Model Tuning Project Overview

Mentum Planet
Customer - ???
Agenda

 Why model tuning is needed

 The process and methodology using Planet
 Some propagation theory
 The scope of the Bharti project (lepton)
 Model Tuning process (best practice)
 Discussion of results (lepton)

2 2006-01-09
RF Propagation Model Tuning 0 dB

RF Model Tuning Theoretically Ideal

Model
 Characterizes how terrain, foliage, building

Percentage of Drive Test Points

density and seasonal changes to the coverage
environment affect RF signal delivery Tuned Model

 Simultaneously considers many variables to

predict the effects on the signals en route from the
base station to the handset and back Un-tuned
Model

 Utilizes model tuning tool, drive test data, test site

configuration data and GIS data in order to tune
one of the available propagation models
 Ensures that RF predictions in the RF planning / - Drive Test minus Prediction
+
optimization tool are accurate
Impact of Model Accuracy
Importance of Model Tuning  Reduce fade margin
 Necessary when:  Increase cell radius
– Existing model accuracy is in question  Minimizes the number of sites for the
– Propagation engine is changed desired coverage
– Frequency band changes  Minimizes CAPEX / OPEX
– Terrain and / or clutter properties change  Improve accuracy of RF tool
– New design or when coverage expansion is simulations and analysis’
required – AFP, Traffic & Capacity Forecast,
Monte Carlo analysis
 Network is tuned for performance and
profitability
3 2006-01-09
Model Tuning with Planet
Model Tuning with Planet EV
Inputs
 Propagation model selected from
the portfolio (CRC Predict, Planet
General Model, Universal or
Wavesight)
 Appropriate GIS data with accurate
surface morphology (clutter) and
elevations data, building data is
necessary for Ray-tracing based
models
 Test site configuration data
 Drive test data

Results
 High quality RF network planning
– Improved KPIs
 Efficient use of RF engineering resources
 Minimize or eliminate RF network
overbuild and under-build

4 2006-01-09
Planet Model Tuning
Attributes
Planet Model Tuning Attributes

6 2006-01-09
 Propagation Model Portfolio
 No single model is suitable for
every application
 Mentum selects the most suitable
model for the applications
– Model selection depends
primarily on the application
and availability of GIS data
 Primary models for macro-cellular
use are CRC Predict and the
Planet
General Model
 CRC Predict is deterministic
– Simulates behavior of the
radio wave as it encounters
terrain features
 Planet General Model is semi-
empirical
– Utilizes measurement data to
define a model path loss
equation
– Flexible hybrid model able to
model various propagation
environments
 Accuracy of the propagation model
ultimately determines the accuracy
of the planning and optimization
tools
8 2006-01-09
Geodata - The Keystone of Network Quality

Accurate & appropriate geodata increases RF network design / optimisation accuracy

 Decreases probability of misplaced and / or ineffectual cell-sites
 Improves accuracy of RF models and all subsequent analysis with RF planning / optimization tools
 Improvement in network QoS
 Mentum geodata is consistently proven more accurate by as much as 20%

Mentum GIS Data

 Planimetric accuracy of vector information relative to appropriate topographic maps
 Coherence, consistency and completeness of the clutter in a sample area relative to test data
– Acquired from latest Landsat 7 satellite imagery
– Derived by skilled Mentum GIS professionals
 The planimetric and altimetric accuracy of a sample of height points relative to appropriate
topographic mapping

10 2006-01-09
Mentum Geodata
Landscapes
Layered 100m DTM, clutter and vector information
Cost-effective solution for large area network
planning and rural planning
Cityscapes
Bundled 25m DTM, clutter, and vector files
Excellent for urban and corridor planning

Skylines
High-resolution (between 1 & 5m)
3D city models can include bald-earth
DTM, building polygons, ortho-rectified photo
and / or canopy DEM layers
3D building visualizations significant for precise
understanding of interference issues

11 2006-01-09
Test Site Configuration Data
 Site location (latitude
and longitude)
 Antenna height, type
and radiation
characteristics
 EiRP (effective isotropic
radiated power)
 Spectrum sweep to
ensure drive test
frequency is clear

13 2006-01-09
Drive Test Data

 Selected test sites for the area are  Test transmitters with omni-antennas
free from near field obstructions radiate a continuous RF signal or
utilize existing sites for radiation
 Test area is divided into two to four
zones  Specialized receiving equipment is
driven in the selected areas for data
 One survey route allocated per zone collection
 Drive route targets all clutter  The data is imported into Planet EV
categories in area and processed to produce calibrated
 Drive test route represents a sample and customised propagation models
of the areas under study which represent the particular RF
environment of the specific location
15 2006-01-09
Planet Tools

 Model Tuning tool

 Model Tuning tool specific

parameters

 Calibrated model parameters

 Survey Prediction Tool

 Propagation Prediction Engine

17 2006-01-09
Planet Tools
Model Tuning tool
 Fully-automated model tuning to maximize prediction accuracy
 Multiple tuning methods
 Comprehensive tuning reports include key statistical information
Model Tuning tool specific parameters
 Automated selection of optimizable parameters based on all the model tuning inputs
 User definable parameters
Calibrated model parameters
 Result of the model tuning process
Survey Prediction Tool
 Facilitates integration of survey data into existing predictions to improve prediction accuracy
 Creates new predictions from survey data or integrates survey data into existing predictions
 Leverages actual data into simulation process
 Improves accuracy of network simulations
 Allows re-use of valuable drive test data
 Compensates for prediction errors caused by inaccurate geodata and propagation models
Propagation Prediction Engine
 Deploys prediction generation algorithm to create predictions

18 2006-01-09
Model Tuning Methodology
Proven Methodology – Seven Steps
Step 1- Project Requirements Development
 Prioritize areas to analyze with special attention to areas lacking drive test data
correlation
 Process inputs, clutter data review, terrain variations and consideration of location
specific effects are examined in detail
 Determine most suitable market modeling strategy

Step 2- Selection of a Suitable Model(s)

 Appropriate type and kind of model depends on
– Propagation tool
– RF environment to model
– Required accuracy
– Available GIS data
– Amount of drive test data the drive test budget can support
 Mentum has in-house world-class GIS lab

Step 3 - Development of Data Collection Methods

 Best representative sites are selected
 Segmented by zones indicating the best test drive path

20 2006-01-09
Proven Methodology – Seven Steps –
cont’d
Step 4 - Data Collection (Drive Testing)
 Actual signal strength is measured along pre-determined routes around the sites

Step 5 - Drive Test Post-processing

 Drive test data is averaged to mitigate fast fading
 Filter out possible drive test errors introduced by the test hardware limitations,
uncharacterized structures like over-passes and tunnels and data influenced by larger
near-field obstructions

Step 6 - Model Tuning

 Models are tuned by utilizing one of the Model Tuning Tools in Mentum Planet
 (PGM example below)

Step 7 - Model Validation

 Calibrated models are tested and evaluated by RF engineers
– Rigorous acceptance process.

21 2006-01-09
PGM Model Tuning Process

PRX = PTX + K1 + K2log(d) + K3log(Heff) +

K4Diffraction + K5log(Heff)log(d) + K6(Hmeff) + KCLUTTER
Where
– PRX Received Signal power in dBm
– PTX Effective radiated signal power in dBm
– D Distance between receiver and transmitter in meters
– Heff Effective base station antenna height in meters
– Diffraction Diffraction loss over path profile between transmitter
and receiver
– Hmeff Effective mobile antenna height in meters
– K1 to K6 Different coefficients of the model
– KCLUTTER Clutter gain/loss in dB

22 2006-01-09
PGM Tuning Regression

23 2006-01-09
PGM Tuning Process Features
– Frequency (effects diffraction co-efficient)
– Earth Curvature (effects diffraction co-efficient)
– Path Clutter
 Useful for modeling tunneling along roads
 Also for effect of clutter adjacent to roads
 Use triangular, 500m.
– Base Height algorithm – Abs Spot Height
 Heff = Hb + (H0b - H0m)

– Bin Size – based on geodata

– Rx mobile height 2m, except in elevated hway(10m)

24 2006-01-09
Customer Project Scope of Work

1. GSM 900 and GSM 1800 Band

2. Urban and Dense Urban Area as one model
3. Use Jakarta as area for data collection
4. Service will be delivered by Mentum Group – Knowledge
transfer and on job training for Customer engineers

25 2006-01-09
 TASK Customer Mentum
Validation of Customer mapping data prior to start of project a a
Radio Propagation presentation a
CW Survey Equipment – Receiver and Transmitter a
CW Survey Equipment – Test Antenna, Tripod, cables and a
connectors
Arranging Clear Test Frequency in the appropriate bands for survey a
Mapping data for the test area a
Survey site selection. a
Acquiring of street maps, atlases, etc. a
Survey route Planning. a a
Vehicle and driver and rigger for tower site a
Obtaining site access, permissions, etc. a
Test equipmen t set up on test site* a a
Equipment set up in vehicle and Collection of CW drive survey data a
Data filtering and management. a
Propagation model tuning with both CW data and test mobile data a
Report writing. a
Propagation model acceptance / val idation. a a
Knowledge transfer for data collection and model tuning a a
Presentation of Final models to Customer a
26 2006-01-09
Defining a Propagation Model
 Each different environment types ideally
should have different propagation model, Class Area Hilly Flat
for every frequency bands
 Dense Urban Flat
 Urban Flat Dense Urban 900
 Urban Hilly 1800
 Suburban Flat
 Rural Flat
 Rural Hilly Urban 900
1800
900
1800
 Good propagation model should be
applicable to all sites within similar
Suburban 900 900
environment, with same geodata source 1800 1800
(same clutter definition)
 Categorized based on geodata
Rural 900 900
availability
Telkomsel’s geodata : 1800 1800

- Geodata 10m : Urban and Dense

Urban Area
- Geodata 20m : Suburban Area
- Geodata 50m : Rural Area

27 2006-01-09
Preparation

1. Area selection
2. Site selection
3. Route selection

28 2006-01-09
Area Selection
Project Scope of Work : create one propagation model for dense urban and
urban area

Defined Class Area

29 2006-01-09
Site Selection
One good propagation model need CW drive test data of minimum 6 sites
following these criteria :
 No obstruction within 50 – 100m from the
sites
 Each site has different height
 Easy access
 Have rooftop for dense / urban
 All site locations are spread out > 1km apart
(not in same direction)

30 2006-01-09
 Proposed Sites

31
2006-01-09

3rd Option
 Proposed Sites List – Site Access Info

32 2006-01-09
Route Selection

 Covers up to 10 km around the site or until SS <= -110 dbm

 Prefers to have overlapped drive test route between each site
 Combination of all sites’ drive test result should cover all the clutter
categories of the geodata used in the project
 Drive test route should be line of sight to transmitter in order to have very
strong signal
 Maintained to ensure sufficient samples in each clutter class
 Try to match 900/1800 routes for comparison and avoid bias in tuning
process.

33 2006-01-09
Automatic Model Tuning (AMT)
 Planet Feature for model tuning process (PGM and CRC
Predict4.0)
– Speeds up the model tuning process
– Greater insight into model tuning performance
– Confidence that the optimum performance has been achieved

 Output of AMT for Planet General Model

– K1, K2, K3, K4 & K5 Optimized Coeficients to minimize RMS
(Root mean square) error between predicted values and
measured data
– Clutter coeficient values that minimize RMS error
– The confidence level assuming a  1dB confidence level for
each of the clutter correction factor estimate

34 2006-01-09
Automatic Model Tuning (AMT)
 Output of AMT for Planet General Model (part2)
– RMS, mean, standard deviation and dynamic range of the error
between the optimized model and the data
– Sensitivity of each optimized co-efficient with respect to the
optimum RMS error when an optimized co-efficient deviates
from the optimum value.
 Output of AMT for CRC Predict 4
– Clutter coefficient values which minimize the RMS and mean
error
– The confidence level assuming a  1dB confidence level for
each of the clutter correction factor estimate.
– RMS, mean, standard deviation and dynamic range of the error
between the optimized model and the data

35 2006-01-09
Model Tuning Process
 Filtering Process
– Removed signal less than 100 meters horizontally from antenna
(avoid vertical radiation patern)
– Kept signal between –40 dBm and –105 dBm (Try to keep as
much valuable signal as possible)
 Notes
– Data was removed manually from site ????l due to local
obstruction on the site. (Difficult to tune data)
– Data was removed manually form site ???? due to elevated
highway
 Averaging the data
– Averaged to 10m to remove fast fading components and to
avoid route bias. (More points in the same position does not
bias the results (i.e. traffic jam)
– Typically half the pixel size

36 2006-01-09
Model Validation Approach
 After filtering and averaging data from 5 of 6 sites,
 Tune a representative model using AMT
 Hold back data from 6th site to test against the tuned
model
 Validation site Mean Error should be <3dB
 Validation site STD deviation of error <8dB
 If validation site passes, model is considered tuned
 If not retune model parameters until validation site
meets validation criteria

37 2006-01-09
Notes on Statistics
 Dark blue is less than one standard deviation from the
mean. For the normal distribution, this accounts for
68.27 % of the set; while two standard deviations from
the mean (medium and dark blue) account for
95.45%; three standard deviations (light, medium, and
dark blue) account for 99.73%; and four standard
deviations account for 99.994%. The two points of the zσ percentage
curve which are one standard deviation from the
1σ 68.27%
mean are also the inflection points.
1.645σ 90%
 The central limit theorem says that the distribution of a
sum of many independent, identically distributed 1.960σ 95%
random variables tends towards the normal 2σ 95.450%
distribution. If a data distribution is approximately 2.576σ 99%
normal then about 68% of the values are within 1
3σ 99.7300%
standard deviation of the mean, about 95% of the
values are within two standard deviations and about 3.2906σ 99.9%
99.7% lie within 3 standard deviations. This is known 4σ 99.993666%
as the 68-95-99.7 rule, or the empirical rule. 5σ 99.99994267%
 For various values of z, the percentage of values
6σ 99.9999998027%
expected to lie in the symmetric confidence interval
(−zσ,zσ) are as follows: 7σ 99.9999999997440%

38 2006-01-09
Model Tuning Comments to
improve accuracy
 Clutter classes/height information
– PGM model will be more accurate if clutter classification by height can
be achieved.
– The current 20m clutter database although very useful in classifying
the building areas and built up areas, has only 2 distinct clutter height
classifications, that is up to 30m for block buildings/dense block
buildings and above 40m for buildings and dense urban high. Since
there is a large variety of building heights in Jakarta area a more
discrete height classification will improve the model accuracy.

 Higher resolution geodata, microcell/mini cell models

– Requires site co-ordinates to be more accurate
– For Macrocells placed at/higher than clutter not required, some
improvement at 5-10m resolution
– For Microcells, sites in the clutter to properly factor for horizontal
diffraction, more advanced model than PGM is required – Suggest
Universal model

39 2006-01-09