5G Coverage Optimization using Lower and Higher

(mmWave) Frequency Bands

A. Alves
Instituto Superior Técnico,
University of Lisbon
Lisbon, Portugal
andre.q.alves@tecnico.ulisboa.pt

Abstract—This paper evaluates 5th Generation (5G) propa- The mobile industry has found benefitial to have separate
gation models for Outdoor and Outdoor-to-Indoor (O2I) tran- pathloss models for Line-of-Sight (LOS) and Non Line-of-
sitions, applying them to an open area and a Manhattan-like Sight (NLOS) conditions, due to the radio channel significant
scenarios, at 3.5 and 28 GHz. In addition, Line-of-Sight (LOS)
condition to a Base Station (BS) is assessed deterministically, and differences. Considering the newly available bands at higher
a novel 3D beamforming antenna model is proposed. The use of frequencies, the LOS and NLOS propagation present even
a massive Multiple Input Multiple Output (mMIMO) antenna more distinct channels, especially at mmWaves [1]. Therefore,
and 3D beamforming allowed an average peak throughput of it is crucial to identify the LOS and NLOS areas to after-
3.23 Gbps at 28 GHz, although the Inter-Site Distance (ISD) wards correctly apply the pathloss model. The most common
must be shorter to ensure seamless coverage. In the presence of
high-rise buildings, it is observed that radiation patterns with approach is to use LOS probability models, i.e., statistical
multiple vertical beams improved indoor coverage in, at least, models to predict the likelihood that a User Equipment (UE)
24.4%. Moreover, data from three measurement campaigns were is in a LOS condition with respect to the Base Station (BS). In
employed to derive two calibrated models. The results show, in [4], LOS/NLOS classification was done visually using Google
the worst case, a Mean Absolute Error (MAE) of 7.51 dB and a Maps. In this paper, one benefits from increasingly available
Root Mean Squared Error (RMSE) of 9.17 dB.
Index Terms—5G, Propagation Models, Line-of-Sight, geodata to deterministically calculate whether or not a UE
Pathloss, Beamforming, Calibration. has LOS to the respective BS, thus ensuring a more reliable
classification.
I. I NTRODUCTION In this work, two propagation models, 3GPP and
The past few years have witnessed a massive prolifera- Millimetre-Wave Based Mobile Radio Access Network for
tion of mobile applications and services, which have caused Fifth Generation Integrated Communications (mmMAGIC),
a tremendous increase in mobile networks traffic. The 5th with different antenna patterns and at FR1 and FR2 fre-
Generation (5G) is expected to benefit from the extension quencies, are applied to testing scenarios in order to assess
of used spectrum in order to improve network capacity and coverage and throughput. Moreover, data from three sets of
achieve high throughputs. The Third Generation Partnership measurement campaigns are used to calibrate the models.
Project (3GPP) defines two major frequency bands on which The paper is organized as follows: in Section II two antenna
5G will operate; Frequency Range (FR) 1, that accommodates mask models are introduced; Section III presents the propa-
the frequencies from 410 MHz to 7125 MHz, and FR2, from gation models and the testing scenarios results; in Section IV
24250 MHz to 52600 MHz, which include Millimeter Wave the Drive Test (DT) data are analyzed and the models are
(mmWave) frequencies. calibrated. Finally, some conclusions are drawn in Section V.
It is then crucial to be aware of propagation characteristics
in these frequency bands in order to properly plan and opti-
mize coverage. Therefore, many measurement campaigns have II. A NTENNA M ASK M ODELS
been performed by several institutions to accurately model
propagation at 5G frequencies. Standard institutions usually The variation of a radio signal strength between the BS and
develop omnidirectional pathloss models, i.e., assuming the a UE is described by propagation models. But as important
use of unity gain antennas [1]. However, concepts such as as quantifying the losses in the propagation channel is to
spatial multiplexing and 3D beamforming are key to enable characterize both link ends, namely the BS antenna mask
5G, hence these path loss models are not usable in directional which results in the transmitting antenna gain.
antenna system analysis unless the antenna patterns are prop- This section starts by presenting the standard 3GPP antenna
erly modeled [1]. The 3GPP antenna pattern model is the most model, which is a simple model and allows a single beam,
common approach, with a horizontally-wide and vertically- either in the horizontal and vertical planes.
narrow single beam [2], [3]. A new approach to beamforming Further on, due to the limitations of the 3GPP model, a
antenna pattern modeling is proposed, enabling the simulation new model was developed in the course of this work, enabling
of antenna patterns with multiple beams and configurations. multiple beams in both horizontal and vertical planes.
A. 3GPP Antenna Model where φAbeam is given by:
This section describes the standard directional antenna pat-
maxrange,H − minrange,H
tern model, proposed by 3GPP to introduce 3D beamforming. φAbeam = (6)
The radiation pattern in the horizontal plane is given by: nH
(  )
φ − φaz
2 and i identifies the respective beam and is given by:
GH (φ) = − min 12 , F BR (1)
φ3dB
i = min{bmax(φ − minrange,H , 0)/φAbeam c, nH − 1} (7)
where φ (∈ [−180◦ , 180◦ ]) is the horizontal angle measured
between the BS boresight and the line in the horizontal plane Similarly, the vertical radiation pattern is obtained as fol-
connecting the UE to the BS, φaz denotes the fixed orientation lows:
angle of BS boresight relative to the x-axis, φ3dB is the  !2 
0
horizontal Half-Power Beamwidth (HPBW), and F BR is the  θ − kV 
front-to-back ratio. GV (θ) = − min 12 , SLL (8)
 θ3dB 
Similarly, the vertical radiation pattern may be obtained as
follows: 0
(  2 ) where kV is the angle at which an vertical beam is centered,
θ − θtilt and is given by:
GV (θ) = − min 12 , SLL (2)
θ3dB  
0 1
where θ (∈ [−90◦ , 90◦ ]) is the vertical angle measured be- kV = minrange,V + θAbeam +j (9)
2
tween the horizon and the line connecting the UE to the BS,
θtilt denotes the antenna tilt and is measured between the where θAbeam is given by:
horizon and the line passing through the peak of the beam,
θ3dB is the vertical HPBW, and SLL is the side-lobe level. maxrange,V − minrange,V
Finally, the 3D antenna gain can be obtained as: θAbeam = (10)
nV
GdBi
BS (θ, φ) = Gm − min{−[GH (φ) + GV (θ)], Gm } (3) and j identifies the respective beam and is given by:
where Gm denotes the peak antenna gain.
j = min{bmax(θ − minrange,V , 0)/θAbeam c, nV − 1} (11)
B. Newly-developed Beamforming Antenna Model
As one of the main features of 5G, massive Multiple Input The 3D antenna gain GdBi BS is obtained from (3). An example
Multiple Output (mMIMO) antennas implement beamforming of the application of this antenna model is shown in Fig. 1.
[5] to form extremely accurate user-level narrow beams. A
mMIMO antenna can improve signal coverage and reduce
interference between cells. This section presents a new ap-
proach in antenna modelling which is compatible with the 5G
requirements and, although a simplified model, closer to the
used 5G Active Antenna Systems (AASs).
In this model, both horizontal and vertical HPBW are the
HPBW per beam. Therefore, if the horizontal and vertical
HPBW read from an antenna datasheet are the total HPBW,
then they must be divided by the number of horizontal and
vertical beams, respectively, before being inputted to the
model. The other model inputs are the number of horizontal Fig. 1. Radiation Pattern (Gm = 25 dBi, 4 hor. beams, 2 vert. beams,
φ3dB = 11.25◦ , θ3dB = 12◦ , hor. and vert. scanning range are [-60◦ ,60◦ ]
(nH ) and vertical (nV ) beams, and the horizontal and vertical and [-15◦ ,15◦ ], respectively.
scanning range of beams, [minrange , maxrange ], which are
centered at the φaz and at the θtilt , respectively.
The horizontal radiation pattern is obtained as follows:
  III. P ROPAGATION F RAMEWORK
0
!2
 φ − kH 
This section starts off by describing the procedure to de-
GH (φ) = − min 12 , F BR (4)
 φ3dB  termine whether a UE is in LOS or NLOS to the respective
0
BS. Afterwards, two propagation models, concerning outdoor
where kH is the angle at which an horizontal beam is centered, propagation and Outdoor-to-Indoor (O2I) transitions, are in-
and is given by: troduced and employed in an open area and a Manhattan-like
testing scenarios. The parameters of these two propagation
 
0 1
kH = minrange,H + φAbeam +i (5) models are given in Table I.
2
A. LOS/NLOS The 3GPP TR 38.901 [3] also has a O2I penetration loss
Most 5G propagation models have different path loss equa- model, which is useful to describe the additional losses that an
tions for LOS and NLOS scenarios. In this paper, the author indoor UE may experience. Thereby, the pathloss experienced
benefits from geodata (namely terrain and buildings [6]) to by an indoor user may be modeled as follows:
deterministically determine whether a UE is in LOS conditions P L = P Lb + P Ltw + P Lin + N (0, σP2 ) (16)
to the BS or not, i.e., whether the 1st Fresnel Ellipsoid radius
is at least 60% clear or not [7]. The radius of the 1st Fresnel where P Lb is the basic outdoor pathloss, where d3D is re-
Ellipsoid in a point of the link between the BS and a UE placed by d3D-out and d3D-in (outdoor and indoor 3D distance,
which is distance z away from one of the terminals (r1e,z ), respectively); P Ltw is the building penetration loss through
being obtained from: the external wall; P Lin is the inside loss dependent on the
s depth into the building; and σP is the standard deviation for
z(d2D − z) the penetration loss. The penetration loss is derived by:
r1e,z = λc (12)
d2D
N 
X Lmaterial,i 
where λc is the wavelength and d2D is the 2D distance be- P Ltw = P Lnpi + 10 log10 pi × 10− 10 (17)
tween the BS and the UE. Afterwards, the difference between i=1
the height of the LOS beam and the geodata height at point z where P Lnpi is an additional loss added to the external wall
is computed: loss to account for non-perpendicular incidence, which is 5 dB
∆hz = hgeodata,z − hray,z (13)
P the 3GPP model, pi is the proportion of the i-th material,
in
where hgeodata,z is the terrain plus buildings height at z and pi = 1, Lmaterial,i = amaterial,i +bmaterial,i ·fc is the penetration
hray,z = mz + b, with m and b being, respectively, the slope loss of material i, where fc is the frequency in GHz, and N is
and intercept of the line equation of the LOS beam between the number of materials. Penetration loss of several materials
BS and UE. may be found in [3].
Finally, the condition which defines whether or not the 1st Two O2I penetration loss models are provided in [3]: a
Fresnel zone obstruction at z is given by: low-loss and a high-loss model, depending on the building
 materials. In this work, an intermediate model is used, where
NLOS if ∆hz > 0 or
 both models contribute with 50% to the total loss.
Link = (∆hz < 0 and |∆hz | < 0.6 · r1e,z ) (14)
 C. mmMAGIC
LOS otherwise

The mmMAGIC models omnidirectional pathloss from 6-
The aforementioned procedure must be run iteratively over 100 GHz [8]. The UMi pathloss also follows the ABG
the link between the BS and a UE. modeling and the distribution of the SF is log-normal. The
B. 3GPP TR 38.901 model parameters are provided in Table I.
The O2I building penetration loss model is basically the
The 3GPP TR 38.901 pathloss models are valid from 0.5-
same as the corresponding model agreed in 3GPP [3]. There
100 GHz. The 3GPP Urban Micro (UMi) and Urban Macro
are, however, a couple of additional terms introduced:
(UMa) models for LOS conditions are breakpoint models, dB
• a log-normal frequency dependent spread σP = 4+kσ fc
i.e., have different equations depending on whether d2D is
0
smaller than the breakpoint distance dBP or not. The breakpoint where kσ is building dependent.
• a term to account for elevation angle loss Lel =
distance can be defined as the distance from the BS where
the 1st Fresnel ellipsoid touches the ground and the Pathloss 20 |θ/90◦ |.
Exponent (PLE) shifts from free space (PLE = 2) to the
D. Testing Scenarios
asymptotic two-ray ground bounce model (PLE = 4) [1]. The
0
LOS breakpoint distance dBP in meters is given by: The propagation models are now applied in several scenar-
0 0 0 ios at 3.5 and 28 GHz, using different antenna patterns, and
dBP = 4hBS hUE /λc coverage and peak throughput [9] are analyzed. It is assumed 2
0
hBS = hBS + hgeodata,BS − hgeodata,UE (15) MIMO layers and no carrier aggregation. At 3.5 GHz, it is used
0 a 100 MHz bandwidth and numerology 2, while at 28 GHz,
hUE = hUE
400 MHz bandwidth and numerology 3 are assumed. The
In NLOS conditions, the 3GPP pathloss models follow used BS antennas are: Kathrein 742212, which is a sectorial
the Alpha-Beta-Gamma (ABG) modeling, with an additional antenna (Gm = 18 dBi, φ3dB = 63◦ , θ3dB = 6.5◦ ), and the
correction term for the UE height. The UMi and UMa model AAU5613, which is a mMIMO antenna with a 8×12 phased
parameters are presented in Table I. array, allowing multiple-beam radiation patterns (Gm = 25
Moreover, the Shadow Fading (SF) for the 3GPP models dBi). Antenna tilts were not object of optimization, and are
follow a log-normal distribution (see standard deviation σSF such that the vertical gain is 6 dB below the maximum in the
in Table I). horizon direction.
 TABLE I
UM A AND UM I PATHLOSS M ODELS

σSF Applicability range

Model P L [dB], fc [GHz], d [m]
[dB] and default values
0
(
P L1 if 10 m ≤ d2D ≤ dBP
P LUMa-LOS = 0
P L2 if dBP < d2D ≤ 5 km

LOS P L1 = 28 + 22 log10 (d3D ) + 20 log10 (fc ) 4 0.5 <fc < 100 GHz
P L2 = 28 + 40 log10 (d3D ) + 20 log10 (fc ) 10 <d2D < 5000 m
3GPP UMa 0
[3] 2 2
− 9 log
 10 ((dBP ) + hef f ) 1.5 ≤hUE ≤ 22.5 m
0

P LUMa-NLOS = max P LUMa-LOS , P LUMa-NLOS hBS = 25 m
NLOS 0 6
P LUMa-NLOS = 13.54 + 39.08 log10 (d3D ) + 20 log10 (fc )
− 0.6
(log10 (hUE − 1.5) 0
P L1 if 10 m ≤ d2D ≤ dBP
P LUMi-LOS = 0
P L2 if dBP < d2D ≤ 5 km

LOS P L1 = 32.4 + 21 log10 (d3D ) + 20 log10 (fc ) 4 0.5 <fc < 100 GHz
P L2 = 32.4 + 40 log10 (d3D ) + 20 log10 (fc ) 10 <d2D < 5000 m
3GPP UMi 0
[3] 2 2
− 9.5 log
 10 ((dBP ) + hef f ) 1.5 ≤hUE ≤ 22.5 m
0

P LUMi-NLOS = max P LUMi-LOS , P LUMi-NLOS hBS = 10 m
NLOS 0 7.82
P LUMi-NLOS = 22.4 + 35.3 log10 (d3D ) + 21.3 log10 (fc )
− 0.3 log10 (hUE − 1.5)
LOS P LUMi-LOS = 32.9 + 19.2 log10 (d3D ) + 20.8 log10 (fc ) 2 6 < fc < 100 GHz
mmMAGIC
1.5 ≤ hUE ≤ 3 m
UMi [8] NLOS P LUMi-NLOS = 31 + 45 log10 (d3D ) + 20 log10 (fc ) 7.82 1.5 ≤ hBS ≤ 10 m

1) Open Area: The open area testing scenario consists of 19 2) Manhattan-like: The Manhattan-like scenario has been
3-sectorized sites (BSs) displayed in a hexagonal grid and flat widely used for multiple radio technologies. A slightly mod-
terrain without obstacles. The UE height is always assumed ified Manhattan-like scenario is used (Fig. 2). Sidewalks (in
to be 1.5 m and outdoor, and LOS and NLOS propagation dark brown) around the buildings (light brown) were added,
are analysed separately. Two different configurations are sim- separating them from the road. Thus, this Manhattan-like
ulated: a UMa scenario and a UMi scenario. scenario consists of a 5x5 building grid, where the building
The UMa scenario is characterized by a 500 m Inter-Site width is 40 m and the street is 20 m wide (16 m for road
Distance (ISD), BS height is 25 m, and their transmitting width and 2 m for each sidewalk width). Buildings are 25 m
power is 40 W. At 3.5 GHz, coverage is very close to 100% height, and are divided in 8 floors considering that each floor
both in LOS and NLOS conditions, whereas at 28 GHz is 3 m height, and the sidewalks are 0.2 m height. The BSs
coverage barely surpasses 20% in NLOS for the best covering height is 10 m, and are located on lampposts in the middle of
antenna pattern (AAU5613, 8 hor. beams and 1 vert. beam, the sidewalks (i.e. 1 m away from the building wall and 1 m
110◦ hor. HPBW, 12◦ hor. HPBW). On the other hand, at away from the road edge), which are assumed to be regularly
28 GHz, it is possible to attain very high throughputs, as the displayed along the sidewalks. For the sake of simplicity, in
AAU5613 reaches an average throughput of 3.23 Gbps in LOS Fig. 2 only the lampposts which give support the the BS are
and 1 Gbps in NLOS and nearly 100 Mbps in the cell edge, represented (black dots). The terrain is assumed to be flat.
outperforming the Kathrein 742212 (3.1 Gbps in LOS and Link budget calculations were performed to determine the
0.22 Gbps in NLOS). maximum ISD to ensure coverage in worst case, which is an
The UMi scenario is characterized by a 200 m ISD, BS indoor UE at 28 GHz. This results in a maximum distance of
height is 10 m, and their transmitting power is 20 W. The around 96 m. Thus, and following the display in [10], there
reduced ISD allows an improved coverage at 28 GHz (52% are 15 3-sectorized BS with an average ISD of 92.5 m, as
for the best antenna pattern) although insufficient, reinforcing shown in Fig. 2.
the idea that for mmWaves, small cells and LOS may be neces- Due to link budget planning, near 100% outdoor and indoor
sary. Again, the AAU5613 antenna outperforms the Kathrein coverage is attained at both 3.5 and 28 GHz. While, at 3.5
742212, and higher throughputs are also achieved, with 3.5 GHz, the average outdoor and indoor peak throughput is
Gbps in LOS and 1.8 Gbps in NLOS, and 835 Mbps in the cell 761 and 349 Mbps, respectively, at 28 GHz the values are
edge for the best pattern. Moreover, when comparing the 3GPP significantly higher. At 28 GHz, the average outdoor and
and the mmMAGIC models, the latter’s 50% percentile NLOS indoor peak throughput is 2.98 and 1.05 Gbps, respectively,
pathloss is 26 dB lower, which is a significant difference. and the outdoor and indoor cell edge throughput peak at 2.2
 Fig. 3. Saldanha Indoor Coverage with Pattern 1.

radiation pattern. With smaller tilt values, the indoor coverage

is smaller, but as the tilt increases the radiation pattern is more
Fig. 2. Manhattan-like scenario. pointed to the buildings hence almost full indoor coverage is
attained. When the tilt is very high, the radiation pattern points
to the ground, deteriorating indoor coverage.
Gbps and 67 Mbps, respectively. The antenna downtilt also
plays an important role in coverage planning. In this work,
the tilt was set so that the antenna radiates 6 dB below the
maximum vertical gain in the horizon in order to improve
outdoor coverage. Therefore, the antenna pattern that shows
better results is the AAU5613 antenna with 8 hor. beams
and 1 vert. beam, 110◦ hor. HPBW, and 12◦ hor. HPBW.
However, as far as indoor UEs are concerned, the radiation
pattern in Fig. 1 shows a performance almost as good as the
aforementioned radiation pattern. Hence, had tilt optimization
been done concerning also indoor UEs, the best antenna
pattern for indoor coverage would be the pattern in Fig. 1.
E. Lisbon
The purpose of this scenario is to demonstrate how areas Fig. 4. Saldanha Indoor Coverage with Pattern 9.
with distinct morphology may require different antenna
parameters for an improved coverage. Amongst the simulated As for Pattern 9, the two scanning ranges provide distinct
scenarios in Lisbon are an area in Saldanha with a large results. The radiation pattern with the range [θtilt − 15◦ ,
building density, with an average height of 20.1 m, and θtilt + 15◦ ] and 0◦ tilt behaves as Pattern 1 (but with 4
a hotel in Picoas which is 92 m tall, and it is considered horizontal beams only) because the lower vertical beam points
that a dedicated cell from a BS is used. The BS is on a to the buildings whereas the upper beam radiates energy to
nearby building, and is at a 26 m height. All simulations are the ”air”. As the tilt increases, the lower beam points more
performed at 3.5 GHz. to the ground while the upper beam does not point yet in
the direction where most buildings would be covered, hence
Saldanha Area indoor coverage deteriorates. When the maximum of the upper
Firstly, the Saldanha area is studied. The first test performed beam radiates at a higher elevation angle, indoor coverage
is the comparison between Pattern 1 and Pattern 9. In the improves again. With a smaller vertical scanning range of
previous testing scenarios the antenna parameters were kept beams, [θtilt − 7.5◦ , θtilt + 7.5◦ ], indoor coverage is almost
constant, namely the scanning ranges of beams and the antenna constant around 90%. The wider the vertical scanning range
tilt, which was set so that the vertical beam radiates 6 dB below of beams, the more dispersed the radiated energy. Thus with
the maximum gain in the horizon direction; however, in this the smaller range, the radiated energy is more focused around
scenario, tilt values from 0◦ to 12◦ are tested, as well as two the tilt angle, and the radiation pattern minimum is higher
different vertical scanning ranges of beams. than with the previous range. There are still the variations that
The results for the Saldanha area are shown in Figures 3 there were in the [θtilt − 15◦ , θtilt + 15◦ ] range, but they are
and 4. For Pattern 1, since it only has one vertical beam, much smaller. Therefore, one concludes that having a wider
changing the vertical scanning range does not affect the range of beams may not always be beneficial due to the high
energy dispersion. In addition, these simulations showed the to provide almost maximum gain across the building, however
difference in the tilt concept which was described above. only in one direction. However, the building height is almost
From the results above, one must find odd that the radia- twice its width, hence covering in the vertical plane is more
tion pattern that provides the highest percentages of indoor important than in the horizontal direction to the overall indoor
coverage is Pattern 1, which has only one vertical beam, coverage. A simulation with Pattern 15 (2 horizontal and 4
when compared to radiation patterns with multiple beams. vertical beams), which is a radiation pattern with 2 horizontal
The explanation is actually a simple one: the eight horizontal beams and 4 vertical beams, improved the indoor coverage
beams of Pattern 1 provide better coverage across the cell’s even more, attaining 97.8%.
service area, which is mainly comprised by indoor areas A top view fooprint of the Reference Signal Received Power
(buildings). With the use of mmWaves, the 5G network will (RSRP) per floor is showed in Figure 6, and it is visible how
be immensely densified with micro and small cells, and some the signal varies from floor to floor1 .
cells will not have to cover larger areas, especially in highly
populated and densely built areas with high-rise buildings and
skyscrapers, where the radiation patterns with multiple vertical
beams will truly improve indoor coverage when compared to
a radiation pattern such as Pattern 1.
It may also be concluded that changing the radiation
pattern per se may not produce the results one could expect.
The analysis must be broader than that: the geometry of the
problem, i.e., the area to be covered and the characteristics of
that area, the BS(s) location and height, the scanning range
of beams, and the antenna tilt.

Picoas Area
In the Picoas area, it is only simulated the indoor coverage
of one particular building. It is a well-known hotel in Lisbon Fig. 6. RSRP per floor of the high-rise building in Picoas for Pattern 9.
and one of the tallest buildings in the city, with a 92 m height,
and the aim is to demonstrate that radiation patterns with IV. C ALIBRATION
multiple beams truly outperform single-beam patterns in such
There is usually some disagreement between the values
scenarios. Assuming an uniform distribution of UEs, the tilt
which result from the use of an empirical model and the
is set so that it points approximately to the middle of the
actual measured data for a specific area. Propagation models
building. The horizontal scanning of beams is [φaz − 25◦ ,
are usually based on extensive measurement campaigns, which
φaz + 25◦ ], and the vertical scanning of beams is [θtilt − 15◦ ,
are strongly dependent on the specific geographical area (e.g.,
θtilt + 15◦ ].
heavy clutter or other topological peculiarities [11]), frequency
The results are presented in Figure 5. As shown in the figure,
band, or weather conditions.
In order to have a better coverage and network planning,
the propagation model parameters must be optimized for each
geographical location or for several similar ones. DT data are
used to compare with the predicted results, and based on that
comparison, calibrate the propagation models using the linear
regression algorithm.
A. Drive Tests
Three DT datasets were used: one in Ciudad Real, Spain,
and the others in Italy (Milan and Ivrea). The Ciudad Real
DT was performed at 3.7 GHz and there are a total of 3 sites
and 4 sectors at a 20 m height. The Milan and Ivrea DTs were
performed at 26 GHz, the data concerns to one sector for each
DT, and the BS are 30 m and 15 m height, respectively. All
samples are outdoor and are taken at a 1.5 m height.
The first step towards models calibration was to obtain
Fig. 5. Picoas Indoor Coverage, with several radiation patterns.
geodata for the aforementioned locations, and thereupon input
it to the LOS algorithm (Section III-A) to separate the data by
from Pattern 1 to Pattern 9 there is a major improvement in the
building indoor coverage, from 58.7% to 83.1%. With such a 1 Due to technical issues which could not be fixed in due time, the building
narrow horizontal scanning range of beams, Pattern 1 is able is not fully represented in the figure.
LOS and NLOS. Fig. 7 exemplifies the procedure described, C. Linear Regression
and visually confirms the outcome of the LOS algorithm. In this work, only the 3GPP UMa model was calibrated. All
BSs heights are much higher than the default values for the
3GPP and mmMAGIC UMi, except in the Ivrea DT. However,
this DT only has 14 LOS samples, which is not enough for
calibration, thus only NLOS calibration would be possible.
All DT samples were found to fall short of the breakpoint
distance. Thus, since the 3GPP UMa model (LOS P L1 and
NLOS) is ABG-type, linear regression was the method chosen
to perform the model calibration. Two calibrated models were
derived:
• A Full-Spectrum model, i.e., for FR1 and FR2, where
all DTs are included in the dataset. Both LOS and NLOS
may be modeled as:
P L = 10α log10 (d3D ) + β + 10γ log10 (fc ) (20)
• A 26 GHz model, where only the Italy DTs are used for
Fig. 7. Milan DT classified in LOS and NLOS. calibration. Modeling is the same as in (20), but since fc
is constant it may be rewritten as in (21), simplifying the
Afterwards, the pathloss is derived from the measured UE regression.
received power, the downlink reference signal power, and the
P L = 10α log10 (d3D ) + ν (21)
3D antenna gain, which is determined using the model in
Section II-B knowing the antennas used for the DTs. where ν = β + 10γ log10 (28).
The Python library scikit-learn with the Sum of Squared
B. Slow Fading Filtering Errors (SSE) as cost function was used to perform the linear
regressions. The results are presented in Table II and compared
Propagation models describe the average pathloss experi- to the original pathloss model.
enced in a mobile radio environment. However, field mea-
surements are instantaneous, hence they include slow fading TABLE II
(shadowing) and fast fading (multipath). Measured data S may M ODEL C ALIBRATION R ESULTS
then be modeled as:
α β γ MAE RMSE R2
3GPP UMa Model
S(t) = M (t) + R(t) (18) LOS 2.2 28 2 21.05 23.42 -
NLOS 3.908 13.54 2 14.48 16.66 -
Full-Spectrum Model
where M is the mean pathloss and R denotes a fading LOS 4.53 20.87 0.40 5.45 7.26 0.50
term. Therefore, before making any comparisons between the NLOS 3.8 24.67 2.46 7.51 9.17 0.58
DT measurements and propagation models, a sliding-window 26 GHz Model
LOS 4.71 -5.45 2 5.0 7.08 0.56
algorithm is employed to filter R out. Fast fading is not NLOS 3.78 31.78 2 7.16 8.71 0.53
considered since it is out of the scope of this work.
The filtering process can be carried out by averaging the
samples over a spatial and temporal range of samples where The results for LOS conditions are far from what was
the mean signal (pathloss) is considered to be constant. Let Si expected, with a high PLE while it should be closer to free
be the ith sample. Hence, the ith estimated mean M̂i is space propagation. This may be explained by the few LOS
samples in the Ciudad Real DT, and whose received power
max(X) appear to be lower than usual, perhaps because of destructive
1 X
M̂i = Sk (19) interference. This leads, on one hand to, a very small γ, and on
|X| the other hand, to a larger α. In the 26 GHz model, the offset
k=min(X)
value is a negative value, nonetheless in ABG modeling, β has
where X = {n : d(Si , Sn ) ≤ L ∧ t(Si , Sn ) ≤ ∆t}. The no physical meaning, being only an optimization parameter.
parameters d(Si , Sn ) and t(Si , Sn ) are the distance and time Generally, Mean Absolute Error (MAE) and Root Mean
interval between samples Si and Sn , respectively; ∆t = L/v Squared Error (RMSE) are improved after the model calibra-
and L define the window size and are the maximum time inter- tion, and the coefficient of determination R2 ranges from 0.50
val and distance, respectively, where the signal is considered to 0.58. There are some associated errors to the calibration
to be sensibly constant, v is the average DT speed. According procedure which may prevent the results to be better: there was
to [12], L must be between 20 and 40λ. some uncertainty relative to the BSs downtilt; the used antenna
model is only a simplified model thus some inaccuracies are [6] P. Vieira, M. P. Queluz, A. Rodrigues, “A dynamic propagation pre-
expected; the LOS algorithm may also have experienced some diction platform over irregular terrain and buildings for wireless com-
munications”, 66th IEEE Vehicular Technology Conference (VTC – Fall
errors due to the buildings data spatial display being slightly 2007), Baltimore, USA, October 2007.
imprecise, and the buildings height in Ciudad Real and Ivrea [7] Carlos Salema, Microwave Radio Links, 1st edition, John Wiley and
had a lower resolution. Sons Inc, 2002.
[8] mmMAGIC D2.2, ”Measurement Results and Final mmMAGIC Channel
If one had access to DTs in the area of Lisbon where geodata Models”, v2.0, ICT-671650, mmMAGIC Project, 2017.
are available, then calibrated models could be derived and [9] 3GPP TS 38.306, User Equipment (UE) radio access capabilities, 3GPP,
used in Section III-E to perform the simulations instead of October 2018.
[10] 3GPP TR 25.942, RF System Scenarios, 3GPP, June 2006.
the propagation models from Section III, thus providing more [11] P. Vieira, M. P. Queluz, A. Rodrigues, “Clustering of scatterers over an
trustworthy results. irregular clutter environment: an extension of COST 273 MIMO channel
model”, 66th IEEE Vehicular Technology Conference (VTC – Fall 2007),
V. C ONCLUSIONS Baltimore, USA, October 2007.
[12] W. C. Y. Lee, ”Estimate of local average power of a mobile radio signal,”
This paper presents a UMa and two UMi 5G propagation in IEEE Transactions on Vehicular Technology, vol. 34, no. 1, pp. 22-27,
models, as well as a O2I penetration loss model. Furthermore, Feb. 1985.
a novel antenna mask model, prepared to deal with 3D
beamforming, is introduced, and a deterministic approach to
determine LOS or NLOS is used. The results for the testing
scenarios show that on one hand, FR1 frequencies provide
a compromise between coverage and throughput, but on the
other hand, higher frequencies (FR2) are the key to unlock
the full potential of 5G, making use of larger bandwidth to
attain throughput up to 3.23 Gbps. In addition, the use of
3D beamforming helps to overcome the higher losses at 5G
frequencies by enabling high-gain directive beams.
Additionally, three measurement campaigns and respective
geodata were analyzed in order to calibrate the UMa model.
The data fading component was removed using a sliding-
window filtering algorithm, and Linear Regression was then
applied to determine the model parameters which minimize
the SSE. Two models were obtained: a full-spectrum model,
with a MAE of 7.51 dB and a RMSE of 9.17 dB (worst case);
and a 28 GHz model , with a MAE of 7.16 dB and a RMSE
of 8.71 dB (worst case). Thus, when performing simulations
at 26 GHz, the latter model should be used since it is based
only on measurements for that frequency, and it provides the
smaller error.
ACKNOWLEDGMENT
An acknowledgment is due to CELFINET and Instituto de
Telecomunicações (IT) for the support to this work.
R EFERENCES
[1] T. S. Rappaport, Y. Xing, G. R. MacCartney, A. F. Molisch, E. Mellios
and J. Zhang, ”Overview of Millimeter Wave Communications for Fifth-
Generation (5G) Wireless Networks—With a Focus on Propagation
Models,” in IEEE Transactions on Antennas and Propagation, vol. 65,
no. 12, pp. 6213-6230, Dec. 2017.
[2] N. Seifi, J. Zhang, R. W. Heath, T. Svensson and M. Coldrey, ”Co-
ordinated 3D Beamforming for Interference Management in Cellular
Networks,” in IEEE Transactions on Wireless Communications, vol. 13,
no. 10, pp. 5396-5410, Oct. 2014.
[3] 3GPP TR 38.901, Study on channel model for frequencies from 0.5 to
100 GHz, 3GPP, October 2019, 16.0.0.
[4] S. Sun, T. A. Thomas, T. S. Rappaport, H. Nguyen, I. Z. Kovacs and I.
Rodriguez, ”Path Loss, Shadow Fading, and Line-of-Sight Probability
Models for 5G Urban Macro-Cellular Scenarios,” 2015 IEEE Globecom
Workshops (GC Wkshps), San Diego, CA, 2015, pp. 1-7.
[5] P. Vieira, M. P. Queluz, A. Rodrigues, “MIMO antenna array impact on
channel capacity for a realistic macro-cellular urban environment”, 68th
IEEE Vehicular Technology Conference (VTC – Fall 2008), Calgary,
Canada, September 2008.