SCF Precision Planning WP FINAL 1

SCF
RELEASE 10.0 scf.io
DOCUMENT
230.10.01
Precision planning for 5G Era
networks with small cells
October 2019
www.scf.io/ www.5gamericas.org www.smallcellforum.org

SMALL CELL FORUM
RELEASE 10.0 scf.io
Small Cell Forum accelerates small cell adoption to drive the

wide-scale deployment of small cells and encourage the delivery of
fully integrated HetNets.
We are a carrier-led organization. This means our operator members establish

requirements that drive the activities and outputs of our technical groups.
We have driven the standardization of key elements of small cell technology

including Iuh, FAPI, nFAPI, SON, services APIs, TR-069 evolution and the
enhancement of the X2 interface.
Today our members are driving solutions that include small cell/Wi-Fi
integration, SON evolution and automation, virtualization of the small cell layer,
driving mass adoption via multi-operator neutral host, ensuring a common
approach to service APIs to drive commercialization and the integration of
small cells into 5G standards evolution.
The Small Cell Forum Release Program has now established business cases
and market drivers for all the main use cases, clarifying market needs and
addressing barriers to deployment for residential, enterprise, rural & remote,
and urban small cells.
The Small Cell Forum Release Program website can be found here: www.scf.io
All content in this document including links and references are for informational
purposes only and is provided ‘as is’ with no warranties whatsoever including
any warranty of merchantability, fitness for any particular purpose, or any
warranty otherwise arising out of any proposal, specification, or sample.
No license, express or implied, to any intellectual property rights is granted or

intended hereby.
©2007-2017 All rights reserved in respect of articles, drawings, photographs

etc published in hardcopy form or made available in electronic form by Small
Cell Forum Ltd anywhere in the world.
If you would like more information about Small Cell Forum or would
like to be included on our mailing list, please contact:
Email [email protected]
Post Small Cell Forum, PO Box 23, GL11 5WA UK
Member Services [email protected]

SCF
RELEASE 10.0 scf.io
About 5G Americas: The Voice of 5G and LTE for the Americas
5G Americas is an industry trade organization composed of leading

telecommunications service providers and manufacturers. The organization’s
mission is to advocate for and foster the advancement and full capabilities
of LTE wireless technology and its evolution beyond to 5G, throughout the
ecosystem’s networks, services, applications and wirelessly connected devices
in the Americas.
5G Americas is invested in developing a connected wireless community

while leading 5G development for all the Americas. The association prepares
numerous white papers and reports on key initiatives, challenges and
opportunities for the industry with a particular focus on those of greatest
importance to North America and Latin America and the Caribbean. The reports
are accessible on the association’s website www.5gamericas.org
5G Americas’ Board of Governors members include: América Móvil, AT&T,

Cable & Wireless, Cisco, CommScope, Entel, Ericsson, HPE, Intel, Kathrein,
Mitel, Nokia, Qualcomm, Sprint, T-Mobile USA and Telefónica.
5G Americas collaborates with government and regulatory bodies, as well as

standards organizations and other associations throughout the world to foster
the successful development of wireless technologies.
• 5G Americas, in partnership with 5G PPP, 5GMF, 5G Forum and IMT-2020,

participates in hosting the Global 5G Event on a bi-annual basis
• 5G Americas is a Market Representation Partner at 3GPP
• 5G Americas has agreements, memberships, relationships and/or

Memorandums of Understanding with: 5G Forum, 5G PPP, ASETA, ASIET,
ATIS, CASETEL, CINTEL, CITEL, ITU, NGMN Alliance, Small Cell Forum
and the UMTS Forum with the purpose of cooperation to improve mobile
communications.
Follow 5G Americas news on Twitter at @5GAmericas and Facebook at

www.facebook.com/5gamericas.
Executive summary
The ever-increasing demand for mobile data is driving network densification with the
deployment of small cells. Although lower cost than macro towers, the compact, low-
power nature of small cells means they also have a smaller serving area. This in turn
means they need to be located closer to demand hotspots in order to be effective in
supplying traffic efficiently and delivering a good return on investment (ROI).
The benefit of any given small cell site essentially boils down to the amount of traffic it
carries, whereas the cost is tied to installation, power, backhaul, rent, etc., and is
largely independent of its utilisation. It follows that the value (or ROI) of a given site
is affected by the precision with which it can be planned and sited. Small cells need to
be located close to areas of high traffic demand that are not well served by the
existing network, but they need to be placed with care. Recommended best practice is
to locate the small cell within 20-40m of the ideal.
A key input to the precision planning process is geolocated measurements of network

quality, which are used to build up a map of how well the existing network is serving
traffic demand. Small cells should be deployed where there is high traffic demand but
low signal quality.
The location accuracy of the quality reports contributes to the overall precision of the
planning process. Averaging of observed time difference of arrival (OTDOA) network
location estimates was found in one study to have a 60m median error, limiting its
applicability to small cell planning. Smarter analysis of the locate data using a machine
learning approach was found to reduce median location error down to 18m, which is
within the required range.
Finally, an example of small cell design is shared. Using data and the goals of
coverage and dominance, small cell placement is determined using an automated
process. The example shared in this document provides improved coverage and
reduced costs over a manual design.
Recommended best practice for precision planning
• For maximum ROI, small cells should be placed as close as possible to

demand peaks; best practice is within 20-40m.
• MNOs would like equipment that estimates location of usage and quality
reports to adopt smarter algorithms such as the machine learning approach
demonstrated. Median locate errors less than 20m are expected for small cell
planning purposes.
• Machine learning models should be part of any small cell design effort.
Different inputs and assumptions will be factors in the resulting models that
are generated.
Report title: Precision Planning for 5G Era Networks with Small Cells
Issue date: 01 October 2019
Version: 230.10.01
Contents
1. Introduction: Why precise location of small cells is
important .....................................................................1
2. Network planning methodology and accuracy
requirements ................................................................3
2.1 Indoor/outdoor differentiation and 3D location .................... 3
2.2 Outdoor planning............................................................. 4
2.3 Indoor planning ............................................................. 14
2.4 Accuracy of network quality data sources for planning ....... 19
3. Impact of siting accuracy on ROI ...............................20
4. Enhancing the accuracy of UE location with machine
learning ......................................................................21
4.1 Types of UE locate ......................................................... 21
4.2 Using machine learning to create ‘enhanced’ timing-based
solutions for improved location accuracy .......................... 22
4.3 Using ML to improve timing-based location techniques ....... 22
4.4 Calculating the relative timing offset between cell sites for
use in location calculations and effective inter-site
synchronization ............................................................. 24
4.5 Machine learning algorithms toward improved accuracy and
results with OTDOA ....................................................... 25
5. Improving small cell placement using machine learning
...................................................................................28
5.1 Factors included in ML model .......................................... 28
5.2 Cost and coverage improvement using ML ........................ 28
6. Acknowledgements ....................................................31
References ............................................................................32
Figures
Figure 1–1 Matching small cell locations to hotspots ........................................1

Figure 2–1 Differentiating between outdoor and indoor payload forecast in 2D ...3
Figure 2–2 Vertical distribution of traffic within buildings .................................4
Figure 2–3 Per building traffic.......................................................................4
Figure 2–4 Los Angeles DVHM (Digital Vector Height Model): submeter vertical
and horizontal digital model of the environment, including elements
that greatly impact the propagation of 5G signals, such as trees and
buildings ....................................................................................5
Figure 2–5 Los Angeles Composite Demand Density Map. Demand is not
uniformly distributed. Hotspots appear as redder areas, each building
Version: 230.10.01
presents a unique demand profile and roads are also characterized by
mobility patterns ........................................................................5
Figure 2–6 Los Angeles RSRP showing macro and small cells (AWS, Tier 1
operator) ...................................................................................6
Figure 2–7 Manhattan RSRP showing macro and small cells (AWS, Tier 1
operator) ...................................................................................7
Figure 2–8 Close up of Manhattan RSRP showing macro and small cells (AWS,
Tier 1 operator). .........................................................................7
Figure 2–9 3D view of Brooklyn DHVM (Digital Vector Height Model): submeter
vertical and horizontal digital model of the environment, including
elements that greatly impact the propagation of 5G signals, such as
trees and buildings......................................................................9
Figure 2–10 Reston, VA RSRP. The interaction of the outdoor signal with the
multiple elements of the environment must be precisely modelled,
specially to understand its impact on indoor coverage at different
heights .................................................................................... 10
Figure 2–11 Automated optimization of 5G cell reach: complex cumulative of
connections at 5G mmWave frequencies (FWA and IAB). Red = lower
latency; Green = higher latency ................................................. 14
Figure 2–12 Indoor capacity coverage map .................................................... 16
Figure 2–13 A 120° sector with seven narrow beams (16° each) ...................... 17
Figure 2–14 Composite coverage of a 120º sector with 35 narrow beams .......... 17
Figure 2–15 Indicative location accuracy for different data sources .................. 19
Figure 3–1 Study of the effect of ROI with distance from the ideal [Source:
Nokia] ..................................................................................... 20
Figure 4–1 Example of GPS accuracy degradation in downtown areas (in this
case Dallas, Tx). Green = over 85% at 40m or better, Yellow /
Orange = 50 to 85% < 40m, Red = less than 50% < 40m ............ 21
Figure 4–2 Dallas metro collection area ....................................................... 22
Figure 4–3 One ‘set’ of locations derived from OTDOA pairs (blue circles), and
the assumed known location (green circle)................................... 23
Figure 4–4 Percentage of time with fewer than N cell pairs producing location
estimates at a given location ...................................................... 24
Figure 4–5 Patented approach toward synchronizing cell sites ........................ 25
Figure 4–6 Location estimation when there are at least 10 cell pairs using simple
averaging, AT&T's machine learning algorithm, and a ‘genie’ (closest
location estimate from the cell pairs to the known location). .......... 26
Figure 4–7 Example of using the improved OTDOA location accuracy in
downtown areas (in this case Dallas, Tx). Green = over 85% at 40m
or better, Yellow / Orange = 50 to 85% < 40m, Red = less than 50%
< 40m. .................................................................................... 27
Figure 5–1 Factors included in ML generation of Sites .................................... 28
Figure 5–2 Manhattan Manual vs. Automated Designs ................................... 29
Figure 5–3 Coverage (RSRP) in Manhattan in Manual vs. Automated Designs ... 29
Figure 5–4 Dominance in Manhattan in Manual vs. Automated Design ............. 30
Version: 230.10.01
1. Introduction: Why precise location of small cells is important
Network planning and optimization is becoming ever more important as small cells,
placed in hotspots, are increasingly used to address capacity needs. But the location of
such small cells is itself critically important. Where small cell placement was off by as
little as half a cell radius from a given hotspot, the result – according to one operator
– was that four times as many small cells were required to carry the same traffic.
Location accuracy is even more challenging when higher frequency bands are used,
reducing the practical serving radius of the cells.
Figure 1–1 illustrates the challenge of locating small cells to align with hotspots of
demand not already served by the existing network, as indicated by the people on the
upper graph. The purple line indicates the potential utilisation of the capacity of a
small cell when placed at a given location.
Figure 1–1 Matching small cell locations to hotspots
A well-placed small cell has its serving area covering locations with high demand. As a
result, most or all of its capacity will be highly utilised. This will represent a good
return on investment for the company deploying the small cell. A poorly located small
cell is too far away from the hotspot; its serving area does not cover an area of high
demand. A further complication is that the efficiency with which any cell site can serve
demand diminishes with distance from the antenna. This is indicated in the diagram by
the fading colour of the serving area. The loftier objective of aligning the most
spectrally efficient parts of the cell serving area to the peaks of demand further
increases the required siting precision.
The size of this serving area is linked closely to the use case of the small cell. In urban
scenarios small cells have to date been used to increase capacity in city squares, in a
series of adjacent bus stops, or around shops, cafés and restaurants.
Getting the location right requires an understanding of the precision required in the
planning process for small cells. This paper considers the accuracy of different parts of
the process. It is arranged as follows:
Version: 230.10.01 1
Section 2 details the network planning process for small cells. This will begin with an
overall understanding of where the network requires additional coverage and capacity.
Depending on the location, the network may best be enhanced through either outdoor
densification or indoor deployment. However, the considerations involved in each case
will be quite different – as described in sections 2.2 and 2.3.
Section 3 shares the results of one study into the impact of siting accuracy on return
on investment (ROI), and recommends best practice accuracy.
A key input to the planning process is an understanding of how well the existing
network is serving demand – in effect the purple residual traffic demand curve in
Figure 1–1. This understanding can be built up from information on geolocated traffic
usage and from quality reports delivered by devices in the network. Such reports are
available from probing network interfaces or from data collected from social media,
mixed with network statistics. The accuracy of the associated locates determines the
accuracy of the demand curve – and therefore the operator’s understanding of where
hotspots really are. Section 4 considers this location accuracy – and, in particular, the
use of machine learning to enhance the location accuracy of traffic usage and quality
reports used in the planning process.
Even with highly accurate planning to identify the optimal location for a small cell,
there are real-world limitations which mean they cannot be used in practice.
Candidate sites for small cells – such as street furniture or indoor mounting – may not
be available in ideal locations. Furthermore, traffic demand distributions vary daily,
weekly and over years, so hotspots may move over time. Although optimisation
through RAN configuration changes can help compensate for suboptimal locations, it
cannot achieve the same performance as using the optimal sites in the first place.
These challenges are discussed further in our paper [SCF195] [1] Small cell siting
challenges and recommendations.
Version: 230.10.01 2
2. Network planning methodology and accuracy requirements
2.1 Indoor/outdoor differentiation and 3D location
A detailed description of the planning of heterogenous networks – networks that

combine small and macro cells – can be found in an accompanying paper [SCF174]
[2] Capacity planning for HetNets. In brief, however, planning starts with an overall
understanding of where the network requires additional coverage and capacity.
Depending on the location, the network may best be enhanced through either outdoor
densification or indoor deployment. Figure 2–1 illustrates different views of network
traffic differentiated into street level and individual buildings.
Figure 2–1 Differentiating between outdoor and indoor payload forecast in 2D
Although 2D is the norm for visualisation of UE location or carried payload,

enhancement into 3D can reveal important insights which may help the operator to
better identify effective small cell placement strategies. For instance, 2D mapping
does not identify traffic variations on different floors of a building. This may lead to an
expectation that multiple street-level small cells may be able to offload the majority of
the indoor traffic. However, this is unlikely due to the physical nature of the overall
site solution, including the antenna radiation pattern used, and could result in small
cells being placed at street level, 5-10m from the ground with a directional antenna.
In these deployments the small cell would not be able to serve traffic on the higher
floors of the building.
A 2D-only view could therefore lead to an excessive number of street-level outdoor

small cells. These would then need to be pruned back after the initial planning is
completed, as part of a design review process.
The use of 3D location and visualisation, by contrast, provides an understanding of the

distribution of traffic within the buildings, as shown in Figure 2–2. This helps
operators decide whether indoor demand could be served using an outdoor-indoor
approach, or whether an indoor deployment is needed.
Version: 230.10.01 3
Figure 2–2 Vertical distribution of traffic within buildings
Supplementary information can be summarised, such as the traffic per building, as

shown in Figure 2–3. This would allow initial business cases to be created, or target
solutions to small groups of floors, rather than the whole building.
Figure 2–3 Per building traffic
Indoor and outdoor planning have different design considerations, which are detailed
in the following sections.
2.2 Outdoor planning
2.2.1 Why does the entire system need to be examined holistically?
If an operator doesn’t build the best possible network – one that not only meets
demand but maximises income and minimises capex and opex – it won’t get the best
ROI. To ensure that it does, certain key elements have to be modelled precisely.
These elements are environment (Figure 2–4), signal (Figure 2–5) and demand
(Figure 2–6). This modelling process is a holistic one, taking data from a wide range of
sources often gathered from a wide area.
Version: 230.10.01 4
Figure 2–4 Los Angeles DVHM (Digital Vector Height Model): submeter vertical and horizontal
digital model of the environment, including elements that greatly impact the
propagation of 5G signals, such as trees and buildings
Figure 2–5 Los Angeles Composite Demand Density Map. Demand is not uniformly
distributed. Hotspots appear as redder areas, each building presents a unique
demand profile and roads are also characterized by mobility patterns
Version: 230.10.01 5
Figure 2–6 Los Angeles RSRP showing macro and small cells (AWS, Tier 1 operator)
This was not always the case. People once designed networks using teams – a
coverage team, a capacity team, and a backhaul team, say. Each would work
independently and feed some information to the other teams. The data input required
was modest, as were the networks, which had a small number of base stations mainly
used to provide voice communications.
Modern-day wireless networks are used not just for the tasks carried out by landline
telephony but those offered by fixed broadband computing. They are influenced by a
large number of independent factors and therefore have to be planned holistically,
using multiple data inputs. Given the vast amount of inputs, the process needs to be
automated – and it can be, using network design software tools.
The new normal
As 5G approaches, HetNets and densification are becoming the new normal for
network rollout. A HetNet auto design process is therefore required to reduce planning
time. The design software that enables this will, of necessity, learn and adapt in a
machine learning process. But it needs something to learn and adapt from. It will need
to include many inputs to inform the decisions that will be made – decisions on sector
count, azimuth, fibre type and siting, street furniture, fronthaul, backhaul (wired or
wireless) and dozens of other considerations that will be part of a design workflow.
But this only begins to illustrate the complexities that are faced when planning a
modern communications network. Obviously since the advent of small cells and the
introduction of a growing number of frequencies (with more – such as mmWave – to
come), the challenges of multiple bands, spectrum and cell placement and
interference also have to be taken into account. IoT, Wi-Fi, LoRa, WiGig and numerous
other protocols, connectivity enablers and access technologies will soon be part of
many operators’ capabilities or requirements.
The logic of HetNets is that they will be made up of many cells of many different sizes
and strengths. Interference will therefore be a constant concern. Even if a macrocell is
Version: 230.10.01 6
many kilometres away, it might cause interference to small cells in the top floor of a
building if it is positioned on a hill or a mountain offering no interruption to signal. At
the same time buildings can block signals, affecting outdoor planning. This means that
scale is also important: the area under consideration could be vast if the relevant data
inputs are dispersed over a large area (Figure 2–7and Figure 2–8).
Figure 2–7 Manhattan RSRP showing macro and small cells (AWS, Tier 1 operator)
Figure 2–8 Close up of Manhattan RSRP showing macro and small cells (AWS, Tier 1
operator).
But a network is only as good as the efficiency of its service to end users. Thus a
planning tool will need to answer many questions – questions like: Where is the
Version: 230.10.01 7
underserved demand? Where is the capacity? How efficient is network performance
during busy hour? Can it be improved? What is the preferred cell radius? Where are
the street poles needed for small cells and how should they be spaced? What is the
signal to noise ratio (SNR)? How will SNR affect cell spectral efficiency – a critical part
of modern network design? Can you extrapolate and predict future demand and
densification roadmaps – and, if so, how far ahead?
There are multiple dimensions that have to be evaluated – and all of these different
dimensions have to go into the planning process as early as possible to guarantee
accuracy and shorten the time to deployment. At the same time, new data will need to
be added over time, although this will of course help to reinforce the model and
benefit planning.
Inputs everywhere
Luckily many data inputs are widely available that can offer useful information. In the
US for example it is possible to connect to the FCC spectrum database to get a
complete picture of US spectrum ownership.
But many other inputs, not directly cell-or signal-specific, can be used to assess user
activity and movement in particular, which will influence cell siting. Most countries
publish the results of their national census audit, offering residential and employment
information along with growth estimates. Traffic estimates, GPS tracks and other data
can help to establish vehicle density at different times of the day. Similarly GPS
tracks, the mobility of social events, and a precise model of the environment can aid
in calculating the mobility and indoor/outdoor status of user distribution.
Social media events can help too. For example, amenities such as shops and
restaurants often involve the most concentrated amount of demand on a mobile
network, and are the locations where the population distribution reaches its highest
density. It is possible to estimate this via the quantity of check-ins found in social
networks such as Facebook and foursquare.
Multiple data sources are useful but they are there to serve a purpose: assessing
efficiency, coverage and capacity, which in turn need to be balanced against ROI.
When an operator decides whether or not to place a site in a given position it has to
balance the cost of putting in a cell and backhaul against demand and likely usage.
What is the time to ROI? When does investment become profit? Network modelling is
about revenue as well as efficiency.
All these factors – and many more – must be considered during the business case
evaluation, since they both inform and define strategy. And that is why it is important
to understand the holistic nature of network planning.
2.2.2 How accurate does the computer model needed to mix powerful
macros with less powerful picos and Wi-Fi APs etc have to be?
The localised signal from a small cell inside a building can be overwhelmed by signals
from more powerful external devices. It is, therefore, important for good indoor
planning to ensure that the outdoor planning isn’t an obstacle. Not just proximity of
macros to small cells but height will be part of this planning. Small cells at lower levels
of a building are less likely to be affected by macro cells. But if a macro itself is at a
great height (on a hill or mountain, say) its effect can be felt tens of kilometres away.
Version: 230.10.01 8
Of course, as small cells become part of the outdoor HetNet, interference in the
outdoor network is also going to be more of a headache for planners. How then does
planning software mitigate the threat of interference?
It will start off with the entire area being analysed. That could be a city of, say, 50 km
across. Next it will create a computer model of the entire area under consideration:
terrain, trees, roads, railroads, and buildings, for example. Buildings in particular are
critical because they can block the signal. This modelling has to be carried out very
precisely: the current state of the art is to model to an accuracy of one metre
vertically and horizontally (Figure 2–9).
Figure 2–9 3D view of Brooklyn DHVM (Digital Vector Height Model): submeter vertical and
horizontal digital model of the environment, including elements that greatly
impact the propagation of 5G signals, such as trees and buildings
Network planning software should provide operators with different options for small
cell placement or changes to the macro network. For example, they could downpower
a macro, say, or reorient it – or reduce the signal in a building. For in-building
coverage, the aim of planning software is to allow you to use a full indoor design – but
it can also help you to minimise cost. Can you, for example, really afford small cells on
each floor and then fibre to each one of them? Could there in fact be minimal use on
specific floors, requiring minimal rollout?
You might then tell the planning tool just to optimise the lower floors and make sure
there’s good coverage from a macro on the lower floors and less interference on the
upper floors. A balance can be struck (Figure 2–10).
Version: 230.10.01 9
Figure 2–10 Reston, VA RSRP. The interaction of the outdoor signal with the multiple elements
of the environment must be precisely modelled, specially to understand its impact
on indoor coverage at different heights
Remember too that small cell planning indoor and outdoor increases the need for
accuracy of signal calculation since errors injected into the signal calculations intensify
when computing the signal-to-interference ratio. If errors are present in the computer
model (such as using very coarse clutter and terrain) this could result in an inability to
predict the signal. And this brings us back to ROI: failure to predict the real return
from a new deployment destroys the value of the planning process.
AI and machine learning will, necessarily, play a part in cell placement. Metaheuristic
algorithms, of the sort used in today’s planning tools, aim to minimise the difference
between input measurement data and predicted data. As happens with machine
learning algorithms, input from previous analyses will help the planning tool decide
what its next move is. For example was a site rotated in the past? What was learned?
Can that experience influence the next situation where site rotation might be an
option?
2.2.3 User activity is hyper-sensitive to network performance. How?

Why?
Correct cell placement is not just about filling in coverage holes. In areas of very high
demand operators will, of course, need more cells – but the signals from some cells
may interfere with the signals from others, and this will influence the throughput users
can get. Therefore an accurate estimation of interference is needed. That requires an
accurate model of how a signal propagates – and, as noted before, that can only be
achieved with a precise model of the environment.
In particular, the signal to noise ratio must also be considered. Strong signal to noise
ratios due to bad placement, interference or other reasons will not just upset end
users. They will require a robust modulation scheme and affect cell spectral efficiency.
Bad cell performance will, in turn, mean less capacity and higher costs. Consumer
distribution greatly affects cell spectral efficiency. A planning tool that mitigates this
Version: 230.10.01 10
will focus on accuracy in signal prediction and consumer distribution as an important
part of the planning process.
The accuracy of the cell spectral efficiency calculation is directly relevant since it
allows the network planning tool to choose the optimal location for new RF
deployments. It is also important therefore to have an accurate model of the user
distribution.
2.2.4 How are outdoor zones of underserved demand identified?
Where are the users? Operators can supply estimates of where they are and how
much demand they are generating of course, but many other supporting sources are
available and useful – social networks for example. Social events from Twitter and
Instagram correlate with actual footfall gathered electronically. By far the most useful
data for estimating population distribution in time are the social networks that allow a
public stream that includes the location and time of an event.
Combining this with a vector land classification (clutter) model means that when
events occur, they can be classified with greater accuracy. For example, did a tweet
occur within a building or a park, or on a ferry?
Twitter features a high volume of such events. Similarly, Instagram and Flickr include
the location of photo uploads and can include EXIF information that allows estimation
of the height of an event.
The more accurate the model the more granular the information. The height of a
building, together with demand on a building-by-building basis, can allow many
different analytics to be performed on this information. Entire cityscapes can be
analysed and the buildings that require the most capacity isolated.
Add the information mentioned earlier – censuses, jobs, traffic, terrain, social media,
and even crime data – and it is possible to build a statistically significant indication of
the population distribution – and thus, combined with operator inputs, actual and
predicted demand. Planning software could even provide a growth model to be used
for a densification roadmap, although the further forward this goes the greater the
margin of error. Nevertheless, a view three years ahead is certainly possible, and with
it a good estimate of when investment becomes profit.
2.2.5 What is the effect of optimal network designs that

simultaneously minimise cost while maximising capacity?
This is a simple question to answer, but making optimal network design actually
happen is difficult. The goal is to place sites with high spectral efficiency in high-
demand areas. However, the set of things you, the operator, can choose from is
massive. You can choose bands. You can choose technologies: Wi-FI, say, may be an
option in places. Do you use wired or wireless backhaul? Are street poles or even
(very cheap) strand-mounting available? Is the spacing of cells precisely modelled –
not just to meet demand but to keep costs down?
Everything is associated with the cost. An operator is trying to maximise spectral

efficiency to get the most capacity but also to maximise the return on investment by
minimising capex and opex.
If you get all that right, then optimal network designs that simultaneously minimise
cost while maximising capacity will, in theory, positively affect equipment purchase,
Version: 230.10.01 11
equipment efficiency, medium and even long-term-planning – and, above all,
customer retention, now and over time.
2.2.6 How do spectrum and multiple bands play a role in controlling

capacity and coverage?
Operators can already choose from a number of bands and often an operator’s needs
will dictate their role. In some areas an operator may be better off using coverage
spectrum like 700MHz and not considering small cells. In other places, where small
cells are needed for capacity reasons, different spectrum will be necessary. There may
also be a trade-off. Eventually mmWave rollout will give operators greater bandwidth
but will make other demands – stricter line of sight requirements for example.
The quantity of spectrum also matters. If the quantity of spectrum is 20MHz and the
cell spectral efficiency is one bit per second per hertz that means 20 megabits per
second. That’s the total capacity of a well-performing site or sector on a site. Multiply
that by the number of cells and you get the true capacity of the network.
But even with adequate spectrum, spectral efficiency and interference matter,
especially when considering ROI. Thus spectrum and multiple bands are useful – but
only if their use is planned correctly.
2.2.7 How do you plan for multi-RAT: mixing different technologies for
different objectives (e.g. Wi-Fi, IoT)?
Everything must be taken into account in the planning process. The whole air interface
is a mixture of different frequencies from lots of different devices, some of them
interfering with one another, some of them isolated and not interfering. But it all has
to be taken into account – because when an operator rolls out its business it needs to
look at every single aspect of it. Again, this is a holistic process.
For example, planning for different objectives can be guided by demand models. High
mobility, low mobility, indoor demand, outdoor demand, type of demand (voice, data
or both), big demand in, say, stadiums or parks, how they are covered and the cost –
all these considerations can inform technology deployment and whether an operator
uses, for example, Wi-Fi, LTE, 5G, 700MHz, 1900MHz, 2500 MHz or 28GHz. If an
operator is building a modern network – as EE (UK) or Verizon (USA) are – it probably
has to consider all bands. All of this has to be predicted within a planning tool –
another aspect of the holistic examination required for accurate network planning.
Relatively new technologies like IoT also have to be considered. IoT today is thought
of as small amounts of information from smart metres, temperature gauges or sensors
on factory equipment, but how will IoT evolve in use and network demand? Will
information be aggregated through gateways? What does this mean for network
design?
And then there’s backhaul. If you’re ten metres from the fibre it’s relatively
inexpensive. If not, what are your options? Satellite or microwave backhaul are
possible, but with 5GNR operators will be able to backhaul wirelessly using WiGIG –
Wi-Fi at 60GHz. It’s unlicensed and therefore quite cheap. How should an operator
plan for that?
Then there’s mmWave – game-changing for fixed wireless and eventually other forms
of wireless: integrated access and backhaul and therefore more bandwidth. The
Version: 230.10.01 12
downside is that it is a short-range line-of-sight-only technology that can be affected
not just by objects but even by rain.
Massive MIMO will offer its own challenges. But all these technologies and frequencies
have to be part of the modelling process.
Finally, of course, the technology mix will also be affected by new or growing user
profiles – travellers by road, rail and air, say. And what will be the requirements of
autonomous vehicles?
2.2.8 How do design and rollout strategies affect ROI?
Once network design software has given an operator all the network modelling
information needed for maximum capacity – and therefore maximum ROI for its opex
– the operator can choose from a number of business plans knowing what (and when)
the ROI is likely to be with only a small possible margin of error. An operator can do
this with the confidence that, assuming the network design software has done its job,
the model is as deployable as it can be because most eventualities that it is possible to
predict have been included in the modelling. The operator knows that, even if one
business model differs slightly from another, it’s going to get a return on this
particular network design.
2.2.9 Why do 5G, and in particular mmWave, change traditional

planning?
5G will bring a number of planning advantages – and challenges. Most notably it will
mean more careful planning – not just because of densification, but because where
everything is on the same band and the small cell is on the same frequency as other
cells, small cells are more likely to experience interference from the macro.
Although it is a part of 5G, mmWave is important in its own right because it gives
operators bandwidth – and when it’s all automated in 3GPP release 16 that means
integrated access and backhaul. That integrated access and backhaul is what changes
everything. mmWave allows operators to create antenna arrays that are the size of
playing cards that can go into a box the size of a paperback book at what is
considered a very low cost and, potentially, access to huge quantities of bandwidth
and 64 users at a given time. The main short-term application for mmWave is fixed
wireless access but that can, and probably will, expand into mobile applications
(Figure 2–11).
Version: 230.10.01 13
Figure 2–11 Automated optimization of 5G cell reach: complex cumulative of connections at
5G mmWave frequencies (FWA and IAB). Red = lower latency; Green = higher
latency
Most importantly of all perhaps, 5G requires not just optimised HetNets but
densification roadmaps, with all the added complexity that implies for cell placement.
2.2.10 Summary of outdoor planning
The first mobile network in London offered had about 18 sites. We can expect city-
wide networks of 2,000 or more sites with multiple bands in the not-too-distant
future, and that number will expand. 5G, in particular, will have an astonishing effect
on demand and capacity but will make ever-more-complex demands on planning. But
planning too will evolve. Planning tools will be soon able to make use of data derived
from drones, context-aware equipment and self-deploying solutions, to name only
three.
Whatever the future of planning brings, however, the underlying opportunity will be
the same: place your cells where the users are – and avoid interference – and you
maximise return on investment.
2.3 Indoor planning
Indoor networks are deployed in a variety of venues that have very few things in
common. Public venues, such as large stadiums, convention centres, airports and
subway stations, have favourable, mostly line-of-sight propagation conditions between
access points and UEs. Enterprise venues – such as hospitals, hotels or corporate
headquarters – will put many obstacles in the propagation path, resulting in mostly
non-line of sight (NLOS) propagation, with reflection and diffraction off nearby
surfaces. Some venues, like shopping malls, have a roughly equal amount of LOS and
NLOS.
The amount of LOS vs NLOS is important if an indoor network is designed for

coverage. A design for coverage specifies a certain signal level over a certain
Version: 230.10.01 14
percentage of a floorplan. For example, a typical coverage design would require an
LTE RSRP (reference signal received power) signal level of -95 dBm over 95% of the
floorplan for a 10 MHz LTE channel. The exact target signal level is a function of
transmit reference signal power and the channel bandwidth. The percentage of the
floorplan that needs to be covered is usually between 90 and 95%; 100% coverage is
unrealistic because it would require too many access points. For coverage then,
venues with mostly NLOS propagation need more access points per unit area than
venues with mostly LOS. For example, a hospital floorplan would require more
antennas than a garage floorplan.
When it comes to designing an indoor network for capacity, it is important to know the
number of UEs and types of traffic they will demand. There are two limits at play here:
a hardware (Layer 2) limit and a physical layer (Layer 1) limit.
In LTE, a hardware (Layer 2) limit for small cells may be up to 32 simultaneous

connections (as an example). However, with the advent of C-RAN and O-RAN
distributed networks, the baseband units, which are capacity sources, are separated
from the amplifier and can be shared with many small cells, thus diminishing the
importance of a hardware limit per individual small cell. On the other hand, physical
layer limitation per cell is still important. Layer 1 capacity calculation implies
calculating the amount of time needed to transmit file(s) to and from UEs during the
peak, or busy hour, usage and comparing it with overall time during the busy hour. If
too many UEs are connected and they all request large file transfers, the needed time
to complete the transmission may be too long for an acceptable user experience. In
areas where we expect heavy data use (such as 4K video files or gaming), more small
cells are needed than in areas where light data use (such as wireless printing or file
exchange) is expected.
A well-planned indoor network designed for capacity should identify areas where users
tend to congregate (for example a lunch cafeteria at the corporate office, hotel bars,
nightclubs or convention conference rooms), the type of traffic expected (streaming
video download, Facebook video upload, email exchange or Twitter posting, say) and
approximate file size download/upload per user during the peak hour. An indoor
network design also needs a breakdown of the percentage of anticipated traffic
between the existing Wi-Fi at the venue and the cellular traffic. The cellular traffic
should also be broken down into MNOs according to their user penetration rate. In
addition, within each MNO, the traffic needs to be broken down between technologies
(3G, 4G and 5G).
When all this is addressed, we can finally place small cells on the floorplan. Care
should be taken to position them near or at areas of heavy traffic, but also near or at
VIP areas, such as C-level suites or penthouse apartments. We should also remember
to cover other areas where traffic is not heavy in order to provide continuous coverage
throughout the venue. Finally, capacity analysis should calculate the percentage of
airtime usage for each small cell/node. The airtime usage can also be interpreted as
‘cell load’. A cell load of 80% means that, on average during peak use, a small cell is
busy transmitting and receiving 80% of the time. A well-designed system should be
designed for 50-80% loading, which represents good utilisation and also leaves room
for growth before an upgrade is needed.
An example of a capacity coverage map is shown in the figure below. Here we

assumed the maximum number of connections per node to be 32, and the target cell
load to be 80%. Areas in dark and light red indicate nodes that experience hardware
failure (more than 32 connections), and physical layer failure (cell load greater than
80%). Nodes that passed are coloured green.
Version: 230.10.01 15
Figure 2–12 Indoor capacity coverage map
Another key performance indicator is the maximum achievable data rate (MADR)
coverage. The MADR is a function of signal to noise and interference ratio (SINR),
which calculates the ratio of the serving cell signal to interference plus noise coming
from non-serving small cells. In LTE networks, which mostly deploy omnidirectional
and wide beam (60-90 degree) directional antennas, interference coming from non-
serving small cells can be significant, which degrades SINR near the cell edge.
Consequently, MADR is also low at the cell edge. With the advent of beamforming
antennas, which will be widely deployed in 5G, the interference will be much reduced,
because narrow beams mean that the target area is narrow, which reduces the
probability that the serving and the interfering signal may arrive at the UE at the same
time.
Here is an example that illustrates this point. Let’s assume that we have an 8x8x2
small cell panel antenna that creates individual 16º narrow beams. To cover a 120º
sector, seven narrow beams are needed:
Version: 230.10.01 16
Figure 2–13 A 120° sector with seven narrow beams (16° each)
These seven beams cover a 120º horizontal sector, but only a 16º vertical sector. If
we want to cover a ±40º vertical sector, then we need five rows of beams, which
means that we need 7x5=35 beams. The composite panel coverage with 35 beams is
shown below:
Figure 2–14 Composite coverage of a 120º sector with 35 narrow beams
If we assume that the antenna array is analogue – which is a fair assumption given
the current state of the art in millimetre wave small cells – then only one beam is
transmitted at a time. If only one beam can be detected as interference at UE (again a
fair assumption given that the beams are narrow), then interference can be detected
only 1/35th of the time. This leads us to the conclusion that interference is reduced by
a factor of 10log(1/35) = -15.4 dB with respect to a full-sector 120-degree antenna.
We can therefore see a significant interference reduction with beamforming.
Two more KPIs – signal strength and MIMO rank – should also be mentioned. For
coverage planning, the main KPI is RSRP for LTE and SS-RSRP (synchronization signal
reference signal received power) for 5GNRs. A MIMO rank indicator shows MIMO
performance throughout the floorplan. For 2x2 MIMO, MIMO rank can be either 1 or 2.
For 4x4 MIMO, the MIMO rank range is 1 through 4. The higher the rank, the better
the MIMO throughput.
Version: 230.10.01 17
Lastly, a 5G network should be designed with backhaul throughput and delay in mind.
A backhaul throughput is a function of capacity and predicted cell load. It tells us how
much throughput is needed to provide connectivity to the outside network. An indoor
network delay is delay caused by active elements and fibre and should be calculated
from small cell to the main closet/MDF. This is important for edge computing, as many
5G applications are sensitive to delay.
To summarize, the important KPIs for indoor network design are
• Capacity and coverage,

• SS-RSRP
• MADR
• SINR
• MIMO rank
• Backhaul throughput requirements
• Indoor network delay
When speaking of millimetre wave indoor propagation, high-frequency signal range is

rather poor, even in LOS conditions, while most indoor materials create a large
penetration loss at that frequency band. Thus, indoor millimetre wave 5G applications
will be limited to opportunistic deployment, rather than a building-wide coverage.
Typical use cases for indoor millimetre wave would be hotspots, coffee shops,
stadiums, convention centres and ballrooms.
Another factor to consider is the impact of the indoor network on the outdoor network.
In most cases, this coverage creep should be kept to a minimum. Many operators
specify the maximum allowed indoor signal level at a certain distance (usually 100
feet) from a building. This limit is not difficult to achieve in newer buildings with solar
glass, which has a high penetration loss for both sunlight and radio waves. However,
buildings with older types of glass, which have far less penetration loss, require careful
network design. A common tactic is to use directional antennas near windows,
pointing inside. Directional antennas with a high front-to-back ratio should be used, to
further decrease the amount of radiation through the back lobes.
However, it should be pointed out that some indoor networks require expansion of the
network outside. Specifically, stadiums experience heavy traffic before and after an
event – at streets leading to a stadium, bus stops, train stations, and parking lots for
example. While in most countries this traffic builds up and subsides 30-45 minutes
before the event, in North America the build-up of traffic can last up to two to three
hours before a sports event, notably when people congregate at a parking lot and stay
there for a while (so-called ‘tailgating’). Therefore, it is a common practice in that part
of the world to extend indoor networks to parking lots surrounding large stadiums, to
address the capacity needs before the game.
Version: 230.10.01 18
2.4 Accuracy of network quality data sources for planning
Figure 2–15 Indicative location accuracy for different data sources
The accuracy of data sources available for planning is illustrated in Figure 2–15. Each
of the data types has its own benefits and drawbacks; hence a pragmatic approach is
required to realise the most appropriate solution. For instance, there is a wealth of cell
statistics, but it cannot be resolved below the serving area of the cell. With
minimisation of drive test data (MDT), the GPS location is reported for a subset of
Android phones including an accuracy flag if the phone considers it to be a ‘stable
location’. 2D and 3D geolocation using network probes has a relatively large variance
of accuracy. Enhancing the accuracy of this geolocation is the subject of section 4 of
this paper. Social media events are often tagged with GPS location. This is fine for
rural or suburban networks but does start to ‘bounce around’ in street canyons of
dense urban areas.
Version: 230.10.01 19
3. Impact of siting accuracy on ROI
Error! Reference source not found.Figure 3–1 shows the ROI (return on
investment) of several different cost profiles of small cell deployment (plotted on the Y
axis) against the distance from the ideal location (on the X axis), as given by a
planning tool. There are 1500 possible candidate sites and 160 ideal locations chosen
to meet the capacity and coverage criteria in this study. The ROI is based on the
principle of income generated by carrying a GB of data – a value which is publicly
estimated in many sources – and the cost of deployment of various site solutions. The
technology deployed in this example is an LTE 5+5w small cell with a relatively small
bandwidth.
Where: ROI = profit / TCO over a five-year timeframe
Figure 3–1 Study of the effect of ROI with distance from the ideal [Source: Nokia]
The ROI is positive for all site solutions as long as the best location is secured during
the build program. If the preferred site is not available, then there is some leeway
before the ROI tends towards zero, but this depends on the cost profile of the site.
However, to maximise the true value of the small cell the accuracy should be within
20-40m of the ideal. This is because there many small cells where the roll-off in ROI is
much steeper than the average shown here and a well-constructed planning process
would want to capture all possible traffic, not just the average case. Another
consideration is that, since this study was completed, the traffic-carrying capability of
small cells has increased significantly. This will help to improve ROI, even with a
falling retail price of a GByte of traffic. However, in all cases the extra cost of diligent
location planning easily pays back.
Version: 230.10.01 20
4. Enhancing the accuracy of UE location with machine learning
Given the importance of obtaining accurate location for small cell planning, we have to
understand the accuracy of the input data. While some mobiles (e.g. Android) may
report accurate location in suburban or rural areas using assisted GPS (A-GPS), this
method does not work well in urban areas. Other timing-based techniques like
standard observed time difference of arrival (OTDOA) are also, typically, not accurate
enough. It is, however, possible to apply a machine learning algorithm approach to
improve the accuracy of OTDOA and other timing-based methods. The results are very
good, achieving typical median location accuracies of about 18-25m in
suburban/urban areas.
4.1 Types of UE locate
There have been several network-based approaches to location over the years, using
signal levels or timing triangulation between cell sites [e.g. Gabe] that could obtain
around 50m accuracy in dense urban areas. However, these approaches are more
likely to have location accuracy of 200m+ in suburban areas and are not sufficient for
most small cell placement strategies in these areas. In suburban and rural areas
handset-based location solutions using A-GPS work well, with accuracies of around 10-
15m. However, neither A-GPS nor other solutions have reliable location accuracies of
much less than 50m in urban areas, since A-GPS from mobile devices is not able to
‘see’ enough satellites to get an accurate location.
Figure 4–1 Example of GPS accuracy degradation in downtown areas (in this case Dallas, Tx).
Green = over 85% at 40m or better, Yellow / Orange = 50 to 85% < 40m, Red =
less than 50% < 40m
One method to potentially overcome the lack of A-GPS in urban areas is to rely on
handset-measured ‘timing measurements’ between cell sites – for example with
Version: 230.10.01 21
OTDOA measurements reported by the mobile stations on request. OTDOA is a
standardized solution supported by nearly all mobile handsets today for emergency-
class location requests (e.g. 911).
4.2 Using machine learning to create ‘enhanced’ timing-based solutions

for improved location accuracy
Standard OTDOA measurements reported have marginal accuracy (more than 60-70m
median accuracy) in multiple environments (urban, suburban, rural). However, we can
show that, with machine learning, accuracy can be significantly improved, yielding
median accuracies about 18-25m across urban, suburban and rural areas.
4.3 Using ML to improve timing-based location techniques
In order to determine the potential accuracy of any solution, measurements from a

large number of handsets must be analysed. For the study used here, the Dallas
metro area was chosen, and several thousand ‘sets’ of OTDOA-derived location
measurements were collected, where a ‘set’ consists of OTDOA location measurements
from multiple pairs of cell IDs.
Figure 4–2 Dallas metro collection area
In addition, for each set of OTDOA location measurements from cell ID pairs, there is
a measured GPS measurement as an assumed known location, as shown by the solid
green circle in Figure 4–3.
Version: 230.10.01 22
Figure 4–3 One ‘set’ of locations derived from OTDOA pairs (blue circles), and the assumed
known location (green circle).
As can be seen, there are a number of ‘location options’ (blue circles), one for each
cell pair. The goal is to find the best estimated location as close to the known location
(green circle), from the original OTDOA measurements (which produced these locate
estimates), and/or by using the original OTDOA measurements and the location
options. One would assume that the more location estimates from OTDOA pairs (blue
circles), the better one’s estimate of the true location. The graph below illustrates how
often fewer than ‘N’ cell pairs (or location estimates) were seen.
Version: 230.10.01 23
Figure 4–4 Percentage of time with fewer than N cell pairs producing location estimates at a
given location
4.4 Calculating the relative timing offset between cell sites for use in
location calculations and effective inter-site synchronization
Before attempting to use OTDOA-reported measurements from cell site pairs for
location estimation, one must first take into account the relative timing offset between
cell site pairs. A model can be created which shows – if one knows the location of a
mobile (from A-GPS, say) – how one can calculate the cell site pair timing offset. This
offset can then be used in subsequent OTDOA location estimations for devices where
no A-GPS is known.
Version: 230.10.01 24
Computed downlink frame timing
difference is determined from site
locations and mobile location
Site B latB/longB
Site A latA/longA
Mobile GPS
latm/longm
The difference between these

Mobile device measures actual
differences is the frame timing
downlink frame timing difference
offset, TB_offset.
Figure 4–5 Patented approach toward synchronizing cell sites
4.5 Machine learning algorithms toward improved accuracy and results

with OTDOA
For any given instant that we collect OTDOA measurements, after, or in conjunction
with, applying the inter-site timing offsets as described in the previous section, we can
apply a variety of machine learning algorithms to estimate the actual location. The
most simplistic algorithm may be to just ‘average’ all the individual location estimates
from all of the cell pairs reporting at a given time. More sophisticated methods may
work on both the OTDOA measurements themselves and the potential location set.
The figure below illustrates location accuracies over the entire test area in Dallas using
simple averaging (green line), and AT&T’s patent-pending machine learning approach
(red line), as compared to a ‘genie location estimate’ (black line), which picks the
single closest location estimate from a given cell pair to the known location. In all
cases, there is no change to the mobile stations and all computations are done in the
cloud.
Version: 230.10.01 25
Figure 4–6 Location estimation when there are at least 10 cell pairs using simple averaging,
AT&T's machine learning algorithm, and a ‘genie’ (closest location estimate from
the cell pairs to the known location).
In this case, instances were chosen where at least 10 cell pairs were measured from
the handset. Notice that the median (50th percentile) accuracy of the ML algorithm
(red curve) is approximately 18m, with an 80th, and 90th percentile around 45m and
80m respectively. Additionally, the experience in comparing the accuracy of the actual
reported OTDOA location from the standard OTDOA system was actually a bit worse
than the ‘simple averaging’ solution shown in the figure above. Consequently, the
accuracy of the ML approach at least improves on ‘simple averaging’.
Taking these results into account and looking again at the downtown Dallas area, but
with enhanced OTDOA, gives us the outcome shown below.
Version: 230.10.01 26
Figure 4–7 Example of using the improved OTDOA location accuracy in downtown areas (in
this case Dallas, Tx). Green = over 85% at 40m or better, Yellow / Orange = 50
to 85% < 40m, Red = less than 50% < 40m.
As can be seen, the location accuracy improvement (Fig. 8 vs. Fig. 1) is dramatic, as it
shows good location accuracy in downtown urban areas now, where A-GPS struggled.
Version: 230.10.01 27
5. Improving small cell placement using machine learning
Given the factors of cost and technology as inputs to small cell placement, planning
and execution of networks becomes complex. Manual planning of build-out plans can
become cumbersome, even to the most knowledgeable and experienced engineers.
Machine learning provides the opportunity to compare multiple scenarios with different
factors to maximize coverage and minimize cost. AT&T has been piloting an effort to
use ML for small cell placement.
5.1 Factors included in ML model
Our models start with the area to be covered (a geographic polygon), the traffic-
weighted location data (ML enhanced per previous section), and the budgetary
constraints associated with that area. Data is added to the model, including terrain
and structures in the area, and the associated clutter classes and heights.
Assumptions are made regarding the antennas (omni-directional, brand, etc.)
Candidates are generated and classified and factored in the following priority order:
1. Existing LTE assets

2. Locations where we have agreement with site owners
3. Intersections identified by open street maps
4. Uniform hexagonal grid
The model uses the goals of coverage and dominance. An outdoor location is
considered covered if the strongest signal at that location has an RSRP greater than -
112 dBm. A location has satisfactory dominance if there is at least 5 dB of separation
between the strongest and third strongest signal at that location. The figure below
shows the factors included in our model
Figure 5–1 Factors included in ML generation of Sites
5.2 Cost and coverage improvement using ML
With the model established and proper data identified and loaded, build-out plans can
be generated. For this example, we used a section of Manhattan. The manual design
identified 185 sites in the polygon that would need to be placed. The ML generated
design created 111 sites. The figure below shows the differences between the two
sites
Version: 230.10.01 28
Figure 5–2 Manhattan Manual vs. Automated Designs
The automated design is able to provide coverage and dominance while reducing the
number of sites required. The reduction of 74 sites provides a savings of 40% - but
the model has also created optimized coverage for this lower cost. The following
figures demonstrate the maintenance of coverage and improvement of dominance in
the manual vs. automated plans.
Figure 5–3 Coverage (RSRP) in Manhattan in Manual vs. Automated Designs
Version: 230.10.01 29
Figure 5–4 Dominance in Manhattan in Manual vs. Automated Design
Small cell design is an emerging application of ML – this example is one of the first
design uses within AT&T. There are significant benefits to using this technology, both
through reduced costs and improved customer experience. These models will
continue to evolve as our assumptions, data, and other inputs continue to improve.
Version: 230.10.01 30
6. Acknowledgements
This work is agreed by and represents the view of all members of the Small Cell
Forum, but we would particularly like to thank the following companies that
contributed content for this paper: AT&T, iBwave, Nokia and Keima.
Version: 230.10.01 31
References
1 [SCF195] ‘Small cell siting challenges and recommendations’, Small Cell Forum
2 [SCF174] ‘Capacity planning for HetNets’, Small Cell Forum
Version: 230.10.01 32

SCF Precision Planning WP FINAL 1

Uploaded by

Copyright:

Available Formats

SCF Precision Planning WP FINAL 1

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

SCF Precision Planning WP FINAL 1

Uploaded by

Copyright:

Available Formats

SCF

RELEASE 10.0 scf.io

www.scf.io/ www.5gamericas.org www.smallcellforum.org

RELEASE 10.0 scf.io

Small Cell Forum accelerates small cell adoption to drive the

We are a carrier-led organization. This means our operator members establish

We have driven the standardization of key elements of small cell technology

No license, express or implied, to any intellectual property rights is granted or

©2007-2017 All rights reserved in respect of articles, drawings, photographs

Post Small Cell Forum, PO Box 23, GL11 5WA UK

Member Services [email protected]

RELEASE 10.0 scf.io

About 5G Americas: The Voice of 5G and LTE for the Americas

5G Americas is an industry trade organization composed of leading

5G Americas is invested in developing a connected wireless community

5G Americas’ Board of Governors members include: América Móvil, AT&T,

5G Americas collaborates with government and regulatory bodies, as well as

• 5G Americas, in partnership with 5G PPP, 5GMF, 5G Forum and IMT-2020,

• 5G Americas is a Market Representation Partner at 3GPP

• 5G Americas has agreements, memberships, relationships and/or

Follow 5G Americas news on Twitter at @5GAmericas and Facebook at

A key input to the precision planning process is geolocated measurements of network

Recommended best practice for precision planning

• For maximum ROI, small cells should be placed as close as possible to

Figure 1–1 Matching small cell locations to hotspots ........................................1

Figure 1–1 Matching small cell locations to hotspots

2.1 Indoor/outdoor differentiation and 3D location

A detailed description of the planning of heterogenous networks – networks that

Figure 2–1 Differentiating between outdoor and indoor payload forecast in 2D

Although 2D is the norm for visualisation of UE location or carried payload,

A 2D-only view could therefore lead to an excessive number of street-level outdoor

The use of 3D location and visualisation, by contrast, provides an understanding of the

Supplementary information can be summarised, such as the traffic per building, as

Figure 2–3 Per building traffic

2.2 Outdoor planning

2.2.1 Why does the entire system need to be examined holistically?

The new normal

2.2.3 User activity is hyper-sensitive to network performance. How?

2.2.4 How are outdoor zones of underserved demand identified?

2.2.5 What is the effect of optimal network designs that

Everything is associated with the cost. An operator is trying to maximise spectral

2.2.6 How do spectrum and multiple bands play a role in controlling

2.2.8 How do design and rollout strategies affect ROI?

2.2.9 Why do 5G, and in particular mmWave, change traditional

2.2.10 Summary of outdoor planning

2.3 Indoor planning

The amount of LOS vs NLOS is important if an indoor network is designed for

In LTE, a hardware (Layer 2) limit for small cells may be up to 32 simultaneous

An example of a capacity coverage map is shown in the figure below. Here we

Figure 2–14 Composite coverage of a 120º sector with 35 narrow beams

To summarize, the important KPIs for indoor network design are

• Capacity and coverage,

When speaking of millimetre wave indoor propagation, high-frequency signal range is

Figure 2–15 Indicative location accuracy for different data sources

Where: ROI = profit / TCO over a five-year timeframe

4.1 Types of UE locate

4.2 Using machine learning to create ‘enhanced’ timing-based solutions

4.3 Using ML to improve timing-based location techniques

In order to determine the potential accuracy of any solution, measurements from a

Figure 4–2 Dallas metro collection area