f0F2 FORECAST 1h IN ADVANCE DURING DISTURBED CONDITIONS BY USING A RECURRENT FUZZY NEURAL NETWORK RF

1
f0F2 FORECAST 1-h IN ADVANCE DURING DISTURBED
CONDITIONS BY USING A RECURRENT FUZZY NEURAL
NETWORK (RFNN)

• Yurdanur Tulunay, Emre Altuntas, Ersin Tulunay,
  Zeynep Kocabas

METU/ODTU, 06531, Ankara, Turkey
2
Introduction

• HF radio communication requires forecasting the
  ionospheric foF2.
• Ionospheric foF2, is a parameter designating the
  MUF.

3
Introduction

• Ionospheric foF2,
• Diurnal, seasonal and solar variability of
• Geomagnetically disturbed conditions, induced by
  Solar Storms, the physics become more complex and
  non-linear.
• The response of ionosphere need to be qualified
  and quantified for,
• system designers, operators and
• users

4
Introduction

• Various models constructed for the forecasting of
  foF2 using different methods,
• mathematical modeling using first phys.
  principles,
• statistical models,
• data driven models.

5
Objective

• A RFNN algorithm to forecast foF2 values 1-h in
  advance at
• Sofia,
• Chilton and
• Juliusruh.
• The RFNN model is tested during
• Halloween storm and
• Nov. 2003 Superstorm.

6
Data

• Data sources
• The Space Physics Interactive Data Resource
  (SPIDR) and
• The CCLRC Rutherford Appleton Laboratory (RAL)
• Complete set of data for both foF2 and Kp indices
  during 2002, 2003 and 2004.

7
VI Station Coordinates
8
Data Coverage
9
Model

• The RFNN model adapted 7
• A fuzzy recurrent neural network model employing
• Dynamical Fuzzy Neural Constrained Optimization
  Method (D-FUNCOM ) optimization algorithm for
  training process.

10
RNNs in General

• Recurrent Neural Networks
• Identification of dynamical systems or
  recognition of temporal sequences
• Closed loop systems, with the feedback paths
  introducing dynamics to the model
• Recurrent (dynamical) neural networks exhibit
  greater prediction capabilities compared to the
  static ones

11
Dynamical Fuzzy NN Model
12
Model

• Fuzzy inference part
• three fuzzy rules
• expert information is employed

13
Model

• The inputs to the fuzzy inference
• Kp(t)
• Kp(t-3h)
• sin( minute of a day) .

14
Model

• 2 recurrent neural networks
• Quiet Neural Network
• Disturbed Neural Network
• Each NN are adapted by the optimization algorithm
  
• qNN for geomagnetically quiet time (Kp3)
• dNN for geomagnetically disturbed time (Kpgt3)

15
Dynamical Neural Network Model
Figure 3. Structure of the recurrent neural
network corresponding to the consequent of the
lth fuzzy rule 7
16
Recurrent Neurons
Figure 2. The generalized Frasconi Gori Soda
neuron (G-FGS) with a single input x(k) 7
17
Model

• Best configuration
• one hidden layer
• with 5 hidden neurons.

18
Model

• The model trained for 40 iterations
• Trained fuzzy NN model simulated using the data
  at the time period of
• Halloween Storm and
• Superstorm of November 2003.

19
Results
Figure 10. Superimposed are the observed and 1-h
in advance forecast values of Chilton foF2
versus the number of hours for the period of 26
Sept-12 Dec 2003
20
Results
Figure 11. Scatter diagram of the observed and
1-h in advance forecast values of Chilton foF2
and the linear best fit for the period of 26
Sept-12 Dec 2003
21
Results
Figure 12. Superimposed are the observed and 1-h
in advance forecast values of Juliusruh foF2
versus the number of hours for the period of 26
Sept-12 Dec 2003
22
Results
Figure 13. Scatter diagram of the observed and
1-h in advance forecast values of Juliusruh foF2
and the linear best fit for the period of 26
Sept-12 Dec 2003
23
Results
Figure 14. Superimposed are the observed and 1-h
in advance forecast values of Sofia foF2 versus
the number of hours for the period of 26 Sept-12
Dec 2003
24
Results
Figure 15. Scatter diagram of the observed and
1-h in advance forecast values of Sofia foF2 and
the linear best fit for the period of 26 Sept-12
Dec 2003
25
Results
Figure 4. Superimposed are observed and 1-h in
advance forecast values of Chilton foF2 versus
the number of hours for the period of 12-24
November 2003
26
Results
Figure 5. Scatter diagram of the observed and 1-h
in advance forecast values of Chilton foF2 and
the linear best fit for the period of 12-24
November 2003
27
Results
Figure 6. Superimposed are the observed and 1-h
in advance forecast values of Juliusruh foF2
versus the number of hours for the period of
12-24 November 2003
28
Results
Scatter diagram of the observed and 1-h in
advance forecast values of Juliusruh foF2 and the
linear best fit for the period of 12-24 November
2003
29
Results
Figure 8. Superimposed are the observed and 1-h
in advance forecast values of Sofia foF2 versus
the number of hours for the period of 12-24
November 2003
30
Results
Figure 9. Scatter diagram of the observed and 1-h
in advance forecast values of Sofia foF2 and the
linear best fit for the period of 12-24 November
2003
31
Performance for Halloween
32
Performance for Superstorm
33
Conclusions

• The performance measure
• observed and forecast values are in good
  agreement (with 3-5 norm. err.).
• At a confidence level 95
• forecast values can be considered significant
  enough for practical applications.

34
References

• Bradley P. A., G. Juchnikowski, H. Rothkaehl,
  I.Stanisiawska (1998), Instantaneous Maps of the
  European Middle and High-Latitude Ionosphere for
  HF propagation assessments, Adv. Space Res.,
  22/6, 861-864
• Cander Lj.R., M.M. Milosavlijevic, S.S.
  Stankovic, and S. Tomasevic (1998), Ionospheric
  Forecasting Technique by Artificial Neural
  Network, Electronics Letters, 34(16), Online No
  19981113, 1573-1574.
• Hornik K., M. Stinchcombe (1989), H. White,
  Multilayer feedforward networks are universal
  approximators, Neural Networks 2, 359-366.
• Lazo B., A. Calzadilla, K. Alazo, M. Rodr iguez,
  J. S. Gonz alez (2004), Regional mapping of F2
  peak plasma frequency by spherical harmonic
  expansion, Advances in Space Research, 33,
  880-883
• Mamdani, E.H. and S. Assilian (1975), An
  experiment in linguistic synthesis with a fuzzy
  logic controller, International Journal of
  Man-Machine Studies, 7(1), 1-13.
• Mastorocostas P. A. and J. B. Theocharis (2000),
  FUNCOM A constrained learning algorithm for
  fuzzy neural networks, Fuzzy Sets Syst., 112,
  no. 1, pp. 126.
• P. A. Mastorocostas and J. B. Theocharis (2002),
  "A Recurrent Fuzzy-Neural Model for DynamicSystem
  Identification", IEEE Transactions on Systems,
  vol. 31, no. 2, pp. 176-190.
• Stanislawska I., T. L. Gulyaeva, D. Altadil, P.
  Bencze, G. Juehnikowski, Yu. N. Korenkov, G. Sole
  and Z. Zbyszyniski (2001), Instantaneous Mapping
  of Ionospheric Characteristics Using 5-Minute
  Measurements for the Day of the Total Solar
  Eclipse of 11 August 1999, Phys. Chem. Earth (C),
  26/ 5, 335-339
• Stanislawska I., D. Bureshova, H. Rothkaehl
  (2004), Stormy ionosphere mapping over Europe,
  Advances in Space Research, 33, 917-919
• Tulunay, E. (1991), Introduction to Neural
  Networks and their Application to Process
  Control, in Neural Networks Advances and
  Applications, edited by E. Gelenbe, 241-273,
  Elsevier Science Publishers B.V., North-Holland. 
• Wang L.-X. and J.M. Mendel (1992),
  Back-propagation fuzzy system as nonlinear
  dynamic system identifers, Proc. IEEE Int. Conf.
  on Fuzzy Systems, 1409-1418.