Académique Documents
Professionnel Documents
Culture Documents
3 INTRODUCTION .................................................................................................. 7
5.1 Methodology.................................................................................................................. 11
6 DESCRIPTION OF RESULTS............................................................................ 20
7 CONCLUSION.................................................................................................... 25
8 APPENDIX I........................................................................................................ 27
10 APPENDIX III................................................................................................... 32
1 Glossary of terms
BCCH Broadcast Control Channel – in this context referring to the entire carrier
containing the BCCH
C/I Carrier to interference
DTM Digital Terrain Model
DTX Discontinuous Transmission
ILSA Intelligent Local Search Algorithm
TCH Traffic Channel – in this context referring to carriers not containing the BCCH
TRX Transmitter/Receiver
Document history
Version Revision Date
1.0 O.Ghawi (Quality Manager) October, 10 2000
2.0 O.Ghawi (Quality Manager) November, 8 2000
3.0 O.Ghawi (Quality Manager) November, 28 2000
4.0 O.Ghawi (Quality Manager) December, 8 2000
5.0 O.Ghawi (Quality Manager) December, 18 2000
6.0 O.Ghawi (Quality Manager) January, 9 2001
7.0 J.-F. Wagen (Consultant) September, 6 2001
2 Executive Summary
The main objective of this case study is to show that using a sophisticated prediction model
reduces cost and saves time in the planning of radio cellular network especially when
adjusting the frequency plans.
In this case study, a frequency-planning tool (ILSA) from AIRCOM, Ltd., was used to
compare frequency plans obtained by using
1. A classical propagation model and
2. The ray-tracing model WaveSight.
The tests presented here were performed on a 4.5 km x 4.5 km area in the city of Paris
comprising 17 sites (36 cells). All data except the predictions have been kindly provided by
the French operator Bouygues Telecom, the buildings database was provided by Istar.
This study demonstrates that using the WaveSight model has the following advantages:
1. The area where interference is unacceptable can be reduced by 80%. This reduction could
be translated directly into an increase of traffic (or revenue) or less “lost traffic”.
2. It could reduce the number of carriers needed to provide the same quality in a radio
network. In the case investigated here, it was possible to reduce from 47 to 40 the number
of necessary carrier. This is significant not only because it can reduce the cost of fine-
tuning the network, but also because extra carriers can be used to increase traffic capacity.
3. WaveSight does not needs any calibration, thus the use of WaveSight saves time,
measurements and provides more realistic prediction.
3 Introduction
Frequency planning is one of the very demanding tasks in planning cellular networks,
especially GSM networks. A better arrangement of frequencies can significantly increase the
service quality an/or ease an increase of network capacity usually required while trying to
minimise the impact on the existing base stations in order to save cost.
This study was conducted to evaluate the impact the advanced propagation prediction model
WaveSight will have on frequency planning.
The GSM radio network performance obtained by using two classical propagation models has
been compared to the results obtained by using the advanced WaveSight model, which
includes the effect of buildings.
All results have been obtained for a set of 36 cells deployed by the French operator Bouygues
Telecom over 17 sites in a 4.5 km x 4.5 km area in the city of Paris. The radio network data
and the geographical databases have been kindly provided by Bouygues Telecom and ISTAR,
respectively. Predictions have been computed using a classical “Okumura-Hata”-type model
(CM) and the WaveSight model (WS) from Wavecall. The frequency plans have been
generated using the commercially available frequency-planning tool ILSA integrated in the
excellent AIRCOM’s ENTERPRISE suit (www.AIRCOM.co.uk).
This case study is structured as follows. The next chapter, Chapter 4, describes the network
configuration used here. Chapter 5 presents our methodology. Chapter 5 also presents the
software tools, the two prediction models (classical: CM and WaveSight: WS) and the chosen
parameters. The main results are then analysed in Chapter 6. Finally, conclusions are provided
in Chapter 7. Additional plots and tables detailing the results are placed in Annexes I to III.
4.1 Sites
The area under study includes 17 sites. These 17 sites are described (location, antenna height,
transmitted power and antenna downtilt are given) in Table 1 (for the so-called “micro-cell”
sites) and in Table 2 (for the “macro-cell” sites).
As mentioned in the introduction, two propagation models have been used: a classical model
(CM) and the WaveSight model (WS).
The classical model (CM) implies two different algorithms: one for micro-cells and another
one for macro-cells. The macro-cell model is based on the ETSI Hata model, and the micro-
cell model is based on a pseudo-ray technique using terrain height, and building outlines.
As usual, the classical model requires measurements to be calibrated. Since this case study
intend to compare generic Classical Model to the WaveSight model we did not used all the
possibilities provided by the AIRCOM’s tool to optimise the calibration of the classical
model. This calibration that has been performed is described in the next section. The
calibration used has not been completely optimised but it provides a very good idea of what
could be obtained in the radio planning of changes in antenna orientation or down-tilt, or of a
new region or when new buildings have grown, or when measurements are not very extensive.
In fact, extensive calibration of a classical model is only useful when optimising the frequency
plan. Any other change in the radio network might affect the calibration. Thus, the calibration
used in this study is claimed to be sufficient. Furthermore, since it is easier and more cost
efficient to optimise a well-planned network and since, it is obviously not possible to take
measurements from base-stations not yet deployed, a perfect calibration of any Classical
Model is not trivial.
The WaveSight model (WS) does not require any calibration. The WaveSight model is a fully
deterministic model based on an efficient implementation of real ray-tracing algorithms. The
same basic principle and the same basic algorithms are used for both macro-cells and micro-
cells. The WaveSight model uses ground height, building outlines and building height to
calculate the predicted field strength. Other radio parameters can be computed by WaveSight
(angle of arrival, delay spread, …) but these are out of the scope of this study.
Accurate terrain data (building and ground) was kindly provided by ISTAR (www.istar.com).
Frequency plans have been computed using the efficient ILSA tools included in the renowned
AIRCOM’s ENTERPRISE suit. More details about the frequency planning algorithms are not
provided here but are available directly from AIRCOM (www.aircom.co.uk).
5.1 Methodology
To evaluate the end effects resulting from the use of two different prediction models, the
following method has been used. The evaluation should include all the major steps involved in
the planning of a complete radio network in a given area. Thus, the steps include coverage
predictions based on geographical and radio network data, computation of the frequency
channel constraints, generation of a frequency plan and final evaluation of the performance of
this frequency plan. Comparisons are made between the results obtained when using a
classical coverage prediction model (CM) and when using the more accurate WaveSight
model (WS). The steps are then as follows:
1. Generate a frequency plan (CM_FP) using the AIRCOM’s ILSA software based on
the predictions computed with the classical prediction models.
2. Use the results of the first step along with WaveSight coverage predictions to compute
the ratio of Carrier/Interference in the network (CM_FP_WS_MDL in Appendix).
3. Generate a frequency plan (WS_FP) using the AIRCOM’s ILSA software based on the
predictions computed with the WaveSight model.
4. Use the results from the previous step along with WaveSight coverage predictions to
compute the ratio of Carrier/Interference in the network (WS_FP_WS_MDL in
Appendix).
Assuming that WaveSight provides more accurate coverage predictions, the comparison of the
interference levels (step 2. and 4. above) indicates the value of using the more realistic
WaveSight model instead of a classical model.
In this study we want to focus on the difference between the uses of two prediction models.
We are not interested in the penalties for misplacing a carrier in two neighbours (see
paragraph 5.8). We expect the frequency planning tool to work at reducing the interference
and not to work at rearranging and/or rejecting frequencies to satisfy neighbours’
relationships. Furthermore, the neighbours’ relationships are usually not very sensitive to the
accuracy of the prediction tools. Thus, we have adjusted the neighbours’ relationship, the
Each carrier must belong to one of the two layers called either the Broadcast Control Channel
(BCCH) layer or the Traffic Channel (TCH) layer. Only one carrier is assigned to the BCCH
layer in each cell. Additional carriers, if any, belong to the TCH layer.
• The prediction for each cell was performed in an area with a 3 km radius for the
macro-cell configuration and a 1 km radius for the micro-cell configuration with
5 m x 5 m resolution.
• No downlink DTX was used.
• The frequency hopping was disabled.
• Traffic data was confidential and thus was not used. While this might be seen as a
major drawback, the goal of this study is to investigate the overall effect of using a
very accurate prediction model. Thus, assuming a uniform traffic simplifies the
analysis without a great loss of generality.
5.5.1 WaveSight
In a macro-cell configuration, the classical model (CM) is taken as the ETSI Hata model
defined by the following formula:
Where:
• The parameters k1, k2, k3, k4, k5, k6, k7 have to be calibrated to reach the lowest
standard deviation between the model prediction and the measurements.
• D is the distance from the base station to the mobile (km).
• Hms is the height of the mobile antenna above ground (m).
• Heff is the effective height of the base station antenna (m) defined as the relative
height to the mobile.
• Diffraction Loss was calculated by the Epstein Peterson method.
• Clutter Loss was not considered and was set here to 0.
Thanks to a very useful feature of the AIRCOM’s ASSET software, the classical model could
be calibrated on a set of measurements provided for all the 10 macro sites considered here.
After calibration we obtained a mean standard deviation of the prediction error of 10 dB when
evaluated over the 10 sites, with a standard deviation of 3.6 dB among the results for each
sites. The mean error was 0 dB when averaged over all sites. The mean error computed over
each site has a standard deviation of 8 dB. More details are available upon request from the
author or from our web site www.wavecall.com.
The parameters of the calibrated macro-cell classical model are shown below in Table 4:
An even better calibration might have been obtained from the ASSET software, for example a
calibration per site could have been performed. However our goal was to obtain a standard
deviation of about 10 dB to simulate the widely accepted performance of a classical model in
non urban area for a cell radius less than 3 km.
The micro-cell model considered here is based on pseudo-ray technique that uses buildings
outlines and a 5 meters resolution digital terrain model (DTM) height. As the area under
investigation is flat, the DTM was not needed.
The Micro-cell classical model employs two different algorithms whether the mobile is in
line-of-sight (LOS) or in non line-of-sight (NLOS) from the base station. In the LOS case, the
path loss is computed by a dual-slope formula. In the NLOS case, the building corner plays
the role of a secondary source. The parameters of the model have been calibrated and are
given in Table 5 and Table 6, respectively.
Again, the calibration used here might have been further optimised, especially if more
measurements had been available. However, the parameters used here are believed to
represent fairly the performance of this type of LOS/NLOS micro-cell model.
Table 7 shows the computation time required to predict coverage over a single cell. It
corresponds to the average found over the 47 cells of the study.
Both prediction models (CM and WS) were executed on a Pentium III-PC 650MHz with 256
MB RAM.
Table 7: Mean computation time for different configurations
Micro-cell Macro-cell
WaveSight model Classical model WaveSight model Classical model
(Pseudo-Ray-Tracing) (Okamura-Hata)
6 min 150 min 25 min 8 min
The best server map displays the colour-coded cell providing the highest carrier power (C)
value in each pixel. The best server coverage map provides the highest C value in each pixel.
The best server coverage map obtained from the classical model (CM) and the WaveSight
model (WS) are shown in
Figure 2, and Figure 3, respectively.
The parameters considered to create the interference array are shown in Table 8 and are
explained below:
Minimum service level: is the minimum service level at which a cell is considered to be a
serving cell. –104 dBm is a rather typical value (the minimum value is –110 dBm).
Maximum timing advance: Is the maximum difference in timing between transmission and
reception. This effectively defines the maximum radius at which a cell will be considered a
best server even if the signal is still good in terms of absolute value. The maximum value is 63
(corresponding to a 35 km radius) which means that no restriction occur here.
Adjacent channel offset: specifies the offset that will be applied to co-channel carriers to
interference value (C/I) to obtain the adjacent channel C/A value. C/A = C/I + Adjacent
channel offset. –18 dB is a typical value.
No traffic data (i.e., a uniform traffic is considered), no Frequency hopping and no downlink
DTX were used.
In this study, both the Worst interferer and the Total interference criteria were considered for
comparisons between the planning using the classical model (CM) and the WaveSight model
(WS).
Poor quality areas are those where the C/I level is less than 12 dB (all non-green area in the
Figures shown in the Appendix).
Figure 1 Best server coverage for the classical model (micro- and macro-cell) predictions.
The coarse accuracy of these predictions is obvious.
Figure 3 Colour code used in the best server coverage maps above.
To create the interference table, the user-friendly “interference table wizard” of ASSET was
appreciated. We used the parameters shown in Table 9.
In this study we assumed one carrier per layer (BCCH, TCH), with the following parameters:
• Entire sites where selected for planning.
• No fix carriers were allowed (a realistic assumption if the frequency plan has no pre-
existing restriction)
The so-called separation rules that dictate the separation between two carriers are defined by
an integer. For example, a separation of 0 means that the carriers could be the same. A
separation of 1 means that the two carriers can be adjacent. Separation numbers for different
cases are shown in Table 10.
• The penalty or relative cost (in $ for example) of not applying to those rules is shown
in Table 11.
• Since the cost of interference alone is substantially low, this will help identify whether
or not the planning tool ISLA applied the rule presented in Table 10. Any relative cost
below 50000$ would indicate that ILSA tool places carriers according to the rules and
works only on reducing the interference.
Table 11 Frequencies separation penalty costs for not applying separation rules.
Carrier Layer Cell Cost Site Cost Neighbour Weight
Cost
BCCH 100000 75000 50000 1
TCH 100000 75000 50000 1
• The ILSA software was allowed to run until the cost of the plan had become
reasonably stable. We obtained an average time of 6 hours per test. Figure 4 shows the
evolution of some ILSA parameters during the frequency planning process.
Figure 4 For information here is the print screen of a graph given by ILSA, plots the iteration
number vs. the cost of the plan (green), the average interference (red), and the Worst
interference (blue). The fast drop of the blue line shows the time when all the separation rules
were met.
6 Description of results
Worst interference: For each 5 m x 5 m area pixel, the carrier with the worst C/I is
determined by taking the stronger interferer generated by the others sites. The result is the so-
called worst C/I.
The worst C/I is meaningful since frequency hopping was not considered here.
Total interference: For each 5 m x 5 m area pixel, the total interference array gives the ratio
of the carrier signal strength within the pixel to the power sum of the interfering signal
strength generated by other sites. The C/I calculated in the total interference array is not
experienced by any subscriber but provides an idea of the interference strength.
The results are detailed in Appendix II, Table 15 ,Table 16 and Table 17.
The goal is to evaluate the performance resulting from the use of either one of the two
prediction models, the classical (CM) and the WaveSight (WS) models. Various interference
environments were considered to obtain results under different test conditions. Three different
tests are documented here. The tests were generated by changing the number of neighbour
relations and the minimum overlapping area size (dictating whether two cells are neighbour or
not):
I.) Maximum of 12 neighbour relations, minimum overlapping area size: 0.01 km2.
II.) Maximum of 10 neighbour relations, minimum overlapping area size: 0.01 km2.
III.) Maximum of 10 neighbour relations, large minimum overlapping area size: 0.122 km2.
I.) In the first test, a maximum of 12 neighbour relations was imposed. With the classical
model, no frequency plan could be found to meet the “Co-Cell, Co-site, Neighbour, 2nd
neighbour” separation rules of (2, 2, 2, 0) (Table 10) when using the 42 available carriers. In
order to guess how many frequencies were needed to comply with the rules, more carriers
were artificially added to the network. It could thus be shown that at least 47 carriers were
required for the classical model to give the same network quality. A frequency plan based on
the WaveSight predictions used only 40 carriers.
Furthermore, Table 12 shows that the classical model leads to 1.63% more bad areas than
when using the WaveSight model. (See also Figure 5 and Figure 6). Roughly speaking, the
use of the WaveSight model could provide a 1-2% increase in traffic revenue, simply by
reducing the area where interference might occur when a classical model is used instead of the
more accurate WaveSight model.
II.) In the second test condition, a maximum of 10 neighbour relations was imposed (instead
of the 12 considered in test I.). Subsequently, frequency plans were found to meet the
separation rules for the two prediction models. However, the frequency plan (CM_FP)
based on the classical model predictions leads to 10 times more bad area than the
frequency plan (WS_FP) based on WaveSight predictions (Table 13, Figure 7 and Figure
8).
III.) The aim of the third test condition was to determine the minimum number of carriers
needed to meet all the rules. The maximum number of neighbour relations was kept at 10, but
the minimum overlapping area to consider two cells as neighbours were increased about 10
times from 0.01 km2 to 0.122 km2. The result was that with a reduced number of cell
neighbours, it is easier for the frequency planning tool to meet rules and stress its works on
reducing the C/I interference level. With both models, only 30 carriers were needed to meet
the frequency planning rules. However, WaveSight gave 5 times less bad area than the
classical model as shown in Table 14.
1800000
1600000
1400000
Planning cost
7 Conclusion
The impact of using a very accurate prediction model such as WaveSight (WS) instead of a
classical model (CM) has been investigated.
The performances of GSM frequency plans resulting from the two prediction models have
been compared using a commercially available radio network planning tool: the user-friendly
ASSET tool from AIRCOM (www.aircom.co.uk). We compared the performance of the two
prediction models (CM and WS), after frequency planning, according to the total size of areas
where interference was unacceptable while sufficient carrier power was available for
coverage. This performance measure indicates the amount of lost traffic. Thus, an operator
can easily translate improvement in this performance measure into revenue gain (for example
a 1% decrease of the area where unacceptable interference occur, means a 1% increase in
revenue if the traffic is uniformly distributed. A non-uniform traffic can increase or decrease
this factor depending on where the interference occur with respect to traffic).
Another performance measure was also used: namely the required number of carrier
frequencies achieving a frequency plan without penalty. This minimum number of carriers
provides a rough estimation of the possible increase in capacity remaining in the network.
The results derived from the two prediction models were compared under the following three
conditions.
I.) Firstly, we assume that up to 12 neighbours were allowed, with 42 carriers available.
II.) Secondly, we decrease to 10 the number of allowed neighbours in order to ease the
feasibility of the frequency plan.
III.) Lastly, we assume that only cells with large overlap (0.122 km2) could be neighbours.
In this case, frequency plans without penalties could be obtained with only 30 carriers
(Table 11).
I.) The results of the first test demonstrated that WaveSight could lead to feasible frequency
plans using down to only 40 carriers. With enough carriers, the frequency-planning tool used
here focused mainly on the interference reduction, which subsequently leads to minimum
penalty cost. Conversely the classical model failed to meet the same rules even with 42
carriers. Other tests have shown that at least 47 carriers were needed to produce a feasible
frequency plan based on the Classical Model based predictions.
II.) The second test showed that WaveSight requires less than 38 carriers to meet the rule,
while the Classical Model requires 39 carriers. Moreover, when both models where able to
apply the rules, WaveSight reduced interference 10 times less than what the Classical
Model see Table 16.
III.) In the third test, only 30 carriers were needed for both models, but the Classical Model
gave an interference area that is 5 times larger than when using WaveSight.
Time efficiency, including run time as well as measurement and calibration time, is also an
important factor. The WaveSight computations do not only need little time to carry out the
predictions, but the WaveSight model does not require performing calibrations against
measurements. The resulting savings are worth to be seriously considered.
8 Appendix I
Figure 11 The vertical pattern for the directional antenna (used in macro-cells)
Figure 12 The horizontal pattern for the directional antenna (used in macro-cells)
9 Appendix II
Table 15 Statistic of the test done where the carriers separation pattern was set at 2 2 2 0, the
maximum number of neighbour relations created for each cell was 12, and an overlap of 400
squares is required between neighbour cells.
Where:
Table 16 statistic of the test done where the carrier separation pattern was set at 2 2 2 0, the
maximum number of neighbour relations created for each cell was 10, and 400 squares were
required for neighbour creations.
Where:
Where:
10 Appendix III
Figure 13 The worst interferer when using WaveSight for planning in test number 19, 1.4% of
bad area <12dB was obtained.
Figure 15 The worst interferer when using the classical model for planning in test number 19,
5.37% of bad area <12dB was obtained.
Figure 16 The total interference when using WaveSight for planning in test number 19, 1.48%
of bad area <12dB was obtained.
Figure 18 The total interference when using the classical model for planning in test number
19, 7.79% of bad area <12dB was obtained.