Automated Short-term Prediction of

journeycartΤεχνίτη Νοημοσύνη και Ρομποτική

15 Οκτ 2013 (πριν από 3 χρόνια και 7 μήνες)

69 εμφανίσεις

http://spaceweather.inf.brad.ac.uk/

Rami Qahwaji

r.s.r.qahwaji@bradford.ac.uk

&

TufanColak

t.colak@bradford.ac.uk


EIMC, University of Bradford

BD71DP, U.K.

Automated Short
-
term Prediction of

Solar Flares using Machine Learning

http://spaceweather.inf.brad.ac.uk/

SIPWORKIII 08/09/06

Organisation of this talk


Objectives & related work


Solar data (features and activities)


Data Association


Machine learning algorithms


Practical results


Conclusions and future work


http://spaceweather.inf.brad.ac.uk/

SIPWORKIII 08/09/06

Objective:



We aim to design an automated system that
could provide short
-
term prediction of solar
flares by establishing a correlation between
sunspots and solar flares using machine
learning.


http://spaceweather.inf.brad.ac.uk/

SIPWORKIII 08/09/06

Related Work


Despite the recent advances in solar imaging, machine
learning has not been widely applied to solar data,
except for verification purposes.


Solar activity (i.e., Wolf Number) was predicted first
by (Calvo

et al.

1995).


(Borda

et al.

2002) described a method for the
automatic detection of solar flares using BP MLP.



MLP, SVM and RBF were used for flares detection in
(Qu

et al.

2003).

http://spaceweather.inf.brad.ac.uk/

SIPWORKIII 08/09/06

Organisation of this talk


Objectives & related work


Solar data (features and activities)


Data Association


Machine learning algorithms


Practical results


Conclusions and future work


http://spaceweather.inf.brad.ac.uk/

SIPWORKIII 08/09/06

Data?


Data from the publicly available National
Geophysical Data Centre (NGDC) sunspot
groups and flares catalogues are used in
our study.


NGDC keeps record of data from several
observatories around the world and holds
one of the most comprehensive publicly
available databases for solar features and
activities.

http://spaceweather.inf.brad.ac.uk/

SIPWORKIII 08/09/06

The NGDC sunspots catalogue


The NGDC sunspot catalogue holds records of
sunspot groups supplying their date, time,
location, physical properties, sunspot area and
classification data.


Two classification systems exist for sunspots:
McIntosh, which depends on the size, shape
and spot density of sunspots, and Mt. Wilson.,
which is based on the distribution of magnetic
polarities within spot groups.

http://spaceweather.inf.brad.ac.uk/

SIPWORKIII 08/09/06

The NGDC Flares catalogue


This catalogue provides information about dates,
starting and ending times for flare eruptions,
location, NOAA number of the corresponding
active region and x
-
ray classification for the
detected flares.


Not all the flares have associated NOAA numbers.
Flares without NOAA numbers are not included in
our study.

http://spaceweather.inf.brad.ac.uk/

SIPWORKIII 08/09/06

Data

http://spaceweather.inf.brad.ac.uk/

SIPWORKIII 08/09/06

Organisation of this talk


Objectives & related work


Solar data (features and activities)


Data Association and prediction model


Machine learning algorithms


Practical results


Conclusions and future work


http://spaceweather.inf.brad.ac.uk/

SIPWORKIII 08/09/06


We’ve investigated all the sunspot groups that were
associated with flares from 01 Jan1992 till 31 Dec 2005.


The degree of association was determined based on the
NOAA region number and the timing information.


A C++ platform that extracts online flares and sunspots info
from NGDC catalogues was created.


Our software has analysed the data related to 29343 flares
and 110241 sunspots and has managed to associate 1425
M and X flares with their corresponding sunspot groups.

Associating Flares and Sunspots

http://spaceweather.inf.brad.ac.uk/

SIPWORKIII 08/09/06

Associating Flares and Sunspots

http://spaceweather.inf.brad.ac.uk/

SIPWORKIII 08/09/06

Flare Prediction

http://spaceweather.inf.brad.ac.uk/

SIPWORKIII 08/09/06

The Theoretical Model

http://spaceweather.inf.brad.ac.uk/

SIPWORKIII 08/09/06

Organisation of this talk


Objectives & related work


Solar data (features and activities)


Data Association


Machine learning algorithms


Practical results


Conclusions and future work


http://spaceweather.inf.brad.ac.uk/

SIPWORKIII 08/09/06


Various neural network topologies, support vector machines
(SVM) and Radial Basis Function Networks (RBFN)

are optimized
and compared.


In our previous work (Qahwaji & Colak, CITSA 2006 and Colak
& Qahwaji, WSC11) the performance of several NN topologies
(i.e., Elman BP, FFBP, cascade FFBP, etc.)
was compared and it
was concluded that CCNN provides better association between
solar flares and sunspot classes.


CCNN and RBFN are used because of their efficient performance
in classification and time
-
series prediction (Frank

et al.

1997).

http://spaceweather.inf.brad.ac.uk/

SIPWORKIII 08/09/06


Thank You for Listening

SVM vs NN?

http://spaceweather.inf.brad.ac.uk/

SIPWORKIII 08/09/06



It is one of the recent trends in machine learning to
compare the performance of SVMs and NNs.


The work reported in (Acir & Guzelis 2004), (Pal &
Mather 2004), (Huang

et al.

2004), and (Distante

et al.

2003) supports this.


Similar performance for SVMs was reported for flares
detection in (Qu

et al.

2003),


http://spaceweather.inf.brad.ac.uk/

SIPWORKIII 08/09/06

Cascade FFBP



In cascade FFBP, the first layer has connecting weights
with the input layer. Each subsequent layer has weights
connecting it to the input layer and all previous layers. .

http://spaceweather.inf.brad.ac.uk/

SIPWORKIII 08/09/06

SVM (Support Vector Machines)

maximises the distance between the closest vectors in both classes
to the hyperplane

http://spaceweather.inf.brad.ac.uk/

SIPWORKIII 08/09/06

Radial Basis Function Networks (RBFN)

http://spaceweather.inf.brad.ac.uk/

SIPWORKIII 08/09/06

Optimising the Learning Algorithms



A learning algorithm provides best generalisation if
it is optimised.


A NN is optimised if the optimum topology, learning
algorithm and learning times are found.


After finding that CCNN provides best performance,
we compared 100 different CCNN topologies.


We found that a CCNN with 6 hidden nodes in the
first layer and 4 hidden nodes in the second layer
gives the best results for CFP and CFTP.


Similar approaches were followed for SVM and
RBNN.

http://spaceweather.inf.brad.ac.uk/

SIPWORKIII 08/09/06

http://spaceweather.inf.brad.ac.uk/

SIPWORKIII 08/09/06

Organisation of this talk


Objectives & related work


Solar data (features and activities)


Data Association


Machine learning algorithms


Practical results


Conclusions and future work


http://spaceweather.inf.brad.ac.uk/

SIPWORKIII 08/09/06


Both NGDC catalogues were used and our software has
analysed the data related to 29343 flares and 110241
sunspots and has managed to associate 1425 M and X
flares with their corresponding sunspot groups.


The total number of samples used for our training set is
2882, where 1425 samples represent sunspots that
produced flares.


The remaining samples represent sunspots that existed
in non
-
flaring days and are not related to any sunspot
groups within the previous flaring sunspot samples.

http://spaceweather.inf.brad.ac.uk/

SIPWORKIII 08/09/06

The Training and Testing Sets


The NN training and testing was carried out based on
the stati
stical Jack
-
knife technique (Fukunaga 1990).


For all the experiments, 80% of the samples are
randomly selected and used for training while the
remaining 20% are used for testing. These experiments
are repeated for number of times and the average is
taken.

http://spaceweather.inf.brad.ac.uk/

SIPWORKIII 08/09/06

Initial Experiments



For each sample, the training vector consists of 5
elements ( 3 for inputs; 2 for outputs).

http://spaceweather.inf.brad.ac.uk/

SIPWORKIII 08/09/06

Initial Experiments


Several experiments based on the Jack
-
knife technique
were carried out and we found that the prediction rate for
flares in the best case scenario was 72.9%.


This indicated that a correlation existed between the input
and output sets. But this value is not high enough to
provide reliable prediction of solar activities.


To improve the learning performance we tried to
associate the classified sunspots with the sunspot cycle.


http://spaceweather.inf.brad.ac.uk/

SIPWORKIII 08/09/06


This seemed logical because the rise and fall of
solar activity coincides with the sunspot cycle (Pap

et al.

1990).


When the solar cycle is at a maximum, plenty of
large active regions exist and many solar flares are
detected. These decreases in number as the Sun
approaches the minimum part of its cycle

(Pap

et
al.

1990).


http://spaceweather.inf.brad.ac.uk/

SIPWORKIII 08/09/06

Solar Cycle and Flares

Science @ NASA,"Solar Minimum Explodes", 9.15.2005

http://spaceweather.inf.brad.ac.uk/

SIPWORKIII 08/09/06

Solar Cycle Modelling
-
Hathawa
y’s Model

a

represents the amplitude and is related to the rise of the cycle minimum,
b

is related to the time in months from minimum to maximum;
c

gives the
asymmetry of the cycle; and
t
o

denotes the starting time

http://spaceweather.inf.brad.ac.uk/

SIPWORKIII 08/09/06


For each sample, the training vector consists of 6 elements
( 4 for inputs; 2 for outputs).


http://spaceweather.inf.brad.ac.uk/

SIPWORKIII 08/09/06


Hence, for
Fkc

sunspot at solar maximum that produced an
M flare, the training vector looks like this:

http://spaceweather.inf.brad.ac.uk/

SIPWORKIII 08/09/06

http://spaceweather.inf.brad.ac.uk/

SIPWORKIII 08/09/06

http://spaceweather.inf.brad.ac.uk/

SIPWORKIII 08/09/06

Organisation of this talk


Objectives & related work


Solar data (features and activities)


Data Association


Machine learning algorithms


Practical results


Conclusions and future work


http://spaceweather.inf.brad.ac.uk/

SIPWORKIII 08/09/06

Conclusions


A fully automated computer platform that could verify this
correlation between sunspot classes and solar flares relation
using machine learning, is designed.


The association and learning softwares will become public
shortly at






Our findings show that there is a direct relation between the
eruptions of flares and certain McIntosh classes of sunspots such
as Ekc, Fki and Fkc. Our findings are in accordance with
(McIntosh 1990), (Warwick 1966), and (Sakurai 1970).

http://spaceweather.inf.bradford.ac.uk
/

http://spaceweather.inf.brad.ac.uk/

SIPWORKIII 08/09/06

http://spaceweather.inf.brad.ac.uk/

SIPWORKIII 08/09/06

A hybrid system, which combines both SVM and CCNN, will
give better results for flare prediction.

http://spaceweather.inf.brad.ac.uk/

SIPWORKIII 08/09/06

Future Work


Apply image segmentation and classification algorithms
to detect sunspots and classify them automatically, so that
the platform is completed.


To track the individual sunspot groups over their lifetime.
The development of the sunspot group can contribute to
the knowledge of the machine learning systems.


Will better prediction be achieved if
the magnetic
configuration of sunspots (Mt. Wilson classification) is
combined with the sunspot area to replace the McIntosh
classification (Sammis, Tang & Zirin, 2000, ApJ)?

http://spaceweather.inf.brad.ac.uk/

SIPWORKIII 08/09/06


To compare our findings with other authors who
tested the correlations of the various McIntosh
classes on flare rates and the applications to solar
flare prediction (e.g. McIntosh 1990; Bornmann &
Shaw 1994, Sol. Phys. 150, p. 127; Gallagher et al.
2002, Sol. Phys. 209, p. 171; Wheatland 2004,

ApJ 609, p. 1134).




http://spaceweather.inf.brad.ac.uk/

SIPWORKIII 08/09/06


Acknowledgment.

This work is supported by an
EPSRC Grant (GR/T17588/01), which is entitled
“Image Processing and Machine Learning
Techniques for Short
-
Term Prediction of Solar
Activity”.