There is no need to state that solar flare
prediction is very important for mankind. Solar activity influences at
nearly all aspects of human life. Energetic solar flare can cause
malfunction of low-Earth orbiting satellites, black-outs in electric power
transmission grids, radar, and navigation, power plant shutdowns, and many
other problems. It is important to be ready to these cases by predicting
solar flares. However solar activity forecast is extremely complex problem.
It is a multi-objective problem with faint correlation between known and
used criteria based on rich variety of gamma-ray, neutron and energetic
charged particle emission, cm radio emission, coronal hole description,
coronal mass ejections (CME), solar proton event (SPE), and others. Current
approaches to predicting of geomagnetic storms reach 20 - 40% accuracy.
We propose a way of solar flare prediction based
on associative self-building neural network that size is limited only PC
processing power and available memory. That means user can train Artificial
Foreteller on a set of solar condition descriptions with as many these
descriptions as computer can process. Each item of the set describes solar
condition before solar flare eruption occurred (for example, one or two
days before the sun flare eruption happened). An item can contain as many
features as a user has or wants to use. Features can be represented as a
set of codes like:
Feature A - A1, A2, ... , An.,where
Feature A is a range of values, and A1, A2, ... , An - set
of sub ranges of values.
Feature B - B1, B2, ... , Bm.
Feature Z - Z1, Z2, ... , Zs.
Based on this set of features user can describe
each case of the sun condition as a sequence of codes:
Case 1 - A1,B2,
... , Zm.
Case 2 - A9, Bw, ... , Zr.
Case m - Ak, B6, ... , Zl
If there is history data about solar flares then
it is possible to generate train set of examples that describe sun
condition before eruption of solar flare. Let say that emanation of solar
flare occurred in a day after described sun condition. A set of cases can
be represented as following:
Train Case 1 - A1, B2, ... , Zm. -
Train Case 2 - A9, Bw, ... , Zr. - positive
Train Case m - Ak, B6, ... , Zl -
negative, - where "positive" means that the description of the
case followed solar flare eruption, and "negative" means that
solar flare eruption did not followed.
A set of cases is used for train and generating of
associative self-building neural network. During train Artificial Foreteller
processes the set to find combinations of values of different features that
are typical for positive or negative train cases. Let's see how Artificial
Foreteller processes a train set is shown below.
These cases are represented in neural network with
corresponding nodes and connections. It will form four main nodes that
represent two positive, and two negative cases. At the same time it will
generate three new nodes that are common for more than one case. One of
them ( A1&B2&C5&G11&K13 ) belongs to two positive cases, one (A10&B21) - belongs to two negative cases, and one node (D11&E13&G11) belongs both positive and negative cases (Case2,
Case3). As a whole, generated network can be represented like sum of
products and dispersions :
((A1 & B2 & C5 & G11 & K13) & D1 & E23 & F43 & L7 & M8 & N22 & O11 )OR((A1 & B2 & C5 & G11 & K13) & (D11 & E13 & G11) & F33 & L8 & M19 & N21 & O17)OR NOT((A10 & B21) & C15 & (D11 & E13 & G11) & F23 & K18 & L5 & M14 & N11 & O2)OR NOT((A10 & B21) & C14 & D8 & E11 & F3 & G1 & K19 & L6 & M4 & N1 & O4)
self-building neural network is not a "black box". The formula
given above can be analyzed to separate most important features from minor
ones. It can be used as a powerful tool for data analysis.
In example shown above the combination of features A1&B2&C5&G11&K13 is common for two cases that describe
solar conditions before eruption of solar flare. Combination of features A10&B21
is common for two cases that describe solar conditions but eruption of
solar flare was not occurred. In that case, Artificial Foreteller will
generate conclusion that solar flare would not happen. And combination of
is common for two cases that belong both positive and negative
descriptions. When Artificial Foreteller predicts situation it assesses
activated nodes and if activated node represents combination of features A1&B2&C5&G11&K13 then its prognosis is positive. If a
node that represents combination of feature's values A10&B21
is activated then Artificial Foreteller predicts that eruption of solar
flare will not happen. In case when a node D11&E13&G11 is activated Artificial Foreteller can
give only assessment of future situation. It can predict that the case can
happen or not with % probability.
who trains the system, can analyze network and decide if it necessary to
keep features D, E, and G, or remove only sub ranges D11, E13,
G11. After that
Artificial Foreteller can be trained again to check if its prediction
becomes better or not. This way user can select most important and
representative features that describe solar condition and remove less
help of generated neural network Artificial Foreteller will predict that
solar flare would happen in a day or not. To have system that will provide
prediction on the second day or later after described solar condition it is
required to generate and train another network. Artificial Foreteller can
have as many network as many problems it should solved.
Artificial Foreteller Structure
Initially network structure
consists of two layers. The lowest layer is a set of sensors. The second
layer consists of receptors connected to the sensors.
Each feature that describes
condition of the sun corresponds to a sensor. Sensors are divided into two
types - qualitative and quantitative. Qualitative sensor transfers only
signals "YES"/"NO" to the receptor that is connected
with the sensor. Signal "YES" activates the receptor.
Quantitative sensor is connected with a set of receptors. Range of values
that sensor receives is divided into a set of sub ranges. Current version
of preprocessor can form from 3 up to 50 sub ranges. Accordingly, quantitative
sensor can combine up to 50 receptors. In that case receptor becomes
activated only if the sensor sends a signal that value corresponds to a sub
range of values that was confronted to the receptor. Nodes of associative
self-building neural network are formed by network builder only during
train neural network is formed. In its structure all derived regularities
(or combination of features) are represented with nodes and their weights.
In a mode of solar flare prediction, entered into Artificial Foreteller
description of solar condition activates part of the neural network.
Activated part of the neural network sends signals about activation into
Artificial Foreteller of Solar Flare where they are assessed. As a result,
Artificial Foreteller of Solar Flare issues prediction if next day will be
solar flare eruption or not.
self-building neural network can be used not only for solar flare
prediction, but for classification and data analysis. Analysis of network
structure is quite important to derive most significant features.
Associative self-building neural network works over and accumulates vast
amount of information about critical solar features.
Artificial Foreteller is a new
powerful system that demonstrates extraordinary ease of use for prediction,
and classification situations. Artificial Foreteller analyses current input
values and predicts future situation with solar flare. That means
Artificial Foreteller tries to predict what will happen, and not what already
happened. As it well known human being can operate with
not more than 7 criteria at the same time. As distinct from human being
Artificial Foreteller is able to operate as many different criteria and
features as it has been trained.
Foreteller can "remember" thousands of solar conditions in its
neural network and use them for analysis of the current one. It forms prediction
based on derived from train data set of regularities and analogies among
analyzed and known situations.
Foreteller was written in VC++ and works under MS Windows98/me/NT/2000/XP.
A virtual memory storage was developed to store neural network and let the
system to create network with size more than size of PC central memory.
With help of virtual memory storage neural network became portable and can
be transferred among different PCs.
Interested parties can contact us at:e-mail