Small-Cell LTE Networks: Evaluating RF Planning Methods

Inês Gonçalves
[email protected]

Instituto Superior Técnico, Lisboa, Portugal

November 2018

Abstract
This work focuses on the study and evaluation of different propagation models and geographical
databases used in the planning of LTE networks. A key aspect of the study includes the comparison
of the distinct types of databases used: 2D, 2.5D and 3D. Two separate studies were performed:
a coverage by area and a point-to-point. The coverage by area analysis used a buffer area, which
recreated the path followed in the walk test, to simulate the RSRP and RSRQ levels obtained with each
model. The point-to-point study used the location of the walk test samples to obtain the equivalent
predicted values. The models simulated are Okumura-Hata, COST 231-Hata, SPM and WinProp UDP.
A detailed analysis of the simulations and their comparison to the results obtained during the walk
test led to several conclusions. Although the Okumura-Hata model obtained the lowest mean absolute
error, it was the WinProp UDP Outdoor model that overall showed the best performance regardless
of the analysis method. SPM 2D was the model that adapted better when the walk test samples had
a RSRP level higher than -85 dBm while COST 231-Hata obtained the best results in the opposite
scenario. A relevant conclusion stresses the importance of accurate drive tests and their validity, as
well as the clear improvement obtained in network planning when using dominant path models.
Keywords: LTE, Propagation Models, Walk Test, Clutter Classes, Ray-Tracing

1. Introduction bile connections but only 24% of the traffic. Accord-

ing to predictions from [1], by 2021, 4G is expected
Over the last decades, fuelled by our need to com- to correspond to 53% of all connections and 79% of
municate and the technological advances, there has the total traffic. As the capacity demand in mobile
been a dramatic change in the way we communicate networks increased, the development of new wire-
that in time altered consumers needs. less communication technologies that solved some of
In the beginning of digital cellular networks the the existing constraints related to radio spectral re-
technologies developed, like the Global System for sources became crucial. Small-cells were presented
Mobile Communications (GSM), were designed to as a promising solution to increase the capacity and
primarily focus on voice traffic. The introduction of coverage of the network by offloading the macro
3rd Generation (3G) networks, such as the Universal layer and increasing the Quality of Service (QoS).
Mobile Telecommunications System (UMTS), insti- One of the tools that contributes to an efficient
gated an increased use of data by enabling the use planning is the use of the correct propagation
of various services like the Global Positioning Sys- model for the area in question and the geograph-
tem (GPS), mobile television and video conferenc- ical database used alongside it. In this thesis differ-
ing. But in order to respond to those ever-changing ent propagation models, used in the planning and
needs, the Long Term Evolution (LTE), commonly deployment of small-cells in a typical urban sce-
known as 4th Generation (4G), was created. It was nario, will be analysed. The objective is to compare
the first cellular communication system designed real data obtained in a walk test with predictions
specifically to support packet-switched data services of different propagation models and geographical
since the beginning. It aims to provide higher data databases, in a LTE network using a radio network
rates, lower latency and better spectral efficiency. planning software. This thesis was developed in col-
Still, tendencies have shown an immeasurable laboration with Nokia Portugal. This partnership
growth of data traffic that clearly dominates in com- was essential, as it allowed for a realistic study using
parison to voice traffic. Although 4G connections data obtained in an actual project.
represented only 26% of mobile connections in 2016,
they already accounted for 69% of mobile data traf-
fic, while 3G connections represented 33% of all mo-

1
2. Fundamental Concepts dling all the radio related protocols. The eNBs are
This section provides an overview of the technolo- linked to each other through a X2 interface and con-
gies referred to in this thesis such as LTE, small- nected to the EPC through the S1 interface. The
cells and the different types of propagation models EPC is responsible for the overall control of the UE
and geographical databases. and establishment of the bearers. A bearer is an IP
packet flow with a defined QoS between the gate-
2.1. LTE way and the UE. The main logical nodes of the EPC
LTE is a wireless communication standard devel- are the PDN Gateway (P-GW), the Serving Gate-
oped by the 3rd Generation Partnership Project way (S-GW) and the Mobility Management Entity
(3GPP), aimed at enhancing the Universal Terres- (MME).
trial Radio Access Network (UTRAN) and optimiz- The multiple access in LTE is based on Orthogonal
ing the radio access architecture. Some of the driv- Frequency Division Multiple Access (OFDMA) for
ing forces for the development of LTE include the the DL and on Single Carrier Frequency Division
increased capability of existing wireline technolo- Multiple Access (SC-FDMA) for the UL, both with
gies, the need for additional wireless capacity, as Cyclic Prefix (CP). The transmission in OFDMA
well as the need for lower cost wireless data deliv- is based on the use of narrow, mutually orthogo-
ery and the existing competition with other wire- nal sub-carriers. In LTE, each sub-carrier is spaced
less technologies. LTE was designed to support 15 kHz regardless of the total transmission band-
only packet-switched services, in contrast to the width. At the sampling instant of a single sub-
circuit-switched data services of the previous gen- carrier, all the others have zero value. The or-
erations. Its objective is to provide continuous IP thogonality achieved by the sub-carrier allows for a
connectivity between the user equipment (UE) and more efficient use of the spectrum, because a band
the packet data network (PDN) without disruptions guard is no longer needed to avoid sub-carrier in-
during mobility. terference. One of the main challenges in OFDMA
The main requirements for this standard included is the high Peak-to-Average Ratio (PAR) of the
high spectral efficiency, high peak data rates, short transmitted signal. This not only limits the trans-
round trip time as well as flexibility in frequency mitted power and reduces coverage but forces the
and bandwidth. LTE should be able to provide use of better power amplifiers. The UL signal is
peak user throughput with a minimum of 100 generated by the users mobile terminal so in this
Mbit/s in the DL and 50 Mbit/s in the UL. LTE sys- scenario a multiple access technique that enables
tems contain two separate categories: the Evolved better power-amplifier efficiency is needed. This
Universal Terrestrial Radio Access Network (E- was solved by choosing the SC-FDMA in the UL.
UTRAN) and the Evolved Packet Core (EPC), as Similarly to OFDMA, SC-FDMA divides the trans-
represented in Figure 1. mission bandwidth into multiple parallel subcarri-
ers maintaining the orthogonality of the subcarri-
ers. However, in SC-FDMA the data symbols are
not directly assigned to each subcarrier indepen-
dently like in OFDMA. Instead, the signal which is
assigned to each subcarrier is a linear combination
of all modulated data symbols transmitted at the
same time instant. [2]
A high degree of spectrum flexibility is one of the
main characteristics of the LTE radio access. LTE
supports two types of duplexing: Time Division Du-
plexing (TDD) and Frequency Division Duplexing
(FDD) although FDD is the most adopted tech-
nique. In TDD, the DL and UL use the same
frequency but transmission takes place in differ-
ent, non-overlapping time slots. Contrarily, FDD
implies that DL and UL transmissions take place
in different, sufficiently separated, frequency bands
but simultaneously.
Figure 1: EPS Architecture (extracted from [3]). LTE uses 3 types of Quadrature Amplitude Modu-
lation (QAM). They are the Quadrature Phase Shift
Keying (QPSK) or 4-QAM, 16-QAM and 64-QAM
The E-UTRAN is a network composed by one
modulation schemes. The QPSK carries 2 bits per
type of node - evolved Node B (eNB). These eNBs
symbol while 16-QAM and 64-QAM carry 4 and
can cover one or more cells and are capable of han-

2
6 bits per modulation symbol, respectively. These ployed in specific areas in which extra capacity
modulation schemes are available in both UL and is known to be needed, for example near a train
DL directions in all devices, except for 64-QAM station or in a city centre. They are also often
in the UL which depends on the UE capability. temporarily deployed during special occasions like
The QPSK is the most robust to interference and sporting events. Adding microcells in the urban
bad channel conditions, but consequently offers the environment allows the operator to subdivide the
lower bit rate. In contrast, the 64-QAM allows the cells, leading to an optimization of the use of the
higher bit rates in LTE, but requires the best con- spectrum and ensuring a better capacity. Micro-
ditions in terms of Signal-to-Interference-plus-Noise cells have a range of a few hundred metres.
Ratio (SINR), by being the most liable to errors due Picocells are small base stations that work very sim-
to interference. ilarly to an Access Point (AP). Picocells have a low
power radiating antenna that connects to the opera-
2.2. Small-Cells
tor’s core network. Picocells can be used in business
The unmeasured usage growth of cellphones, tablets environments, shopping centres and airports, for ex-
and data-hungry applications has caused an ex- ample. They are also useful in high-rise buildings,
ponential growth of traffic in mobile networks, as where the macro cell signal strength decreases with
stated before. Operators have met this challenge height. The main advantage is that they are con-
by increasing capacity with new radio spectrum, siderably cheaper and since they are placed indoor,
adding multi-antenna techniques and implement- they effectively increase indoor radio coverage and
ing more efficient modulation and coding schemes. capacity. These small-cells have a range of approx-
However, these measures alone are insufficient in imately 100 metres.
extremely crowded environments and at cell edges Femtocells were proposed to extend the concept of
where performance can significantly degrade. Com- picocells to home networks, in particular. They
plementing macro networks with small-cells is an ef- have limited output power, between 10 and 20 dBm,
fective way to extend coverage and increase capac- and capacity as they serve 4 to 6 users. These type
ity indoors and outdoors, in public spaces, offices or of small-cells can be deployed in a variety of sce-
residences. narios. A typical deployment is in a rural envi-
Small-cells are operator-controlled, low-powered ra- ronment, where a femtocell enables access to mo-
dio access nodes, used to complement mobile ser- bile network for users which are typically under a
vices that is served by macro cells. They are primar- coverage hole but with a reliable Digital Subscriber
ily deployed to increase capacity in hot spots with Line (DSL) connection. As it happens homes in ru-
high user demand and to fill in areas not covered ral areas usually lack the macro cell coverage, but
by the macro network both outdoors and indoors. some already have access to high data rate Inter-
Small-cells also improve network performance and net, making it an ideal scenario to deploy femto-
service quality by offloading from the large macro cells. Another use case case for femtocells is an
cells. A combination of macro cells and small-cells enterprise environment, where the low capacity of
is usually called as a Heterogeneous Network that femtocells is typically enough to cover a number of
provides increased bitrates per unit area. There are offices. The major advantage of femtocells is the
different types of small-cells which can be grouped fact they are directly installed by the end users in-
according to cell size: femtocells, picocells and mi- side their homes and have a maximum range of ap-
crocells, as represented in Figure 2. proximately 30 metres. This will ensure good cover-
age for subscribers. For operators, femtocells are a
cheap solution because they are paid by customers.

2.3. Propagation Models and Databases

The overall objective of mobile operators is to sat-
isfy users needs with adequate coverage and QoS.
The performance of wireless communication sys-
tems depends in a fundamental way on the mobile
radio channel. Consequently, predicting the propa-
gation characteristics between the transmitters and
Figure 2: Cell Sizes of Different Technologies (ex- receivers is crucial for the design and installation of
tracted from [6]). a radio network communications system. All com-
putations and simulation procedures depend on re-
Operators began to install smaller outdoor base liable propagation models. A propagation model
stations called microcells as new data services that is a formulation that intends to characterise and
required higher coverage, emerged. These are de- predict radio wave propagation. The basis for any

3
propagation model is a database, which describes used in construction) to simulate the shadowing ef-
the propagation environment. Several propagation fect that a signal experiences to provide a correlated
models have been developed and proposed for cel- spatial variation of the path loss.
lular systems operating in different environments Dominant Path Models (DPM) are another type of
(outdoor, urban, suburban, rural, and indoor) but propagation models widely used in radio network
there are generally three different types of propaga- planning. A DPM does not rely only on the direct
tion models: empirical models, deterministic mod- ray (like empirical models) and it does not consider
els and semi-empirical models. hundreds of rays for a single radio link (like ray
An empirical model is based on observations and tracing), as pictured in Figure 3.
measurements. These are useful to study the gen-
eral behaviour or to enable a rough estimation of the
number of required cells in a large area. The classi-
fication of empirical models can be further divided
into time dispersive and non-time dispersive. Time
dispersive provides information about time disper-
sive characteristics of the channel, like the multi-
path delay spread of the channel. Non-time disper-
sive models consider various parameters, such as
distance, antenna heights, frequency and transmit- Figure 3: Empirical models (left), Ray Tracing
ter power to predict average path loss. (middle) and DPM (right) (extracted from [5]).
The Okumura Model and the Okumura-Hata Model
Instead it focuses on the most dominant path be-
were some of the first empirical models developed
tween transmitter and receiver. The Urban Dom-
that served as groundwork to many others. The
inant Path (UDP) model is one of the dominant
COST-231 Hata Model, based on the previous ones,
path models that was developed for the prediction
is another commonly used empirical model.
of the field strength specifically in urban environ-
The Standard Propagation Model (SPM), used in
ment. The algorithm of DPMs can be resumed in
radio network planning software, is a propagation
two different stages. First, the dominant paths have
model based on formulas of Okumura and Hata that
to be determined and then the prediction of the
was adapted to perform signal coverage predictions
path loss for each path must be calculated. The
in the frequency range of 150 MHz to 3500 MHz and
dominant path must lead via convex corners from
for distances between 1 km to 20 km. The SPM is
the transmitter to the receiver. For the determina-
best applied on mobile technologies such as LTE
tion of the path, a tree with all convex corners is
and LTE-A.
computed starting with the corners visible from the
Semi-empirical are based on a combination of mea-
transmitter.
surements and theory and attempt to interpret field
As mentioned above, propagation models per-
data based on theoretical principles. The first site-
form their predictions using information from ter-
specific propagation models were semi-empirical
rain databases. The environment to be analysed
and were based on detailed terrain characteristics
and predicted is stored in a database file. But de-
extracted along the individual propagation paths
pending on the type of scenario (indoor, urban or
between transmitter and receiver. These models use
rural), the resources available and the overall level
low resolution geographic data but still manage to
of accuracy intended for each project a different de-
obtain reasonably fair results for coverage predic-
scription of the relevant obstacles is necessary.
tions in urban areas. They are, however inefficient
The most rudimentary type of database used nowa-
when it comes to indoor coverage estimates in dense
days is a 2D database. It is composed by a height
urban areas. The COST 231 Walfisch-Ikegami is
file or Digital Terrain Model (DTM) and a clut-
considered a semi-empirical model used to predict
ter file. The DTM consists in a digital model that
urban environments. This model is an extension of
represents the surface of the Earth, excluding veg-
COST-231 Hata model and it can be used for fre-
etation as well as buildings and man-made struc-
quencies above 2000 MHz.
tures. It contains the altitude values of the region
Deterministic models are numerical methods that
and as such is characterised by a latitude Y, a lon-
simulate radio waves propagation by reproducing
gitude X and a specific resolution. The resolution
the physical propagation phenomena. These mod-
is linked to the retrieval of the topographic data ac-
els use a specific location for the transmitter and
cording to a given scale. The clutter file used in a
the receiver to provide a reliable and thorough es-
2D database describes the land use classification of
timation of the path losses and the channel charac-
the area in analysis including any features that im-
teristics. They rely on terrain descriptive features
pact the radio wave propagation. The clutter types
(altitude, the geometry of the buildings, materials
or so-called clutter classes include natural environ-

4
ments: river, forest, ocean or man-made: industrial, given by:
suburban, dense urban, to name a few. The appro-
priate number of clutter classes depends on the geo- P L = 69.55 + 26.16log10 (f ) − 13.82log10 (hT )+
graphic area. The combination of the DTM and the [44.9 − 6.55log10 (hT )]log10 (d) − H(hR , f )
clutter types result in a 2D database because the al- (1)
titude only refers to the height of the terrain and
no clutter height information is usually provided. where:
The propagation models that use 2D databases are f: frequency in MHz;
empirical ones. In empirical models the path loss is HT : transmitter height in metres;
calculated for the direct path between the transmit- d: distance in kilometres;
ter and receiver and the topography of the terrain H(hR ,f): correction factor for the mobile antenna
is only used to calculate obstructions. According to height.
the given clutter category of the area, an attenua- The correction factor is computed differently ac-
tion or gain is added to the total path loss. cording to the size of the area in question. In this
Another type of database used is a 2.5D database. case:
While it also consists of a DTM and clutter cate- 
8.29(log10 1.54hR )2 − 1.1,

gories it includes the average height of each scenario 

f ≤ 200M Hz
by introducing the clutter heights. Each specific H(hR , f ) = (2)
 3.2(log10 11.75hR )2 − 4.97,
category will have a preassigned average height, for 

f ≥ 400M Hz
example a low density urban area will have an av-
erage height of 20 metres and an area classified as a where:
village might have an average height of 6 metres. If f from 150 MHz to 1500 MHz;
a higher accuracy is necessary, subcategories with a hT from 30 m to 200 m;
specific height assigned to it can be created within
hM from 1 m to 10 m;
a clutter class. This method provides with a bet-
d from 1 km to 20 km.
ter correspondence with the real environment. For
This model will only be applied with a 2D
example, a scenario labelled as dense urban can be
database as it is empirical.
divided in two, one with an average height of 15
metres and the other with 20 metres. The COST 231 Hata Model will also be used.
This model extends the previous one to cover the
The most comprehensive and detailed type of frequency range of 1500 MHz to 2000 MHz. The
database is a 3D database. Like previous databases other validity conditions are maintained.
it uses a DTM and clutter types (it may or may The path loss is given by:
not include the clutter heights) but incorporates the
height and contour of the buildings (3D Building P L = 46.3 + 33.9log10 (f ) − 13.82log10 (hT )
Vectors). When using a 3D database, each of the +[44.9 − 6.55log10 (hT )]log10 (d) − H(hR , f ) + Cm
buildings will have its own height and shape repre- (3)
sented as close to reality as possible, unlike previous
databases. where:

8.29(log10 1.54hR )2 − 1.1,
3. Methodology 

f ≤ 300M Hz

H(hR , f ) = (4)
This sector presents the methodology used in this 
 3.2(log10 11.75hR )2 − 4.97,
f ≥ 300M Hz

work.

3.1. Propagation Models Used 

0dB, medium-sized city or suburban areas
Cm =
The existing LTE network, operating at 1800 MHz, 3dB, metropolitan areas
was configured in the 9955 RNP tool. All the nec- (5)
essary information regarding each site, transmitter Similarly, to the Okumura-Hata model, the pre-
and cell was included to obtain accurate predic- dictions using the COST 231 Hata Model will only
tions. use a 2D database given its empirical nature.
The first propagation model tested is Okumura- The SPM will also be tested. The model is based
Hata. This model established empirical mathemat- on Equation 6 and each of the K parameters can
ical relationships which described the graphical in- be adjusted according to specific conditions such
formation of the Okumura model. The equation as terrain profile, diffraction mechanisms, morphol-
describing the average path loss in an urban area is ogy of clutter classes and the effective height of the

5
transmitting and receiving antennas. [4] A buffer area was used in the coverage by area
analysis. This buffer area was created by analysing
PR = PT [K1 + K2 log10 (d) + K3 log10 (hT )+ and connecting each position of the walk test to
K4 .Dif f ractionLoss + K5 log10 (d).log10 (hT )+ create a path. After the path was created, it was
K6 hR + K7 log10 (hR ) + KCLU T T ER .fCLU T T ER transformed in an area using QGIS.
By using the buffer most of the area analysed is
+KHILL ] categorized as Open Area which corresponds to the
(6) actual streets. Considering 9955 adds a specific loss
value according to the clutter class, if this correc-
where:
tion was not performed, the values of the predic-
PR : received power in dBm; tions would be over-estimated because they would
PT : eirp in dBm; include the loss value associated to each building or
hT : effective height of the transmitter in metres; other obstacles.
hR : effective height of the receiver antenna in It is relevant to add that any errors found while
metres; analysing the walk test points were also corrected.
d: distance between the receiver and the trans- This means any walk test location that was shown
mitter in metres; to be in an indoor situation was placed in the near-
Diffraction Loss: losses due to diffraction over an est possible outdoor position, since it is known the
obstructed path in dB; walk test only took place outdoors.
fCLUTTER : weighted average losses due to the The data obtained in the walk test was also treated
clutter. to eliminate any inconsistencies. If at a given point,
The specific recommended adjustment coeffi- with the same coordinates, different signal values
cients will not be presented as they are classified by were recorded by the same transmitter an average
Nokia as internal use only. The SPM will be used value was then used.
with the 2D and 2.5D database and the formula The measurements recorded during the walk test in-
will be adapted for each one according to Nokia’s clude the Reference Signal Received Power (RSRP)
internal guidelines. For the coverage prediction us- and Reference Signal Received Quality (RSRQ).
ing a 2D database, 9955 computes the total path The RSRP is the average of power levels received
loss using the topographical information of the ter- across all reference signal symbols within the con-
rain provided in the DTM files and the clutter class sidered measurement frequency bandwidth. The
of the area. Each clutter class will have assigned RSRQ is the ratio of RSRP and the E-UTRA Car-
a specific loss value that is then added to the total rier Received Signal Strength Indicator (RSSI). The
path loss equation. The loss values used correspond E-UTRA carrier RSSI comprises the total received
to Nokia internal guidelines. wideband power observed by the UE. The signal
To compute the predictions with the 2.5D database, values recorded during the walk test correspond to
9955 will use the path loss formula and added losses the signal level from the best server.
according to the clutter class like in the previous To sum up the propagation models simulated with
case but it will also consider an average height for the corresponding databases are listed in Table 1.
each clutter class, on top of the terrain topography.
The specific height given to each clutter class was
also defined after analysing Nokias internal guide- Table 1: Propagation Models and Databases Tested
lines.
Model 2D 2.5D 3D
The main difference between a 2.5D and a 3D
Okumura - Hata X
database consists in the building vector file. The
COST 231 - Hata X
vector file describes each building height, shape and
SPM X X
orientation in detail. Another propagation model
WinProp UDP Outdoor X
tested was the WinProp UDP. This model is known
WinProp UDP Indoor X
to be a trade-off between computation complexity
and accuracy. The 3D building database was pre-
processed prior to the prediction computation, us-
ing Wallman a module that is part of the WinProp 4. Results
wireless network planning software package. This section presents the results of the different
analysis performed in this work.
3.2. Data Adjustments
Some corrections had to be implemented to reduce 4.1. Coverage Results
any inconsistencies in the data obtained in the walk In an initial stage a coverage analysis by area was
test and to make it comparable to the results of the performed to compare the Cumulative Distribution
predictions obtained with 9955. Functions (CDF) of the several models with the

6
 RSRP Coverage
120%
walk test resuls.
100%
In Figure 4 the RSRP results of the walk test are
80%
presented. The scale goes from RSRP levels higher

CDF [%]
60%
or equal to -55 dBm to RSRP levels higher or equal
40%
to -115 dBm, with a step of -5 dBm.
20%

0%

-102.95

-106.05

-109.15

-112.25

-115.35

-118.45
-89

-120
-62.65

-65.75

-68.85

-71.95

-75.05

-78.15

-81.25

-84.35

-87.45

-90.55

-93.65

-96.75

-99.85
-61.1

-64.2

-67.3

-70.4

-73.5

-76.6

-79.7

-82.8

-85.9

-92.1

-95.2

-98.3

-101.4

-104.5

-107.6

-110.7

-113.8

-116.9
RSRP Level [dBm]

Walk Test Okumura-Hata COST 231-Hata SPM 2D

SPM 2.5D WinProp UDP Outdoor WinProp UDP Indoor

Figure 6: Comparison of RSRP Coverage Predic-

tions

tained values more aligned with the measurements

retrieved during the walk test. Both WinProp UDP
models predicted -65.5 dBm as the highest RSRP
Figure 4: RSRP Results Obtained in the Walk Test
level in the buffer area.
The numerical results of the RSRP levels ob- The results can also be interpreted by separat-
tained in the walk test are presented in Figure 5. ing the results by power level. In the lower range
of signal power the WinProp UDP Indoor is the
Walk Test model that reaches an RSRP Level closest to the
120 one obtained in the walk test. In this case for a
100
CDF value of 95 % UDP Indoor obtained -104.25
80
CDF [%]

60
dBm while the walk test measurements resulted in
40 -106.57 dBm.
20 As the signal power increases, the WinProp UDP
0
Outdoor and the COST 231-Hata model also tend
-66.65
-70.4

-76.65
-78.35
-79.9
-81.5125
-82.8
-84.4
-85.9
-87.3
-88.8
-90.2
-91.95
-93.65
-95.17
-96.95
-98.45
-73
-75

-100.1
-101.57
-103.23
-104.57
-106.2
-107.95
-110.05
-118.9

to approximate more to the walk test results. When

RSRP Level [dBm] the walk test obtained -98.85 dBm for a CDF value
Walk Test of 75 %, the result of the COST 231-Hata model
Figure 5: CDF of the Walk Test RSRP Results was -95.7 dBm while UDP achieved -95.25 dBm and
-96.3 dBm, for the outdoor and indoor mode respec-
The highest value of RSRP measured throughout tively. When the signal power increases and crosses
the route was -60.8 dBm. This value was found only the -85 dBm mark, the Okumura-Hata starts to ob-
once among all the samples and it corresponds to tain much better results than the COST 231-Hata
0.024 % of all the valid measurements. The graph model. For a CDF value of 25 % the walk test
shows that as the RSRP level decreases, more sam- obtained an RSRP level of -82.55 dBm while the
ples are included. According to this results, 50 % Okumura-Hata also attained -82.55 dBm and the
of the samples have a RSRP level greater than or COST 231-Hata result was-87.55 dBm In the higher
equal to -90.8 dBm. The lowest value of RSRP level range of signal power both WinProp UDP models
measured during the walk test was -118.9 dBm. closely align with the walk test results, as well as
The results obtained from the predictions with the SPM 2D until around -72 dBm.
the propagation models were all performed consid- The same analysis was performed for the RSRQ
ering the same buffer area. In Figure 6 a graph measurements. The results of the RSRQ levels ob-
depicting the RSRP coverage results of every prop- tained in the walk test were also added to 9955 and
agation model is presented. The SPM 2.5D appears are represented in Figure 7. The scale portrayed
to be highly optimistic with a CDF curve distant in the figure ranges from lower RSRQ values such
from the curve of the walk test results. The SPM as -22 dB to higher ones like -6 dB. Although the
2.5D simulated values as high as -58.8 dBm while RSRQ levels measured in the walk test did not reach
the highest values obtained during the walk test a value lower than -17.2 dB, the scale includes some
were -60.8 dBm. In comparison the highest value lower values to account for the results obtained in
calculated by SPM but with a 2D database was - the predictions of the different models.
63.8 dBm. The highest value obtained by the simu- The numerical results of the RSRQ levels obtained
lation of Okumura-Hata model was -67.7 dBm and in the walk test are presented in Figure 8. As seen in
with COST 231-Hata was -72.6 dBm. The simula- the graph, the results are presented using the CDF
tions with WinProp UDP model show that it ob- in the vertical axis and the RSRQ level measured in

7
 walk test clearly obtained results with a higher
RSRQ level. The walk test measurements of the
RSRQ range from -17.2 dB to -5.8 dB. The model
that obtains a highest value of RSRQ is the Win-
Prop UDP with -11.5 dB for both the indoor and
the outdoor mode. With the other 4 models the
highest value of RSRQ is around -13 dB.

4.2. Validation of Results

A statistical analysis of the accuracy of each model
studied is necessary. The following statistical per-
Figure 7: RSRQ Results Obtained in the Walk Test formance metrics were used: Mean Absolute Error
(MAE) and Standard Deviation (SD). These met-
dB in the horizontal axis. The graph shows that the rics were calculated in linear units and converted
back to dB.
Walk Test
N
120 1 X M
M AE = |S − SiP | (7)
100 N i=1 i
80 CDF [%]
60 v
40
u
u1 X N

20 SD = t (|S M − SiP | − µ)2 (8)

N i=1 i
0
-6.1
-6.55
-6.825
-7.05
-7.187
-7.34
-7.466
-7.65

-8.383
-8.675
-8.925
-7.9
-8.1

-9.2
-9.55
-9.89
-10.13
-10.46
-10.87
-11.15
-11.57
-12.1
-12.7
-13.3
-14.3
-17.2

where µ is the absolute error.

RSRQ Level [dB]
The MAE and SD of the RSRP point-to-point
results are presented in Table 2. According to the
Figure 8: CDF of the Walk Test RSRQ Results
results in the table the propagation model that ob-
tains the best approximation to the walk test is
highest value of RSRQ measured throughout the
Okumura-Hata model with a MAE of 13.533 dB.
route was -6 dB, which corresponds to 0.253 % of
all the values obtained. By observing the graph one
concludes that 50 % of the samples have a RSRQ
level greater than or equal to -8.95 dB. The lowest Table 2: Statistical Analysis of RSRP Results
value of RSRQ level measured during the walk test
was -17.2 dB. Each of these measurements corre- Model MAE (dB) SD (dB)
spond to the signal received from the best server at Okumura - Hata 13.533 17.333
that moment and exact location. COST 231 - Hata 14.190 21.188
In order to compare the RSRQ results obtained SPM 2D 17.953 21.507
by every model, a graph with the cumulative dis- SPM 2.5D 29.213 33.087
tribution function of the results is shown in Figure WinProp UDP Outdoor 15.621 24.086
9. WinProp UDP Indoor 26.466 38.950
RSRQ Coverage
120%
100% In Table 3, the statistical analysis of the RSRP
80%
CDF [%]

60% results is performed but with a variation. In this

40%
20%
case, the samples that had an absolute error higher
0% than or equal to 30 dB were excluded. This analy-
-5.55
-6.15
-6.75
-7.35
-7.95
-8.55
-9.15
-9.75
-10.35
-10.95
-11.55
-12.15
-12.75
-13.4
-14.7

-17.2
-17.8
-18.4
-19
-19.6
-20.2
-20.8
-21.4
-16

-22

sis attempts to reduce the effect of the outliers on

RSRQ Level [dB] the overall results. According to these results, the
Okumura-Hata COST 231-Hata SPM 2D model that obtained the lowest MAE is now Win-
SPM 2.5D WinProp UDP Indoor WinProp UDP Outdoor Prop UDP Outdoor with 12.427 dB. WinProp UDP
Walk Test Indoor also obtained a much lower MAE than in the
Figure 9: Comparison of RSRQ Coverage Predic- previous case.
tions The results were also divided according to power
level in order to better analyse the general be-
As opposed to the RSRP results, none of the haviour of each model. In Table 4 only walk test
models used for the RSRQ coverage prediction im- samples with a RSRP value lower than or equal to
mediately stands out as the best fit, seeing as the -85 dBm were considered. On the other hand, in

8
Table 3: Statistical Analysis of RSRP Results ex- Table 6: Statistical Analysis of RSRQ Results
cluding samples with an AE higher than 30 dB
Model MAE (dB) SD (dB)
Model MAE (dB) SD (dB) Okumura Hata 7.343 4.733
Okumura - Hata 13.486 17.149 COST 231 Hata 7.406 4.859
COST 231 - Hata 12.659 17.958 SPM 2D 7.573 5.052
SPM 2D 17.660 20.877 SPM 2.5D 8.127 5.700
SPM 2.5D 22.833 23.866 WinProp UDP Outdoor 6.319 3.634
WinProp UDP Outdoor 12.427 17.519 WinProp UDP Indoor 6.326 3.594
WinProp UDP Indoor 13.712 18.553

is much lower than in the RSRP case. The model

Table 4: Statistical Analysis of RSRP Results- that obtained the highest MAE was SPM 2.5D with
Walk Test Samples ≤ - 85 dBm 8.127 dB, which is a difference of only 1.808 dB.
Model MAE (dB) SD (dB) 5. Conclusions
Okumura - Hata 13.800 16.584 Initially, the work proposed in this thesis aspired
COST 231 - Hata 9.182 11.280 to compare the results of the same propagation
SPM 2D 19.522 22.195 model using different types of databases and dif-
SPM 2.5D 30.867 33.769 ferent propagation models using the same type of
WinProp UDP Outdoor 11.955 15.338 geographical databases. However, the analysis of
WinProp UDP Indoor 11.031 16.534 several propagation models was found to be more
difficult than expected.
The validity of the walk test is another factor to
Table 5: Statistical Analysis of RSRP Results - be taken into consideration. There are several pos-
Walk Test Samples > - 85 dBm sible sources of errors. Apart from the mobile ter-
minal error (which depending on the vendor might
Model MAE (dB) SD (dB) be up to 3 dB), positioning also suffers from in-
Okumura - Hata 12.927 18.346 consistencies, which added another error factor by
COST 231 - Hata 17.858 23.433 requiring artificial correction of measurements that
SPM 2D 8.384 11.104 had their GPS locations registered indoor when it
SPM 2.5D 14.606 15.636 is known all walk tests were outdoor. Also, it is
WinProp UDP Outdoor 18.942 26.439 important to note that the database of the sites’
WinProp UDP Indoor 31.270 41.399 location is prone to errors, since the exact location
of the antennas is most of the time unknown and
assumed to be the geographical location given for
the site.
Table 5 only the samples with a RSRP level higher Moving into the analysis of the actual results,
than -85 dBm were used. a first outlook showed Okumura-Hata and COST
The propagation model that achieved an overall 231-Hata obtained the lowest values of MAE, which
lower MAE in the lower power scenario was COST is somehow surprising considering that these propa-
231-Hata with a MAE of 9.182 dB. It is worth men- gation models are the most generic ones. Okumura-
tioning that WinProp UDP Indoor had a clear im- Hata obtained a MAE of 13.533 dB, while COST
provement of its error. In the low power scenario its 231-Hata obtained 14.190 dB of MAE. Both mod-
MAE was 11.031 dB whereas in the general analy- els are deterministic and based on the Okumura
sis its MAE was 26.466 dB. For the higher power model. As such, these models are ideal for urban
signal scenario, the model that achieved the lowest cities with many low buildings. Antwerp, as it hap-
MAE was SPM 2D with 8.384 dB. pens, falls particularly well into this scenario, es-
An equivalent analysis of the point-to-point com- pecially in the city centre. The citys architecture
parison of the RSRQ results was elaborated. The is very regular with perpendicular streets and most
most relevant results of all the analysis are depicted buildings of similar height, particularly in the area
in Table 6. where the walk test took place. The good perfor-
According to these results, the propagation mance of these models can partly be justified by
model that obtains the lowest MAE is WinProp these factors. In addition, these measurements were
UDP Indoor with 6.319 dB. In this comparison how- performed outdoor and at ground level which corre-
ever, the difference between the MAE of each model sponds to the conditions these models are applied.

9
 Still regarding RSRP, going into more detail, it tions where the existing macro network does not
was decided to avoid samples with an absolute error provide enough radio quality to allow higher order
higher than 30 dB and those samples were excluded modulation schemes. However, it has not been pos-
in a second analysis. This procedure intended to sible to assess the behaviour of these propagation
mitigate the effect of outliers, as the logarithmic models applied to outdoor small-cells, traditionally
measurement has a lot more impact in the over- deployed at street level at heights no higher than
all statistical distribution of the results. Therefore, 2 floors and where most of the radio propagation
this analysis aims to provide a wider perspective of happens below rooftop level.
the results. Regarding future work, it would be interesting
According to this analysis, the model that ob- to validate some of these results by recreating this
tained the lowest value of MAE was the WinProp study in other urban areas or even expand it to dif-
UDP Outdoor, with a value of 12.427 dB. ferent scenarios and assess its validity. An equally
RSRP was also analysed according to the range valid experiment would be to recreate this exact
level. The MAE and SD were calculated for two study but changing the terminal and increasing the
different RSRP intervals. In the first case samples route so as to cover more clutter classes, across a
of the walk test that had RSRP lower or equal to wider area to better complement the measurement
-85 dBm were analysed separately. This scenario scenario and increase the amount of measured sam-
attempts to interpret the results of the different ples.
models when the mobile terminal is located at a
References
considerable distance from the base station. The
[1] Cisco, ”Cisco Visual Networking Index: Global
model that obtained the lowest MAE in this case
Mobile Data Traffic Forecast Update, 2016-
was COST 231-Hata. The second scenario only
2021”, White Paper, Cisco Systems, USA,
analyses walk test samples with a power level higher
February 2017.
than -85 dBm. In this situation, when the receiver
is located closer to the base station the model that [2] B. Hanta. ”SC-FDMA and LTE ul Physical
obtained the lowest MAE of all the different simula- Layer Design,”. Seminar LTE: Der Mobilfunk
tions was SPM 2D, which also registered the lowest der Zukunft, University of Erlangen-Nuremberg,
standard deviation. LMK, 2009.
The model that overall seems to show a best per-
formance regardless the method of analysis is Win- [3] H. Holma and A. Toskala. LTE for UMTS-
Prop UDP Outdoor. This model presents almost OFDMA and SC-FDMA based Radio Access.
the best MAE when all samples are analysed to- John Wiley & Sons, Chichester, 2009.
gether and a good match when splitting the RSRP [4] S. I. Popoola and O. F. Oseni. Performance
into different categories, although for higher RSRP evaluation of radio propagation models on gsm
levels the MAE is slighlty degraded. In the area network in urban area of lagos, nigeria. Inter-
analysis section, WinProp UDP Outdoor is also the national Journal of Scientific & Engineering Re-
model that more consistently approximates to the search, 5(6):1212–1217, 2014.
results measured in the walk test.
When looking to RSRQ, it shows less variation [5] R. Wahl, G. Wölfle, P. Wertz, P. Wildbolz,
than the RSRP. A possible cause for this is the fact and F. Landstorfer. ”Dominant Path Predic-
the RSRQ measures the quality of the received ref- tion Model for Urban Scenarios,”. 14th IST
erence signal over the RSSI, which is a measure of Mobile and Wireless Communications Summit,
the entire bandwidth highly associated with other Dresden (Germany), 2005.
cells interference. In this case both numerator and
[6] J. Zhang and G. De la Roche. Femtocells: tech-
denominator are equally affected by the error. In
nologies and deployment. John Wiley & Sons,
the first analysis of the RSRQ results, the model
Chichester, 2010.
that obtained the lowest error was WinProp UDP
Outdoor with a MAE of 6.319 dB.
Another important aspect is the fact that, in-
cidentally, all base stations were located above
rooftop level within the scenario under analysis.
This is a typical scenario when a macro network
requires capacity expansion and the proposed solu-
tion is to deploy small-cells. Under these conditions
the traditional approach is to plan small-cells for
coverage holes that the existing macro sites can-
not cover or to improve network quality in loca-

10