Recognition of the Operational States in Electric Arc Furnaces

D. Raisz M. Sakulin H. Renner Y. Tehlivets

Department of Power Systems Department for Electrical Power Systems
University of Technology and Economics Budapest University of Technology Graz
Egry J. u. 18, 1111 Budapest, Hungary Inffeldgasse 18, 8010 Graz, Austria
Sakulin@ifea.tu-graz.ac.at

Abstract: For the optimization of the operation of electric arc furnaces Now the power and the arc lengths can be increased, so that
(EAFs) it is important that the actual operational state of the furnace can be
the crater is widened and the risk of electrode break will be
quickly and exactly determined. This paper presents a new approach that
allows to track the melting process. This method uses a neural network in lowered. In the following, attention has to be focused on the
order to classify the dynamic characteristics, and is compared in this paper bottom of the furnace which could be damaged if an
with other methods, like the smoothened standard deviation of arc voltages electrode reaches it at a time, when there is not enough melt.
and the partial harmonic distortion approaches. Finally, an application
In this case (Fig. 1.c) the melting power also has to be
example for the introduced procedure is shown.
reduced. In the main phase of the melting (Fig. 1.d) the scrap
Keywords: Electric Arc Furnace, Neural Networks slides down from the furnace lining into the steel bath. When
there is no scrap at the wall in the regions near the electrodes
(Fig. 1.e), the radiation of the arcs could again cause a quick
I. INTRODUCTION wear of the furnace lining. To avoid this, short arc lengths
have to be set, and the arc has to be covered by foaming slag.
The melting process in an AC EAF is shown in Fig.1. below: (Fig. 1.f)
The last stage of the melting process is not shown in Fig. 1.
Ignition of the Burning craters into the Electrodes at the bottom
arcs iron scrap of the furnace During the “refining” phase the fluid steel is heated up to
a b c 1630-1640 °C, non-metallic impurities are removed by means
of oxygen blasting, and finally, the chemical composition is
controlled and corrected, if necessary.

An optimization problem arises related to the operation of the

furnace, where conflicting objectives have to be taken into
account:

− On the one hand, the melting power should be kept as

Main melting Melting the remaining Arcs burning (in foam) high as possible in order to obtain lowest possible “tap-
phase filling on the steel bath to-tap” times, and thus achieve high productivity.
d e f
− On the other hand, the furnace lining should be
protected, because fettling (repairing) decreases
productivity and increases the specific costs of the steel.

− The electrode consumption (wear) is proportional to the

square current. Therefore, higher power causes more
electrode consumption, and that again increases the
specific costs.
Fig. 1. Melting process in an EAF (see also [1])
The various strategies of furnace operation could be
compared economically, if these objectives would be
After charging the arcs are burning very close to the furnace weighted with their costs (or earns, respectively), and
roof; - thus, melting power has to be reduced in order to measured (or simulated) over a longer period of time, say one
protect it from the radiation of the arcs. After 1 or 2 minutes year. In that comparison additional objectives could be
the arcs have already burnt a crater and keep on penetrating included, e.g. power quality factors like flicker, THD,
into the iron scrap. (Fig. 1.b). unbalance, or a complex power factor described in [2]. These
objectives would then have to be weighted according to the
actual preferences, e.g. the penalties for violating the limits.

The main task is, of course, to determine the optimal way of

furnace operation, i.e. the one to achieve the highest profit.
This rather involving problem leads to the following
questions:

− How can be found out, whether the arc already burns

on the steel bath or not?
− How can be determined the point of time, from which
the furnace lining is exposed to arc radiation?
− How can be determined to which extent the arc is
covered by foaming slag?

The aim of this paper is to present and introduce methods that

help in answering these questions by means of electric
parameters, and thus contribute to the solution of the
optimization problem described above.

II. INDICATORS FOR THE PROGRESS OF MELTING

Smoothened standard deviations

of arc voltages Fig. 2. Smoothened standard deviation of the arc voltage
One can state in general, that the running of the furnace is This index shows the progress of melting very well. (Fig. 2.
much more unquiet and uneven at the beginning stage of the shows a heat of three baskets.)
melting process than at the end (before tapping). Scrap piece The standard deviations of the arc voltage in the times near
fall-downs can produce short-cuts or arc blow-outs; the arc is 500 sec and 1100 sec are about 10 V, this means that the
extremely unstable. The base point of the arc permanently corresponding arc length variations are about 1 cm.
“leaps” from one scrap peak to another looking for better Therefore it can be assumed, that in these time periods the arc
burning conditions. This causes rapid changes in the arc already burns on the surface of the steel bath.
length, and thus results in strong, stochastic voltage (A similar index is presented in [3].)
fluctuations. Hence the amplitudes of arc voltages, currents,
and active powers of the arcs fluctuate very strong as well. Partial harmonic distortions of arc voltages
These variations become smaller when the load is almost
entirely molten. From measured data it can be seen, that at the beginning of
An index to measure these fluctuations is the standard the melting the time function of the instantaneous arc voltage
deviation of the rms values of the arc voltage. is similar to a square wave with strong stochastic signals
Let UPh be the rms of the arc voltage in phase Ph. The added (very noisy square wave), and that the positive and
standard deviation is computed over short time intervals: negative halves are different in form and length. The physical
reasons of this are discussed in [1].
1/ 2 In [1] the partial harmonic distortions are used as measuring
( )
U Ph ,std k =
1 50
50 − 1 i =1
(
U Ph ,k ,i − U Ph ,k )
2
(1) indices of these properties of the arc voltage.

The partial distortion computed from the odd-harmonics of

where the arc voltage is defined as follows:
U Ph ,std (k ) represents the standard deviation in the k-th
time interval, U 2B3 + U 2B5 +
kB = (3)
U Ph ,k ,i is the rms value in the i-th cycle of the k-th U 2B1 + U 2B2 + U 2B3 +
second, (0 i 50)
U Ph ,k is the average of the rms values in time interval k: while the partial distortion computed from the even-
harmonics of the arc voltage can be computed from:

1 50
U Ph ,k = U Ph ,k ,i (2) U 2B2 + U 2B4 +
50 i =1 kC = (4)
U 2B1 + U 2B2 + U 2B3 +
The time function of this value is shown in Fig. 2.
In the upper equations UBi are the i-th order Fourier- assumed that uncovered furnace lining spots exist in the
coefficients of the arc voltage. proximity of the electrode.
Fig. 3. and Fig. 4. show the partial distortions as defined The odd partial distortion is a measurement of the sinusoidal
above. nature of the arc voltage. Values below 18 % show that the
arc is covered by foaming slag.

III. CLASSIFICATION OF THE DYNAMIC ARC

CHARACTERISTICS WITH A NEURAL NET

The theory of neural networks is described among others in

[4]. They can be used e.g. for solving classification problems.
In this paper a neural net is used for classifying the dynamic
uarc(i) characteristics, three of which are shown in the figures
below.
Fig. 5. is typical for the firing and penetrating melt-down
phase (as in Fig. 1.a and 1.b):

Fig. 3. Partial distortion kC of the arc voltage (even

harmonics)

Fig. 5. Srap

Characteristics like the one in Fig. 6. can be observed when

the arcs already burn on the fluid steel:

Fig. 4. Partial distortion kB of the arc voltage (odd

harmonics)

The value of kC shows the difference between cathode and

anode half waves, and it is close to zero when the arc voltage
cycle is symmetric, i.e. when the arc behavior is the same in
the positive and negative half.
The decrease of kC below 2.5 % indicates (according to [1])
that there is no cold scrap under the appropriate electrode. In
the main melting phase charge slides down from the furnace
Fig. 6. Liquid Steel Bath
wall into the steel bath and hence kC is greater in that time.
But if it stays below 2.5 % for about 2 minutes, it can be
 The characteristic in Fig. 7. is typical for the arc covered by
foaming slag and represents a very even arc behavior:
Before the neural network can be used as a classificator, it
has to be trained with several input characteristics.
First 60 characteristics per minute of one heat are selected as
a training set. Then the desired output has to be defined for
every member of this training set. This is done not only by
visually observing them (since the training set consists of
about 1900 characteristics), but also by means of the
estimations that can be made based on the indices shown in
the previous chapter. The goal is that the output has to be
between 0 and 1: the network should output larger values
(near 1) during the first, uneven stage of melting and smaller
(near 0) values after reaching the molten steel.
The output of the trained net and some of the corresponding
charakteristics are shown in Fig. 8:
Fig. 7. Foam Slag

Fig. 8. Output of the trained neural network

When observing the outputs of the three different methods This approach would also involve the estimation of the
described, it can be seen that they are quite similar, but there remaining time (TR) needed to melt the scrap between the
are also remarkable differences. Perhaps the most detailed electrode and the furnace wall. However, this estimation
information can be obtained from the partial distortions, but would concern a much smaller time interval of approximately
the neural network seems to give the clearest distinction of 1-2 minutes and could, therefore, be more precise.
the operating state of the furnace.
The proposed neural network could be used in this approach
to indicate the arriving at the steel bath. Since the change of
IV. APPLICATION EXAMPLE the transformer step has strong influence on every phase, the
average of the network outputs in the three phases should be
The requirements from the operation and the complex used. This average network output is shown in Fig. 9 together
optimization task resulting from them were described in the with the actual transformer steps that were set during the first
first chapter. heat measured:

An essential part of the optimization of the furnace operation

is the optimization of the control strategy, i.e. the fastest
possible adaptation of the controll set values to the actual
state of the melting process.
Due to the mechanical inertia the electrode control is not able
to follow the arc length variations or to compensate them. Its
purpose is to keep an average distance between the electrode
tip and the charge. Therefore, the optimization deals only
with the determination of the set values for this average
distance.

In the case of the examined furnace the set value is

determined mainly by the tap position of the furnace
transformer, and can be slightly modified by a multiplier.
At the beginning of the melt-down the furnace is run with
transformer tap 10. Approximately, 1 minute after the start
the tap position is changed to 13, which means greater
melting power and longer arcs.
The next change of the transformer tap position from 13 to 12 Fig. 9. First Heat
aims the protection of the furnace lining. This change should
be made when there’s no scrap material left between (Transformer step 11 is reserved for the interventions of the
electrode and furnace wall. smelter.)
Unfortunately, there is no method to recognize this point of Based on the previous considerations a threshold for the net
time by means of electrical parameters, since in the last output of approx. 0.3 can be defined with following purpose:
melting phase the remaining scrap isn’t being melt under the If the average net output becomes smaller than this value, the
electrodes directly by the arc, but is smelting directly in the remaining time could be calculated and after that time the tap
hot liquid bath. changing should be performed.
That’s why a time factor is involved: from the weight of a Figures 9-12 demonstrate the possible usage of the proposed
basket the energy needed to melt it is approximated. Two method concerning tap change times. Without knowing the
minutes before reaching this energy the tap change is exact value of TR, one can see the differences between the
performed. proposed method and the tap change strategy being used.
The approximated melting time of the examined EAF had to Assuming TR being always zero, the first tap change from 13
be corrected during the measurement in 17 from 18 baskets, to 12 during the first heat (Fig. 9.) would then occur
the correction times were between +1 and –3 minutes. (It’s immediately after 400 sec, i.e. approx. 1 minute earlier than
remarkable that the correction times were always positive with the actual method, and this would mean a much better
during the last basket of a heat, and negative before. This protection of the furnace lining. The second tap change from
could indicate that the recharging is performed deliberately 13 to 12 during the second basket would take place approx. 1
before melting the total amount of charge in a basket. This minute later, thus causing a faster melting, and a higher
would mean an effective way for protecting the furnace productivity.
lining, but only during the first baskets.) Assuming TR=1 min. (always the same), there wouldn’t be
any difference between the the times of the first tap position
Another approach would be to perform the tap change when changes in the first basket, but the time differences in the
the arcs have already burnt a certain amount of time on the second basket would be increased.
molten steel surface.
 Similar comparisons can be made observing data recorded
during further heats shown in figures 10-12.
The exact determination of the threshold and of the remaining
time should be considered in a future work, where parameters
like temperatures and the foaming slag index kB also should
be taken into account.

V. CONCLUSIONS

A neural network based determination of the actual melting

state was proposed. The results obtained are compared with
those gained from the smoothened standard deviations
method and the partial harmonic distortions method. This
comparison shows, that the developed neural network is a
promising tool and could be used further for setting the
furnace transformer step in order to cope with the conflicting
Fig. 10. Second Heat objectives of high melting power and the protection of the
furnace lining.

VI. REFERENCES

[1] Grigat, R.R.: Messung und Modellbildung

elektrischer Lichtbogengrößen in Drehstrom-
Lichtbogenöfen, Dissertation zur Erlangung des
akad. Grades Dr-Ing., Universität der Bundeswehr
Hamburg, Hamburg, 1986
[2] Dán, A.M., Mohácsi, A.: Computer Simulation of a
Three Phase A.C. Electric Arc Furnace and ist
Reactive Power Compensation, Proceedings of IEEE
ICHPS VI, Bologna, 1994
[3] Knoop, M., Köhle, S.: Schwankung der elektrischen
Größen von Drehstrom-Lichtbogenöfen,
Elektrowärme International, 54 (1996) B 1, March,
pp. 32-39
[7] Nauck, D.,Klawonn, F.,Kruse, R.: Neuronale Netze
Fig. 11. Third Heat und Fuzzy-Systeme, Grundlagen des
Konnektionismus, Neuronaler Fuzzy-Systeme und
Kopplung mit wissensbasierten Methoden,Vieweg &
Sohn GmbH, Braunschweig/Wiesbaden, 1996

VII. BIOGRAPHIES

David Raisz was born in Budapest, Hungary in 1977. He graduated from the
Department of Power Systems of the University of Technology and
Economics Budapest, Hungary, in June 2000 after a one year research stay at
Department for Electrical Power Systems of University of Technology Graz,
Austria.
He is now working towards his Ph.D. degree in Electrical Engineering.

Manfred Sakulin was born in Bad Ischl, Austria, in 1944. He is professor

and senior researcher at the Department for Electrical Power Systems at
Graz, University of Technology. His main activities are in the fields of
energy efficiency, new technologies and power quality.
Fig. 12. Fourth Heat