Download as pdf or txt
Download as pdf or txt
You are on page 1of 6

Modeling of the AC-EAF process using computational intelligence methods

Tadeusz Wieczorek, Krystian Mączka


Silesian University of Technology, Department of Electrotechnology
8 Krasińskiego Str., 40-019 Katowice, E-mail: tadeusz.wieczorek@polsl.pl

I. INTRODUCTION highest possible power with a low variance. The amount of


electrical energy transferred to the steel melt should be
Electric arc furnaces (EAFs) are widely used in steelmaking optimally maintained during the entire melting and refining
and in smelting of nonferrous metals. The EAF is the central process, with the optimal distribution of radiated heat within
process of the so-called mini-mills, which produce steel the furnace.
mainly from scrap. Typical EAFs operate at power levels
from 10MVA to 130MVA, which is directly related to
production throughput. The production of steel by EAFs has
steadily expanded during the past few decades as well as
furnace capacity and power level. The electric-arc
steelmaking process is usually carried out in three stages: the
melt process (EAF), the ladle furnace process (LHF or VAD)
and the casting process (the second and third stage of the
process will not be considered in the paper).
Scrap is melted in AC electric arc furnaces, so called the
UHP furnaces, which serve as melting units only, because the
liquefying must be done in a time as short as possible. Scrap,
alloys and fluxes are charged into the UHP furnace by baskets
or by means of charging barrows. After charging, the furnace
is closed to begin with the meltdown period (fig. 1). The Fig. 1. EAF – General scheme of furnace
liquefying of the charged materials is performed by electric
energy but this process stage might be supported by the use of Control of the EAF process is carried out by
wall and door burners in order to deliver energy to areas determination of electrical working points (set points). It takes
which cannot be reached by the electric arc. The process stage place through step-wise adjustment of the output voltage of
is repeated by discharging additional baskets into the furnace the furnace transformer and through an adjustment of the
and melting them until enough liquid steel is available. After control value. The determination of the set points is based on
meltdown, a sample is taken from the furnace and the practical experience with the EAF processes. These methods
temperature is measured. The steel sample is analyzed in the of determining the set points are never completely accurate
laboratory and according to the results of the analysis, some and have got a static character. A number of non-linear
fluxes and alloys are added into the furnace or into the tapping influences and the temporary change in the process must be
stream during the tapping process. After that the liquid steel is taken into account. Due to the variability and complexity of
heated up to the required tapping temperature and additional the EAF process, accurate optimization must be based on
metallurgical tasks, such as decarburization by oxygen actual operating data, which are often noisy, and requiring
blowing or some kind of metallurgical treatment depend on significant pre-processing. The steelmaking process is
the steel grade to be produced are performed. Before tapping continually changing, as the furnace wears, procedures
of steel to the ladle, the deslagging of the bath in the arc change, and raw materials change. This requires such a data-
furnace is done, through the opened furnace door. After based modeling system which is adaptive. Computational
deslagging the liquid steel is tapped into the steel ladle, and
intelligence1 methods (CI) can take over these difficulties,
the EAF process starts again.
and therefore the methods are used for solving this kind of
The basic operating principle of this furnace, whether of the
modeling problems [13], [14].
AC or the DC type, is that an electric arc is struck from one or
The paper provides methods for modeling of the EAF
more electrodes to a metallic charge, which is usually steel
steelmaking process, particularly intelligent modeling. We
scrap. With this type of charge, the operation of the furnace is
will provide a literary review of the publications on the
unstable, especially during the initial meltdown period, that
problem of CI application to the modeling of the steelmaking
can result in significant electrical disturbances. Due to
electric-arc furnaces (section 2). Section 3 presents our
stochastic behavior of the EAF load and severity of the
approach to find a good model of the process, including data
disturbances, there has been an ongoing need for accurate and
preprocessing and feature extraction. Final section 4 presents
reliable characteristic predictors. The prediction problem is
discussion on obtained results and conclusions about building
further complicated by the facts that the EAF characteristics
precise models.
are essentially nonlinear and the fluctuations of the voltage
amplitude are not only random but depend on the wearing of
the furnace and its mode of operation. The known 1 IEEE Computational Intelligence Society defines its subjects of interest as
conventional control and optimization methods are not able to neural networks, fuzzy systems and evolutionary computation, including swarm
solve this task. It is very important to control the EAF at the intelligence [4].
II. MODELING OF AN ELECTRIC ARC FURNACE PROCESS allowing it to optimize the power and provide for maximum
arc power by maintaining a stable long arc. The overall goal
The literature on the modeling and control of EAF of such an optimization is the dynamic generation of reference
processes is very extensive due to both the commercial value values for the electrode position such that at any instant, a
of obtained practical results and the variety of engineering maximum level of active power can be transferred into the
applications of electric arc energy. Early papers discuss furnace with an optimum distribution of radiated heat within
simple models involving only some chosen groups of the furnace [6]. Additionally the cooling segment
phenomena concerned with EAF, e.g. electric equivalent temperatures are continuously monitored, a corresponding
circuits, heat and mass transfer, chemical reactions. A closed temperature distribution profile is generated in the neural
loop control system individually adjusts the three electrode model, and the intensity of the arc(s) closest to the endangered
heights above the steel surface (equals the arc lengths) so that spot is reduced by moving the relevant electrode. Only after
a constant arc voltage or a constant system impedance is this control span is exhausted, the tap on the transformer
maintained. lowered – and raised again as soon as the wall panel
Considering the interactive three phase system, and the temperature permits. This added load control results not only
dynamic operation of the three arcs, it becomes clear that a set in an optimized power transfer to the furnace, reducing the
of static setpoints can never yield an optimal energy efficient tap-to-tap time, but also prevents premature refractory wear
process. Some papers discuss the same methods but for the and cooling panel damage, significantly increasing the
purposes of dynamic control [1], [3]. Attempts to model the availability of the furnace and thus its productivity. Perhaps
dynamic asymmetries of the three arcs have largely failed – most important, the ratio of productivity to power-on time can
mainly because the three arcs form a coupled system, and all be improved.
electrical parameters fluctuate with the dynamic behavior of Pappe et.al. described a system in which an analytical
the arc – and thus the vast majority of AC EAFs are controlled model was combined with a neural network to form a hybrid
in the static way. model. Implementation of the system at Krupp’s EAF plant
But classical modeling based on differential or integral reduced energy and raw materials requirements [7].
equations, equilibrium chemistry, and material and energy An intelligent optimal control system used in the
balances does not lead to required results. Precise modeling of steelmaking process of the EAF is presented by Gao et.al. A
three, asymmetrical arcs with process-dependent, nonlinear hybrid real-time intelligent control strategy was set up to
characteristics (such as highly variable reactances) is not realize the close-loop control of the arc furnace [5]. First, the
possible because number of nonlinear influences and systems based on expert knowledge were proposed in China
temporary changes in the process must be taken into account. by Shanghai University, in 1992 as well as an intelligent
Therefore, an interest in computational intelligence methods control of electrode and adaptive predictive control were
has grown. presented by Beijing University, in 1994. All the solutions
W.E. Staib first applied the methods, developing so-called were invented for AC EAF. In the Gao’s paper, a set-point
“Intelligent Arc Furnace (IAF™)” in 1992 [10]. Staib intelligent control system for UHP DC EAF is presented.
assumed that control over electric arcs is crucial for effective Electric energy had been reduced by 10%, electrode
furnace operation. Existing electrode regulators are based on consumption by 28%, tap-to-tap time reduced by 10 min.
an impedance-model and enable many facets of furnace to be The neural network supported system for optimal control of
controlled. Therefore, he researched the application of energy input into the EAF has been described by [6]. The
Artificial Neural Networks (ANN) to closed-loop control of system developed by Siemens AG is based on neural
an electrode positioning subsystem. Some rule-based systems electrode and energy control. It has been installed in
have been developed to establish optimum set points during Stahlwerk Bous GmbH (Germany), in Benteler AG,
the heat. The set points are given to the existing regulator of Stahlwerk Lingen (Germany), Krupp Nirosta Bochum
the furnace, tracing its operating program. Application of (Germany). The neural network based control system, called
Artificial Neural Networks could overcome all the limitations SIMELT®NEC brings reduction of costs and increasing of
of traditional regulators. For proper learning of a neural production. The key of the solution is a complete electrical
system, large amounts of accurate data are necessary. The three-phase model of the furnace supported by neural
IAF™ uses three neural models, these are: the Regulator networks. The existing electrode control system is extended
Emulator, the Furnace Emulator and the Neural Furnace by a neural optimization control system that provides the
Controller [13]. dynamic corrections for the set point of the electrode control
Results of applications of the IAF™ systems in American system. When the optimization system is activated, instead of
steelworks were very positive [2]. Therefore, European the set points being taken from a table as a function of the
companies intensified research on CI methods for the EAF given transformer tap, they are transmitted from the model in
process. Many papers on optimization of EAFs with the use of real time. Siemens solution utilizes a hybrid model in the
neural networks were published, e.g. [6]. Sesselmann et.al. [9] optimization concept. This intelligent approach has the
developed a system that constantly optimizes the set point for advantage of combining a physical model of the three-phase
the electrode control so that the three-phase EAF can operate AC electric-arc furnace with a neural network capable of
with the maximum constant energy consumption. The learning.
developed optimization method was based on an adaptive ANNs based models for prediction of the electric energy
neural network. The system provided an adaptive control consumption and end-of-heat in electric arc furnaces were
which learned the characteristics of the furnace as it operated, implemented by [1]. Besides the approach using a time-series
model, a new dynamic process model based on series of signals being stored in the database which describe actual
interconnected ANNs has been developed. A new method, process of metal scrap meltdown in the electric arc furnace.
called by authors, “error-feed-through” technique has been There are, amongst other: primary and secondary voltage,
used. Improving process control and gains in energy, time and primary and secondary current, transformer tap position,
productivity have been obtained for the arc furnace in electric and gas energy consumption, power coefficient,
ProfilARBED-Esch-Belval. positions of electrodes etc. Each signal is sampled once per
Sadeghian and Lavers [8] propose and apply feedforward second. Due to such a great amount of data, proper analysis
neuro-fuzzy multi-step predictors. Based on the principles of and preprocessing of data have to be done.
neural networks and fuzzy logic inference mechanism, the
proposed feedforward predictor implements the fuzzy rule- A. Building data verification models
base using data clustering techniques, and adjusts the
parameters of the rules using neural network adaptive All computational intelligence systems need
capabilities. The network learns to predict the v-i confirmed data for a proper running. False data can have a
characteristics of the particular EAF, whose measurements negative impact on CI systems predictions. In the real
have been used to build the neuro-fuzzy network. Attempts industrial data bases, there are many incorrect and missing
are also made to show how accurate and fast EAF prediction values or other false registered data. Therefore, verification of
can be accomplished using adaptive fuzzy logic techniques. real process data is very important.
The underlying idea is that by identifying the proper During analyzing of the collected data, it was noticed
input/output structure, and using a suitable learning algorithm, that some values of two variables: oxygen consumption and
a well-structured feedforward neuro-fuzzy system has the burners gas consumption in some cases diverged much from
capability of predicting the dynamics of the electric arc their mean values. Therefore, it was necessary to build such a
furnace. The implementation of the feedforward neuro-fuzzy model which could verify correctness of this data. Melting
network together with its training and validation using the metal scrap in the electric arc furnace is performed according
actual recorded data are highlighted [8]. The CI methods were to the melting program which is based on the consumption of
used for predictive control of an electric arc furnace off-gas electric energy.
process. In the literature, there is a discussion on a chaotic The first step was to identify the duration of
EAF model. The concept of chaotic parameters, such as particular steps in the melting program (fig. 2). In each step,
chaotic resistance, inductance or admittance has been also burners and oxygen lances operate with the known (specified
proposed for the characterization of arc furnace operation. by producer) flow.
Number of authors who have modeled the EAF steelmaking
process have obtained applicable practical results. Basing on
the presented state of the art we conclude, that intelligent
modeling using CI methods seems to be an ideal solution for
complex and non-linearly related EAF process. The EAF
requires a controlling, electrical model of the furnace that is
continuously adapted to the changing conditions. Adaptive
control is the only way for this class of nonlinear, dynamic
systems, for which determination of the uncertainty in the
dynamics is either unknown or impossible.

III. EXPERIMENT AND RESULTS

For verification of the possibility of CI application for the


EAF process modeling, data of about 2700 melts have been
collected, for some grades of steel, made in one of the Polish Fig. 2. Electric energy consumption during EAF process
steelworks.
The first task was to build models of lances oxygen Identification of duration of individual steps made it possible
consumption and burners gas consumption during the melting to calculate oxygen and gas consumption. This approach was
process. The aim of the second task – intelligent modeling – called by the authors energy curve algorithm.
was to predict the electric energy needed for the whole The second estimation method for the consumption
process of melting metal scrap and to predict the electric of oxygen lances was based on the observation that oxygen
energy needed for the last step – refining. As a neural model injection has an impact on stabilizing electric arc. That is
of the electric energy calculation, a multilayer perceptron has caused by the fact that oxygen injection to the EAF makes
been used. slag more foamy, which covers the arc. The covered arc burns
Data used for computations was collected in the real process in a more stable way. The increase of plasma gases also
of steel production in one of the Polish steelworks during EAF stabilizes the arc. In the figure 3, secondary current with the
process. In this steelworks, for controlling of process periods of calm arc (SpT1, SpT2, SpT3) is presented.
parameters, programmable logic controllers are used (Simatic-
5 and Simatic-7). Signals from the controllers are stored in
MS SQL 2005 Server database. There are over two hundred
furnace. By identification of duration of the calm arc period, it
was possible to calculate consumption of oxygen during the
process. This approach was called by the authors low variance
algorithm. Table I presents the results for above mentioned
methods of calculating gas and oxygen consumption, where
quality of the model means (MSE)-1.

Table I
RESULTS OF THE OBTAINED DATA VERIFICATION MODELS

Model Quality of the Standard


model deviation
Oxygen - energy curve
algorithm 0,867 0,23
Oxygen - low variance
model 0,807 0,33
Fig. 3. Secondary current of the firs electrode during EAF process Burners gas - energy
curve algorithm 0,733 0,34
Identification of calm arc periods turned out not to be easy
because the melting in the EAF is a very unstable process. As The greatest error was made by model which was estimating
it is shown in figure 3, values of current are very irregular. To burners gas consumption. However, during examining all
simplify the analysis of these irregular values, simple moving results, it is necessary to take error distribution into
average (SMA, eq. 1) of the following n=50 data points was consideration. As to estimating gas consumption (fig. 5), in
used: most cases (about 1400 instances), error range of the model
was between -20% and -30%, which it could mean that
specified by the producer values of the burners flows, which
xi + xi+1 + xi+2 ... + xi+n
SMA = (1) were used for calculation, were underestimated. Results of
n the model can be easily corrected by the use of proper values
of the flows.
Figure 4 presents the chart of secondary current after applying
SMA transformation. After that operation, smoothed data was
used for estimation of current values irregularity during
melting process.

Fig. 5. Mean error distribution for the energy curve algorithm for burners gas
consumption

The error distribution for models estimating oxygen


Fig. 4. Secondary current after applying SMA transformation. consumption shows (fig. 6). Results of these two models are
similar. However, model based on low variance algorithm has
The index, which can estimate variability is running variance got a greater error than the one based on energy curve
(eq. 2). In a similar way, like for SMA, running variance of algorithm. This can be explained by the fact that there is a
the following n=50 data points was calculated. possibility of existence a delay between the start of oxygen
injection and the start of calm arc period, which was not taken
into consideration in this approach.
( xi − xn ) 2 + ( xi +1 − xn ) 2 + ... + ( xi + n − xn ) 2
RV = (2)
n
After calculating RV for the whole dataset, a threshold
δ = 0,25 was defined on the basis of the analysis of running
variance charts. It is recognized that below the threshold, calm
arc period begins. It means that oxygen was injected into the
were performed in the ten fold cross-validation. The dataset
consisted of 24 features describing melting process such as:
the time durations of individual basket charge meltdown, time
durations of melting breaks, the amount of oxygen, carbon,
gas, lime used in the process, weight of each basket, weight of
tapped steel, liquid steel temperature.
For that experiment, multilayer perceptron (MLP)
was chosen. Various network configurations were tested in
order to obtain the least prediction error.
Various amounts of neurons in the hidden layer and various
activation functions were tested. The set of five best networks
and their results for predicting electric energy needed for the
whole process is presented in table II.

Table II
RESULTS OF PREDICTION OF ELECTRIC ENERGY
CONSUMPTION FOR THE WHOLE PROCESS
Fig. 6. Mean error distribution for the energy curve algorithm and the low
variance algorithm for lances oxygen consumption Type of Quality Quality Activation Activation
network (learning) (testing) Function Function
(hidden layer) (output layer)
The models presented in this chapter contributed to MLP 0,926 0,903 Hyperbolic
Linear
verification of correctness of the two uncertain variables, 24-13-1 ±0,002 ±0,002 tangent
which was extremely important before the next phase of MLP 0,930 0,900 Hyperbolic
Exponential
calculations – learning neural network. In addition, features 24-11-1 ±0,001 ±0,002 tangent
MLP 0,920 0,897 Hyperbolic
which were extracted during that calculations enabled to 24-8-1 ±0,002 ±0,002 tangent
Exponential
obtain additional information about a course of the process MLP 0,918 0,897
such as: time durations of every melting program steps and Logistic Linear
24-7-1 ±0,002 ±0,002
time durations of calm arc periods. The extracted features MLP 0,921 0,895
Exponential
Hyperbolic
were used for learning the neural network. 24-8-1 ±0,002 ±0,002 tangent

B. Predicting energy consumption by the use of neural According to the melting program, the end of melting of each
networks basket should always be finished for the same value of unit
consumption of electric energy. Therefore, an attempt was
Correct prediction of electric energy consumption made to train neural network to predict electric energy needed
can contribute to reduction of melting time. If an operator of only for the refining. The obtained results for this approach
the furnace knew the energy which must be used to melt scrap are presented in table III.
metal completely and which acquire proper temperature of the
Table III
liquid metal bath, he could finish the process earlier without
RESULTS OF PREDICTION OF ELECTRIC ENERGY
wasting the time and energy. CONSUMPTION FOR THE REFINING
To properly predict the electric energy, needed in the process,
the neural network was supposed to learn from the subset of Type of Quality Quality Activation Activation
melts, which were the most typical. To select this subset of network (learning) (testing) Function Function
typical melts by rejection the cases in which some values were (hidden layer) (output layer)
MLP 0,965 0,962
extremely different from mean values, outliers analysis was 24-11-1 ± 0,001 ± 0,001
Exponential Linear
done. This operation was performer by rejection of all melts MLP 0,969 0,961
in which values of some features were greater than outmax or Exponential Linear
24-12-1 ± 0,001 ± 0,001
less than outmin (eq. 3): MLP 0,971 0,961
Exponential Exponential
24-9-1 ± 0,001 ± 0,001
MLP 0,967 0,961 Hyperbolic
out max = Q3 + 1,5 ⋅ IQR, 24-5-1 ± 0,001 ± 0,001 tangent
Linear
(3)
out min = Q1 − 1,5 ⋅ IQR MLP 0,970 0,961
Logistic Exponential
24-11-1 ± 0,001 ± 0,001

where, Q1 and Q3 are first and third quartiles and IQR=Q3-Q1. Neural networks used to predict electric energy,
After this preprocessing, about 1200 instances were rejected. needed for the whole process, achieved greater error than the
Another rejection criterion was an inconsistency with ones which predicted energy needed only for the refining. It
the melting program which is based on the consumption of means that analyzed features, which describe melting process,
electric energy. In this step of preprocessing, 1200 instances are more correlated with energy of the refining than with
were rejected because of major exception to the process rules. energy of the whole process.
Finally, for training the neural network, about 1300 instances
were chosen. The dataset was divided into two subsets:
training and testing subset. Learning and testing processes
IV. DISCUSSION OF THE OBTAINED RESULTS AND CONCLUSIONS
[5] Gao X., Li S., Chai T., Shao C., Wang X., Set point intelligent optimal
control of electric arc furnace. Proc. of 2nd Asian Control Conf., Seul,
Modeling and control of complex, modern, industrial 1997, pp. 763-766.
processes, as EAF steelmaking process, often exceeds the
capabilities of conventional modeling techniques, based on [6] Gerling R., Louis T., Schmeiduch G., Sesselmann R., Sieber A., Optimizing
the melting process at AC-EAF with neural networks. Metall. Jg.53. No 7-
differential or integral equations, equilibrium chemistry, and 8, 1999, pp.410-418.
material and energy balances. Nonlinear transformations,
parallel processing and adaptive capabilities make the CI [7]. Pappe T., Obradovic D., Schlang M., Neural networks: reducing energy
and raw materials requirements. Simens Review, Fall 1995, pp.24-27.
methods the most promising for modeling such processes.
Both verification of values which are extremely different [8] Sadeghian A. R., Lavers J. D., Application of Feedfoward Neuro-Fuzzy
from the mean ones and supplement of the missing values are Networks for Current Prediction in Electric Arc Furnaces. Proc. of IJCNN
vol.4, 2000, 420-428.
very significant for CI systems, which can not obtain
satisfactory results if the data is vague or missing. [9] Sesselmann R., Wahlers F-J., Zocher H., Poppe T., Optimization of the
Models presented in the paper enabled verification of real electrode control system with neural network. V European Electric Steel
industrial data coming from the EAF process and supplement Congress, Paris, (1995) pp. 260-271.
of missing values with the values calculated by the model. [10] Staib W.E., The intelligent arc furnace. Steel Technology Int., Sterling Publ.
Achieving additional knowledge about the process by London, (1993), pp.103-107.
feature extraction allowed to extend the analysis of the
[11] Wieczorek T., Pyka M., Hybrid expert system for arc-electric steelmaking
process and new features such as: time durations of every process monitoring. Acta Metallurgica Slovaca, 10, 2004, pp.408-412.
melting program steps and duration of calm arc periods were
used to train CI systems. In that case, it was a MLP neural [12] Wieczorek T., Pyka M., Neural modeling of the arc-electric steelmaking
network. process. Proc. of IXth Int. Conf. „Research in Electrotechnology and
Applied Informatics” Katowice 2005, pp.105-108.
The number of rejected instances from the analysis points
that it is necessary to build another neural network model [13] Wieczorek T.: Intelligent control of the electric-arc steelmaking process
which will properly predict process parameters for the melts, using artificial neural networks. Computer Methods in Material Science,
vol.6, No 1, 2006, 9-14.
which in this model were classified as untypical.
Every heat in the EAF is different from another one. [14] Wieczorek T., Blachnik M., Mączka K., Building model for time reduction
Therefore, the system should be very flexible and adaptive. of steel scrap meltdown in the electric arc furnace. Lecture Notes in
Effective application of an intelligent system demands Computer Science, Springer-Verlag Berlin Heidelberg, 2008.
understanding of the steelmaker’s needs, analysis of the noise
level present in the data transferred from the EAF furnace,
and implementation and evaluation of several process models
and optimization approaches. Therefore, the intelligent
systems are often difficult to apply. Such advantages like
nonlinearity, online adaptability, generality and distributed
parallel processing ability make technology the most
developing.
Moreover, practical application of CI methods for control
and modeling of EAF processes is determined by evaluation
of costs and benefits. While making decision about control
system, we have to take into considerations the following:
furnace equipment, metallurgical procedures, scrap and other
charge materials characteristics as well as shift work.

REFERENCES

[1] Baumert J-C., Engel R., Weiler C., Dynamic modeling of the electric arc
furnace process using artificial neural networks. La Revue de Metallurgie-
CIT, no.10, 2002, pp. 839-849.

[2] Bliss N. G, Gilbertson G J, Sweet J.E., Neural networks for the high
voltage EAF. V European Electric Steel Congress, Paris, (1995) pp. 240-
247.

[3] Boulet B., Lalli G., Ajersch M. (2003), Modeling and control of an electric
arc furnace. Proc. of the American Control Conf., Denver, Colorado,
IEEE Press, pp. 3060-3064.

[4] Duch W., What is computational intelligence and where it is going. In:
Duch W., Mandziuk J. (Eds). Challenges for Computational Intelligence.
Springer-Verlag Berlin Heidelberg 2007, 9-21.

You might also like