©2019 IEEE. Personal use of this material is permitted.

Permission from IEEE must be obtained for all other uses, in any current or future media, including
reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or
reuse of any copyrighted component of this work in other works. DOI: 10.1109/VTCSpring.2019.8746551

3.5 GHz Coverage Assessment with a 5G Testbed

Adrian Schumacher*# , Ruben Merz* and Andreas Burg#
*
Enterprise Architecture & Innovation, Swisscom Ltd., CH-3050 Bern, Switzerland
{adrian.schumacher,ruben.merz}@swisscom.com
#
Telecommunications Circuits Laboratory, EPFL, CH-1015 Lausanne, Switzerland
andreas.burg@epfl.ch

deterministic models (e.g., ray tracing) may yield accurate

Abstract—Today, cellular networks have saturated frequencies results, they require a high degree of detailed description
below 3 GHz. Because of increasing capacity requirements, 5th and calibration of the environment and are computationally
generation (5G) mobile networks target the 3.5 GHz band (3.4 to
arXiv:2105.06812v1 [cs.NI] 14 May 2021

3.8 GHz). Despite its expected wide usage, there is little empirical expensive. Stochastic models (e.g., COST 207) employ
path loss data and mobile radio network planning experience random variables and do not require much information about
for the 3.5 GHz band available. This paper presents the results the environment, but also cannot provide high-accuracy
of rural, suburban, and urban measurement campaigns using a path loss predictions. Empirical (e.g., Hata-Okumura) and
pre-standard 5G prototype testbed operating at 3.5 GHz, with standardized models are based on measurements and have been
outdoor as well as outdoor-to-indoor scenarios. Based on the
measurement results, path loss models are evaluated, which are widely adopted to predict a mean path loss as a function of
essential for network planning. distance, frequency, and some further parameters specific to
Index Terms—5G, sub-6 GHz, 3.5 GHz, propagation measure- the environment or infrastructure. They require also only a few
ments, beamforming parameters and can provide an acceptable prediction accuracy,
better than the very general and simplified path loss model (1).
I. I NTRODUCTION For these reasons, radio network planners in the industry often
The global demand for higher-capacity mobile Internet ser- use empirical path loss models for network planning simulations.
vices drives the continuous evolution of mobile (cellular) tech-
A. Overview of Empirical Path Loss Models
nologies. Up until the 4th generation mobile network (4G,
also commonly known as Long Term Evolution (LTE)), only Because most mobile communications applications use fre-
frequency bands up to around 2.7 GHz have been used (with a quencies up to around 2 GHz, the empirical path loss models
few exceptions). To carry the required increase in data capacity, were optimized to cover only these frequencies. Examples are
more bandwidth in higher frequency spectrum is required. (see also Table I): Hata-Okumura [2], COST 231 Hata [3],
CEPT/ECC identified1 the 3400–3800 MHz frequency band to COST 231 Walfish-Ikegami (COST 231 WI) [3], and Erceg [4].
be harmonized in Europe for 5G usage, along with the 24.25– Path loss models valid for the 3.5 GHz band are the Stanford
27.5 GHz band. National regulators already have made, or are in University Interim (SUI) IEEE 802.16 model [5] which is an
the process of making the 3.4–3.8 GHz (referred to as 3.5 GHz) extension to the Erceg model, the ECC-33 model [6] which
band available for mobile network operators. For instance, in extrapolates the Hata-Okumura model, the WINNER II models
Switzerland, an auction for this band was conducted during [7], the ITU-R P.1411-9 models [8], as well as the 3GPP models
the first quarter of 2019. With 5G, we refer to the Release 15 described in [9].
New Radio (NR) standard by the 3rd Generation Partnership When designing wireless networks, it is obvious that underes-
Project (3GPP). 5G NR is frequency band agnostic and could timating the path loss can lead to coverage holes. However, also
be deployed on legacy bands (below 3 GHz), but also in the overestimating the path loss is undesirable because this leads to
milimeter-wave (mmWave) spectrum. As a compromise to still severe inter-cell interference issues. For an accurate prediction,
have fairly good propagation properties, but also allow for existing models need to be evaluated for the environments in
wider carrier bandwidths, the 3.5 GHz band is of particular which they shall be used and adjusted if necessary. Additionally,
interest among the bands below 6 GHz (sub-6 GHz). Propagation penetration of signals from outdoor to indoor is depending
channels for mobile use in the 3.5 GHz band and above, have heavily on the building materials and is varying a lot, even within
mainly been studied in the context of WiMAX (Worldwide the same cell coverage area. While wooden and older houses –
interoperability of Microwave Access) and with less carrier Rec. ITU-R P.2109-0 [10] refers to ‘traditional buildings’ – tend
bandwidth compared to what 5G will be able to use (up to to have a small penetration loss, modern ‘thermally-efficient’
200 MHz carrier bandwidth for sub-6 GHz). It also needs to be buildings with infrared light reflecting low emissivity (low-e)
considered that an increase in bandwidth requires the same factor coated glass windows impose high losses.
for a power increase, to maintain the same coverage area. In general the path loss (in dB) can be expressed according to
For radio network planning, propagation models are used, the simplified model in [1, eq. (1.12)]:
which can be categorized into deterministic, stochastic, PL (d) = A0 + 10γ log10 (d/d0 ) + χσ d > d0 , (1)
empirical, and standardized models [1, Chapter 4]. While
where A0 represents the deterministic path loss component at
1 https://www.cept.org/ecc/topics/spectrum-for-wireless-broadband-5g the reference distance d0 in meters. The path loss exponent is
 TABLE I TABLE II
S ELECTED E MPIRICAL PATH L OSS M ODELS C OMPARISON OF PATH L OSS PARAMETERS FOR (1) AT 3.5 GH Z
Name Frequency Range Distance Range Terrain Location UE Height Distance γ σ
Hata-Okumura 150 − 1500 MHz 1 − 20 km [m] [m] [dB]
COST 231 Hata 1500 − 2000 MHz 1 − 20 km rural Cambridge, UK [12] 6 250-2000 2.7 10
COST 231 Walfish-Ikegami 800 − 2000 MHz 0.02 − 5 km rural Piemonte, Italy [11] 2 1000-10000 2.5 8.9
Erceg ≈ 2000 MHz 0.1 − 8 km suburban Cambridge, UK [12] 6 250-2000 2.13 11.1
SUI IEEE 802.16 1 − 4 GHz 0.1 − 8 km suburban Ghent, Belgium [14] 2.5 30-1500 4.9 7.7
ECC-33 3.4 − 3.8 GHz 1 − 10 km suburban Shanghai, China [15] 3 300-1800 3.6 9.5
WINNER II 2 − 6 GHz 0.05 − 5 km urban Cambridge, UK [12] 6 250-2000 2.3 11.7
ITU-R P.1411-9 0.3 − 100 GHz 0.055 − 1.2 km urban United Kingdom [13] 2.5 100-2000 4.3 7.5
3GPP 0.5 − 100 GHz 0.01 − 5∗ km
∗ 10 km for RMa LOS C. Contribution and Outline
denoted γ and d is the distance in meters between base station The related works listed above show that the path loss char-
(BS) and user equipment (UE). Finally, χσ is the stochastic acteristics (path loss exponent, standard deviation) vary a lot
shadow fading component in dB with a zero-mean log-normal depending on the environment. Because it is difficult to decide for
distribution and standard deviation σ. The above mentioned a specific path loss model and do the network planning accord-
empirical models can also be expressed according to (1) with ingly, we conducted extensive measurements in Switzerland in a
additional terms depending on, e.g., the operating frequency, rural, suburban, and urban environment. The measurement setup
BS and/or UE antenna height, etc. [11, eqs. (3)–(12)]. and environments are described in Section II. Contrary to most
prior works, a beamforming BS antenna was used and parallel
B. Related Work measurements on a live network in legacy frequency bands were
Before 5G, there was already an interest in the 3.5 GHz band conducted for comparison and validation, and more realistic UE
in the early 2000’s with the radio access technology WiMAX for antenna heights for mobile cellular applications were used. The
fixed wireless access. Therefore, several measurement results are obtained results are compared against the 3GPP, WINNER II,
available for the frequency range 3.4–3.8 GHz, e.g., [11], [12], and SUI path loss models, and described in Section III (ITU-R
[13], [14], [15]. The corresponding measurement campaigns ei- P.1411-9 was excluded due to the specificity to environments).
ther used a WiMAX system with a signal bandwidth of 3.5 MHz, We find that most models overestimate the path loss, and for
or a continuous wave (CW) signal. Therefore, no conclusions can every scenario, a different model predicts the path loss with the
be drawn confidently for a wide bandwidth such as 100 MHz that least error. Finally, the paper is concluded in Section IV.
can be used for 5G. The BS antenna height varied between 15–
II. M EASUREMENT S ETUP AND E NVIRONMENT
36 m, while the UE antenna height varied between 2–10 m above
ground. Some comparisons have been done with the models A. Measurement Method
Erceg, COST 231 WI, COST 231 Hata, SUI, and ECC-33, but The propagation path loss in the downlink is defined as the
none with 3GPP’s path loss models. All publications calculated radio frequency (RF) attenuation PL in dB of a transmitted signal
a path loss exponent and standard deviation according to the when it arrives at the receiver:
simplified path loss model from (1), which are summarized
in Table II (only the lowest UE height is considered in the PL = PT + GT (φ, θ) + GR − PR [dB], (2)
table). The variations in these parameters (e.g., γ for a suburban where PT is the BS transmitted power in dBm and GT (φ, θ)
terrain ranges from 2.13 to 4.9 and for urban terrain from 2.3 to the transmitter antenna gain in dB as a function of azimuth
4.3) indicate that there are many influences on the propagation angle φ and elevation angle θ. Because the UE antenna is
channel that may depend on the geographic region, construction omnidirectional, we can simplify the receiver antenna gain as
material, and vegetation which in parts has also been confirmed a constant GR in dB. Finally, PR is the local mean received
in [16]. While [12] suggests the ECC-33 model for urban and power in dBm.
the SUI (terrain B) for suburban environments, [14] found that For the analysis, all measurement parameters were geographi-
the Erceg (terrain C) fits best, but underpredicts the measured cally binned in a two-dimensional square 5 m grid, thus removing
path loss. For rural environments, [11] and [12] show that the fast fading effects according to the Lee sampling criterion [18],
best fitting models SUI (terrain B & C) and COST 231 Hata and to prevent a bias from temporal influences due to, e.g., stops
overestimate the measured path loss. Because the path loss model at red traffic lights. For each bin, the median value was computed.
with the lowest prediction error varies, it is difficult to set on a For the log-distance path loss plots, also the median value was
specific model for network planning purposes. calculated for each 5 m distance bin.
Regarding outdoor-to-indoor propagation, measurements
were presented in [17], along with few outdoor measurements. B. 5G Testbed
A difference of only 10 dB more attenuation was found for a For conducting measurements as close to 5G as possible, we
modern building compared to an old building. According to the employed a 5G testbed (similar to [19]). It consists of a BS
authors, this stems from different wall thicknesses and building unit, an active antenna system (AAS) connected via fiber to the
material. Therefore we conclude that the modern building was BS, and in our case two UEs (see Fig. 1a for a picture of the
not equipped with low-e coated windows. AAS and one UE). The center frequency for which test-licenses
 Azimuth Elevation
90 90
120 0 60 120 0 60
-10 -10
150 -20 30 150 -20 30
-30 -30
-40 -40
180 -50 0 180 -50 0

(a) 5G Testbed (b) Rural Deployment 210 330 210 330

240 300 240 300

270 270
Macro Antenna, 0.8/2.1 GHz Beamforming Antenna, 3.5 GHz
Fig. 2. Antenna pattern envelope for a macro antenna and the AAS.
and scanner antennas were mounted 16 cm over the vehicle’s
roof to limit its influence on the antenna characteristic, which
resulted in 2.1 m above ground. When manually pushing the UE
around, the antennas were at a height of 1.4 m above ground,
(c) Suburban Deployment (d) Urban Deployment corresponding to the approximate height at which an adult is
Fig. 1. (a) Shows the 5G testbed AAS on a mast and the UE in the field, (b) usually holding a smartphone.
the map of the rural deployment with a red star indicating the AAS location
with the 120◦ sector, and the LOS/NLOS classification. Similarly, (c) and (d) C. Comparison With 4G/LTE
show the map of the suburban and urban deployments, respectively.
For the analysis of the 5G cell, it is interesting to have a
were available, was 3.55 GHz. The time division duplex (TDD) comparable 4G (LTE) cell available. Thus, the test deployments
and orthogonal frequency division multiplexing (OFDM) signal were chosen to have not only an existing macro site antenna mast
occupies a bandwidth of 80 MHz. Further details can be found for installation of the 5G AAS, but also to have a live LTE signal
in [19, Tab. 1] and the references therein. Contrary to the beam on a legacy frequency band available.
grid described in [19], we were using a configuration with the 48 For the measurements, a mobile network scanner (Rohde &
different dual-polarized beams arranged in a grid of three rows Schwarz® TSMW) with a laptop running Nemo Outdoor has
with 16 beams each (visible in Fig. 2). The antenna gain and been used to log propagation channel metrics based on the
resulting coverage remains approximately the same as with the LTE downlink signal. The RF was locked to one frequency
configuration described in [19] with 27 dBi, and 120◦ in azimuth band (800 MHz or 2.1 GHz) and all RSRP values per cell were
and 30◦ in elevation, respectively. The two identical UEs have recorded. For the analysis, only the values of the cell(s) of interest
eight antenna ports connected to an eight element omnidirec- were used. The 2×2 MIMO antennas (Rohde & Schwarz®
tional cross-polarized antenna, each element with a gain of 6 dBi. TSMW-Z8) which were used, emulate omnidirectional smart-
As in [19], the Mobility Reference Signal Received Power phone antennas to give realistic results. Together with the cable
(MRSRP) for each of the 48 beams has been logged and losses, their gain is considered with -3 dBi each (note that smart-
used for the analysis. The Mobility Reference Signal (MRS) phone antennas are usually integrated and partially covered,
is a pilot signal that is used for subframe synchronization leading to negative antenna gains). On the LTE base station side,
and identification of downlink beams. It is transmitted every the transmit power is corrected with the feeder cable losses in
5 ms and cycles through all 48 beams over a period of 20 ms. order to have the transmit power PT with respect to the antenna
Although there are significant differences such as the subcarrier ports. Again, a directive antenna gain GT (φ, θ) is considered (see
bandwidth, it can be used similarly to the LTE Reference Signal Fig. 2), depending on the relative location of the testbed UE to the
Received Power (RSRP). For each period, the highest of the 48 LTE BS antenna. The total measurement uncertainty is 4.5 dB.
MRSRP values is used as PR . The cable loss between the UE’s The network scanner has been fitted inside the 5G testbed UE,
RF frontend and the UE’s antenna can be considered as part of the and the 2×2 MIMO antennas could also be mounted close to the
constant GR , which then becomes 4 dBi. On the transmitter side, 5G antennas of the UE, without shielding them.
we can directly use the configured power PT of the AAS, because
the radio chains and power amplifiers are directly connected D. Rural Area Environment
with the antenna elements. However, depending on where the The rural deployment was located close to a village (Meikirch)
UE is located with respect to the AAS, a directive antenna north-west of Bern, Switzerland. The selected live macro site
gain GT (φ, θ) is considered. Because this is a beamforming only had an LTE 800 MHz cell with a 10 MHz carrier directed
antenna with a grid of beams, an artificial envelope pattern can over flat open fields towards a village at 1–1.5 km distance. The
be calculated by cycling through all beams sequentially, then 5G AAS was installed right below the live macro antenna at a
virtually overlay them, and taking the maximum at each angle, height of 12.4 m above ground, with the same azimuth angle as
see Fig. 2. The total measurement uncertainty is 6.25 dB. the macro antenna. The vertical separation of the macro antenna
The UE has been moved around with a van, or where driving to the testbed AAS was 2.1 m, which resulted in a mean elevation
was not possible, pushed by hand. When using the van, the UE angle difference of 0.22◦ seen from the UE antenna over all
measurement sample locations. The maximum used transmit TABLE III
power for 3.5 GHz was 1560 WERP . E STIMATED PATH L OSS M ODEL PARAMETERS FOR (1)
The measurements took place in mid-July 2018 during a sunny Frequency Rural LOS/NLOS Suburban Urban
and clear week with an air temperature of around 22◦ C. γ χσ [dB] γ χσ [dB] γ χσ [dB]
3.55 GHz 2.3 / 3.1 5.1 / 9.4 2.9 6.9 4.8 7.1
E. Suburban Area Environment 2.1 GHz – – 3.7 7.0 – –
800 MHz 2.8 / 3.4 5.9 / 7.5 – – – –
The suburban deployment was done in Ittigen, just north-east
of Bern, Switzerland, situated on a slightly ascending slope. γ is estimated with the reference distance d0 at 100 m, and the
For comparison, the signal from an existing LTE macro site at standard deviation χσ is calculated according to the model (1).
2.1 GHz with a 20 MHz carrier was used. The testbed AAS has The estimated values are listed in Table III. A comparison with
been mounted on a smaller extra mast next to the LTE macro site the models SUI IEEE 802.16 (A, B, C), ECC-33, WINNER II
mast due to space limitations on the latter. The height of the AAS (C1, C2, D1), and 3GPP (RMa, UMa) was performed. In general,
was 24.5 m above ground. The horizontal separation between the ECC-33 model overestimates the path loss and is not well
the antennas was 3.8 m, while the vertical separation was 8.5 m. suited because it was derived by fitting curves for distances of 1 to
The mean angular separation of both antennas seen from the UE 10 km. For the few models that agree the most with the respective
over all measurement sample locations is 0.232◦ in azimuth and measurements, prediction error statistics have been computed.
0.819◦ in elevation and is regarded as negligible. The maximum The mean prediction error µe indicates an over- (positive value)
transmit power for 3.5 GHz had to be limited due to stringent non- or under-prediction (negative value) with standard deviation σe ,
ionizing radiation (NIR) regulations in Switzerland. Without and the root mean square of the prediction error RMSE represents
lowering the transmit power on the live macro cells, the allowed the general metric for comparison of how well the models fit
maximum transmit power for 3.5 GHz was 384 WERP . the measurements. These were calculated in log-domain [dB].
The measurements took place end of April and beginning of
May 2018, with sunny and also overcast days, but no rain. The A. Rural Environment
air temperature was around 15–20◦ C.
As expected, this environment provides a good share of line-
F. Urban Area Environment of-sight (LOS) opportunities, in fact 42 % of the measurement
For an urban deployment, a suitable and existing live macro samples are estimated to be in LOS condition. The large share of
site was selected in the city of Zurich, Switzerland. The AAS LOS makes it necessary to analyze LOS and NLOS separately.
has been mounted just below a macro sector antenna on the The classification for LOS/NLOS has been done by manually
same mast at a height of 29.4 m above ground, overlooking a defining five coordinate polygons. All measurement samples
flat part of the city. Unfortunately, due to technical issues, no within these five polygons were categorized as LOS, the others
measurements of an LTE legacy band could be taken for this as NLOS. The measured path loss for 3.5 GHz is shown in
scenario. Also here, the NIR regulations required that we had Fig. 3 (LOS marked with circles) with predictions from the
to lower the maximum transmit power. By slightly reducing the models 3GPP Rural Macro (RMa) NLOS and the WINNER II
power on a few live macro cells on this site, we were allowed to D1 (rural macro) NLOS. The free space path loss (FSPL) is also
use the same maximum 384 WERP for the 3.5 GHz 5G cell as shown as a reference. Variations in the path loss from 600 m
in the suburban deployment. to 2 km (see shaded area) show the changes between LOS and
The measurements took place middle of August 2018 during NLOS environments. All error statistics are listed in Table IV.
a sunny week with an air temperature of around 25◦ C. Both models overestimate the path loss by 14.9 dB and 9.8 dB,
respectively, and the WINNER II D1 rural NLOS model fits
G. Outdoor-to-Indoor Scenario best, although the measurements are often well outside of its
Additionally to outdoor drive and walk tests, indoor mea- 8 dB standard deviation. The shadow fading distribution is
surements were collected at four buildings of the suburban shown in Fig. 7 in the top two distributions for LOS and NLOS.
deployment and two buildings of the urban deployment. The goal
160
was to assess the penetration loss into buildings for frequencies Measurement 3.55 GHz Mix of distinct
above 3 GHz. Measurements were started inside, just behind a LS Model 3.55 GHz LOS and NLOS
140 3GPP RMa NLOS
window with line-of-sight to the 5G AAS. The UE was then WINNER II D1 NLOS
ditto, = 8 dB
slowly and steadily moved further inside the building, either
Path Loss [dB]

FSPL @ 3.55 GHz

120 Measurement 800 MHz
until the connection dropped or the back side of the building was FSPL @ 800 MHz
reached. Outdoor reference measurements were either taken just
100
in front of the building or on a terrace on top of the building.
III. M EASUREMENT R ESULTS 80

For all environments described in the previous section, a

60
total of 267 data sets were analyzed. Each site deployment was 10 1 10 2 10 3 10 4
analyzed separately, and the path loss was obtained according to Distance from Base Station Antenna to Mobile Station Antenna [m]
(2). Using a least squares (LS) regression, the path loss exponent Fig. 3. Rural environment measurements compared to models.
 160 160
Measurement 3.55 GHz
LS Model 3.55 GHz
140 3GPP UMa NLOS
Path Loss [dB] for 3.5 GHz

140
ditto, = 6 dB
SUI IEEE 802.16 A

Path Loss [dB]

FSPL @ 3.55 GHz
120 120

100 100

Rural Measurements 3.5 GHz/800 MHz

80 Linear fit, = 15 dB, = 8.3 dB 80
Suburban Measurements 3.5 GHz/2.1 GHz
Linear fit, = 5.8 dB, = 8.1 dB
Courtyard LOS Street
60 60
50 60 70 80 90 100 110 120 130 140 150 10 1 10 2 10 3
Path Loss [dB] for 800 MHz and 2.1 GHz Distance from Base Station Antenna to Mobile Station Antenna [m]
Fig. 4. Path loss for 3.5 GHz vs. 800 MHz and 2.1 GHz at each measurement Fig. 6. Urban environment measurements compared to models.
location; validates the theory of 12.8 dB and 4.4 dB path loss difference.
160 Rural LOS
= 5.1

140
Rural NLOS

Probability Density
= 9.4
Path Loss [dB]

120

Measurement 3.55 GHz Suburban

LS Model 3.55 GHz = 6.9
100 3GPP RMa NLOS
SUI IEEE 802.16 C
ditto, = 10 dB
80 FSPL @ 3.55 GHz Urban
Measurement 2.1 GHz = 7.1
FSPL @ 2.1 GHz
60
10 2 10 3 -25 -20 -15 -10 -5 0 5 10 15 20 25
Distance from Base Station Antenna to Mobile Station Antenna [m] Shadow Fading Distribution [dB]
Fig. 5. Suburban environment measurements compared to models. Fig. 7. Distribution of the measured shadow fading components. The solid
curve represents the Gaussian fit for each distribution and is an indication if
the shadow fading has a log-normal distribution.
For comparison, the measured path loss for the 800 MHz
LTE signal is also shown in Fig. 3, with the corresponding 3.83 dB, respectively. These results are also listed in Table IV.
FSPL as a reference. Furthermore, for each 5 m measurement The shadow fading with a close to normal distribution is shown
grid location, both path losses are plotted in Fig. 4. Assuming in Fig. 7 on the third row.
only a frequency dependent offset, a line with a slope The available measurement data on 2.1 GHz also allows us to
of 10 dB/10 dB can be fitted. This offset in theory is do comparisons. The measured path loss at 2.1 GHz is plotted
20 log10 (3.5 GHz/0.8 GHz) = 12.8 dB, but measurements in Fig. 5 and a comparison with 3.5 GHz is shown in Fig. 4.
show 15 dB with a standard deviation of 8.3 dB. The 2.2 dB Fitting a slope of 10 dB/10 dB results in an offset of 5.8 dB
residual is within the measurement uncertainty. with a standard deviation of 8.1 dB. In theory, this frequency
B. Suburban Environment dependent offset is 20 log10 (3.5 GHz/2.1 GHz) = 4.4 dB. The
1.4 dB residual is within the measurement uncertainty.
As explained in Section II-E, parallel measurements at
2.1 GHz have been conducted for this campaign. During valida-
C. Urban Environment
tion of the data, two anomalies were corrected: the first stemmed
from a loss of connectivity because of a large apartment complex In our measurement data from the urban deployment, we have
affecting signal propagation. The affected street segment was 27 % of the samples in LOS conditions. This may seem a lot, but
excluded from the analysis; the second stemmed from the local considering a dense deployment with inter-site distances of a
geography, resulting in a much higher than expected signal few hundred meters and only considering outdoor coverage, this
strength. Fitting the data to a two-ray model [20] using the becomes realistic. The main LOS contributions come from one
delay spread information from the network scanner allowed us long street at the azimuth angle of the antenna, between 110-
to explain this anomaly. The corrected overall path loss versus 210 m (see the gray-shaded area in Fig. 6). Another area with
distance is shown in Fig. 5. The environment does not offer many close to LOS path loss has been identified as the courtyard of the
LOS opportunities. We could therefore categorize LOS areas by city block on which the antenna was mounted (south-west corner,
manually defining coordinate polygons. As a result, 14 % of all overlooking the city block towards north-east). Regarding the
samples fall in such a polygon. comparison with empirical models, the two that fit the best, are
Comparing empirical models with the measurements at the 3GPP Urban Macro (UMa) NLOS (RMSE = 6.84 dB) and the
3.5 GHz shows the best agreement with the SUI Terrain C model SUI Terrain A (RMSE = 9.78 dB). The 3GPP UMa model slightly
(RMSE = 4.87 dB), and the 3GPP RMa NLOS (RMSE = 5.29 dB) overestimates the path loss by 2.1 dB, and the SUI Terrain A
model. Both models overestimate the path loss by 2.11 dB and model underestimates the path loss by -8.3 dB. Also these results
 1 scenario, and 14.9 dB for the rural scenario. Furthermore, the
Bldg 1
Bldg 2 excessive attenuation of around 10 dB per coating layer of low-e
0.8 Bldg 3, low-e glass
Bldg 4, low-e glass non-coated windows has an impact on outdoor-to-indoor penetration as the
Bldg 5, ground floor
Bldg 5, 13th floor
1 coated number of newly built and renovated houses grows. Finally, path
0.6
layer
ECDF

Bldg 6, 6th floor loss exponents are slightly lower with a beamforming antenna at
Bldg 6, 12th floor
2 or more
0.4 coated layers
3.5 GHz, compared to a conventional macro antenna at 2.1 GHz
and 800 MHz.
0.2
R EFERENCES
0 [1] B. Clerckx and C. Oestges, MIMO Wireless Networks: Channels,
-60 -50 -40 -30 -20 -10 0 10 Techniques and Standards for Multi-Antenna, Multi-User and Multi-Cell
RSRP Difference to Oudoor [dB] Systems. Academic Press, Jan. 2013.
Fig. 8. Outdoor-to-indoor penetration loss shown as empirical CDF of the [2] M. Hata, “Empirical formula for propagation loss in land mobile radio
received power difference from indoor compared to outdoor. services,” IEEE Transactions on Vehicular Technology, vol. 29, no. 3,
pp. 317–325, Aug. 1980.
TABLE IV [3] “COST Action 231: Digital mobile radio towards future generation
PATH L OSS M ODEL P REDICTION E RROR S TATISTICS systems: Final Report.” European Commission, Brussels, Belgium, Tech.
Environment Rural Suburban Urban Rep. EUR 18957, Apr. 1999.
Model RMa WINNER2 RMa SUI C UMa SUI A [4] V. Erceg, L. J. Greenstein et al., “An empirically based path loss
model for wireless channels in suburban environments,” IEEE Journal
µe [dB] 14.9 9.82 3.83 2.11 2.07 -8.3 on Selected Areas in Communications, vol. 17, no. 7, pp. 1205–1211,
σe [dB] 10.8 7.53 3.66 4.4 6.54 5.18 Jul. 1999.
RMSE 18.4 12.4 5.29 4.87 6.84 9.78 [5] IEEE, “Channel Models for Fixed Wireless Applications,” IEEE, Tech.
Rep. IEEE 802.16.3c-01/29r4, Jul. 2001.
are listed in Table IV. The shadow fading distribution is shown [6] “The analysis of the coexistence of FWA cells in the 3.4 - 3.8
GHz band,” Electronic Communication Committee (ECC) within the
at the bottom of Fig. 7 and conforms to the normal distribution. European Conference of Postal and Telecommunications Administration
(CEPT), Tech. Rep. ECC Report 33, May 2003.
D. Outdoor-to-Indoor Scenario [7] Pekka Kyösti, Juha Meinilä et al., “WINNER II Channel Models,” Eu-
ropean Commission, Tech. Rep. D1.1.2 V1.2, IST-4-027756 WINNER
The indoor RSRP measurements are normalized with the II Deliverable, Feb. 2008.
corresponding reference outdoor RSRP and the empirical cu- [8] “Propagation data and prediction methods for the planning of short-range
outdoor radiocommunication systems and radio local area networks
mulative distribution function (CDF) is plotted, see Fig. 8. The in the frequency range 300 MHz to 100 GHz,” ITU-R, Tech. Rep.
buildings 1–4 are from the suburban deployment, the buildings Recommendation P.1411-9, Jun. 2017.
5 & 6 from the urban deployment. The measurements clearly [9] 3GPP, “Study on channel model for frequencies from 0.5 to 100 GHz,”
3rd Generation Partnership Project (3GPP), Technical Report (TR)
show which buildings are new or with coated low-e windows, 38.901, 06 2018, version 15.0.0.
and which ones use regular windows: building 1 is an older office [10] “Prediction of Building Entry Loss,” ITU-R, Tech. Rep. Recommenda-
building with regular double glazing windows. Building 2 is a tion P.2109, Jun. 2017.
[11] P. Imperatore, E. Salvadori, and I. Chlamtac, “Path Loss Measurements
home for elderly people, also with uncoated double glazing win- at 3.5 GHz: A Trial Test WiMAX Based in Rural Environment,” in 2007
dows. Building 3 is a new apartment building, with triple glazing 3rd International Conference on Testbeds and Research Infrastructure
low-e windows, causing around 30 dB of additional loss. Build- for the Development of Networks and Communities, May 2007, pp. 1–8.
[12] V. S. Abhayawardhana, I. J. Wassell et al., “Comparison of empirical
ing 4 is a newer office building with triple glazing windows and propagation path loss models for fixed wireless access systems,” in 2005
one coated layer. Building 5 is a renovated 14 floor office build- IEEE 61st Vehicular Technology Conference, vol. 1, May 2005, pp. 73–
ing. The measurements from the ground floor behind large single 77 Vol. 1.
[13] M. C. Walden and F. J. Rowsell, “Urban propagation measurements and
layer shop-windows show the lower attenuation than from the statistical path loss model at 3.5 GHz,” in 2005 IEEE Antennas and
13th floor behind multiple glazing low-e windows. Building 6 is a Propagation Society International Symposium, vol. 1A, Jul. 2005, pp.
12 floor renovated office building with low-e windows. The mea- 363–366 Vol. 1A.
[14] W. Joseph, L. Roelens, and L. Martens, “Path Loss Model for Wireless
surements from the 6th and 12th floor show about the same atten- Applications at 3500 MHz,” in 2006 IEEE Antennas and Propagation
uation, only walls inside the building cause different variations Society International Symposium, Jul. 2006, pp. 4751–4754.
in the path loss. In general, we can say that newer or renovated [15] S. Kun, W. Ping, and L. Yingze, “Path loss models for suburban scenario
at 2.3GHz, 2.6GHz and 3.5GHz,” in 2008 8th International Symposium
buildings with coated low-e windows add 10–30 dB of additional on Antennas, Propagation and EM Theory, Nov. 2008, pp. 438–441.
penetration loss compared with buildings with regular windows. [16] M. Riback, J. Medbo et al., “Carrier Frequency Effects on Path Loss,”
in 2006 IEEE 63rd Vehicular Technology Conference, vol. 6, May 2006,
IV. C ONCLUSIONS pp. 2717–2721.
[17] I. Rodriguez, H. C. Nguyen et al., “Path loss validation for urban micro
Measurements from extensive rural, suburban, and urban cell scenarios at 3.5 GHz compared to 1.9 GHz,” in 2013 IEEE Global
measurement campaigns with a 5G testbed operating in the Communications Conference (GLOBECOM), Dec. 2013, pp. 3942–3947.
[18] W. Lee and Y. Yeh, “On the Estimation of the Second-Order Statistics of
3.5 GHz band have been analyzed and compared with predictions Log Normal Fading in Mobile Radio Environment,” IEEE Transactions
from empirical path loss models (3GPP, WINNER II, SUI). on Communications, vol. 22, no. 6, pp. 869–873, Jun. 1974.
Almost all models tend to overestimate the path loss. Even [19] B. Halvarsson, A. Simonsson et al., “5G NR Testbed 3.5 GHz Coverage
Results,” in 2018 IEEE 87th Vehicular Technology Conference (VTC
though no model gives the least error in all environments, the Spring), Jun. 2018, pp. 1–5.
3GPP group of models always ranked among the top two, with an [20] A. Goldsmith, Wireless Communications. Cambridge: Cambridge
over prediction of 2.1 dB and 3.8 dB for the urban and suburban University Press, 2005.