کاربرد شبکه عصبی موجک با الگوریتم آموزش بهینه سازی انبوه ذرات در مدل سازی تغییرات زمانی محتوای الکترون کلی یون سپهر

نوع مقاله: مقاله پژوهشی

نویسندگان

1 استادیار، گروه مهندسی نقشه برداری، دانشکده مهندسی علوم زمین، دانشگاه صنعتی اراک

2 دانشیار، گروه مهندسی ژئودزی، دانشکده مهندسی نقشه برداری، دانشگاه صنعتی خواجه نصیرالدین طوسی

10.22131/sepehr.2020.38603

چکیده

در این مقاله از ترکیب شبکههای عصبی موجک (WNNs) به همراه الگوریتم آموزش بهینهسازی انبوه ذرات (PSO) جهت مدلسازی تغییرات زمانی محتوای الکترون کلی (TEC) یونسپهر در منطقه ایران استفاده شده است. چهار ترکیب از تعداد مشاهدات ورودی مختلف جهت تست روش، مورد ارزیابی قرار گرفته است. تعداد مشاهدات ورودی انتخاب شده جهت آموزش شبکه عصبی موجک با الگوریتم PSO به ترتیب 25، 20، 15 و 10 ایستگاه از شبکه مبنای ژئودینامیک ایران (IPGN) میباشند. در هر چهار حالت تعداد پنج ایستگاه با توزیع مناسب در گستره جغرافیایی ایران به عنوان ایستگاههای آزمون در نظر گرفته شدهاند. شاخصهای آماری خطای نسبی، خطای مطلق و ضریب همبستگی جهت ارزیابی مدل شبکه عصبی موجک مورد استفاده قرار گرفته است. نتایج حاصل از مدل پیشنهادی این مقاله با TEC حاصل از مشاهدات GPS به عنوان مرجع اصلی و مدل جهانی یونسپهر 2016 (IRI-2016) مقایسه شده است. میانگین خطای نسبی محاسبه شده در 5 ایستگاه آزمون برای شبکه عصبی موجک با 25 ایستگاه آموزش برابر با 43/13%، با 20 ایستگاه آموزش برابر با 73/13%، با 15 ایستگاه آموزش برابر با 05/15% و با 10 ایستگاه آموزش برابر با 17/28% تعیین شده است. همچنین میانگین مقادیر ضریب همبستگی محاسبه شده در پنج ایستگاه آزمون برای شبکه عصبی موجک با 25 ایستگاه آموزش برابر با 9768/0، با 20 ایستگاه آموزش برابر با 9545/0، با 15 ایستگاه آموزش برابر با 9376/0 و با 10 ایستگاه آموزش برابر با 7569/0 محاسبه شده است. نتایج این مقاله نشان میدهد که مدل شبکه عصبی موجک با الگوریتم آموزش PSO یک مدل قابل اعتماد جهت پیشبینی تغییرات زمانی یونسپهر در منطقه ایران است. این مدل میتواند یک جایگزین بسیار مطمئن برای مدل مرجع جهانی یونسپهر در ایران باشد.

کلیدواژه‌ها


عنوان مقاله [English]

Application of wavelet neural network with Particle Swarm Optimization (PSO) learning algorithm in modeling of ionospheric total electron content temporal variations

نویسندگان [English]

  • Mir Reza Ghaffari Razin 1
  • Behzad Vosooghi 2
1 Department of Geo-science Engineering, Arak University of Technology, Arak, Iran
2 Department of geodesy and geomatics engineering, K. N.Toosi University of Technology, Tehran, Iran
چکیده [English]

Extended Abstract
Introduction
Development of reliable models for estimation and prediction of changes inTotal Electron Content (TEC) of the ionosphere is still considered to be a real challenge for geodesists and geophysicists. This ispartly due to the nonlinear behavior of the physical and geophysical parameters affecting the TEC variations, as well as the difficulty in accurate measurement of some of these parameters. Due to its specific nature, as well as its physical and geophysical properties, quantity of TEC hasspatio-temporal variations, which can be attributable to daily, and seasonal variations, various anomalies, or periods of solar activity. Total Electron Content is the quantity which can be used to study ionospheric activities, as well as the spatio-temporal variations in electron density of this layer. In fact, TEC is the total number of free electrons in the path between the satellite and the receiver in a one square meter column. The measurement unit of TEC is TECU, which is equivalent to 1016electrons/m2. Due to inappropriate spatial distribution of GPS receivers and their limited number, as well as observationaldiscontinuity in the time domain, TEC values and electron density obtained from theGPS measurements will be spatiallyand temporallyconstrained. In order to calculate TEC value in areas lacking observation or appropriatestation distribution, TEC value obtained from GPS measurements must be interpolated or extrapolated in a suitable manner.
 
Materials and Methods
By combining wavelet localization features with standard neural networks, Wavelet Neural Networks (WNN) have emerged as a new mathematical method for modeling and predicting the behavior of different phenomena.In WNNs, the output parameter is usually calculated by the following equation:
(1)               
 
 
wherex is the inputobservations vector,  is a the multi-variablewavelet whichcan be calculated by the tensor productof m (basic function of single variable wavelets), ë is the number of neurons in the hiddenlayer, and ù shows the network weight. Unlike the Backpropagation (BP) algorithm, PSO is a global search algorithm that can optimize the initial weights and introduce the appropriate structure for the network. Equations used in this algorithm are as follows:
                                                                                                                                                                                                    (2)
 
                                                                                                                                                                                                         (3)
In which, shows the initial weight, represents the particle’s velocity i in repetition t, c1 and c2, indicate the particle acceleration coefficients,  is the current position of particle i in repetition t and gbest represents the best particle position. The present study took advantage of a smoothing algorithm to determine STEC observations. Observed STEC values are as follows:
                                                                                                                                                                                                           (4)
 
To obtain TEC value along the zenith, the following mapping function can be used:
                                                                                                                                                                                                        (5)
Which we will have:
                                                                                                                                                                                                          (6)
Elev. in relation (6) is the satellite’s elevation angle.
 
Results and Discussion
Observations of 37 Iranian GeodynamicNetworkson 2012.08.11 (DAY 224) were used to evaluate the efficiency of WNN and PSO training algorithm in modeling and predictingspatio-temporal variations of TEC in Iran. Of the 37 stations, 5 were used as test stations, 2 were used to evaluate the wavelet neural network, and the rest were used to train the network. Four different combinations of input observations are examined in this paper. Number of input observations selected from the Iranian Permanent Geodynamic Network(IPGN) to train the WNN using PSO algorithm was25, 20, 15 and 10, respectively.Table 1 shows the characteristics of different combinations evaluated in this paper.
Table 1. Characteristics of the observations used in the different combinationsevaluated
 
 
To evaluate the accuracy of the results obtained from IRI and WNN model, all results were compared with TEC observations obtained from GPS. Table 2 shows the correlation coefficient for different scenarios.
 
Table 2. correlation coefficient for different scenarios
 
 
According to Table (2), the first scenario in WNN method with GPS hasthe highest correlation coefficient. Even when the number of observations in the databasedecreases in the third scenario, theWNN method still has a higher correlation coefficient compared to the IRI2012 model. In the fourth scenario, the correlation coefficient for WNN method is reduced to some degree. The average relative and absolute error values at the 5 test stations were calculated for the four different scenarios and presented in Table3.
 
Table 3. Comparison of mean relative error and absolute error values
at 5 test stations for four different scenarios.
 
 
Statistical analysis of relative and absolute error showssuperiority of WNN method in TEC modeling as compared to the IRI2012.
 
Conclusion
To model total electron content of the ionosphere, 4 combinations of observations were evaluated. 25, 20, 15 and 10 stations were used to train the wavelet neural network. 300, 240, 180, and 120 observations(latitude and longitude, observation time)were considered in the database, respectively.Results of the analysis indicated that with a decrease in the number of observations in the database, the absolute and relative error increase, while correlation coefficient decreases. This decrease was not evident before 180 observations, but relative and absolute errorreached up to twice their values with 120 observations. It should be noted that even with 120 observations (10 stations for training), results of the wavelet neural network model are more accurate than the results of the IRI2012 model.

کلیدواژه‌ها [English]

  • TEC
  • Wavelet Neural Network
  • PSO algorithm
  • GPS
1. Amerian, Y., Hossainali, M., Voosoghi, B., Ghaffari Razin, M. R., 2010, Tomographic Reconstruction of the Ionospheric Électron Density in term of Wavelets. International Journal of Aerospace science and Technologie

2. Cander, R. Artificial neural network applications in ionospheric studies, Annali di Geofisica, Vol.5-6, 757-766, 1998.

3. Chen, Y.h., 2000, evolving wavelet neural networks for system identification. Proceeding of International Conference on Electrical Engineering, 2000, pp. 279–282.

4. Ciraolo, L., Azpilicueta, F., Brunini, C., Meza, A., Radicella, S.M., 2007, Calibration errors on experimental Slant Total Electron Content (TEC) determined with GPS. Journal of Geodesy 81 (2), 111–120.

5. Ghaffari Razin, M.R., 2015a, Development and analysis of 3D ionosphere modeling using base functions and GPS data over Iran. ActaGeodGeophys, DOI 10.1007/s40328-015-0113-9 Volume 51, Issue 1 ,pp 95-111.

6. Ghaffari Razin., M.R., Voosoghi, B., 2016a, Regional ionosphere modeling using spherical cap harmonics and empirical orthogonal functions over Iran. ActaGeodGeophys,  DOI 10.1007/s40328-016-0162-8.

7. Ghaffari Razin, M.R., Voosoghi, B., 2016b, Regional application of multi-layer artificial neural networks in 3-D ionosphere tomography. Advances in Space Research, http://dx.doi.org/10.1016/j.asr.2016.04.029.

8. Ghaffari Razin, M. R., Voosoghi, B., Mohammadzadeh, A., 2015b, Efficiency of artificial neural networks in map of total electron content over Iran. ActaGeodGeophys, DOI 10.1007/s40328-015-0143-3.

9. Ghaffari Razin, M.R., Voosoghi, B., 2016c, Modeling of ionosphere time series using wavelet neural networks (case study: N-W of Iran), Advances in Space Research. doi: http://dx.doi.org/10.1016/j.asr. 2016.04.006.

10. Ghaffari Razin, M.R., Voosoghi, B., 2016d, Wavelet neural networks using particle swarm optimization training in modeling regional ionospheric total electron content, Journal of Atmospheric and Solar–Terrestrial Physics, http://dx.doi.org/10.1016/j.jastp.2016.09.005, 149 (2016) 21–30.

11. Habarulema, J.B., McKinnell, L.A., Cilliers, P.J., 2007, Prediction of Global Positioning System total electron content using neural networks over South Africa. J. Atmos. Sol. Terr. Phys. 69 (15), 1842–1850.

12. Habarulema, J.B., McKinnell, L.-A., Cilliers, P.J., Opperman, B.D.L. 2009, Application of neural networks to South African GPS TEC modelling. Adv. Space Res., 43(11), 1711–1720. doi:10.1016/j.asr.2008.08.020, 2009.

13. Hirooka, S., K. Hattori, and T. Takeda 2011, Numerical validations of neural-network-based ionospheric tomography for disturbed ionospheric conditions and sparse data, Radio Sci., 46, RS0F05, doi: 10.1029/2011RS004760.

14. Haykin. S. 1994, Neural Networks, a comprehensive Foundation, Macmillan College Publishing Company, New York, 1994.

15. Leandro, R.F., Santos, M.C., 2007, A neural network approach for regional vertical total electron content modelling. Stud. Geophys.Geod. 51 (2), 279–292.

16. Mars, P., J.R. Chen, and R. Nambiar. 1996, Learning Algorithms: Theory and Applications in Signal Processing, Control and Communications, CRC Press, Boca Raton, Florida, 1996.

17. Moon, Y., 2004, Evaluation of 2-dimensional ionosphere models for national and regional GPS networks in Canada, Master’s thesis, Univ. of Calgary, Calgary, Alberta, Canada.

18. Orus, R., 2005, Improvement of global ionospheric VTEC maps by using Kriging interpolation technique, J. Atmos. Sol. Terr. Phys., 67, 1598–1609.

19. Rodrigo F Leandro., 2007, A New Technique to TEC Regional Modeling using a Neural Network. Geodetic Research Laboratory, Department of Geodesy and Geomatics Engineering, University of New Brunswick, Fredericton, Canada.

20. Sayin, I., F. Arikan, and O. Arikan, 2008, Regional TEC mapping with random field priors and Kriging, Radio Sci., 43, RS5012, doi: 10.1029/2007RS003786.

21. Seeber, G., 2003, satellite Geodesy: Foundations. Methods and Applications, Walter de Gruyter.Berlin and New York, 531.

22. Tulunay, E., Senalp, E.T., Radicella, S.M., Tulanay, Y., 2006, Forecasting total electron content maps by neural network technique. Radio Sci. 41, doi:10.1029/2005RS003285.

23. Wielgosz, P., D. Brzezinska, and I. Kashani, 2003, Regional ionosphere mapping with Kriging and multiquadratic method, J. Global Pos. Syst., 2, 48–55.

24. Yilmaz, A., K. E. Akdogan, and M. Gurun, 2009, Regional TEC mapping using neural networks, Radio Sci., 44, RS3007, doi:10.1029/2008RS004049.

25. Zhang, Q., Benveniste, A., 1992, Wavelet Networks. IEEE Trans. Neural Networks 3 (6) 889–898.