Swarm Intelligence Optimization of Lee Radio-wave Propagation Model for GSM Networks in Irbid M. S. H. Al Salameh 1 M. M. Al-Zu'bi Department of Electrical Engineering Research Assistant American University of Madaba Jordan University of Science and Technology King s Highway, Madaba, Jordan Irbid 22110, Jordan m.salameh@aum.edu.jo Abstract Measurements for GSM cellular phone networks (the 1800 MHz as well as the 900 MHz bands) in different areas of Irbid city, Jordan, were carried out, for over a year and under varying weather conditions, by the authors. To find a radiofrequency propagation model that can correctly predict the propagation in this environment, various path loss models are compared with the measurements. These models include: Lee, COST-231 Hata, COST-231 WI, and Egli models. The results show that the COST-231 Hata model is the most accurate model whereas Lee model is the least accurate model for Irbid city. To enhance the accuracy of the least accurate model for this area, i.e., Lee model, this paper suggests a path loss model based on optimizing Lee model using the swarm intelligence optimization (PSO) technique. It is worth noting that optimization of Lee model for Irbid measurements using the least squares method produced the same results as the PSO optimization; therefore, only the PSO results will be presented here. The accuracy of the optimized Lee model is verified by comparison with measurements in other locations in Irbid. Furthermore, measurements made in Amman city in Jordan confirm the usefulness and validity of the measurements and predictions of Irbid city. The root mean square error (RMSE) between the measured and predicted values for the proposed model is significantly improved by up to 37 db compared with Lee model. Keywords Lee model, Wireless; Measurements; Path loss; Radiowave; Swarm intelligence. I. INTRODUCTION Cellular phone networks are the most widely used systems all over the world. In order to efficiently plan new communication networks and improve the existing networks, accurate radio frequency (RF) path loss models are required. Although various models are available, only some of these models will match the area considered because every model was derived from measurements performed for certain area conditions. Therefore, we need to determine the appropriate model that can accurately predict the path losses in Jordan. 1 Different path loss models were investigated and then modified to match different environmental conditions in the world. For example, Lee model was optimized for global 1 On leave from Jordan University of Science and Technology, Irbid 22110, Jordan, salameh@just.edu.jo system for cellular communications at 850 MHz frequency band using the least squares (LS) method [1], where the radio frequency measurements were collected using commercial measuring equipments in suburban and urban areas with flat terrains in Florida, USA. In [2], Lee model was calibrated by the least squares method for Jiza town in Jordan based on data supplied by mobile operators in Jordan. Another study was conducted on cellular communications at 900 MHz for different areas in Istanbul in Turkey where the Bertoni- Walfisch model was optimized using the mean square error (MSE) method [3]. Hata model was optimized for mobile communications using the least squares method for suburban areas within Cyberjaya and Putrajaya areas in Selangor state in Malaysia [4]. Different path loss models, including Lee model, were compared for mobile communications in Kuala Lumpur, Malaysia [5]. Okumura Model was optimized by the use of the regression fitting method and measured data for communication network in urban area in Kuala Lumpur in Malaysia [6]. Hata model was tuned and fitted to measurements using the mean square error method for cellular communications in Brno area in Czech Republic [7]. In [8], the COST-231 Walfisch-Ikegami (WI) path loss model was tuned by the particle swarm optimization (PSO) method for communication networks in the south-western part of Amman, Jordan. The COST-231 Hata model was optimized for communication networks in Banciao city, Taiwan, by fitting this model with measured data using the dual least-squares approach [9]. Also, Okumura model was found to be suitable for cellular communication systems in the suburban area Pathum Thani of Thailand [10]. This paper is intended to find a radiofrequency propagation model that can correctly predict the path loss in the environment of Irbid city and extend this model to other areas in Jordan. To that end, various path loss models are compared with extensive measurements, at both bands of GSM cellular communications (the 1800 MHz as well as the 900 MHz bands), performed by the authors in different areas of Irbid. These models include: Lee, COST-231 Hata, COST- 231 WI, and Egli models. The results show that the COST-231 Hata model is the most accurate model whereas Lee model is the least accurate model for Irbid city. To improve the accuracy of the least accurate model, i.e., Lee model, for Jordan environment, this paper suggests a path loss model based on optimizing Lee model using the swarm intelligence ISBN: 978-1-61804-285-9 207
optimization technique. The accuracy of the optimized Lee model is verified by comparison with measurements in other locations in Irbid. Furthermore, measurements made in Amman city in Jordan confirm the usefulness and validity of the measurements and predictions of Irbid city. The root mean square error (RMSE) between the measured and predicted values for the proposed model, is significantly improved by up to 37 db compared with the Lee model. II. PATH LOSS MODELS AND THE DRIVE TEST The path loss models investigated here are COST-231 WI [11, 12], Lee [13], [14], [15], COST-231 Hata [7, 9], and Egli models [5]. More detailed information about these models can be found in [7], [9], [11],[12], [5], [13], [14], [15]. In this paper, the received power from the base station is measured for different distances. After that, the path loss in db is calculated from the measured received power using the equation [11]: L = P t + G t + G r -P r- L t - L r (1) Where, P t is the transmitter power, P r is the received power, G t is the gain of the transmitting antenna, G r is the gain of the receiving antenna, L t is the feeder losses of the transmitter (e.g., connector losses and cables), and L r is the feeder losses of the receiver (e.g., cables and body losses of the car used in the drive test). Here, these parameters have the values: P t = 42 dbm, G t = 18 db, G r = 2.15 db, L t = 2 db, and L r =8 db. The loss L r is due to the penetration loss through the car body with an average value of 8 db; based on experiments. This value of L r is similar to the value provided by [16], [17]. The values of L t, P t, and G t were obtained from the operators of the cellular communications in Jordan. The parameters used in this paper have the following meanings: h b is the transmitter antenna height in meters, f c is the operating frequency in MHz, d is the distance between the transmitter and the receiver in km, h m is the height of the receiving antenna in meters. The measurements have been performed, for over a year, by using RF measuring tools while driving a car (drive test) on many paths in Irbid city around the GSM (Global System for Mobile Communications) cellular phone base stations at the 1800 MHz and the 900 MHz frequency bands. The drive test is performed to measure the received signal strengths from the base stations, from which the path losses are calculated by equation (1). The measuring tools consist of TEMS (Test Mobile System) RF measuring software [18], GPS receiver, Laptop, and mobile phone. The mobile phone is equipped with RF measuring firmware [19] in order to extract the received RF signal strengths and send these readings to the laptop. The measured data involve the received signal strength levels of the serving base stations for each ARFCN (Absolute Radio Frequency Channel Number) scanned channel, cell-id, and mobile station (MS) location coordinates. III. OPTIMIZING THE PROPAGATION MODEL A. General Form of the Path Loss Model The purpose of the optimization process is to improve the accuracy of the path loss model in order to fit the area considered. To that end, the general form of the path loss according to Lee model is expressed as follows [14, 15]: L = A + Blog(d) + 10 nlog f c - 10 log(α) (2) 900 Where f c is the operating frequency in MHz, d is the distance between the transmitter and receiver in km, A is the path loss at reference distance, B is the slope of path loss curve in db/decade, the parameter n is the frequency path loss exponent, and the factor α involves correction factors for the heights and gains of the transmitter and receiver antennas. More details on equation (2) can be found in [14] and [15]. Here, the main focus is on how to utilize the general form of the Lee path loss, equation (2), in order to obtain an optimized equation that matches our measurements. In other words, we need to find the optimum values of A, B, and n such that the predictions of equation (2) become as close as possible to the average of the measurements made in Irbid. The particle swarm optimization (PSO) technique is used to find the optimum values of the three parameters: A, B, and n. B. Swarm Intelligence Optimization The particle swarm optimization (PSO) algorithm is a global optimization method for the solution of non-linear problems [20]. The idea is related to the swarm intelligence of organisms such as swarms of bees, flocks of birds, schools of fish, and colonies of ants. Each particle in the swarm of the PSO method involves position vector (x), velocity vector (v) and personal best vector (i.e., best previous position encountered by each particle) and its fitness value. In the PSO algorithm, the initial velocity and initial position vectors of each particle are randomly assigned in n-dimensional search space. Each particle moves (i.e., modifies its velocity and position) according to its best previous experience in order to reach the best possible solution. The best solution reached by each particle is called the personal or local best (X pbest ). The best solution obtained among all the swarm of particles is called the global best (X gbest ). Each particle updates its velocity and position according to the following equations [21]: v n+1 id = C [ ω v n id + c 1 r n 1d (X - X n id) + c 2 r n 2d (X - X n id) ] (3) Where, X n+1 id = X n id + v n+1 id t (4) - v n+1 id, v n id: Velocity component along d th coordinate of i th particle at the (n+1) th and n th iterations, respectively. - X n+1 id, X n id : d th coordinate of i th particle at the (n+1) th and n th iterations, respectively. - X n pbest id : Personal best position along d th coordinate of i th particle at n th iteration. - X n gbest d : Global best position along d th coordinate at n th iteration. - t: Time step value; usually chosen to be 1 s. ISBN: 978-1-61804-285-9 208
New Developments in Circuits, Systems, Signal Processing, Communications and Computers - i = 1,..,Np, and Np is the number of particles in the swarm - d= 1,,Nd, where Nd is the search space dimension The inertia weight ω and the convergence factor C in equation (3) are as follows [21]: ω = ωmax (ωmax - ωmin ) C= Nj 2 j '2 - α - )α2-4α' (5) (6) Where the maximum number of iterations is Nj, and the current iteration number is j. The parameter values in this paper are, ωmin= 0, ωmax= 1; these values are chosen in order to obtain largest convergence speed of PSO, where the experiments showed that the best value of ωmax falls between 0.8 and 1.2. The acceleration constants c1 and c2 are used to increase the new velocity towards the best solutions. The parameter α= c1+c2, and c1, c2 are chosen to be: c1=2, c2=2, accordingly α=4, C=1. The experiments showed that, for best performance, c1=c2=2. The random numbers, rn1d and rn2d, are uniformly distributed between 0 and 1. Each iteration of the PSO algorithm includes evaluating the fitness value for each particle, where this value replaces the personal best value if the currently obtained fitness value is better than the personal best value. Similarly, the best fitness value obtained by the swarm replaces the global best value if the currently obtained fitness value among all particles is better than the global best value. cellular base stations in Irbid, for which the measurements were performed, are shown in Fig. 2. Fig. 3 compares the existing path loss models, the measured path loss data, and the proposed optimized Lee model for base station 6, sector 1. The results of the other base stations 1, 2, 3, 6, 8, 10 are summarized in Table 1 which also shows the values of the optimized Lee model parameters, i.e., A, B, and n of equation (2). Table 1 shows significant improvement of the optimized Lee model as compared with Lee model. The improvement ranges between 17.6 db for base station 10 sector 3 to 35 db for base station 6 sector 1. It is to be noted here that the optimized Lee model was extracted from the measurements performed in these base stations, i.e., base stations 1, 2, 3, 6, 8, and 10. The measurements in the remaining base stations, i.e., base stations 4, 5, 7, and 9 are used to verify the optimized model in Irbid city, as shown in Table 2 and Fig. 4. Fig. 4 compares the existing path loss models, the measured path loss data, and the proposed optimized Lee model for base station 9, sector 2 which shows significant improvement in the accuracy of the optimized Lee model in comparison with the Lee model. The accuracy enhancement, as can be seen in Table 2, ranges between 14.5 db for base station 9 sector 4 to 37.2 db for base station 7 sector 2. Thus, the overall improvement for Irbid city associated with the proposed model compared with the Lee model ranges between 14.5 db for base station 9 sector 4 to 37.2 db for base station 7 sector 2. IV. RESULTS The parameters of the optimized Lee model are determined by means of the particle swarm optimization (PSO) method. The fitness function of the PSO method is considered to be the root mean square error RMSE given by equation (7) below [5]. In other words, this root mean square error RMSE is used to evaluate the accuracy of the path loss models as compared with the measurements. 2 3 456 [Lmi - Li] RMSE =) N-1 (7) Where, N is the number of path loss data points, Lmi is the measured value of the path loss at position i in db, and Li is the predicted path loss at position i in db. Due to the very large number of measured data samples for the site of each base station and in order to eliminate the effects of fast fading, the measured data were averaged over every 1 m of the path between the transmitting base station and the mobile receiver. The swarm size, of the PSO method, is chosen in this paper to be 10; based on many trials to obtain accurate results. The optimal solution is reached, on average, after 25 iterations, for all the base stations. As an example, Fig. 1 shows the convergence of the fitness function (RMSE value) for base station 1, sector 1, where the RMSE decreases rapidly from 120 db to 7.9 db after about 25 iterations. The locations of the ISBN: 978-1-61804-285-9 Fig. 1: Solution convergence of the PSO algorithm for base station 1. Other base stations have almost similar convergence. Fig. 2: GSM Cellular Base Stations under study in Irbid City. 209
From Table 1 and Table 2, it is clear that the proposed optimized Lee model has the best average RMSE values compared with the other examined models. Substituting the average values of the optimized parameters (A, B, and n) from the last row of Table 1 into equation (2), the proposed optimized Lee model can be written as follows: L= 124.64 + 23.2log(d) + 26.94log fc - 10 log(α) (8) 900 Fig. 3: Path loss vs. transmitter to receiver (Tx-Rx) distance in kilometres for base station 6 (BS6). Table 1: The root mean square errors (RMSE) of the models compared with measurements, in addition to the optimized Lee model parameters for the base stations used to build the optimized Lee model. Info BS Sector No. 1 10 2 3 8 6 Average Root Mean Square Error (RMSE) db Optimized Lee Hata Egli Lee 8.2861 27.2091 13.8918 22.2750 7.9004 32.9196 11.3244 27.7726 10.0668 27.9662 14.1491 22.8772 5.4516 31.3264 8.7931 25.4180 13.2322 30.8646 16.3376 25.7568 7.0682 7.4808 6.2116 7.4931 8.8160 8.3021 8.0540 6.8632 5.7996 5.8158 10.4525 6.5744 10.4085 8.01538 30.9001 36.8599 38.5488 35.6255 35.6336 32.4521 36.9021 39.0120 32.1626 40.7997 33.0037 39.5954 33.4385 34.1789 9.1672 7.8158 7.9642 9.6318 12.0170 10.5352 8.5430 9.9891 11.7094 7.0539 12.7395 7.9736 11.8711 10.6393 25.3527 31.2849 32.7864 29.7714 29.8440 26.7466 30.8139 32.9072 26.6179 34.5448 27.3376 33.2684 27.5894 28.4980 WI 10.0340 13.2643 10.7640 8.7172 14.3335 7.79778 10.6795 10.7139 9.7865 12.0280 9.8155 10.8794 13.1312 10.9171 12.5532 11.9023 11.2226 11.2466 11.0992 Optimized Lee model Parameters A B n 122.4484 125.8005 122.3311 122.9291 119.7959 124.9121 130.6544 131.5981 124.0177 121.1776 122.9000 128.7991 121.4220 118.9979 135.3671 120.0212 128.3335 121.9733 124.6377 24.6108 19.3576 27.8632 25.5341 23.2512 28.4576 26.0445 25.3403 20.7344 15.7935 23.8026 27.9144 13.918 12.8889 32.1887 19.9881 24.6807 25.1409 23.1950 634 2.7512 2.8075 2.8592 2.5480 2.7671 2.5 2.7602 2.7943 2.7016 2.7 2.7349 2.9 938 ISBN: 978-1-61804-285-9 210
Table 2: The RMSE values of the models compared with the measurements for the base station sites which were used to validate the optimized Lee model. BS 7 4 5 9 Info Sector No. Average Root Mean Square Error (RMSE) db Optimized Lee Lee Hata Egli 11.4694 8.1812 9.3195 10.4311 11.4253 8.7615 8.7154 9.7138 9.5448 1058 7.5550 9.5604 8.9551 7.2975 13.3604 9.7931 42.4688 45.4074 44.5295 44.8020 41.6590 39.9167 42.0617 35.6519 30.1527 40.8144 37.5223 28.5422 44.1671 39.6647 27.8148 39.0117 12.3352 12.4173 10.6993 9.9025 10.6715 6.5816 10.1256 10.6986 13.2996 10.5086 9.2033 14.1265 11.5784 9.4433 17.0125 11.2403 36.0470 38.3363 37.9189 38.4007 35.7684 33.8914 35.4723 30.0986 24.9481 35.0020 31.7698 23.1041 37.5854 33.5609 22.7166 32.9747 WI 12.9086 11.5059 11.0074 18.3551 16.8768 13.7637 16.1118 14.4432 11.7719 18.0704 15.1924 11.4169 16.2953 13.0553 14.0077 14.3411 Fig. 4: Path loss models compared with measurements vs. transmitter to receiver (Tx-Rx) distance in kilometres for base station 9 (BS9) sector 2, used to verify the optimized Lee model. ISBN: 978-1-61804-285-9 211
V. CONCLUSIONS This study proposes an optimized Lee path loss model for Irbid city in Jordan. The Lee model is optimized using the swarm intelligence technique and the measurements at several GSM base stations performed by the authors for over a year under varying weather conditions. It is worth mentioning that optimization of Lee model for Irbid measurements using the least squares method produced the same results as the PSO optimization; therefore, only the PSO results were presented in this paper. The validity of the optimized Lee model is demonstrated by comparison with measurements at other base stations. The proposed model showed high degree of agreement with the measurements where the accuracy of the proposed Lee model is improved by up to 37 db compared with the Lee model. Also, measurements performed in Amman city supports the results presented in this paper, and thus this optimized model is expected to suit other areas in Jordan. The proposed model may help in the design and expansion of the cellular communication networks. REFERENCES 1. N. Mijatovic, I. Kostanic, G. Evans, Use of scanning receivers for RF coverage analysis and propagation model optimization in GSM networks, 14th European Wireless Conference, Prague, Czech Republic, June 2008, pp. 1-6. 2. L.A. Nissirat, M. Ismail, M. Nisirat, M. Singh, Lee s path loss model calibration and prediction for Jiza town, south of Amman city, Jordan at 900 MHz, IEEE International RF and Microwave Conference (RFM), Seremban, Negeri Sembilan, Dec. 2011, pp. 412-415. 3. B.Y. Hanci, I.H. Cavdar, Mobile radio propagation measurements and tuning the path loss model in urban areas at GSM-900 band in Istanbul turkey, IEEE 60th Vehicular Technology Conference, Los Angeles, USA, Sept. 2004, vol. 1, pp. 139-143. 4. R. Mardeni, K.F. Kwan, Optimization of Hata propagation prediction model in suburban area in Malaysia, Progress In Electromagnetic Research, vol. 13, pp. 91-106, 2010. 5. J. Chebil, A.K. Lwas, M.R. Islam, A. Zyoud, Comparison of empirical propagation path loss models for mobile communications in the suburban area of Kuala Lumpur, 4th International Conference On Mechatronics (ICOM), Kuala Lumpur, Malaysia, May 2011, pp.1-5. 6. R. Mardeni, Lee Yih Pey., Path loss model development for urban outdoor coverage of code division multiple access (CDMA) system in Malaysia, International Conference on Microwave and Millimeter Wave Technology (ICMMT), Chengdu, China, May 2010, pp. 441-444. 7. L. Klozar, J. Prokopec, Propagation path loss models for mobile communication, In Proceedings of 21st International Conference Radioelektronika, Brno, Czech Republic, April 2011, vol. 48, pp. 1-4. 8. A. Tahat, M. Taha, Statistical tuning of Walfisch-Ikegami propagation model using particle swarm optimization, IEEE 19th Symposium on Communications and Vehicular Technology in the Benelux (SCVT), Eindhoven, Netherlands, Nov. 2012, pp. 1-6. 9. Y.H. Chen, K.L. Hsieh, A dual least-square approach of tuning optimal propagation model for existing 3G radio network, IEEE 63rd Vehicular Technology Conference (VTC), Melbourne, Australia, May 2006, vol. 6, pp. 2942-2946. 10. W. Bhupuok, K. Dejhan, A new method for prediction 3G path loss propagation in suburban of Thailand, The International Conference on Electrical Engineering/ Electronics, Computer, Telecommunications and Information Technology (ECTI-CON), Phetchaburi, Thailand, May 2012, pp. 1-4. 11. S. R. Saunders, Alejandro Aragón-Zavala., Antennas and propagation for wireless communication system, 2nd ed. Wiley, 2007. 12. M.V.S.N. Prasad, S. Gupta, M.M. Gupta, Comparison of 1.8 GHz cellular outdoor measurements with AWAS electromagnetic code and conventional models over urban and suburban regions of northern India, IEEE Antennas and Propagation Magazine, vol. 53, pp. 76-85, 2011. 13. Lee WCY. Mobile communications engineering: theory and applications. 2nd ed. New York: McGraw-Hill; 1997: 689. 14. Gordon LS. Principles of mobile communication. 3rd ed. Germany: Springer; 2012: 819. 15. Alshami M, Arslan T, Thompson J, Erdogan AT. Frequency analysis of path loss models on WIMAX. 3rd Computer Science and Electronic Engineering Conference (CEEC); 2011 July 13-14; Colchester, UK. 1-6. 16. H. Holma, A. Toskala, WCDMA for UMTS: HSPA evolution and LTE, 5th ed. John Wiley & Sons, 2010. 17. H. Holma, A. Toskala, LTE for UMTS: OFDMA and SC-FDMA based radio access, UK: John Wiley & Sons, 2009. 18. TEMS Investigation [Online]. Available: http://www.ascom.com/nt/en/ index-nt/tems-products-3/tems-investigation-5.htm#overview 19. TEMS Pocket [Online]. Available: http://www.ascom.com/nt/en/indexnt/tems-products-3/tems-pocket-5.htm#overview 20. J. Kennedy, R. Eberhart, Particle swarm optimization, IEEE International Conference on Neural Networks, Perth, WA, Nov./Dec. 1995, vol. 4, pp. 1942-1948. 21. Q. Bai., Analysis of particle swarm optimization algorithm, Computer and Information Science, vol. 3, pp. 180-184, 2010. ISBN: 978-1-61804-285-9 212