AND HIROSE: FADING CHANNEL PREDICTION BASED ON COMBINATION OF CVNNs AND CZT
[8] T. Eyceoz, A. Duel-Hallen, and H. Hallen, “Deterministic channel mod-
eling and long range prediction of fast fading mobile radio channels,”
IEEE Trans. Commun. Lett., vol. 2, no. 9, pp. 254–256, Sep. 1998.
[9] A. Arredondo, K. Dandekar, and G. Xu, “Vector channel modeling
and prediction for the improvement of downlink received power,” IEEE
Trans. Commun., vol. 50, no. 7, pp. 1121–1129, Jul. 2002.
[10] M. Sternad and D. Aronsson, “Channel estimation and prediction for
adaptive OFDM downlinks,” in Proc. IEEE Veh. Technol. Conf., vol. 2.
Oct. 2003, pp. 1283–1287.
[11] W. Liu, L.-L. Yang, and L. Hanzo, “Recurrent neural network based
narrowband channel prediction,” in Proc. IEEE Veh. Technol. Conf.,
vol. 5. May 2006, pp. 2173–2177.
[12] C. Potter, G. K. Venayagmoorthy, and K. Kosbar, “RNN based MIMO
channel prediction,” Signal Process., vol. 90, pp. 440–450, Feb. 2010.
[13] J. Hwang and J. Winters, “Sinusoidal modeling and prediction of fast
fading processes,” in Proc. IEEE Globecom, Nov. 1998, pp. 892–896.
[14] J. Andersen, J. Jensen, S. Jensen, and F. Frederiksen, “Prediction of
future fading based on past measurements,” in Proc. IEEE 50th Veh.
Technol. Conf., 1999, pp. 151–155.
[15] L. R. Rabiner, R. W. Schafer, and C. Rader, “The chirp
z-transform algorithm,” IEEE Trans. Audio Electroacoust., vol. 17, no. 2,
pp. 86–92, Jun. 1969.
[16] S. Tan and A. Hirose, “Low-calculation-cost fading channel prediction
using chirp z-transform,” Electron. Lett., vol. 45, no. 8, pp. 418–420,
Apr. 2009.
[17] S. Ozawa, S. Tan, and A. Hirose, “Channel prediction experiment based
on linear prediction in frequency domain,” in Proc. Asia-Pacific Microw.
Conf., Dec. 2010, pp. 1280–1283.
[18] S. Ozawa, S. Tan, and A. Hirose, “Errors in channel prediction based on
linear prediction in the frequency domain: A combination of frequency-
domain and time-domain techniques,” URSI Radio Sci. Bull., vol. 337,
pp. 25–29, Apr. 2011.
[19] H. Matsui and A. Hirose, “Nonlinear prediction of frequency-domain
channel parameters for channel prediction in fading and fast doppler-
shift change environment,” in Proc. Int. Symp. Antennas Propag., 2012,
pp. 1132–1135.
[20] A. Hirose, Complex-Valued Neural Networks, 2nd ed. New York, NY,
USA: Springer-Verlag, 2012.
[21] S. Chen, S. McLaughlin, and B. Mulgrew, “Complex-valued radial basis
function network, Part I: Network architecture and learning algorithms,”
Signal Process., vol. 35, no. 1, pp. 19–31, 1994.
[22] S. Chen, S. McLaughlin, and B. Mulgrew, “Complex-valued radial basis
function network, Part II: Application to digital communications channel
equalisation,” Signal Process., vol. 36, no. 2, pp. 175–188, 1994.
[23] A. Hirose, “Applications of complex-valued neural networks to coherent
optical computing using phase-sensitive detection scheme,” Inf. Sci.
Appl., vol. 2, no. 2, pp. 103–117, 1994.
Authorized licensed use limited to: Rohde & Schwarz. Downloaded on February 18,2020 at 13:00:00 UTC from IEEE Xplore. Restrictions apply.
1695
[24] A. Hirose and S. Yoshida, “Generalization characteristics of complex-
valued feedforward neural networks in relation to signal coherence,”
IEEE Trans. Neural Netw. Learn. Syst., vol. 23, no. 4, pp. 541–551,
Apr. 2012.
[25] G. Kechriotis and E. S. Manolakos, “Training fully recurrent neural
networks with complex weights,” IEEE Trans. Circuits Syst., Analog
Digital Signal Process., vol. 41, no. 3, pp. 235–238, Mar. 1994.
Tianben Ding received the B.Eng. degree from
Yokohama National University, Kanagawa, Japan, in
2012. He is currently pursuing the M.S. degree in
electrical engineering and information systems from
the University of Tokyo, Tokyo, Japan.
His current research interests include neural net-
works and wireless electronics.
Akira Hirose (F’13) received the Ph.D. degree in
electronic engineering from the University of Tokyo,
Tokyo, Japan, in 1991.
He was a Research Associate with the Research
Center for Advanced Science and Technology
(RCAST), University of Tokyo, in 1987. In 1991,
he was an Instructor with RCAST. From 1993
to 1995, he was with the University of Tokyo,
Institute for Neuroinformatics, University of Bonn,
Bonn, Germany. He is currently a Professor with
the Department of Electrical Engineering and Infor-
mation Systems, University of Tokyo. His current research interests include
neural networks and wireless electronics.
Dr. Hirose is a Senior Member of the Institute of Electronics, Information,
and Communication Engineers (IEICE) and a member of the Japanese Neural
Network Society (JNNS). He served as the Editor-in-Chief of the IEICE
Transactions on Electronics, an Associate Editor of journals such as the
IEEE T RANSACTIONS ON N EURAL N ETWORKS and the IEEE G EOSCIENCE
AND R EMOTE S ENSING N EWSLETTER, the Chair of the Neurocomputing
Technical Group, IEICE, and the General Chair of the 2013 Asia-Pacific
Conference on Synthetic Aperture Radar, Tsukuba, Japan. He serves as a
member of the IEEE Computational Intelligence Society Neural Networks
Technical Committee, the Founding Chair of the Complex-Valued Neural
Network Task Force in NNTC, the Governing Board Member of the Asia-
Pacific Neural Network Assembly, the Co-Editor-in-Chief of the Journal of
IEICE, the Vice President of the IEICE Electronics Society, and the President
of the JNNS.