0% found this document useful (0 votes)
30 views10 pages

Path Loss Modelling Based On Path Profile in Urban

1. The document presents a path loss model based on path profile for 5G systems in urban environments. It combines a log-distance path loss model for line-of-sight propagation and a machine learning model for non-line-of-sight cases. 2. Principal component analysis is used to extract relevant features from path profile attributes for non-line-of-sight cases, and polynomial regression is used to construct the path loss model. 3. Simulation results show the proposed model outperforms conventional models, reducing prediction error standard deviation by 22.2-37.2% dB, and field measurements show a reduction of 3.33-6.13 dB compared to conventional models.

Uploaded by

Mongi Ben Ameur
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
30 views10 pages

Path Loss Modelling Based On Path Profile in Urban

1. The document presents a path loss model based on path profile for 5G systems in urban environments. It combines a log-distance path loss model for line-of-sight propagation and a machine learning model for non-line-of-sight cases. 2. Principal component analysis is used to extract relevant features from path profile attributes for non-line-of-sight cases, and polynomial regression is used to construct the path loss model. 3. Simulation results show the proposed model outperforms conventional models, reducing prediction error standard deviation by 22.2-37.2% dB, and field measurements show a reduction of 3.33-6.13 dB compared to conventional models.

Uploaded by

Mongi Ben Ameur
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 10

Received: 6 April 2021 Revised: 9 August 2021 Accepted: 23 February 2022 IET Communications

DOI: 10.1049/cmu2.12369

ORIGINAL RESEARCH PAPER

Path loss modelling based on path profile in urban propagation


environments

Rong-Terng Juang

Department of Electronic Engineering, Feng Chia Abstract


University, ROC, Taichung, Taichung, Taiwan
This paper presents a path loss model based on path profile in urban propagation envi-
ronments for 5G systems. Although deep learning approaches are indeed powerful in
Correspondence
Rong-Terng Juang, Department of Electronic tasks involving prediction or classification, they often lack transparency and suffer from
Engineering, Feng Chia University, Taiwan, ROC. high computational complexity. The proposed model combines the log-distance path loss
Email: rtjuang@mail.fcu.edu.tw
model for line-of-sight propagation scenarios and a machine-learning-based model for
non-line-of-sight (NLOS) cases. This paper uses the principal component analysis algo-
rithm to extract relevant features out of some selected attributes of the path profile for
NLOS cases. Then, the path loss model can be constructed based on the approach of
polynomial regression. Simulation results show that the proposed model outperforms the
conventional models when operating in the 3.5 GHz frequency band. The standard devi-
ation of prediction error was reduced by about 22.2–37.2% dB when compared to the
conventional models. Furthermore, the prediction performance was also evaluated in a
non-standalone 5G New Radio network in the urban environment of Taipei city. The real-
world measurements show that the standard deviation of prediction error can be reduced
by 3.33–6.13 dB when compared to the conventional models.

1 INTRODUCTION Meanwhile, many empirical path loss models have been pro-
posed by prior researchers for 5G systems. Sun et al. presented
The emerging fifth-generation (5G) mobile communication sys- the comparison of the alpha-beta-gamma (ABG) and close-in
tems are expected to bring a complete revolution in applications (CI) path loss models using measured data and ray-tracing tech-
and experiences [1]. Mobile operators worldwide are racing to nique from 2 GHz to 73 GHz for 5G networks [5]. The authors
roll out 5G services. However, among the technical challenges, found that the physically-based CI model yielded smaller pre-
accurate channel models to predict path loss are vital for the diction errors, with a standard deviation ranging from about
design of 5G communication systems. 6 dB to 12 dB [5]. To improve the ABG model, a weighted
Among the existing path loss prediction methods, ray-tracing ABG model was proposed in [6], which suitably integrated
technique is recognized as an approach with high accuracy [2]. or combined different available datasets by considering differ-
It is a site-specific, deterministic channel model based on the ent weighting policies. It was reported that the weighted ABG
theory of geometrical optics, by which the electric field is dis- model could obtain better results in terms of model accu-
cretised into rays that are launched from the transmitter (Tx) in racy, with a standard deviation of prediction errors ranging
every direction and tracked as they propagate through the geo- from 1.2 dB to 12.5 dB [6]. To predict the millimetre-wave
metric environment [3]. With attractive accuracy, ray-tracing suf- path loss, probabilistic models were presented in [7]. Based on
fers from high computational complexity in the calculation of real-world 28 GHz and 73 GHz measurements, the authors
the shoot and bouncing rays. Therefore, hardware accelerators, obtained the line-of-sight (LOS) probabilities from ray-tracing
such as Graphics Processing Unit (GPU), Application Specific techniques and proposed a hybrid path loss model, which was
Integrated Circuit (ASIC), and Field Programmable Gate Array a weighted sum of LOS and non-LOS (NLOS) propagation
(FPGA), have been adopted in implementing ray-tracing [4]. losses [7].

This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is
properly cited.
© 2022 The Authors. IET Communications published by John Wiley & Sons Ltd on behalf of The Institution of Engineering and Technology

IET Commun. 2022;16:685–694. wileyonlinelibrary.com/iet-com 685


686 JUANG

The aforementioned empirical models were built based 2. The advantages of the proposed machine-learning-based
on the traditional statistical approaches. Recently, machine- method, including avoiding complex feature engineering, uti-
learning-based methods have been used in path loss modelling lizing low complexity linear regression etc., are explored;
[8]. Thrane et al. proposed a path loss prediction method 3. Simulation results as well as field measurements in a 5G New
based on deep learning techniques utilizing satellite images Radio (NR) network are shown.
aided by a simple path loss model [9]. Also, a deep learn-
ing approach to model radio propagation behaviours from The rest of this paper is organized as follows. Section 2
satellite images was proposed in [10], where the path loss expo- reviews some conventional path loss models. Section 3 presents
nent and shadowing factor for the entire coverage area were a detailed description of the proposed method. Subsequently,
estimated. Furthermore, an enhanced convolutional neural Section 4 explores the prediction performance based on numer-
network (CNN)-based path loss modelling was presented for ical simulations. Section 5 examines the performance in a real-
millimetre-wave propagations [11]. It was reported that the world 5G system in urban Taipei city. Finally, Section 6 presents
root-mean-square error of the prediction could be reduced to some concluding remarks.
8.59 dB [11]. In addition to point-to-point path loss prediction,
Ahmadien et al. proposed a method to evaluate path loss distri-
bution in the coverage area directly from satellite images based 2 CONVENTIONAL PATH LOSS
on deep learning approaches [12]. Although deep learning MODELS
approaches are indeed powerful in tasks involving prediction
or classification, they often lack transparency. The black-box 2.1 Walfisch–Ikegmi model
nature of deep learning approaches decreases the insight and
trust people have in the optimized models. Also, they often The Walfisch–Ikegami model computes the path loss by consid-
require huge amount of training data and high computational ering both LOS and NLOS conditions. The path loss in decibels
capabilities. is given by:
The Taiwan government completed its initial 5G spectrum
auction at the beginning of 2020. The spectrum on offer cov- PLWI = LFS + Lrts + Lmsd , (1)
ered 270 MHz in the 3.5 GHz band, 2,500 MHz in the 28 GHz
band, and 20 MHz in the 1.8 GHz band. Telecom operators are where LFS is the free-space path loss, Lrts is the roof-top-to-
currently rolling out 5G services. In practice, mobile operators street diffraction and scatter loss, and Lmsd is the multiscreen
tend to roll out the initial deployments in urban cities. There- diffraction loss [13, 14]. These loss terms are all functions of
fore, this paper focuses on path loss prediction in urban envi- the carrier frequency. In addition, they relate to some geomet-
ronments. ric parameters. For instance, LFS relates to the distance between
Among many path loss models, the COST231 Walfisch– the transmitter and receiver (Rx) pair; Lrts relates to the average
Ikegami model is one of the most widely used models to pre- street width, the road orientation angle, and the average build-
dict path loss in urban environments [13, 14]. It considers ing height; and Lmsd relates to the average building separation,
the obstruction of buildings in the propagation path. Specifi- the height from the average building level to the Tx antenna,
cally, the Walfisch–Ikegami model is a hybrid model combin- and a factor associated with diffraction loss. Note that the oper-
ing diffraction down to street level and some empirical correc- ation frequency for the Walfisch–Ikegami model ranges from
tion factors. Inspired by the Walfisch–Ikegami model, this paper 800 MHz to 2 GHz, although most of the 5G frequency bands
proposes a path loss model based on the profile along the direct are out of this range.
propagation path in urban environments for 5G cellular com-
munication systems. Following the machine learning method-
ology [8], this paper firstly selects some potential attributes of 2.2 ABG model
the path profile based on intuition. Then the essential features
are extracted out of the attributes using the principal compo- The ABG model is a simple extension to the alpha-beta (AB)
nent analysis (PCA) algorithm [15, 16], which is a process of model used in the 3rd Generation Partnership Project (3GPP)
computing the principal components. The dimensions of the by adding a frequency-dependent parameter [5, 17]. It is also
original data can then be reduced by ignoring some insignificant one of the standard 3GPP models and is currently widely used
components. Subsequently, the extracted features are used to in 5G applications [6]. The ABG model can be expressed as [5]:
model the path loss using the approach of least-square regres-
sion. The proposed method has the advantage of transparency
and low computational complexity, in contrast to the deep learn- PL ABG = 10𝛼log10 r + 𝛽 + 10𝛾log10 fc , (2)
ing approaches. The main contributions of this paper are as fol-
lows. where 𝛼 is the path loss exponent associated with the Tx–Rx 3D
distance r, 𝛽 is an offset term, and 𝛾 represents the dependence
1. An improved path loss model utilizing path profile is pre- of path loss on the logarithm of the carrier frequency fc . The
sented for 5G mobile communication systems in urban optimal model parameters are typically achieved by performing
propagation environment; linear regression with measurement data [18].
JUANG 687

FIGURE 1 Illustration of the Tx–Rx pair and its path profile

FIGURE 2 Flowchart of the proposed path loss modelling FIGURE 3 Some real-world path profiles in urban Taipei city. The
subfigures on the left-hand side are profiles with a Tx–Rx separation distance
of about 250 m, those on the right-hand side are with a distance of about
450 m, and the upper subfigures have higher path losses compared to the lower
2.3 CI model ones

The CI free-space reference distance model with a 1 m reference


distance is given by [5]: TABLE 1 Structure of the training dataset

( ) Input data Target path loss


PLCI = LFS fc , d0 + 10𝛼log10 (r∕d0 ) (3)
r1 , a1,1 , a1,2 , … , a1,9 L1

where d0 is the close-in free-space reference distance, and r2 , a2,1 , a2,2 , … , a2,9 L2
LFS ( fc , d0 ) is the free-space path loss in decibels at the carrier ⋮ ⋮
frequency fc and at a Tx–Rx separation distance of d0 . The opti- rM , aM ,1 , aM ,2 , … , aM ,9 LM
mal model parameters are typically estimated by performing lin-
ear regression with measurement data [18].

features are used to model the path loss using the approach of
3 PROPOSED PATH LOSS MODEL least-square regression. A detailed description of the proposed
method is given below.
In urban environments, the direct connection line between
the Tx and Rx is very likely to be blocked by buildings.
Unlike the ray-tracing model which performs the shooting 3.1 Calculation of profile attributes
and bouncing of rays at the cost of high computation com-
plexity, the proposed method predicts the path loss based Some real-world path profiles in urban Taipei city are shown in
on the geometric information along the direct path between Figure 3, where those on the left-hand side are profiles with a
the Tx–Rx pair. Figure 1 shows the conceptual overview of Tx–Rx separation distance of about 250 m, those on the right-
the Tx–Rx pair and the path profile along the direct con- hand side are with a distance of about 450 m, and the upper
nection line, where hTx is the Tx height, d is the horizontal subfigures have higher path loss values compared to the lower
1∕2
distance, and r = [d 2 + (hTx − 1.5)2 ] is the 3D distance ones. Specifically, based on the ray-tracing model, the path loss
between the Tx–Rx pair. The obstruction of buildings results values for the upper subfigures are 149 dB and 155 dB, respec-
in excessive path loss, in addition to free-space loss. There- tively, and those for the lower ones are 89 dB and 127 dB,
fore, this paper incorporates the profile features in path loss respectively. This figure reveals that different path profiles lead
modelling. to different path losses. Thus, based on visual inspection of the
Based on the machine learning methodology [8, 19], Figure 2 path profiles, this paper subjectively selects nine attributes as
shows the flowchart of the proposed path loss modelling. It follows.
starts by collecting the records of path loss and the associ-
ated path profiles. Then the attributes of the profiles are cal- 1. Average building height: the obstruction of buildings results
culated. Subsequently, the essential features are extracted out in propagation loss; therefore, this attribute is the arithmetic
of the attributes using the PCA algorithm. Lastly, the dominant mean of building heights along the profile.
688 JUANG

2. Standard deviation of building height: this attribute refers to Since there are M records in the training dataset, the attributes
the standard deviation of building heights along the profile. can be expressed in a matrix form as:
3. Normalized mean value of building distance: analogously to
the two aforementioned attributes which take the statistics ⎡ ã 1 ⎤
in the vertical axis, this attribute and the subsequent one take ⎢ ⎥
the statistics in the horizontal axis. This attribute is the arith- ⎢ ã 2 ⎥
A = ⎢ ⎥. (6)
metic mean of the distances from the Tx to the buildings ⎢⋮⎥
along the profile. The mean value is normalized by the Tx– ⎢ ⎥
Rx separation distance. ⎣ã M ⎦
4. Normalized standard deviation of building distance: this
attribute refers to the standard deviation of the distances The PCA algorithm computes the eigenvectors and the
from the Tx to the buildings along the profile. The standard corresponding eigenvalues of the covariance matrix of
deviation is normalized by the Tx–Rx separation distance. A. Denote the resulting eigenvalues in descending order
5. Building density: in addition to the above two statistics in as {𝜆1 , 𝜆2 , … , 𝜆N }and the corresponding eigenvectors as
the horizontal axis, this attribute describes the percentage of {v1 , v2 , … , vN }. Then the PCA algorithm selects the top 𝜏 prin-
building along the profile. cipal eigenvalues (𝜏 < N ) such that the ratio
6. Average building width: this attribute refers to the arithmetic
𝜆1 + 𝜆2 + ⋯ + 𝜆𝜏
mean of the building width along the profile. 𝜌= , (7)
7. Average street width: this attribute is the arithmetic mean of 𝜆1 + 𝜆2 + ⋯ + 𝜆𝜏 + 𝜆𝜏+1 ⋯ + 𝜆N
the street width along the profile.
8. Distance to the nearest building from the Rx: in general, the is higher than a threshold to keep sufficient relevant informa-
path loss tends to be high when a high building is close to tion. According to [20], the ratio should be greater than 0.9
the Rx. This attribute refers to the distance from the Rx to to retain most of the information in the dataset. Finally, the
the nearest building along the profile. original attributes are projected to the 𝜏-dimensional feature
space constructed by the selected eigenvectors. For instance,
9. Height of the nearest building from the Rx: this attribute is ∑3 ∑N
the height of the nearest building along the profile from the assume 𝜏 is selected as three because 𝜌 = ( i=1 𝜆i )∕( i=1 𝜆i )
Rx. is greater than 0.9, then the dimension of the original data can
be reduced from N to three by projecting the data to a 3-D
Given the definitions above, the profile attributions can be feature space constructed by v1 , v2 , and v3 . Specifically, the 𝜏
calculated accordingly. Denote (ai,1 , ai,2 , … , ai,9 ) as the nine features, (𝜒i,1 , 𝜒i,2 , … , 𝜒i,𝜏 ), extracted from the path profile of
selected attributes for the ith Tx–Rx pair, and ri and Li as the the ith Tx–Rx pair can be expressed as:
associated distance and path loss, respectively. Table 1 shows
the structure of the training dataset, with a size of M records. fi = ã i V, 1 ≤ i ≤ M , (8)
The objective of the machine learning here is to find the rela-
tionship between the input data and the target path loss in the where fi = [ 𝜒i,1 𝜒i,2 ⋯ 𝜒i,𝜏 ] and V = [v1 v2 ⋯ v𝜏 ] .
table. The distance ri is a well-known path loss feature, while
the feature extraction of the attributes is described in the subse-
quent section. 3.3 Regression of path loss model

The features extracted above are used to model the path loss.
The proposed model combines LOS and NLOS propagation
3.2 Extraction of essential features scenarios. For LOS cases, where the radiation from the Tx is not
blocked by buildings, the path loss is determined using the ABG
Without complex feature engineering, there is no guarantee that
model. For single frequency scenarios, 𝛾 in (2) is set to two, thus
all the aforementioned attributes are mutually independent and
the ABG model reverts to the AB model used in 3GPP and
highly correlated to path loss. Therefore, this paper uses the
WINNER II channel models [5, 21] and has the form:
PCA algorithm to extract essential features out of the attributes.
In general, the selected attributes for the ith Tx–Rx pair can be
PL AB = 10𝛼log10 r + 𝛽. (9)
expressed in a vector form as:
[ ] This paper assumes the excessive path loss in addition to the
ai = ai,1 ai,2 ⋯ ai,N , (4)
free-space propagation is caused by the obstruction of build-
ings. Therefore, the path loss model for NLOS cases considers
where N is the number of selected attributes. The logarithm of
the Tx–Rx separation distance and the extracted features in (8).
ai can be expressed as:
Since plain linear models usually suffer from under-fitting prob-
[ ] lems, this paper increases the model capability by adding powers
ã i = 10log10 ai,1 10log10 ai,2 ⋯ 10log10 ai,N . (5) of the original features as new features. As a result, the proposed
JUANG 689

path loss model for NLOS cases is a high order equation and
can be expressed as:
𝜏𝜅

K ∑
𝜃𝜅, j f j𝜅 ,
K
PL ABS = 10𝛼log10 r + 𝛽 + (10)
𝜅=1 j =1

where 𝛼, 𝛽 and 𝜃𝜅, j are the model parameters to be determined,


f j is the j-th dominant feature of the path profile, K is a hyper-
parameter of the machine learning approach, that is, the poly-
nomial regression [22], and 𝜏𝜅 is a value to be determined such
that the ratio
𝜆1𝜅 + 𝜆2𝜅 + ⋯ + 𝜆𝜏𝜅𝜅
𝜌𝜅 = , (11)
𝜆1𝜅 + 𝜆2𝜅 + ⋯ + 𝜆𝜏𝜅𝜅 + 𝜆𝜏𝜅 𝜅
⋯ + 𝜆N
𝜅 +1

is higher than 0.9. Here, 𝜌1 and 𝜏1 are equivalent to 𝜌 and 𝜏 in


(7), respectively. The hyper-parameter of the polynomial regres-
sion, that is, the degree of the polynomial, cannot be learned
but is typically tuned manually by trial and error. Specifically,
FIGURE 4 Top view of the simulation environment which covers an area
this paper sets K = 2 and the proposed path loss model can of 900 m × 900 m. Buildings are represented by coloured polygons. Seven base
be expressed as a quadratic polynomial with multiple indepen- stations with omnidirectional antennae are designated by the encircled crosses
dent variables: (⊕)
𝜏1
∑ 𝜏2

2
PL ABS = 10𝛼log10 r + 𝛽 + 𝜃1, j f j + 𝜃2, j f j2 , (12)
j =1 j =1 [ ]T
𝜽= 𝛼 𝛽 𝜃1,1 ⋯ 𝜃1,𝜏1 𝜃2,1 ⋯ 𝜃2,𝜏2
2
where PL ABS is the dependent variable we want to predict (16)
while 10log10 r, f j and f j2 are the independent variables we use
to predict the value of the dependent variable. This equation In (15), 𝜒i,2 j = 𝜒i, j × 𝜒i, j is the quadratic form of the jth
is still considered as a linear model as the coefficients associ- extracted feature for the ith Tx–Rx pair. As a result, the opti-
ated with the independent variables are still linear. The term mal parameter that minimizes the fitting error is given by [6]:
f j2 = f j × f j is only one of the independent variables and
the equation is quadratic in nature. Note that since the hyper-
parameter K is set as two, 𝜏1 and 𝜏2 are the two parameters ( )−1
∑𝜏1 ∑N 𝜽 OPT = FT F FT L. (17)
to be determined such that 𝜌1 = ( j =1 𝜆 j )∕( j =1 𝜆 j ) and
∑𝜏2 2 ∑N 2 Given the path loss features of the ith Tx–Rx pair as ̃f i =
𝜌2 = ( j =1 𝜆 j )∕( j =1 𝜆 j ) in (11) are both greater than 0.9.
To achieve the optimal model parameters based on the least- [10log10 ri 1 𝜒i,1 ⋯ 𝜒i,𝜏1 𝜒i,1
2
⋯ 𝜒i,𝜏
2
], the optimal path loss
2
square regression, (12) can be expressed in a matrix form as: prediction is obtained from

L̂ i = ̃f i 𝜽 OPT (18)
L = F𝜽, (13)

where 4 SIMULATION RESULTS


[ ]T A software package, SignalPro, by EDX Wireless, Inc., was used
L = L1 L2 ⋯ LM , (14)
to conduct the simulations. It provides a set of simulation tools
for wireless communication systems. To facilitate simulations
with site-specific geographic information, SignalPro supports
⎡ 10log10 r1 1 𝜒1,1 ⋯ 𝜒1,𝜏1 𝜒1,1
2
⋯ 2 ⎤
𝜒1,𝜏 the import of digital maps. In this paper, a 3D map of urban
⎢ 2

⎢ ⎥ Taipei city was imported to the software package. Figure 4
⎢ 10log10 r2 1 𝜒2,1 ⋯ 𝜒2,𝜏1 𝜒2,1
2
⋯ 𝜒2,𝜏 ⎥
2
shows the top view of the simulation environment, which cen-
2
F=⎢ ⎥,
⎢ ⎥ tred at latitude 25.057457◦ N and longitude 121.534808◦ E
⎢ ⋮ ⋮ ⋮ ⋱ ⋮ ⋮ ⋱ ⋮ ⎥
⎢ ⎥ and covered an area of 900 m x 900 m. The polygons in the
⎢ ⎥ figure represent buildings with different heights. The mean and
⎣10log10 rM 1 𝜒M ,1 ⋯ 𝜒M ,𝜏1 𝜒M
2
,1 ⋯ 𝜒M
2
,𝜏 ⎦
2
standard deviation of building heights were 15.7 m and 8.7 m,
(15) respectively. The maximum and minimum building heights were
690 JUANG

FIGURE 5 Information ratios of the principal components. The FIGURE 6 Prediction performance of the proposed path loss model and
cumulated information ratio of the top four components reaches 95% the conventional models in simulations

ber of the selected attributes N is nine and the number of the


46 m and 4 m, respectively. Seven base stations with omnidi- top principal eigenvalues 𝜏 is four. In addition, the path loss pre-
rectional antennae were deployed and are designated by encir- diction in (18) requires 3 + 9 = 12 multiplication operations
cled crosses (⊕) in Figure 4. In the simulations, the effective and eight addition operations, because 𝜏1 = 4 and 𝜏2 = 3.
isotropic radiated power and the height of the base stations were On the contrary, the deep-learning-based approaches [9, 12]
set to 50 dBm and 30 m, respectively. Receivers with a height of required high computational complexity. They all incorporated
1.5 m were placed on the roads in Figure 4. Distances between CNN to extract path loss features from aerial images. Ignoring
the Tx and Rx pairs ranged from 10 m to 1000 m. The ray- other hidden layers in the deep neural networks, the CNN alone
tracing model was applied for the path loss calculations at the consumed high computing power. For example, the size of the
3.5 GHz frequency band. As a result, the received power lev- grey-scaled input image in [9] was 256 × 256 pixels, which cov-
els at all the Rx locations were recorded. Because the maximum ered an area of 185 × 185 m2 enclosed by the Tx–Rx pair, and
distance between Tx-Rx pairs was 1 Km, the Tx is referred to as the first convolutional layer had 200 5 × 5 filters. Excluding the
a macro-cell base station here to distinguish from the micro-cell processes of padding, activation, and maximum polling, the first
base station in the latter. CNN layer alone in [9] required 252 × 252 × 25 × 200 ≈
At the cost of high computational complexity, the perfor- 3 × 108 multiplications and 252 × 252 × 24 × 200 ≈ 3 ×
mance of the ray-tracing model was satisfactory when compared 108 additions, which are much higher than the proposed model.
to the field measurements [23, 24]. In this work, the simulation Figure 6 shows the prediction performance of the proposed
results using the ray-tracing model were taken as the ground path loss model and the conventional models, that is, the ABG
truth for the path loss predictions. In the simulations, 25,093 model and the CI model. Note that the Walfisch–Ikegami model
records were collected, where each record contained the receiv- is excluded here because the operation frequency of 3.5 GHz
ing power and the path profile of the specific Tx–Rx pair. All is out of its frequency range. In the figure, black dots are the
the records were used to calculate the profile attributes. Then ground truth values whereas red dots are the predicted values
the dominant features were extracted using (8). Figure 5 shows by using the proposed model. The conventional models in the
the information ratios of the extracted features, where the ratio figure exhibit good matches with the mean values of path loss,
for the jth principle component is defined as: while the variations between the perditions and the ground truth
values are large.
𝜆j Define prediction error as the difference between the pre-
𝜂j = ,1 ≤ j ≤ N . (19)
𝜆1 + 𝜆2 + ⋯ + 𝜆N dicted value and the ground-truth value, that is:

Accordingly, this paper selects 𝜏1 = 4 and 𝜏2 = 3 such that ei = L̂ i − Li , 1 ≤ i ≤ M . (20)


𝜌1 = 0.95 and 𝜌2 = 0.97, respectively.
To evaluate the computational complexity of the proposed Then, the standard deviation of the prediction error is given
machine-learning-based path loss model, this paper counts the by:
numbers of multiplication and addition operations in the path
loss prediction process. Here, the feature extraction in (8) √
∑M
requires 9 × 4 = 36 multiplication operations and 8 × 4 = i=1
(ei − 𝜇 )2
32 addition operations for each Tx–Rx pair, because the num- STD = , (21)
M
JUANG 691

TABLE 2 Simulation results of path loss prediction error for the TABLE 3 Simulation results of path loss prediction error for the
macro-cell base station operating in the 3.5 GHz band at 30 m height micro-cell base station operating in the 3.5 GHz band at 30 m height

Mean error Standard deviation Mean error Standard deviation

ABG model 0.49 dB 17.50 dB ABG model −0.02 dB 12.75 dB


CI model 0.32 dB 17.79 dB CI model 0.55 dB 13.89 dB
Proposed model 0.00 dB 12.28 dB Proposed model 0.0 dB 9.92 dB
Improve over ABG model – 5.22 dB (29.9%) Improve over ABG model – 2.83 dB (22.2%)
Improve over CI model – 5.51 dB (30.1%) Improve over CI model – 3.97 dB (28.6%)

TABLE 4 Simulation results of path loss prediction error for the


micro-cell base station operating at the 3.5 GHz band and 20 m height

Mean error Standard deviation

ABG model 0.09 dB 14.13 dB


CI model 0.56 dB 15.01 dB
Proposed model 0.00 dB 9.42 dB
Improve over ABG model – 4.71 dB (33.3%)
Improve over CI model – 5.59 dB (37.2%)

TABLE 5 Simulation results of path loss prediction error for the


micro-cell base station operating in the 1.8 GHz band at 30 m height

Mean error Standard deviation

ABG model 0.43 dB 13.20 dB


FIGURE 7 The cumulative distribution of the absolute error of path loss
prediction in simulations CI model 0.64 dB 14.65 dB
Walfisch–Ikegami Model −10.67 dB 14.09 dB
∑M Proposed model 0.00 dB 10.95 dB
where 𝜇 = i=1 ei ∕M is the mean error. Table 2 shows the
Improve over ABG model – 2.25 dB (17.1%)
statistics of the prediction errors. In terms of mean error, all the
models showed good performance with only slight variances. In Improve over CI model – 3.7 dB (25.3%)
terms of the standard deviation, however, the proposed model Improve over Walfisch–Ikegami model – 3.14 dB (22.3%)
outperformed the ABG and the CI models by reductions of
about 5.22 dB and 5.51 dB, respectively. Furthermore, Figure 7
shows the cumulative distribution of the absolute error, |ei |.
Treat the absolute error as a continuous random variable E was limited to 500 m. Thus, the base station here is referred
with density function fE (𝜀) ≥ 0 for 𝜀 ≥ 0, and is 0 otherwise, to as a micro-cell base station. Based on the aforementioned
then the cumulative distribution function of the absolute error simulation settings, Tables 3–6 show the statistics of the pre-
𝜀
is defined as FE (𝜀) ∶= P (E ≤ 𝜀) = ∫0 f (t )dt , which represents diction errors. Table 3 reveals the errors when the Tx was
the probability that E will take a value less than or equal to 𝜀. operating in the 3.5 GHz frequency band at 30 m height. The
The vertical axis in Figure 7 represents FE (𝜀) while the horizon- proposed model reduced the standard deviation of prediction
tal axis stands for the threshold 𝜀. In the figure, it is apparent error by 22.2 % and 28.6 % when compared to the ABG model
that the proposed model outperforms the conventional models. and the CI model, respectively. Table 4 reveals the prediction
At the 67th percentile, it reduced the error by about 26.6% from errors when the Tx was operating in the 3.5 GHz frequency
15.45 dB to 11.33 dB when compared to the ABG model, and band at 20 m height. At a lower Tx height, the obstruction
by about 28.7% from 15.90 dB to 11.33 dB when compared to of buildings to wave propagation becomes more significant,
the CI model. therefore the proposed model is superior to the conventional
To further explore its reliability, the proposed model was models. Compared to the ABG model and the CI model, the
applied to another simulation environment, which was centred proposed model reduced the standard deviation by 33.3 % and
at latitude 25.060559◦ N and longitude 121.529321◦ E. Readers 37.2 %, respectively.
can use Google Maps or OpenStreetMap to see more details To include Walfisch–Ikegami model, Tables 5 and 6 show
about the regional information. One base station was deployed the prediction errors when operating in the 1.8 GHz band.
at the centre of the simulation environment, and the cell radius Table 5 reveals the prediction errors when the Tx was at
692 JUANG

TABLE 6 Simulation results of path loss prediction error for the


micro-cell base station operating in the 1.8 GHz band at 20 m height

Mean error Standard deviation

ABG model 0.41 dB 13.52 dB


CI model 0.69 dB 14.96 dB
Walfisch–Ikegami Model −6.76 dB 13.95 dB
Proposed model 0.00 dB 10.20 dB
Improve over ABG model – 3.32 dB (24.5%)
Improve over CI model – 4.76 dB (31.9%)
Improve over Walfisch–Ikegami model – 3.75 dB (26.9%)

FIGURE 9 Prediction performance of the proposed path loss model and


the conventional models in measurements

TABLE 7 Path loss prediction error in a real-world 5G system

Mean error Standard deviation

ABG model 0.98 dB 11.78 dB


CI model −0.97 dB 14.58 dB
Proposed model 0.41 dB 8.45 dB
Improve over ABG model – 3.33 dB (28.3%)
Improve over CI model – 6.13 dB (42.0%)

FIGURE 8 Top view of the measurement environment, where the


polygons, triangles, and red dots represent the buildings, cell sectors, and
receivers, respectively
in a vehicle driving along selected routes. The measurement
reports include the reference symbol received power (RSRP),
30 m height. The Walfisch–Ikegami model had a high value in E-UTRAN cell identity (ECI), and the vehicle’s GNSS coordi-
terms of mean error because its parameters were not adjusted nates. Then, the path losses between the Tx–Rx pairs were cal-
to fit the ray-tracing data. However, the mean error can be culated according to the link budgets.
compensated by adding a constant offset. In terms of standard There were 12,616 records collected during the measurement
deviation, the proposed model outperformed the ABG, CI, campaigns. Figure 9 shows the prediction performance of the
and Walfisch–Ikegami models by reductions of 17.1 %, 25.3 proposed path loss model and the conventional models. The
%, and 22.3 %, respectively. Table 6 reveals the prediction black dots are the measurements, whereas the red dots are the
errors when the Tx was at 20 m height. The proposed model values predicted using the proposed model. The blue and the
reduced the standard deviation by 24.5 %, 31.9 %, and 26.9 %, green curves represent the predictions using the ABG and CI
when compared to the ABG, CI, and Walfisch–Ikegami models, models, respectively. Table 7 shows the statistics of prediction
respectively. error resulting from the measurements. The conventional mod-
els and the proposed model show good matches in terms of
the mean prediction error. In terms of the standard deviation,
5 VERIFYING PERFORMANCE USING however, the proposed model outperformed the ABG model
5G NR MEASUREMENT DATA and the CI model by reductions of about 3.33 dB and 6.13 dB,
respectively. Furthermore, Figure 10 shows the cumulative dis-
In addition to simulations, the proposed path loss model was tribution of the absolute error of path loss prediction. The pro-
further applied to a real-world non-standalone 5G NR system posed model outperforms the conventional models. At the 67th
in urban Taipei city. As shown in Figure 8, the measurement percentile, it reduced the error by about 29.7%, from 10.10 dB
environment covered an area of 1 Km x 1 Km. Nine base sta- to 7.10 dB, when compared to the ABG model, and by about
tions operating in the 2.1 GHz frequency band were included 40.5%, from 11.95 dB to 7.10 dB, when compared to the CI
in the measurement area. Measurement equipment was placed model.
JUANG 693

DATA AVAILABILITY STATEMENT


Research data are not shared.

ORCID
Rong-Terng Juang https://orcid.org/0000-0002-9965-2396

REFERENCES
1. Vannithamby, R., Talwar, S.: Towards 5G: Applications, Requirements and
Candidate Technologies. Wiley, Chichester (2017)
2. Mbugua, A.W., Chen, Y., Raschkowski, L., et al.: Review on ray tracing
channel simulation accuracy in Sub-6 GHz outdoor deployment scenarios.
IEEE Open J. Antennas Propag. 69, 4808–4815 (2020). https://doi.org/
10.1109/TAP.2021.3060051
3. Kasdorf, S., Troksa, B., Key, C., et al.: Advancing accuracy of shooting
and bouncing rays method for ray-tracing propagation modeling based
on novel approaches to ray cone angle calculation. IEEE Trans. Anten-
nas Propag. 69, 4808–4815, (2021) (https://doi.org/10.1109/TAP.2021.
3060051)
4. Chen, Y., Yin, R., Gao, B., et al.: Ray tracing on single FPGA. In: IEEE
FIGURE 10 The cumulative distribution of the absolute error of path Asia-Pacific Conference on Image Processing, Electronics and Computers
loss prediction in measurements (IPEC). IEEE, Piscataway (2021)
5. Sun, S., Rappaport, T.S., Thomas, T.A., et al.: Investigation of prediction
accuracy, sensitivity, and parameter stability of large-scale propagation path
6 CONCLUSIONS loss models for 5G wireless communications. IEEE Trans. Veh. Technol.
65(5), 2843–2860 (2016)
6. Casillas-Perez, D., Camacho-Gómez, C., Jiménez-Fernandez, S., et al.:
To facilitate the deployments of 5G networks in urban environ- Weighted ABG: A general framework for optimal combination of
ments, this paper has proposed a machine-learning-based path ABG path-loss propagation models. IEEE Access 8, 101758–101769
loss model utilizing the building profile along the direct prop- (2020)
agation path. Firstly, some attributes of the path profile poten- 7. Samimi, M.K., Rappaport, T.S., MacCartney, G.R.: Probabilistic omnidi-
tially related to path loss were selected. Then the essential fea- rectional path loss models for millimeter-wave outdoor communications.
IEEE Wireless Commun. Lett. 4(4), 357–360 (2015)
tures were extracted out of the attributes using the PCA algo- 8. Zhang, Y., Wen, J., Yang, G., et al.: Path loss prediction based on machine
rithm. Subsequently, the extracted features were used to predict learning: Principle, method, and data expansion. Appl. Sci. 9, 1908 (2019).
the path loss based on the approach of polynomial regression. https://doi.org/10.3390/app9091908
Simulation results showed that the proposed model outper- 9. Thrane, J., Zibar, D., Christiansen, H.L.: Model-aided deep learning
forms the conventional models. The standard deviation of pre- method for path loss prediction in mobile communication systems at 2.6
GHz. IEEE Access 8, 7925–7936 (2020)
diction error was reduced by about 22.2–37.2% when compared 10. Ates, H.F., Hashir, S.M., Baykas, T., et al.: Path loss exponent and shad-
to the conventional models. Furthermore, the proposed model owing factor prediction from satellite images using deep learning. IEEE
was also applied to a non-standalone 5G NR system in urban Access 7, 101366–101375 (2019)
Taipei city. The measurement results showed that the stan- 11. Cheng, H., Ma, S., Lee, H.: CNN-based mmWave path loss modeling for
dard deviation of prediction error was reduced by 3.33–6.13 dB fixed wireless access in suburban scenarios. IEEE Antennas Propag. Lett.
19(10), 1694–1698 (2020)
when compared to the conventional models. Unlike deep learn- 12. Ahmadien, O., Ates, H.F., Baykas, T., et al.: Predicting path loss distribu-
ing approaches, the proposed machine-learning-based model is tion of an area from satellite images using deep learning. IEEE Access 8,
explainable and has the advantage of computational simplicity. 64982–64991 (2020)
The methodology of the proposed path loss modelling can be 13. Saunders, S.R.: Antennas and Propagation for Wireless Communication
further applied to other scenarios for the upcoming 5G deploy- Systems. Wiley, Chichester (1999)
14. Har, D., Watson, A.M., Chadney, A.G.: Comment on diffraction loss of
ments by selecting proper attributes of the path profile. rooftop-to-street in COST 231-Walfisch–Ikegami model. IEEE Trans.
Veh. Technol. 48(5), 1451–1452 (1999)
ACKNOWLEDGEMENTS 15. Getnet, A.: Improving path loss prediction in WCDMA network using
This work was supported by the Ministry of Science and Tech- principal component analysis. In: IST-Africa Conference & Exhibition.
nology (MOST), Taiwan, under Grant No. 109-2222-E-035- IEEE, Piscataway (2013)
16. Liu, B., Gong, C., Xu, Z.: Correlation analysis and path loss prediction for
003-MY2 and by the Taiwan Mobile Co., Ltd., Taiwan. optical wireless scattering communication over broad spectra. In: IEEE
International Conference on Communications (ICC). IEEE, Piscataway
CONFLICT OF INTEREST (2018)
No conflict of interest. 17. 5G; Study on channel model for frequencies from 0.5 to 100 GHz
(3GPP TR 38.901 version 15.0.0 Release 15). ETSI TR 138 901, July
2018. Available: https://www.etsi.org/deliver/etsi_tr/138900_138999/
PERMISSION STATEMENT TO 138901/15.00.00_60/tr_138901v150000p.pdf
REPRODUCE THE MATERIALS FROM THE 18. Sun, S., Rappaport, T.S., Rangan, S., et al.: Propagation path loss mod-
OTHER SOURCES els for 5G urban micro- and macro-cellular scenarios. In: IEEE Vehicular
None. Technology Conference. IEEE, Piscataway (2016)
694 JUANG

19. Piacentini, M., Rinaldi, F.: Path loss prediction in urban environment using 24. Zhu, M., Singh, A., Tufvesson, F.: Measurement based ray launching
learning machines and dimensionality reduction techniques. Computa- for analysis of outdoor propagation. In: 6th European Conference on
tional Manage. Sci. 8, 371–385 (2011) Antennas and Propagation (EUCAP), pp. 3332–3336. IEEE, Piscataway
20. Alpaydin, E., Introduction to Machine Learning, 2nd ed. The MIT Press, (2012)
Cambridge (2014)
21. Pekka, K., Juha, M., Lassi, H., et al.: WINNER II channel models’. Euro-
pean Commission, IST-4-027756-WINNER, Tech. Rep. D1.1.2, Sep. 2007
22. Hany, U., Akter, L.: Local polynomial regression based path loss estimation
forweighted centroid localization of endoscopic capsule. In: International How to cite this article: Juang, R.-T.: Path loss
Conference on Control, Electronics, Renewable Energy and Communica- modelling based on path profile in urban propagation
tions (ICCEREC). IEEE, Piscataway (2017) environments. IET Commun. 16, 685–694 (2022).
23. Ai, B., Guan, K., He, R., et al.: On indoor millimeter wave massive MIMO
channels: Measurement and simulation. IEEE J. Sel. Areas Commun.
https://doi.org/10.1049/cmu2.12369
35(7), 1678–1690 (2017)

You might also like

pFad - Phonifier reborn

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.

Note: This service is not intended for secure transactions such as banking, social media, email, or purchasing. Use at your own risk. We assume no liability whatsoever for broken pages.


Alternative Proxies:

Alternative Proxy

pFad Proxy

pFad v3 Proxy

pFad v4 Proxy