Small-Cell LTE Networks: Evaluating RF Planning Methods

Small-Cell LTE Networks: Evaluating RF Planning Methods

Ines [email protected]

Instituto Superior Tecnico, Lisboa, Portugal

November 2018

Abstract

This work focuses on the study and evaluation of different propagation models and geographicaldatabases used in the planning of LTE networks. A key aspect of the study includes the comparisonof 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, whichrecreated the path followed in the walk test, to simulate the RSRP and RSRQ levels obtained with eachmodel. The point-to-point study used the location of the walk test samples to obtain the equivalentpredicted 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 walktest led to several conclusions. Although the Okumura-Hata model obtained the lowest mean absoluteerror, it was the WinProp UDP Outdoor model that overall showed the best performance regardlessof the analysis method. SPM 2D was the model that adapted better when the walk test samples hada RSRP level higher than -85 dBm while COST 231-Hata obtained the best results in the oppositescenario. A relevant conclusion stresses the importance of accurate drive tests and their validity, aswell 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

Over the last decades, fuelled by our need to com-municate and the technological advances, there hasbeen a dramatic change in the way we communicatethat in time altered consumers needs.In the beginning of digital cellular networks thetechnologies developed, like the Global System forMobile Communications (GSM), were designed toprimarily focus on voice traffic. The introduction of3rd Generation (3G) networks, such as the UniversalMobile Telecommunications System (UMTS), insti-gated an increased use of data by enabling the useof various services like the Global Positioning Sys-tem (GPS), mobile television and video conferenc-ing. But in order to respond to those ever-changingneeds, the Long Term Evolution (LTE), commonlyknown as 4th Generation (4G), was created. It wasthe first cellular communication system designedspecifically to support packet-switched data servicessince the beginning. It aims to provide higher datarates, lower latency and better spectral efficiency.Still, tendencies have shown an immeasurablegrowth of data traffic that clearly dominates in com-parison to voice traffic. Although 4G connectionsrepresented 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-

bile connections but only 24% of the traffic. Accord-ing to predictions from [1], by 2021, 4G is expectedto correspond to 53% of all connections and 79% ofthe total traffic. As the capacity demand in mobilenetworks increased, the development of new wire-less communication technologies that solved some ofthe existing constraints related to radio spectral re-sources became crucial. Small-cells were presentedas a promising solution to increase the capacity andcoverage of the network by offloading the macrolayer and increasing the Quality of Service (QoS).One of the tools that contributes to an efficientplanning is the use of the correct propagationmodel for the area in question and the geograph-ical database used alongside it. In this thesis differ-ent propagation models, used in the planning anddeployment of small-cells in a typical urban sce-nario, will be analysed. The objective is to comparereal data obtained in a walk test with predictionsof different propagation models and geographicaldatabases, in a LTE network using a radio networkplanning software. This thesis was developed in col-laboration with Nokia Portugal. This partnershipwas essential, as it allowed for a realistic study usingdata obtained in an actual project.

1

2. Fundamental Concepts

This section provides an overview of the technolo-gies referred to in this thesis such as LTE, small-cells and the different types of propagation modelsand geographical databases.

2.1. LTE

LTE is a wireless communication standard devel-oped by the 3rd Generation Partnership Project(3GPP), aimed at enhancing the Universal Terres-trial Radio Access Network (UTRAN) and optimiz-ing the radio access architecture. Some of the driv-ing forces for the development of LTE include theincreased capability of existing wireline technolo-gies, the need for additional wireless capacity, aswell as the need for lower cost wireless data deliv-ery and the existing competition with other wire-less technologies. LTE was designed to supportonly packet-switched services, in contrast to thecircuit-switched data services of the previous gen-erations. Its objective is to provide continuous IPconnectivity between the user equipment (UE) andthe packet data network (PDN) without disruptionsduring mobility.The main requirements for this standard includedhigh spectral efficiency, high peak data rates, shortround trip time as well as flexibility in frequencyand bandwidth. LTE should be able to providepeak user throughput with a minimum of 100Mbit/s in the DL and 50 Mbit/s in the UL. LTE sys-tems contain two separate categories: the EvolvedUniversal Terrestrial Radio Access Network (E-UTRAN) and the Evolved Packet Core (EPC), asrepresented in Figure 1.

Figure 1: EPS Architecture (extracted from [3]).

The E-UTRAN is a network composed by onetype of node - evolved Node B (eNB). These eNBscan cover one or more cells and are capable of han-

dling all the radio related protocols. The eNBs arelinked to each other through a X2 interface and con-nected to the EPC through the S1 interface. TheEPC is responsible for the overall control of the UEand establishment of the bearers. A bearer is an IPpacket flow with a defined QoS between the gate-way and the UE. The main logical nodes of the EPCare the PDN Gateway (P-GW), the Serving Gate-way (S-GW) and the Mobility Management Entity(MME).The multiple access in LTE is based on OrthogonalFrequency Division Multiple Access (OFDMA) forthe DL and on Single Carrier Frequency DivisionMultiple Access (SC-FDMA) for the UL, both withCyclic Prefix (CP). The transmission in OFDMAis based on the use of narrow, mutually orthogo-nal sub-carriers. In LTE, each sub-carrier is spaced15 kHz regardless of the total transmission band-width. At the sampling instant of a single sub-carrier, all the others have zero value. The or-thogonality achieved by the sub-carrier allows for amore efficient use of the spectrum, because a bandguard is no longer needed to avoid sub-carrier in-terference. One of the main challenges in OFDMAis the high Peak-to-Average Ratio (PAR) of thetransmitted signal. This not only limits the trans-mitted power and reduces coverage but forces theuse of better power amplifiers. The UL signal isgenerated by the users mobile terminal so in thisscenario a multiple access technique that enablesbetter power-amplifier efficiency is needed. Thiswas solved by choosing the SC-FDMA in the UL.Similarly to OFDMA, SC-FDMA divides the trans-mission bandwidth into multiple parallel subcarri-ers maintaining the orthogonality of the subcarri-ers. However, in SC-FDMA the data symbols arenot directly assigned to each subcarrier indepen-dently like in OFDMA. Instead, the signal which isassigned to each subcarrier is a linear combinationof all modulated data symbols transmitted at thesame time instant. [2]A high degree of spectrum flexibility is one of themain characteristics of the LTE radio access. LTEsupports 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 samefrequency but transmission takes place in differ-ent, non-overlapping time slots. Contrarily, FDDimplies that DL and UL transmissions take placein different, sufficiently separated, frequency bandsbut simultaneously.LTE uses 3 types of Quadrature Amplitude Modu-lation (QAM). They are the Quadrature Phase ShiftKeying (QPSK) or 4-QAM, 16-QAM and 64-QAMmodulation schemes. The QPSK carries 2 bits persymbol while 16-QAM and 64-QAM carry 4 and

2

6 bits per modulation symbol, respectively. Thesemodulation schemes are available in both UL andDL directions in all devices, except for 64-QAMin the UL which depends on the UE capability.The QPSK is the most robust to interference andbad channel conditions, but consequently offers thelower bit rate. In contrast, the 64-QAM allows thehigher bit rates in LTE, but requires the best con-ditions in terms of Signal-to-Interference-plus-NoiseRatio (SINR), by being the most liable to errors dueto interference.

2.2. Small-Cells

The unmeasured usage growth of cellphones, tabletsand data-hungry applications has caused an ex-ponential growth of traffic in mobile networks, asstated before. Operators have met this challengeby increasing capacity with new radio spectrum,adding multi-antenna techniques and implement-ing more efficient modulation and coding schemes.However, these measures alone are insufficient inextremely crowded environments and at cell edgeswhere performance can significantly degrade. Com-plementing macro networks with small-cells is an ef-fective way to extend coverage and increase capac-ity indoors and outdoors, in public spaces, offices orresidences.Small-cells are operator-controlled, low-powered ra-dio access nodes, used to complement mobile ser-vices that is served by macro cells. They are primar-ily deployed to increase capacity in hot spots withhigh user demand and to fill in areas not coveredby the macro network both outdoors and indoors.Small-cells also improve network performance andservice quality by offloading from the large macrocells. A combination of macro cells and small-cellsis usually called as a Heterogeneous Network thatprovides increased bitrates per unit area. There aredifferent types of small-cells which can be groupedaccording to cell size: femtocells, picocells and mi-crocells, as represented in Figure 2.

Figure 2: Cell Sizes of Different Technologies (ex-tracted from [6]).

Operators began to install smaller outdoor basestations called microcells as new data services thatrequired higher coverage, emerged. These are de-

ployed in specific areas in which extra capacityis known to be needed, for example near a trainstation or in a city centre. They are also oftentemporarily deployed during special occasions likesporting events. Adding microcells in the urbanenvironment allows the operator to subdivide thecells, leading to an optimization of the use of thespectrum and ensuring a better capacity. Micro-cells have a range of a few hundred metres.Picocells are small base stations that work very sim-ilarly to an Access Point (AP). Picocells have a lowpower radiating antenna that connects to the opera-tor’s core network. Picocells can be used in businessenvironments, shopping centres and airports, for ex-ample. They are also useful in high-rise buildings,where the macro cell signal strength decreases withheight. The main advantage is that they are con-siderably cheaper and since they are placed indoor,they effectively increase indoor radio coverage andcapacity. These small-cells have a range of approx-imately 100 metres.Femtocells were proposed to extend the concept ofpicocells to home networks, in particular. Theyhave limited output power, between 10 and 20 dBm,and capacity as they serve 4 to 6 users. These typeof small-cells can be deployed in a variety of sce-narios. A typical deployment is in a rural envi-ronment, where a femtocell enables access to mo-bile network for users which are typically under acoverage hole but with a reliable Digital SubscriberLine (DSL) connection. As it happens homes in ru-ral areas usually lack the macro cell coverage, butsome already have access to high data rate Inter-net, making it an ideal scenario to deploy femto-cells. Another use case case for femtocells is anenterprise environment, where the low capacity offemtocells is typically enough to cover a number ofoffices. The major advantage of femtocells is thefact they are directly installed by the end users in-side their homes and have a maximum range of ap-proximately 30 metres. This will ensure good cover-age for subscribers. For operators, femtocells are acheap 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 mobileradio channel. Consequently, predicting the propa-gation characteristics between the transmitters andreceivers is crucial for the design and installation ofa radio network communications system. All com-putations and simulation procedures depend on re-liable propagation models. A propagation modelis a formulation that intends to characterise andpredict radio wave propagation. The basis for any

3

propagation model is a database, which describesthe propagation environment. Several propagationmodels have been developed and proposed for cel-lular systems operating in different environments(outdoor, urban, suburban, rural, and indoor) butthere are generally three different types of propaga-tion models: empirical models, deterministic mod-els and semi-empirical models.An empirical model is based on observations andmeasurements. These are useful to study the gen-eral behaviour or to enable a rough estimation of thenumber of required cells in a large area. The classi-fication of empirical models can be further dividedinto time dispersive and non-time dispersive. Timedispersive 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 asdistance, antenna heights, frequency and transmit-ter power to predict average path loss.The Okumura Model and the Okumura-Hata Modelwere some of the first empirical models developedthat served as groundwork to many others. TheCOST-231 Hata Model, based on the previous ones,is another commonly used empirical model.The Standard Propagation Model (SPM), used inradio network planning software, is a propagationmodel based on formulas of Okumura and Hata thatwas adapted to perform signal coverage predictionsin the frequency range of 150 MHz to 3500 MHz andfor distances between 1 km to 20 km. The SPM isbest applied on mobile technologies such as LTEand LTE-A.Semi-empirical are based on a combination of mea-surements and theory and attempt to interpret fielddata based on theoretical principles. The first site-specific propagation models were semi-empiricaland were based on detailed terrain characteristicsextracted along the individual propagation pathsbetween transmitter and receiver. These models uselow resolution geographic data but still manage toobtain reasonably fair results for coverage predic-tions in urban areas. They are, however inefficientwhen it comes to indoor coverage estimates in denseurban areas. The COST 231 Walfisch-Ikegami isconsidered a semi-empirical model used to predicturban environments. This model is an extension ofCOST-231 Hata model and it can be used for fre-quencies above 2000 MHz.Deterministic models are numerical methods thatsimulate radio waves propagation by reproducingthe physical propagation phenomena. These mod-els use a specific location for the transmitter andthe receiver to provide a reliable and thorough es-timation of the path losses and the channel charac-teristics. They rely on terrain descriptive features(altitude, the geometry of the buildings, materials

used in construction) to simulate the shadowing ef-fect that a signal experiences to provide a correlatedspatial variation of the path loss.Dominant Path Models (DPM) are another type ofpropagation models widely used in radio networkplanning. A DPM does not rely only on the directray (like empirical models) and it does not considerhundreds of rays for a single radio link (like raytracing), as pictured in Figure 3.

Figure 3: Empirical models (left), Ray Tracing(middle) and DPM (right) (extracted from [5]).

Instead it focuses on the most dominant path be-tween transmitter and receiver. The Urban Dom-inant Path (UDP) model is one of the dominantpath models that was developed for the predictionof the field strength specifically in urban environ-ment. The algorithm of DPMs can be resumed intwo different stages. First, the dominant paths haveto be determined and then the prediction of thepath loss for each path must be calculated. Thedominant path must lead via convex corners fromthe transmitter to the receiver. For the determina-tion of the path, a tree with all convex corners iscomputed starting with the corners visible from thetransmitter.

As mentioned above, propagation models per-form their predictions using information from ter-rain databases. The environment to be analysedand predicted is stored in a database file. But de-pending on the type of scenario (indoor, urban orrural), the resources available and the overall levelof accuracy intended for each project a different de-scription of the relevant obstacles is necessary.The most rudimentary type of database used nowa-days is a 2D database. It is composed by a heightfile or Digital Terrain Model (DTM) and a clut-ter file. The DTM consists in a digital model thatrepresents the surface of the Earth, excluding veg-etation as well as buildings and man-made struc-tures. It contains the altitude values of the regionand as such is characterised by a latitude Y, a lon-gitude X and a specific resolution. The resolutionis linked to the retrieval of the topographic data ac-cording to a given scale. The clutter file used in a2D database describes the land use classification ofthe area in analysis including any features that im-pact the radio wave propagation. The clutter typesor so-called clutter classes include natural environ-

4

ments: river, forest, ocean or man-made: industrial,suburban, dense urban, to name a few. The appro-priate number of clutter classes depends on the geo-graphic area. The combination of the DTM and theclutter types result in a 2D database because the al-titude only refers to the height of the terrain andno clutter height information is usually provided.The propagation models that use 2D databases areempirical ones. In empirical models the path loss iscalculated for the direct path between the transmit-ter and receiver and the topography of the terrainis only used to calculate obstructions. According tothe given clutter category of the area, an attenua-tion or gain is added to the total path loss.

Another type of database used is a 2.5D database.While it also consists of a DTM and clutter cate-gories it includes the average height of each scenarioby introducing the clutter heights. Each specificcategory will have a preassigned average height, forexample a low density urban area will have an av-erage height of 20 metres and an area classified as avillage might have an average height of 6 metres. Ifa higher accuracy is necessary, subcategories with aspecific height assigned to it can be created withina clutter class. This method provides with a bet-ter correspondence with the real environment. Forexample, a scenario labelled as dense urban can bedivided in two, one with an average height of 15metres and the other with 20 metres.

The most comprehensive and detailed type ofdatabase is a 3D database. Like previous databasesit uses a DTM and clutter types (it may or maynot include the clutter heights) but incorporates theheight and contour of the buildings (3D BuildingVectors). When using a 3D database, each of thebuildings will have its own height and shape repre-sented as close to reality as possible, unlike previousdatabases.

3. Methodology

This sector presents the methodology used in thiswork.

3.1. Propagation Models Used

The existing LTE network, operating at 1800 MHz,was configured in the 9955 RNP tool. All the nec-essary information regarding each site, transmitterand cell was included to obtain accurate predic-tions.The first propagation model tested is Okumura-Hata. This model established empirical mathemat-ical relationships which described the graphical in-formation of the Okumura model. The equationdescribing the average path loss in an urban area is

given by:

PL = 69.55 + 26.16log10(f)− 13.82log10(hT )+

[44.9− 6.55log10(hT )]log10(d)−H(hR, f)

(1)

where:

f: frequency in MHz;

HT: transmitter height in metres;

d: distance in kilometres;

H(hR,f): correction factor for the mobile antennaheight.

The correction factor is computed differently ac-cording to the size of the area in question. In thiscase:

H(hR, f) =

8.29(log101.54hR)2 − 1.1,f ≤ 200MHz3.2(log1011.75hR)2 − 4.97,f ≥ 400MHz

(2)

where:

f from 150 MHz to 1500 MHz;

hT from 30 m to 200 m;

hM from 1 m to 10 m;

d from 1 km to 20 km.

This model will only be applied with a 2Ddatabase as it is empirical.

The COST 231 Hata Model will also be used.This model extends the previous one to cover thefrequency range of 1500 MHz to 2000 MHz. Theother validity conditions are maintained.

The path loss is given by:

PL = 46.3 + 33.9log10(f)− 13.82log10(hT )

+[44.9− 6.55log10(hT )]log10(d)−H(hR, f) + Cm

(3)

where:

H(hR, f) =

8.29(log101.54hR)2 − 1.1,f ≤ 300MHz3.2(log1011.75hR)2 − 4.97,f ≥ 300MHz

(4)

Cm =

{0dB,medium-sized city or suburban areas3dB,metropolitan areas

(5)

Similarly, to the Okumura-Hata model, the pre-dictions using the COST 231 Hata Model will onlyuse a 2D database given its empirical nature.

The SPM will also be tested. The model is basedon Equation 6 and each of the K parameters canbe adjusted according to specific conditions suchas terrain profile, diffraction mechanisms, morphol-ogy of clutter classes and the effective height of the

5

transmitting and receiving antennas. [4]

PR = PT [K1 +K2log10(d) +K3log10(hT )+

K4.DiffractionLoss+K5log10(d).log10(hT )+

K6hR +K7log10(hR) +KCLUTTER.fCLUTTER

+KHILL]

(6)

where:PR: received power in dBm;PT: eirp in dBm;hT: effective height of the transmitter in metres;hR: effective height of the receiver antenna in

metres;d: distance between the receiver and the trans-

mitter in metres;Diffraction Loss: losses due to diffraction over an

obstructed path in dB;fCLUTTER: weighted average losses due to the

clutter.The specific recommended adjustment coeffi-

cients will not be presented as they are classified byNokia as internal use only. The SPM will be usedwith the 2D and 2.5D database and the formulawill be adapted for each one according to Nokia’sinternal guidelines. For the coverage prediction us-ing a 2D database, 9955 computes the total pathloss using the topographical information of the ter-rain provided in the DTM files and the clutter classof the area. Each clutter class will have assigneda specific loss value that is then added to the totalpath loss equation. The loss values used correspondto Nokia internal guidelines.To compute the predictions with the 2.5D database,9955 will use the path loss formula and added lossesaccording to the clutter class like in the previouscase but it will also consider an average height foreach clutter class, on top of the terrain topography.The specific height given to each clutter class wasalso defined after analysing Nokias internal guide-lines.The main difference between a 2.5D and a 3Ddatabase consists in the building vector file. Thevector file describes each building height, shape andorientation in detail. Another propagation modeltested was the WinProp UDP. This model is knownto be a trade-off between computation complexityand accuracy. The 3D building database was pre-processed prior to the prediction computation, us-ing Wallman a module that is part of the WinPropwireless network planning software package.

3.2. Data AdjustmentsSome corrections had to be implemented to reduceany inconsistencies in the data obtained in the walktest and to make it comparable to the results of thepredictions obtained with 9955.

A buffer area was used in the coverage by areaanalysis. This buffer area was created by analysingand connecting each position of the walk test tocreate a path. After the path was created, it wastransformed in an area using QGIS.By using the buffer most of the area analysed iscategorized as Open Area which corresponds to theactual streets. Considering 9955 adds a specific lossvalue according to the clutter class, if this correc-tion was not performed, the values of the predic-tions would be over-estimated because they wouldinclude the loss value associated to each building orother obstacles.It is relevant to add that any errors found whileanalysing the walk test points were also corrected.This means any walk test location that was shownto be in an indoor situation was placed in the near-est possible outdoor position, since it is known thewalk test only took place outdoors.The data obtained in the walk test was also treatedto eliminate any inconsistencies. If at a given point,with the same coordinates, different signal valueswere recorded by the same transmitter an averagevalue was then used.The measurements recorded during the walk test in-clude the Reference Signal Received Power (RSRP)and Reference Signal Received Quality (RSRQ).The RSRP is the average of power levels receivedacross all reference signal symbols within the con-sidered measurement frequency bandwidth. TheRSRQ is the ratio of RSRP and the E-UTRA Car-rier Received Signal Strength Indicator (RSSI). TheE-UTRA carrier RSSI comprises the total receivedwideband power observed by the UE. The signalvalues recorded during the walk test correspond tothe signal level from the best server.To sum up the propagation models simulated withthe corresponding databases are listed in Table 1.

Table 1: Propagation Models and Databases Tested

Model 2D 2.5D 3DOkumura - Hata XCOST 231 - Hata XSPM X XWinProp UDP Outdoor XWinProp UDP Indoor X

4. ResultsThis section presents the results of the differentanalysis performed in this work.

4.1. Coverage ResultsIn an initial stage a coverage analysis by area wasperformed to compare the Cumulative DistributionFunctions (CDF) of the several models with the

6

walk test resuls.In Figure 4 the RSRP results of the walk test are

presented. The scale goes from RSRP levels higheror equal to -55 dBm to RSRP levels higher or equalto -115 dBm, with a step of -5 dBm.

Figure 4: RSRP Results Obtained in the Walk Test

The numerical results of the RSRP levels ob-tained in the walk test are presented in Figure 5.

0

20

40

60

80

100

120

-11

8.9

-11

0.0

5

-10

7.9

5

-10

6.2

-10

4.5

7

-10

3.2

3

-10

1.5

7

-10

0.1

-98

.45

-96

.95

-95

.17

-93

.65

-91

.95

-90

.2

-88

.8

-87

.3

-85

.9

-84

.4

-82

.8

-81

.51

25

-79

.9

-78

.35

-76

.65

-75

-73

-70

.4

-66

.65

CD

F [%

]

RSRP Level [dBm]

Walk Test

Walk Test

Figure 5: CDF of the Walk Test RSRP Results

The highest value of RSRP measured throughoutthe route was -60.8 dBm. This value was found onlyonce among all the samples and it corresponds to0.024 % of all the valid measurements. The graphshows that as the RSRP level decreases, more sam-ples are included. According to this results, 50 %of the samples have a RSRP level greater than orequal to -90.8 dBm. The lowest value of RSRP levelmeasured during the walk test was -118.9 dBm.

The results obtained from the predictions withthe propagation models were all performed consid-ering the same buffer area. In Figure 6 a graphdepicting the RSRP coverage results of every prop-agation model is presented. The SPM 2.5D appearsto be highly optimistic with a CDF curve distantfrom the curve of the walk test results. The SPM2.5D simulated values as high as -58.8 dBm whilethe highest values obtained during the walk testwere -60.8 dBm. In comparison the highest valuecalculated by SPM but with a 2D database was -63.8 dBm. The highest value obtained by the simu-lation of Okumura-Hata model was -67.7 dBm andwith COST 231-Hata was -72.6 dBm. The simula-tions with WinProp UDP model show that it ob-

0%

20%

40%

60%

80%

100%

120%

-12

0-1

18

.45

-11

6.9

-11

5.3

5-1

13

.8-1

12

.25

-11

0.7

-10

9.1

5-1

07

.6-1

06

.05

-10

4.5

-10

2.9

5-1

01

.4-9

9.8

5-9

8.3

-96

.75

-95

.2-9

3.6

5-9

2.1

-90

.55

-89

-87

.45

-85

.9-8

4.3

5-8

2.8

-81

.25

-79

.7-7

8.1

5-7

6.6

-75

.05

-73

.5-7

1.9

5-7

0.4

-68

.85

-67

.3-6

5.7

5-6

4.2

-62

.65

-61

.1

CD

F [%

]

RSRP Level [dBm]

RSRP Coverage

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 measurementsretrieved during the walk test. Both WinProp UDPmodels predicted -65.5 dBm as the highest RSRPlevel in the buffer area.

The results can also be interpreted by separat-ing the results by power level. In the lower rangeof signal power the WinProp UDP Indoor is themodel that reaches an RSRP Level closest to theone obtained in the walk test. In this case for aCDF value of 95 % UDP Indoor obtained -104.25dBm while the walk test measurements resulted in-106.57 dBm.

As the signal power increases, the WinProp UDPOutdoor and the COST 231-Hata model also tendto approximate more to the walk test results. Whenthe walk test obtained -98.85 dBm for a CDF valueof 75 %, the result of the COST 231-Hata modelwas -95.7 dBm while UDP achieved -95.25 dBm and-96.3 dBm, for the outdoor and indoor mode respec-tively. When the signal power increases and crossesthe -85 dBm mark, the Okumura-Hata starts to ob-tain much better results than the COST 231-Hatamodel. For a CDF value of 25 % the walk testobtained an RSRP level of -82.55 dBm while theOkumura-Hata also attained -82.55 dBm and theCOST 231-Hata result was-87.55 dBm In the higherrange of signal power both WinProp UDP modelsclosely align with the walk test results, as well asthe SPM 2D until around -72 dBm.

The same analysis was performed for the RSRQmeasurements. The results of the RSRQ levels ob-tained in the walk test were also added to 9955 andare represented in Figure 7. The scale portrayedin the figure ranges from lower RSRQ values suchas -22 dB to higher ones like -6 dB. Although theRSRQ levels measured in the walk test did not reacha value lower than -17.2 dB, the scale includes somelower values to account for the results obtained inthe predictions of the different models.The numerical results of the RSRQ levels obtainedin the walk test are presented in Figure 8. As seen inthe graph, the results are presented using the CDFin the vertical axis and the RSRQ level measured in

7

Figure 7: RSRQ Results Obtained in the Walk Test

dB in the horizontal axis. The graph shows that the

0

20

40

60

80

100

120-1

7.2

-14

.3

-13

.3

-12

.7

-12

.1

-11

.57

-11

.15

-10

.87

-10

.46

-10

.13

-9.8

9

-9.5

5

-9.2

-8.9

25

-8.6

75

-8.3

83

-8.1

-7.9

-7.6

5

-7.4

66

-7.3

4

-7.1

87

-7.0

5

-6.8

25

-6.5

5

-6.1

CD

F [%

]

RSRQ Level [dB]

Walk Test

Figure 8: CDF of the Walk Test RSRQ Results

highest value of RSRQ measured throughout theroute was -6 dB, which corresponds to 0.253 % ofall the values obtained. By observing the graph oneconcludes that 50 % of the samples have a RSRQlevel greater than or equal to -8.95 dB. The lowestvalue of RSRQ level measured during the walk testwas -17.2 dB. Each of these measurements corre-spond to the signal received from the best server atthat moment and exact location.

In order to compare the RSRQ results obtainedby every model, a graph with the cumulative dis-tribution function of the results is shown in Figure9.

0%

20%

40%

60%

80%

100%

120%

-22

-21

.4

-20

.8

-20

.2

-19

.6

-19

-18

.4

-17

.8

-17

.2

-16

-14

.7

-13

.4

-12

.75

-12

.15

-11

.55

-10

.95

-10

.35

-9.7

5

-9.1

5

-8.5

5

-7.9

5

-7.3

5

-6.7

5

-6.1

5

-5.5

5

CD

F [%

]

RSRQ Level [dB]

RSRQ Coverage

Okumura-Hata COST 231-Hata SPM 2D

SPM 2.5D WinProp UDP Indoor WinProp UDP Outdoor

Walk Test

Figure 9: Comparison of RSRQ Coverage Predic-tions

As opposed to the RSRP results, none of themodels used for the RSRQ coverage prediction im-mediately stands out as the best fit, seeing as the

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

4.2. Validation of Results

A statistical analysis of the accuracy of each modelstudied is necessary. The following statistical per-formance metrics were used: Mean Absolute Error(MAE) and Standard Deviation (SD). These met-rics were calculated in linear units and convertedback to dB.

MAE =1

N

N∑i=1

|SMi − SP

i | (7)

SD =

√√√√ 1

N

N∑i=1

(|SMi − SP

i | − µ)2 (8)

where µ is the absolute error.

The MAE and SD of the RSRP point-to-pointresults are presented in Table 2. According to theresults in the table the propagation model that ob-tains the best approximation to the walk test isOkumura-Hata model with a MAE of 13.533 dB.

Table 2: Statistical Analysis of RSRP Results

Model MAE (dB) SD (dB)Okumura - Hata 13.533 17.333COST 231 - Hata 14.190 21.188SPM 2D 17.953 21.507SPM 2.5D 29.213 33.087WinProp UDP Outdoor 15.621 24.086WinProp UDP Indoor 26.466 38.950

In Table 3, the statistical analysis of the RSRPresults is performed but with a variation. In thiscase, the samples that had an absolute error higherthan or equal to 30 dB were excluded. This analy-sis attempts to reduce the effect of the outliers onthe overall results. According to these results, themodel that obtained the lowest MAE is now Win-Prop UDP Outdoor with 12.427 dB. WinProp UDPIndoor also obtained a much lower MAE than in theprevious case.

The results were also divided according to powerlevel in order to better analyse the general be-haviour of each model. In Table 4 only walk testsamples with a RSRP value lower than or equal to-85 dBm were considered. On the other hand, in

8

Table 3: Statistical Analysis of RSRP Results ex-cluding samples with an AE higher than 30 dB


Table 4: Statistical Analysis of RSRP Results-Walk Test Samples ≤ - 85 dBm


Table 5: Statistical Analysis of RSRP Results -Walk Test Samples > - 85 dBm


Table 5 only the samples with a RSRP level higherthan -85 dBm were used.

The propagation model that achieved an overalllower MAE in the lower power scenario was COST231-Hata with a MAE of 9.182 dB. It is worth men-tioning that WinProp UDP Indoor had a clear im-provement of its error. In the low power scenario itsMAE was 11.031 dB whereas in the general analy-sis its MAE was 26.466 dB. For the higher powersignal scenario, the model that achieved the lowestMAE was SPM 2D with 8.384 dB.

An equivalent analysis of the point-to-point com-parison of the RSRQ results was elaborated. Themost relevant results of all the analysis are depictedin Table 6.

According to these results, the propagationmodel that obtains the lowest MAE is WinPropUDP Indoor with 6.319 dB. In this comparison how-ever, the difference between the MAE of each model

Table 6: Statistical Analysis of RSRQ Results

Model MAE (dB) SD (dB)Okumura Hata 7.343 4.733COST 231 Hata 7.406 4.859SPM 2D 7.573 5.052SPM 2.5D 8.127 5.700WinProp UDP Outdoor 6.319 3.634WinProp UDP Indoor 6.326 3.594

is much lower than in the RSRP case. The modelthat obtained the highest MAE was SPM 2.5D with8.127 dB, which is a difference of only 1.808 dB.

5. Conclusions

Initially, the work proposed in this thesis aspiredto compare the results of the same propagationmodel using different types of databases and dif-ferent propagation models using the same type ofgeographical databases. However, the analysis ofseveral propagation models was found to be moredifficult than expected.

The validity of the walk test is another factor tobe taken into consideration. There are several pos-sible sources of errors. Apart from the mobile ter-minal error (which depending on the vendor mightbe up to 3 dB), positioning also suffers from in-consistencies, which added another error factor byrequiring artificial correction of measurements thathad their GPS locations registered indoor when itis known all walk tests were outdoor. Also, it isimportant to note that the database of the sites’location is prone to errors, since the exact locationof the antennas is most of the time unknown andassumed to be the geographical location given forthe site.

Moving into the analysis of the actual results,a first outlook showed Okumura-Hata and COST231-Hata obtained the lowest values of MAE, whichis somehow surprising considering that these propa-gation models are the most generic ones. Okumura-Hata obtained a MAE of 13.533 dB, while COST231-Hata obtained 14.190 dB of MAE. Both mod-els are deterministic and based on the Okumuramodel. As such, these models are ideal for urbancities with many low buildings. Antwerp, as it hap-pens, falls particularly well into this scenario, es-pecially in the city centre. The citys architectureis very regular with perpendicular streets and mostbuildings of similar height, particularly in the areawhere the walk test took place. The good perfor-mance of these models can partly be justified bythese factors. In addition, these measurements wereperformed outdoor and at ground level which corre-sponds to the conditions these models are applied.

9

Still regarding RSRP, going into more detail, itwas decided to avoid samples with an absolute errorhigher than 30 dB and those samples were excludedin a second analysis. This procedure intended tomitigate the effect of outliers, as the logarithmicmeasurement has a lot more impact in the over-all statistical distribution of the results. Therefore,this analysis aims to provide a wider perspective ofthe results.

According to this analysis, the model that ob-tained the lowest value of MAE was the WinPropUDP Outdoor, with a value of 12.427 dB.

RSRP was also analysed according to the rangelevel. The MAE and SD were calculated for twodifferent RSRP intervals. In the first case samplesof the walk test that had RSRP lower or equal to-85 dBm were analysed separately. This scenarioattempts to interpret the results of the differentmodels when the mobile terminal is located at aconsiderable distance from the base station. Themodel that obtained the lowest MAE in this casewas COST 231-Hata. The second scenario onlyanalyses walk test samples with a power level higherthan -85 dBm. In this situation, when the receiveris located closer to the base station the model thatobtained the lowest MAE of all the different simula-tions was SPM 2D, which also registered the loweststandard deviation.

The model that overall seems to show a best per-formance regardless the method of analysis is Win-Prop UDP Outdoor. This model presents almostthe best MAE when all samples are analysed to-gether and a good match when splitting the RSRPinto different categories, although for higher RSRPlevels the MAE is slighlty degraded. In the areaanalysis section, WinProp UDP Outdoor is also themodel that more consistently approximates to theresults measured in the walk test.

When looking to RSRQ, it shows less variationthan the RSRP. A possible cause for this is the factthe RSRQ measures the quality of the received ref-erence signal over the RSSI, which is a measure ofthe entire bandwidth highly associated with othercells interference. In this case both numerator anddenominator are equally affected by the error. Inthe first analysis of the RSRQ results, the modelthat obtained the lowest error was WinProp UDPOutdoor with a MAE of 6.319 dB.

Another important aspect is the fact that, in-cidentally, all base stations were located aboverooftop level within the scenario under analysis.This is a typical scenario when a macro networkrequires capacity expansion and the proposed solu-tion is to deploy small-cells. Under these conditionsthe traditional approach is to plan small-cells forcoverage holes that the existing macro sites can-not cover or to improve network quality in loca-

tions where the existing macro network does notprovide enough radio quality to allow higher ordermodulation schemes. However, it has not been pos-sible to assess the behaviour of these propagationmodels applied to outdoor small-cells, traditionallydeployed at street level at heights no higher than2 floors and where most of the radio propagationhappens below rooftop level.

Regarding future work, it would be interestingto validate some of these results by recreating thisstudy in other urban areas or even expand it to dif-ferent scenarios and assess its validity. An equallyvalid experiment would be to recreate this exactstudy but changing the terminal and increasing theroute so as to cover more clutter classes, across awider area to better complement the measurementscenario and increase the amount of measured sam-ples.

References[1] Cisco, ”Cisco Visual Networking Index: Global

Mobile Data Traffic Forecast Update, 2016-2021”, White Paper, Cisco Systems, USA,February 2017.

[2] B. Hanta. ”SC-FDMA and LTE ul PhysicalLayer Design,”. Seminar LTE: Der Mobilfunkder Zukunft, University of Erlangen-Nuremberg,LMK, 2009.

[3] H. Holma and A. Toskala. LTE for UMTS-OFDMA and SC-FDMA based Radio Access.John Wiley & Sons, Chichester, 2009.

[4] S. I. Popoola and O. F. Oseni. Performanceevaluation of radio propagation models on gsmnetwork in urban area of lagos, nigeria. Inter-national Journal of Scientific & Engineering Re-search, 5(6):1212–1217, 2014.

[5] R. Wahl, G. Wolfle, P. Wertz, P. Wildbolz,and F. Landstorfer. ”Dominant Path Predic-tion Model for Urban Scenarios,”. 14th ISTMobile and Wireless Communications Summit,Dresden (Germany), 2005.

[6] J. Zhang and G. De la Roche. Femtocells: tech-nologies and deployment. John Wiley & Sons,Chichester, 2010.

10

Small-Cell LTE Networks: Evaluating RF Planning Methods

Documents

Transcript of Small-Cell LTE Networks: Evaluating RF Planning Methods