- Email: [email protected]
Online estimation of electric arc furnace tap temperature by using fuzzy neural networks
ARTICLE IN PRESS
Engineering Applications of Artificial Intelligence 21 (2008) 1001–1012 www.elsevier.com/locate/engappai
Online estimation of electric arc furnace tap temperature by using fuzzy neural networks Jose´ Manuel Mesa Ferna´ndez, Valeriano A´lvarez Cabal, Vicente Rodrı´ guez Montequin, Joaquı´ n Villanueva Balsera University of Oviedo, Project Engineering Area, c/Independencia, 13, 33004 Oviedo, Asturias, Spain Received 2 May 2007; received in revised form 3 October 2007; accepted 20 November 2007 Available online 8 January 2008
Abstract Industrial factories require continuous analysis and reengineering over its production processes but always keeping a severe control of material costs and operation emissions. In the past electric steel mills have been subject of some operation models developed in order to improve the control of the arc furnace by means of mathematical techniques and, later on, with finite elements technique (FEM). However, these models have not reached the expected results and applicability. In this case, a model has been developed that allows improving the control through a better prediction of the final temperature and, as consequence, to reduce the consumption of energy in the electric arc furnace. Required information for this new model will be obtained gathering knowledge collected up from data obtained of a certain electric furnace and also considering the plant operators and technicians experience. The model has been constructed by using neural networks as classifier, and with a final fuzzy inference function to return a predicted temperature value. r 2007 Elsevier Ltd. All rights reserved. Keywords: Energy consumption; Electric arc furnace; Neural networks; Fuzzy inference function; Temperature prediction
1. Introduction Nowadays there are two main industrial processes to produce steel. The first one, which is known as integrated steel plant, produces steel by refining iron ore in several steps. This ore-based process uses a blast furnace. The other one, steel-making from scrap metals involves melting scrap metal, removing impurities and casting it into the desired shapes. Although originally the steel production in the electric arc furnaces (EAFs) was applied mainly to the special steel grades, the situation has changed with taps size increase and the high productivity that has been reached progressively. This has allowed significant cost reduction, diminishing consumption of energy, electrodes and refractory. At present, electric furnace combined with secondary metallurgy allows to make a very important part of the worldwide steel production on the basis of the massive recycling Corresponding author. Tel.: +34 985 10 43 48; fax: +34 985 10 42 56.
E-mail address: [email protected] (J.M.M. Mesa Ferna´ndez). 0952-1976/$ - see front matter r 2007 Elsevier Ltd. All rights reserved. doi:10.1016/j.engappai.2007.11.008
of the iron scrap. In fact, the share of electric furnace steelmaking technology has reached 33% in world crude steel production and this trend continues although permanently altered by international scrap market. This trend is due of both the capability of this route to produce steel at low cost per unit and to constant improvements in steel quality. According to several studies (Birat, 2000), in an immediate future this tendency will remain, predicting an increase from present 33% to 40% by the year 2010 (Table 1). 1.1. Electric arc furnace process overview EAFs are used to melt scrap by means of electric energy and oxygen. Production of steel from scrap can also be economical on a smaller scale. EAF technology is most often used within steel minimills, whose basic design includes one or two EAFs, a continuous casting and a rolling mill. EAFs produce steel by melting scrap using electric supply as the main energy source. Scrap is generated from the scrapping of capital equipment or as a by-product of the steel-making process. Small quantities
ARTICLE IN PRESS 1002
J.M.M. Mesa Ferna´ndez et al. / Engineering Applications of Artificial Intelligence 21 (2008) 1001–1012
of direct reduced iron and pig iron are used as scrap substitutes or in the higher quality steel. By oxygen blowing, the concentrations of carbon and silicon can be reduced, while providing an additional energy source. The EAF operating cycle is called the tap-to-tap and starts when the furnace roof is initially open allowing to charge the scrap. The solid scrap is charged on top of liquid steel (a ‘‘hot heel’’), which was retained from the previous tap. Then electrodes are lowered into the furnace and the heat melts the charge due to electric resistance of the metal and radiation from the arc. Then electrodes are moved down into the furnace and the heat melts the charge due to electric resistance of the metal and radiation from the arc. Chemical energy is also supplied to melt the furnace charge. Oxygen is blown into the molten steel through one or more lances. The meltdown period is followed by another period in which the impurity levels in the melt are decreased. Also in this period the foamy slag layer is formed and its function
Table 1 Steel production prevision
World steel production (MTm/year) BOF (basic oxygen furnace) (%) EAF (electric arc furnace) (%) Another processes (%)
2000
2010
760 57 33 10
850 50 40 5.5
(+20/ 30) (+5/ 0) (+5/ 0) (+5/ 0)
is to absorb those impurities (adding lime and dolomite to modify its composition) and to shield furnace walls. Foaming occurs by injecting graphite into the slag layer and due to gas bubbles moving through. Once the scrap charge is fully melted, flat bath conditions are reached. At this point, a tap temperature and a steel sample are taken. Once the expected steel composition and temperature are achieved to further processing (secondary steel-making) the tap-hole is opened, the furnace is tilted, and the steel poured into a ladle. 1.2. Improving EAF productivity In a complex industrial system as an EAF, the ways to improve this process are diverse, such as selection and treatment of materials, the redesign of the facilities and new kinds of energetic contributions to the process. The evolution of electric furnaces is reflected mainly in the progressive reduction of specific consumption of energy, tap-to-tap times (Fig. 1) (Astigarraga, 1998) and improvement of the metallic yield. A deep study, comparing energy consumed currently in the production of steel with theoretically needed energy, was made by the Carnegie Mellon University (Pittsburg, USA) for the US Department of Energy (Fruehan et al., 2000). Though the direct comparison of different EAFs is difficult due to the operational differences, the disparity between the theoretical energy and the consumed energy contributes to the idea of potential improvement of these
Fig. 1. EAF process time evolution.
ARTICLE IN PRESS J.M.M. Mesa Ferna´ndez et al. / Engineering Applications of Artificial Intelligence 21 (2008) 1001–1012
facilities which could be about 25% of the energy consumption per ton. Table 2, extracted from Steel Industry Technology Roadmap (Energetics, 2003) and published by the American Iron and Steel Institute, lists four major steel industry unit operations and their estimated yield losses. It presents the targets for reducing these losses through a research and development program and shows the reduction in energy intensity resulting from achieving these targets. One of the pathways proposed to reach this yield improvement in EAF operation is the optimization of process sequencing by mean or artificial intelligent techniques. Important effort has been directed towards further productivity improvements under EAF process. Much of this has focused on developing alternative energy sources to reduce the high cost of electrical energy, which means 15% of EAF operation costs (Fig. 2). Improvements involving the use of cheaper forms of energy such as coal and oxygen; however, it has been achieved limited success
Table 2 Steel industry yield losses and targets Unit operation
Estimated yield loss (%)
Yield loss target (% reduction)
Energy savings targeta (MJ)
Iron makingb BOF steel-makingc EAF steel-makingc Finishing operationsd Applications and material propertiese
2–6 7–9 6–8 1 19
25 33 33 33 50
100–200 200–300 200–300 100 1800
a
Reduction in energy intensity that will result from achieving the corresponding yield loss target. b Includes tapping, metal handling, skimming, and desulphurization. c Includes ladle refining and casting. d Includes hot and cold rolling, coating, scarfing, etc. e Based on 14 million tons of prompt scrap plus 5 million tons reduced production resulting from improved properties.
1003
because of the cost of additional equipment (Mapelli and Baragiola, 2006). Another two different types of costs that have impact on the total cost of EAF operation can be identified: the cost of the feed materials and the cost implication of not reaching control objectives. Feed materials are very conditioned by the final quality of steel to obtain (scrap substitutes are mainly used in the production of higher quality products) and the steel cost due to materials is mainly determined by the price of scrap in the world-wide market. Processing control optimization is another very important way for the reduction of the energy consumption. Since no suitable plant model was available, the operations in the furnace were only based on empirical knowledge. There are a large number of traditionally manually controlled variables. The furnace operator, in accordance with his own experience and in his particular way of working, has been taking the decisions. For example, he was deciding if it was necessary to inject more or less oxygen, coal or HBI (hot briquetted iron), or even stopping the process and measuring the steel temperature. The automation of these variables may result in EAFs being operated more efficiently. However, the development of a mathematical model of the process is extremely problematic because of the extreme conditions in the furnace. The mathematical representation of the physical and chemical phenomena (electrical arc, materials flow, chemical reactions, etc.) that take place inside the furnaces needs, for its resolution, some hypotheses and simplifications that can limit its application enormously. For a long time, operations in the furnace were only based on empirical knowledge. On the other hand, the considerable design and operative differences in facilities make the generalization of any model very complicated. These difficulties cause less advances in process control than in other aspects. The furnace model development by means of classic mathematics or finite elements techniques (FEMs) to improve the control of the EAF and the minimization of energy consumption have not reached the expected success. After the appearance of the artificial intelligence techniques in the early 1990s, as new way of learning and modeling from data, these techniques have became the principal line of investigation in process control, in the metallurgical sector, as it shows the numerous projects developed by steel industry. The application of these techniques has turned out to be significant for physical deformations (Montequin et al., 2002; Ortega, 2003), but it still is in development for thermal models. 1.3. Proposed solution
Fig. 2. Operation cost rates in an EAF.
This paper proposes to optimize furnace control by means of a neural network approach, which improves the yield of EAF process by reducing electric energy consumption tap-to-tap times.
ARTICLE IN PRESS 1004
J.M.M. Mesa Ferna´ndez et al. / Engineering Applications of Artificial Intelligence 21 (2008) 1001–1012
Initially, a study of available process information will be done and control systems of EAF will be modified to allow storing all that information or variables that could be considered relevant. Also, an audio system will be installed for the acquisition of sound environment of furnace because several studies have shown that it can provide information about the heat state. Later, a process of data evaluation will be carried out, filtering incomplete or invalid cases. From that database, a selection of variables of the neural models will be made, including the output variable. The above-mentioned models will be trained by a wide dataset and validated by new data proceeding from furnace and not used in its training. The aim is that the final model is completely operative and can be used in EAF plant as main decision criteria and allow to determine when heat is ready to continue to the following phase of the steel production. 2. Data acquisition and problem analysis The development of a prediction model based on mathematical representation of physical and chemical processes in EAF furnaces needs, for its resolution, some hypotheses and simplifications that avoid its applicability in furnace operations as was mentioned in Section 1. An alternative way to improve furnace control is to collect, analyze and use empirical process data and, by means of knowledge-based techniques, build a tap prediction model. Data acquisition, modeling methodology and mathematical techniques used to construct that neural model are described in the next sections. 2.1. Data acquisition and pre-processing Modeling from empirical data requires the correct gathering of the most important process data, collected with enough accuracy and frequency. Measurements in EAFs are limited and quite affected by noise due to the environment conditions. Continuous changes in processes like feed materials alterations, different energy supplies or operation circumstances, require a model capable to adapt to these changes. The process data collected in a DC EAF for this purpose and stored in about 250 variables can be put into the next categories:
foaming, was detected in previous researches (Buydens et al., 1998; Vidacak et al., 2002; Yi and Kim, 2002). Those works studied especially the foaming slag phase and their conclusions were that the frequency range between 100 and 150 Hz was the best to detect that phase. In this work, the frequency is extended up to 20,000 Hz. The sound signal was processed by fast Fourier transformade (FFT) in order to store the amplitude values in the database. Data were cleaned and pre-processed using classical statistics, Sammon projections and other methods of visualization. A considerable number of incomplete or no valid cases were found in this process. 2.2. Output variable selection One important decision in model development is the selection of the prediction variable. That parameter must allow estimating in a most precise way when the steel is ready to go on refining phase. Furnace energy supply comes from the electrical power through the electrodes and from the chemical energy due to the combustion of the natural gas, oxidation of the coal and other elements. In addition, there are different losses of energy as thermal (through the refrigeration system, the refractory or exhaust gases), electrical, etc. Therefore, the power consumption, absolute or specific (by unit of raw material) could be the model estimation variable. Nevertheless, except the electrical consumption, whose final value is available, other parameters are not directly measurable, what would make difficult the validation of the results of the model. On the other hand, process time is another variable with which the furnace yield is usually quantified and it could also be the purpose of model prediction. Nevertheless, it was not used because the furnace stops for material load, whose duration is quite irregular, make its use difficult. This paper is focused in one of these variables, the steel tap temperature. To reach a valid temperature value is the main criteria (besides the carbon content) to stop the heat, but it is difficult to get an accurate measurement because:
variables related with raw materials (scrap, hbi, coal, etc.); time event and duration variables; electrical consumption and other chemical energy supplies like injected oxygen; the sound captured by a microphone staying next to the furnace; and temperature measurements.
Slag foaming is a steel-making process that has been shown to improve the efficiency of EAF plants (Karr and Wilson, 2005; Wilson et al., 2004). The connection between level sound and some phases or events of process, like slag
There is not a continuous sensor inside the furnace. Temperatures are measured with a pyrometer with a limited precision. There is not a uniform temperature in the furnace because its size, the difficult flow of materials and convection heat transmission. There is a lack of accuracy because of the own measurement method or operators errors.
The target temperature to melt materials, in the studied DC electrical arc furnace, is around 1620 1C although that value could be changed due to subsequent treatments. To overcome this temperature value supposes a waste of energy. Nevertheless, energy lost in cooling during the measurement delay is significant when the temperature is lower than target and furnace must be switched on again.
ARTICLE IN PRESS J.M.M. Mesa Ferna´ndez et al. / Engineering Applications of Artificial Intelligence 21 (2008) 1001–1012
In both cases there are energy losses and the tap-to-tap times are longer and therefore the process yield is reduced. A better temperature prediction model based in process parameters could improve a lot the process efficiency. One important decision in the model developed is to determine what should be the most appropriate prediction variable. That parameter must allow estimating more precisely when the steel is ready to go ahead with the refining phase. Furnace energy supply comes from the electrical power through the electrodes and from the chemical energy due the combustion of the natural gas, oxidation of the coal and other elements. In addition, there are different losses of energy as thermal (through the refrigeration system, the refractory or exhaust gases), electrical, etc. Therefore, the power consumption, absolute or specific (by unit of raw material) could be the model estimation variable. Nevertheless, except the electrical consumption, whose final value could be known, other parameters are not directly measurable, what would make difficult to validate the results produced by the model. On the other hand, process time is one more variable with which the furnace yield is usually quantified and it could also be the value that should be determined by the model. However, it was not used because the furnace stops for material load, whose duration is quite irregular, making its use tough. This paper is focused in one of these variables, the steel tap temperature. A primary criterion is to reach a valid
1005
temperature value (besides the carbon content) to stop the heat, but is arduous to obtain an accurate measurement because:
There is not a continuous sensor inside the furnace. Temperatures are measured with a pyrometer with a limited precision. There is not a uniform temperature in the furnace because its size, the materials flow and convection heat transmission. There is a lack of accuracy because of the own measurement method or operators errors.
The target temperature to melt materials, in the analyzed DC electrical arc furnace, is around 1620 1C although that value could be modified due subsequent treatments. To overcome this temperature value implies a waste of energy. Nevertheless, energy lost in cooling during the measurement delay is significant when the temperature is lower than the target and furnace must be switched on again. In both cases, there are energy losses, the tap-to-tap times are longer and therefore the process yield is reduced. A better temperature prediction model based in parameters of the process could improve a lot/could derive to an important improvement of the process efficiency. Fig. 3 shows the increase of the average electric consumption (E) when it is needed a second or third temperature measurement (a second or third stopping of furnace therefore).
Fig. 3. Electric consumption comparative for different number of temperature measurements.
ARTICLE IN PRESS 1006
J.M.M. Mesa Ferna´ndez et al. / Engineering Applications of Artificial Intelligence 21 (2008) 1001–1012
Fig. 4. Consecutive temperature difference.
The first temperature measurements were considered not valid to go through refining stage and these cases represent around 45% of the heats. Only a reduction of this rate can be a significant process improvement. Just a reduction of this rate can be a significant process improvement. A model capable to predict the tap temperature during the heat would be essential to introduce corrective actions immediately, with the consequent energy saving and optimization of the process. Also, the operator would be provided with a less subjective reason of process evolution based in acquired data to stop the heating and to measure tap temperature.
adjusted for its use in the studied furnace and therefore preliminary relevant variables for prediction were selected. Neural networks are successfully used for modeling, classifying and controlling complex systems, and a large number of training algorithms have been developed. Good performance to data noise, the possibility of training again to fit properly in changeable conditions and its capability to detect non-lineal relations become neural networks in an adequate choice to develop a temperature prediction model. Projects on EAF control have focused mainly on electrode control, where neural networks were used alone or in combination with other techniques. In a first stage, the method to develop a predictive temperature model was based in Perceptron multilayer neural networks (Rosenblatt, 1958) and can be summarized as follows:
2.3. Temperature data treatment As it was previously mentioned, in this research the tap temperature was selected as target variable of the prediction model to develop and it requires a special attention. The study of collected temperature data showed that variable contains very low values (down 1550 1C) and very high ones (up 1700 1C). Low values are due to stop the heat process before reaching the target value, and high ones, when the required temperature value is widely exceeded, are due to delay unnecessary process stopping. In both cases, the energy consumption is increased. Also, many temperature measurement errors were detected in data cleaning due to the difficult measurement conditions. In some cases, the temperature difference between two consecutive measurements was too high in a small period of time (a 50 1C temperature change in 1 or 2 min as shown in Fig. 4) or temperature increased when the furnace was off. Temperature values were also contrasted too with data from ladle furnace where another temperature measurement is realized at the starting point of the process. A number of cases with unexpected temperature values higher in ladle than EAF were detected and rejected. 3. Tap temperature modeling A statistical regression model developed by S. Ko¨hle and his team in BFI (Betriebsforschungsinstitu, Du¨sseldorf, Germany) was used as starting point (Ko¨hle, 2002). This model estimates the specific energy required in EAFs and was based in data collected from 50 EAFs. This regression model is applicable to any DC or AC furnaces and it was
As starting point and taking the Ko¨hle formulae as a reference model, the main process parameters were considered as initial model inputs. Then several neural networks topologies were trained with BackPropMomentum algorithm (Rumelhart et al., 1994), using different training parameters in order to reach the best model accuracy. The results were evaluated by using the quadratic medium error to test patterns. The contribution of each variable into the model was evaluated using a sensitivity analysis. The next step was to add new inputs to improve model accuracy. A SOM network (self organizing map) (Kohonen, 2006) was trained verifying that the new set of variables is capable to categorize the output variable.
Neural networks must be trained with a representative process dataset before they can generalize to new data and can securely take action within a controller. Several nets, with a typical multilayer Perceptron topology but changing the number of input and hidden neurons, were trained with a dataset corresponding to more than 2500 heats.
Table 3 Neural model variables t tfu tv tpaf Ch Cal Dol HBI O2 O2eaf C RO2 C E Eeaf fa fb
Total time Melting time Idle time Refining time Scrap Lime Dolomite Hot briquetted iron Total injected oxygen Specific oxygen injected at refining Injected carbon Injected oxygen and carbon rate Electric power Specific electric power at refining Mean amplitude (frequencies 2116–2248 Hz) Mean amplitude (frequencies 10,252–10,436 Hz)
ARTICLE IN PRESS J.M.M. Mesa Ferna´ndez et al. / Engineering Applications of Artificial Intelligence 21 (2008) 1001–1012
The input neurons number in the best model trained was 16 (Table 3), including time variables (tfu, tv, t and tpaf), raw materials (Ch, Cal, Dol, HBI), oxygen and coal injection (O2, O2eaf , C, RO2 C ), electric consumption and sound amplitude in two frequencies (fa, fb). The success percentage in training were over 80% of with an absolute error 25 1C and over 75% with a tolerance of 20 1C. In the temperature, SOM map shown in Fig. 5 (corresponding to the inputs of the selected model) four groups can be identified corresponding to heats with a different behavior. Number 1 group (highest temperature values) is related with high oxygen–coal rates and the sound levels (fb frequency) in those heats are higher than usually. It could be a sign of wrong coal injection. Number 2 group (part of lowest temperature heats) is classified by the two sound variables (fa and fb) and the oxygen injection in the last process phase (O2eaf ). A change in raw materials (low scrap and dolomite charge and high cal and HBI values) seems to be the cause of number 3 group. The last group, number 4, (middle-high temperature values) could be related to a high oxygen injection, especially during refining, and a high electrical consumption. Fig. 6 presents the behavior of the neural model for 100 cases and it is possible to observe very good predictions, but considerable errors in some cases too. However, the neural model accuracy is much better than the results achieved by using statistical regression model, which was taken as reference. In order to get a better accuracy and to reduce the bigger errors which occurred mainly in the extremes of work interval, a new approach was tried.
1007
To take advantage of the good behavior of neural networks as classifier, a model with different structure was tried, which is showed in Fig. 7. A first block, constituted by a neuronal network multilayer, classifies the value of tap temperature. Each one of the output neurons corresponds with a certain category that covers a range of temperature. It could be considered that a greater number of neurons in the exit layer would increase model accuracy. The obtained classification usually will not have a clear winning neuron, but the solution will be a set of membership values. The use of many output neurons would make more difficult the creation of inference functions to get a temperature prediction. On the other hand, a greater number of output neurons increase the complexity of the neuronal network and can reduce its ability to generalize, mainly when the number of training patterns is limited as in this case (Zhang, 2000). Another aspect to consider for the selection of output categories are temperature ranges usually used in plant. The proposed model have five output neurons, labeled as MA, A, N, B and MB. They correspond, respectively, to very high, high, medium, low and very low temperature. That classification will assign a value between 0 and 1 to each one of the outputs indicating membership degree of an input vector to each category. As it was mentioned before, the second block transforms these membership degrees in a value of prediction of temperature. The work range of temperature extends from 1550 to 1750 1C. Outside this interval it would be enough if
Fig. 5. SOM map of the selected temperature model.
ARTICLE IN PRESS 1008
J.M.M. Mesa Ferna´ndez et al. / Engineering Applications of Artificial Intelligence 21 (2008) 1001–1012
Fig. 6. Neural model predictions.
Fig. 7. Neural classifier model scheme.
the temperature in each was classified as ‘‘very low’’ or ‘‘very high’’ because they would be far to target temperature (around 1620 1C).
Several membership functions, with triangular and trapezoidal shapes and with different intervals and slopes, were tried to calculate the degree of membership to each
ARTICLE IN PRESS J.M.M. Mesa Ferna´ndez et al. / Engineering Applications of Artificial Intelligence 21 (2008) 1001–1012
1009
Fig. 8. Selected membership function.
temperature category. Some of the studied alternatives were:
distribute uniformly the number of patterns along work temperature interval; extend the work interval, diminishing the slope of triangles that define the membership to each category; use a trapezoidal function around the objective value of temperature (1620 1C); define different triangle slopes, higher for the central zone and lower in the ends where the number of data is smaller.
Fig. 8 shows the selected function, which fits better to the criteria used in EAF by technicians. Selected membership function is composed of triangular functions according to the expression that is in Fig. 8. Outside the ends of work interval the output value would be 1 in ‘‘very low’’ or ‘‘very high’’ groups according to it corresponds. Using that selected membership function, patterns for neural network training (75%) and model testing (25%). were generated. New neural network has the same input variables as the previous predictive model but hidden neurons number was modified. This allows evaluating if the neuronal network works better as classifier or as predictive function. A very important aspect in model structure, because it determines its results enormously, is how to convert these
Fig. 9. Inference function selected.
five classification values in a temperature prediction. For the selection of the inference function in order to turn the five values of the output neurons to a temperature prediction were tried several rules like select only the winner neuron value, the two best neuron values (only if they are corresponding to contiguous temperature groups), a weigh of all the output values, etc. The best results were achieved by the winner neuron value when it is greater than 80% of sum or, in the opposite case, weighing up the two best values but only if they are contiguous (Fig. 9). Trying to use more complex rules to weigh up the influence of three or more neurons give worse results.
ARTICLE IN PRESS 1010
J.M.M. Mesa Ferna´ndez et al. / Engineering Applications of Artificial Intelligence 21 (2008) 1001–1012
Fig. 10. Actual versus estimated temperatures.
The results achieved with the prediction model were around 80% of accuracy for an absolute error under 20 1C and a 90% of accuracy for a relative tolerance of a 25 1C. Real measurements versus model predictions are shown in Fig. 10. When modeling the temperature, prediction should not be unbiased. As it was already discussed, the accuracy of the temperature measurements is low due to the difficult operation in that hard environment and the lack of representativeness of a single measure in a unique point. In this condition, if the predicted value is under the actual temperature, efficiency gets worse but process is safe and it is guaranteed that all the materials are melted. However, an estimation of the temperature over the actual value when temperature is under target could cause part of the material to solidify. Unmelted steel scraps could cause different problems in the installation as, for instance, to obstruct melted steel flow or to clog up the tapping hole but also human hazards. Considering the risk, model has been intentionally shifted to keep mean in a safer region. Standard error deviation is around 14 1C and, as it happened with the previous neuronal models, the greater errors are located at the ends of the work interval. These errors are due to the small number of data existing in these zones and they correspond with the taps that have had more difficulties during the process. The precision reached by artificial intelligent modeling techniques are very high as regards the previous statistical regression model. The improvement in temperature prediction with the neural working as a classifier is significant and could be raised by introducing some improve inference
rules or modifying the membership functions in plant testing (Fig. 11). Continuous prediction during the process would allow obtaining a better estimation of the temperature at the moment of the decision to stop the process and to pour melting steel in ladle. The obtained measurement would have to validate model prediction. It is not required to obtain the exact value of temperature measurement, but an approximated value would be enough, because this measurement has its own error. With the objective to validate the results of the model, a simulation of operation was made as it would take place in the plant. Off-line simulation was made by introducing continuous data to neural classifier model. The output temperature values were compared with the measurements from EAF and some samples are showed in Fig. 12. It is possible to observe in these samples the good behavior of prediction model, even when several measurements were taken and, therefore, the furnace was stopped for some time. Neural network model will be installed next in plant control for the accomplishment of an online tests stage and, after the required adjustments, to be incorporated later in EAF control. 4. Conclusions Process modeling in electrical arc furnaces makes by means of classic methods with a suitable precision it is very difficult. In this work, techniques that extract the knowledge of empirical data gathered of plants in operation have
ARTICLE IN PRESS J.M.M. Mesa Ferna´ndez et al. / Engineering Applications of Artificial Intelligence 21 (2008) 1001–1012
Fig. 11. Model results comparison.
Fig. 12. Simulation off-line samples.
1011
ARTICLE IN PRESS 1012
J.M.M. Mesa Ferna´ndez et al. / Engineering Applications of Artificial Intelligence 21 (2008) 1001–1012
been used. The measurement or storage errors and incomplete cases found in the gathered data required an intense treatment and filtering of data. The high tolerance to the noise of neuronal networks causes that this artificial intelligence technique is suitable for modeling this process. Developed neuronal networks used as predictive functions improve significantly the results obtained with previous models based on statistical regression, like the model of Ko¨hle. Finally, neural networks are used as classifier in combination with membership functions and inference rules to develop a hybrid temperature prediction model. As a consequence of this structure was obtained an improvement in precision reached with regards to the previous predictive neural network model and the decrease of the errors at the ends of the temperature prediction interval. On-line estimation of tap temperature will help plant operator to monitor process and to know when target tap temperature is reached with the consequent energy saving and optimization of the process. References Astigarraga, J., 1998. Hornos De Arco Para Fusio´n De Acero. McGrawHill. Birat, J.P., 2000. A futures study analysis of the technological evolution of the EAF by 2010. Revue de Metallurgie-Cahiers D Informations Techniques 97 (11), 1347–1363. Buydens, J.M., Nyssen, P., Marique, C., Salamone, P., 1998. The dynamic control of the slag foaming operation in the electric arc furnace. Revue de Metallurgie-Cahiers D Informations Techniques 95 (4), 501–509. Energetics, Inc., 2003. Steel industry technology roadmap. Barriers and pathways for yield improvements American iron and steel institute.
Fruehan, R.J., Fortini, O., Paxton, H.W., and Brindle, R., 2000. Theoretical Minimum Energies to Produce Steel for Selected Conditions. Energetics, Inc., Columbia, MD, US Department of Energy Office of Industrial Technologies Washington, DC. Karr, C.L., Wilson, E.L., 2005. Improved electric arc furnace operation via implementation of a geno-fuzzy control system. Materials and Manufacturing Processes 20 (3), 381–405. Ko¨hle, S., 2002. Recent improvements in modeling energy consumption of electric arc furnaces. In: Proceedings of the 7th European Electric Steelmaking Conference. Kohonen, T., 2006. Self-organizing neural projections. Neural Networks 19 (6–7), 723–733. Mapelli, C., Baragiola, S., 2006. Evaluation of energy and exergy performances in EAF during melting and refining period. Ironmaking & Steelmaking 33 (5), 379–388. Montequin, V.R., Fernandez, F.O., Martinez, N.D., Rodriguez, J.A.G., 2002. Using artificial intelligence to model heavy plate mill rolling. JOM—Journal of the Minerals Metals & Materials Society 54 (7), 46–50. Ortega, F., 2003. Data based modelling for prediction and control of internal cleanliness in steel strips. Thermec’2003, parts 1–5, 426-4, 3897–3902. Rosenblatt, F., 1958. The Perceptron—a probabilistic model for information-storage and organization in the brain. Psychological Review 65 (6), 386–408. Rumelhart, D.E., Widrow, B., Lehr, M.A., 1994. The basic ideas in neural networks. Communications of the ACM 37 (3), 87–92. Vidacak, B., Arvanitidis, I., Jonsson, P.G., Sjoberg, P., 2002. Observation on foaming of EAF slags in the production of stainless steel. Scandinavian Journal of Metallurgy 31 (5), 321–327. Wilson, E.L., Karr, C.L., Bennett, J.P., 2004. An adaptive, intelligent control system for slag foaming. Applied Intelligence 20 (2), 165–177. Yi, S.H., Kim, S.M., 2002. Effect of surface-active species on the foam stability in the EAF slag system. Scandinavian Journal of Metallurgy 31 (2), 148–152. Zhang, G.Q.P., 2000. Neural networks for classification: a survey. IEEE Transactions on Systems Man and Cybernetics Part C—Applications and Reviews 30 (4), 451–462.
Engineering Applications of Artificial Intelligence 21 (2008) 1001–1012 www.elsevier.com/locate/engappai
Online estimation of electric arc furnace tap temperature by using fuzzy neural networks Jose´ Manuel Mesa Ferna´ndez, Valeriano A´lvarez Cabal, Vicente Rodrı´ guez Montequin, Joaquı´ n Villanueva Balsera University of Oviedo, Project Engineering Area, c/Independencia, 13, 33004 Oviedo, Asturias, Spain Received 2 May 2007; received in revised form 3 October 2007; accepted 20 November 2007 Available online 8 January 2008
Abstract Industrial factories require continuous analysis and reengineering over its production processes but always keeping a severe control of material costs and operation emissions. In the past electric steel mills have been subject of some operation models developed in order to improve the control of the arc furnace by means of mathematical techniques and, later on, with finite elements technique (FEM). However, these models have not reached the expected results and applicability. In this case, a model has been developed that allows improving the control through a better prediction of the final temperature and, as consequence, to reduce the consumption of energy in the electric arc furnace. Required information for this new model will be obtained gathering knowledge collected up from data obtained of a certain electric furnace and also considering the plant operators and technicians experience. The model has been constructed by using neural networks as classifier, and with a final fuzzy inference function to return a predicted temperature value. r 2007 Elsevier Ltd. All rights reserved. Keywords: Energy consumption; Electric arc furnace; Neural networks; Fuzzy inference function; Temperature prediction
1. Introduction Nowadays there are two main industrial processes to produce steel. The first one, which is known as integrated steel plant, produces steel by refining iron ore in several steps. This ore-based process uses a blast furnace. The other one, steel-making from scrap metals involves melting scrap metal, removing impurities and casting it into the desired shapes. Although originally the steel production in the electric arc furnaces (EAFs) was applied mainly to the special steel grades, the situation has changed with taps size increase and the high productivity that has been reached progressively. This has allowed significant cost reduction, diminishing consumption of energy, electrodes and refractory. At present, electric furnace combined with secondary metallurgy allows to make a very important part of the worldwide steel production on the basis of the massive recycling Corresponding author. Tel.: +34 985 10 43 48; fax: +34 985 10 42 56.
E-mail address: [email protected] (J.M.M. Mesa Ferna´ndez). 0952-1976/$ - see front matter r 2007 Elsevier Ltd. All rights reserved. doi:10.1016/j.engappai.2007.11.008
of the iron scrap. In fact, the share of electric furnace steelmaking technology has reached 33% in world crude steel production and this trend continues although permanently altered by international scrap market. This trend is due of both the capability of this route to produce steel at low cost per unit and to constant improvements in steel quality. According to several studies (Birat, 2000), in an immediate future this tendency will remain, predicting an increase from present 33% to 40% by the year 2010 (Table 1). 1.1. Electric arc furnace process overview EAFs are used to melt scrap by means of electric energy and oxygen. Production of steel from scrap can also be economical on a smaller scale. EAF technology is most often used within steel minimills, whose basic design includes one or two EAFs, a continuous casting and a rolling mill. EAFs produce steel by melting scrap using electric supply as the main energy source. Scrap is generated from the scrapping of capital equipment or as a by-product of the steel-making process. Small quantities
ARTICLE IN PRESS 1002
J.M.M. Mesa Ferna´ndez et al. / Engineering Applications of Artificial Intelligence 21 (2008) 1001–1012
of direct reduced iron and pig iron are used as scrap substitutes or in the higher quality steel. By oxygen blowing, the concentrations of carbon and silicon can be reduced, while providing an additional energy source. The EAF operating cycle is called the tap-to-tap and starts when the furnace roof is initially open allowing to charge the scrap. The solid scrap is charged on top of liquid steel (a ‘‘hot heel’’), which was retained from the previous tap. Then electrodes are lowered into the furnace and the heat melts the charge due to electric resistance of the metal and radiation from the arc. Then electrodes are moved down into the furnace and the heat melts the charge due to electric resistance of the metal and radiation from the arc. Chemical energy is also supplied to melt the furnace charge. Oxygen is blown into the molten steel through one or more lances. The meltdown period is followed by another period in which the impurity levels in the melt are decreased. Also in this period the foamy slag layer is formed and its function
Table 1 Steel production prevision
World steel production (MTm/year) BOF (basic oxygen furnace) (%) EAF (electric arc furnace) (%) Another processes (%)
2000
2010
760 57 33 10
850 50 40 5.5
(+20/ 30) (+5/ 0) (+5/ 0) (+5/ 0)
is to absorb those impurities (adding lime and dolomite to modify its composition) and to shield furnace walls. Foaming occurs by injecting graphite into the slag layer and due to gas bubbles moving through. Once the scrap charge is fully melted, flat bath conditions are reached. At this point, a tap temperature and a steel sample are taken. Once the expected steel composition and temperature are achieved to further processing (secondary steel-making) the tap-hole is opened, the furnace is tilted, and the steel poured into a ladle. 1.2. Improving EAF productivity In a complex industrial system as an EAF, the ways to improve this process are diverse, such as selection and treatment of materials, the redesign of the facilities and new kinds of energetic contributions to the process. The evolution of electric furnaces is reflected mainly in the progressive reduction of specific consumption of energy, tap-to-tap times (Fig. 1) (Astigarraga, 1998) and improvement of the metallic yield. A deep study, comparing energy consumed currently in the production of steel with theoretically needed energy, was made by the Carnegie Mellon University (Pittsburg, USA) for the US Department of Energy (Fruehan et al., 2000). Though the direct comparison of different EAFs is difficult due to the operational differences, the disparity between the theoretical energy and the consumed energy contributes to the idea of potential improvement of these
Fig. 1. EAF process time evolution.
ARTICLE IN PRESS J.M.M. Mesa Ferna´ndez et al. / Engineering Applications of Artificial Intelligence 21 (2008) 1001–1012
facilities which could be about 25% of the energy consumption per ton. Table 2, extracted from Steel Industry Technology Roadmap (Energetics, 2003) and published by the American Iron and Steel Institute, lists four major steel industry unit operations and their estimated yield losses. It presents the targets for reducing these losses through a research and development program and shows the reduction in energy intensity resulting from achieving these targets. One of the pathways proposed to reach this yield improvement in EAF operation is the optimization of process sequencing by mean or artificial intelligent techniques. Important effort has been directed towards further productivity improvements under EAF process. Much of this has focused on developing alternative energy sources to reduce the high cost of electrical energy, which means 15% of EAF operation costs (Fig. 2). Improvements involving the use of cheaper forms of energy such as coal and oxygen; however, it has been achieved limited success
Table 2 Steel industry yield losses and targets Unit operation
Estimated yield loss (%)
Yield loss target (% reduction)
Energy savings targeta (MJ)
Iron makingb BOF steel-makingc EAF steel-makingc Finishing operationsd Applications and material propertiese
2–6 7–9 6–8 1 19
25 33 33 33 50
100–200 200–300 200–300 100 1800
a
Reduction in energy intensity that will result from achieving the corresponding yield loss target. b Includes tapping, metal handling, skimming, and desulphurization. c Includes ladle refining and casting. d Includes hot and cold rolling, coating, scarfing, etc. e Based on 14 million tons of prompt scrap plus 5 million tons reduced production resulting from improved properties.
1003
because of the cost of additional equipment (Mapelli and Baragiola, 2006). Another two different types of costs that have impact on the total cost of EAF operation can be identified: the cost of the feed materials and the cost implication of not reaching control objectives. Feed materials are very conditioned by the final quality of steel to obtain (scrap substitutes are mainly used in the production of higher quality products) and the steel cost due to materials is mainly determined by the price of scrap in the world-wide market. Processing control optimization is another very important way for the reduction of the energy consumption. Since no suitable plant model was available, the operations in the furnace were only based on empirical knowledge. There are a large number of traditionally manually controlled variables. The furnace operator, in accordance with his own experience and in his particular way of working, has been taking the decisions. For example, he was deciding if it was necessary to inject more or less oxygen, coal or HBI (hot briquetted iron), or even stopping the process and measuring the steel temperature. The automation of these variables may result in EAFs being operated more efficiently. However, the development of a mathematical model of the process is extremely problematic because of the extreme conditions in the furnace. The mathematical representation of the physical and chemical phenomena (electrical arc, materials flow, chemical reactions, etc.) that take place inside the furnaces needs, for its resolution, some hypotheses and simplifications that can limit its application enormously. For a long time, operations in the furnace were only based on empirical knowledge. On the other hand, the considerable design and operative differences in facilities make the generalization of any model very complicated. These difficulties cause less advances in process control than in other aspects. The furnace model development by means of classic mathematics or finite elements techniques (FEMs) to improve the control of the EAF and the minimization of energy consumption have not reached the expected success. After the appearance of the artificial intelligence techniques in the early 1990s, as new way of learning and modeling from data, these techniques have became the principal line of investigation in process control, in the metallurgical sector, as it shows the numerous projects developed by steel industry. The application of these techniques has turned out to be significant for physical deformations (Montequin et al., 2002; Ortega, 2003), but it still is in development for thermal models. 1.3. Proposed solution
Fig. 2. Operation cost rates in an EAF.
This paper proposes to optimize furnace control by means of a neural network approach, which improves the yield of EAF process by reducing electric energy consumption tap-to-tap times.
ARTICLE IN PRESS 1004
J.M.M. Mesa Ferna´ndez et al. / Engineering Applications of Artificial Intelligence 21 (2008) 1001–1012
Initially, a study of available process information will be done and control systems of EAF will be modified to allow storing all that information or variables that could be considered relevant. Also, an audio system will be installed for the acquisition of sound environment of furnace because several studies have shown that it can provide information about the heat state. Later, a process of data evaluation will be carried out, filtering incomplete or invalid cases. From that database, a selection of variables of the neural models will be made, including the output variable. The above-mentioned models will be trained by a wide dataset and validated by new data proceeding from furnace and not used in its training. The aim is that the final model is completely operative and can be used in EAF plant as main decision criteria and allow to determine when heat is ready to continue to the following phase of the steel production. 2. Data acquisition and problem analysis The development of a prediction model based on mathematical representation of physical and chemical processes in EAF furnaces needs, for its resolution, some hypotheses and simplifications that avoid its applicability in furnace operations as was mentioned in Section 1. An alternative way to improve furnace control is to collect, analyze and use empirical process data and, by means of knowledge-based techniques, build a tap prediction model. Data acquisition, modeling methodology and mathematical techniques used to construct that neural model are described in the next sections. 2.1. Data acquisition and pre-processing Modeling from empirical data requires the correct gathering of the most important process data, collected with enough accuracy and frequency. Measurements in EAFs are limited and quite affected by noise due to the environment conditions. Continuous changes in processes like feed materials alterations, different energy supplies or operation circumstances, require a model capable to adapt to these changes. The process data collected in a DC EAF for this purpose and stored in about 250 variables can be put into the next categories:
foaming, was detected in previous researches (Buydens et al., 1998; Vidacak et al., 2002; Yi and Kim, 2002). Those works studied especially the foaming slag phase and their conclusions were that the frequency range between 100 and 150 Hz was the best to detect that phase. In this work, the frequency is extended up to 20,000 Hz. The sound signal was processed by fast Fourier transformade (FFT) in order to store the amplitude values in the database. Data were cleaned and pre-processed using classical statistics, Sammon projections and other methods of visualization. A considerable number of incomplete or no valid cases were found in this process. 2.2. Output variable selection One important decision in model development is the selection of the prediction variable. That parameter must allow estimating in a most precise way when the steel is ready to go on refining phase. Furnace energy supply comes from the electrical power through the electrodes and from the chemical energy due to the combustion of the natural gas, oxidation of the coal and other elements. In addition, there are different losses of energy as thermal (through the refrigeration system, the refractory or exhaust gases), electrical, etc. Therefore, the power consumption, absolute or specific (by unit of raw material) could be the model estimation variable. Nevertheless, except the electrical consumption, whose final value is available, other parameters are not directly measurable, what would make difficult the validation of the results of the model. On the other hand, process time is another variable with which the furnace yield is usually quantified and it could also be the purpose of model prediction. Nevertheless, it was not used because the furnace stops for material load, whose duration is quite irregular, make its use difficult. This paper is focused in one of these variables, the steel tap temperature. To reach a valid temperature value is the main criteria (besides the carbon content) to stop the heat, but it is difficult to get an accurate measurement because:
variables related with raw materials (scrap, hbi, coal, etc.); time event and duration variables; electrical consumption and other chemical energy supplies like injected oxygen; the sound captured by a microphone staying next to the furnace; and temperature measurements.
Slag foaming is a steel-making process that has been shown to improve the efficiency of EAF plants (Karr and Wilson, 2005; Wilson et al., 2004). The connection between level sound and some phases or events of process, like slag
There is not a continuous sensor inside the furnace. Temperatures are measured with a pyrometer with a limited precision. There is not a uniform temperature in the furnace because its size, the difficult flow of materials and convection heat transmission. There is a lack of accuracy because of the own measurement method or operators errors.
The target temperature to melt materials, in the studied DC electrical arc furnace, is around 1620 1C although that value could be changed due to subsequent treatments. To overcome this temperature value supposes a waste of energy. Nevertheless, energy lost in cooling during the measurement delay is significant when the temperature is lower than target and furnace must be switched on again.
ARTICLE IN PRESS J.M.M. Mesa Ferna´ndez et al. / Engineering Applications of Artificial Intelligence 21 (2008) 1001–1012
In both cases there are energy losses and the tap-to-tap times are longer and therefore the process yield is reduced. A better temperature prediction model based in process parameters could improve a lot the process efficiency. One important decision in the model developed is to determine what should be the most appropriate prediction variable. That parameter must allow estimating more precisely when the steel is ready to go ahead with the refining phase. Furnace energy supply comes from the electrical power through the electrodes and from the chemical energy due the combustion of the natural gas, oxidation of the coal and other elements. In addition, there are different losses of energy as thermal (through the refrigeration system, the refractory or exhaust gases), electrical, etc. Therefore, the power consumption, absolute or specific (by unit of raw material) could be the model estimation variable. Nevertheless, except the electrical consumption, whose final value could be known, other parameters are not directly measurable, what would make difficult to validate the results produced by the model. On the other hand, process time is one more variable with which the furnace yield is usually quantified and it could also be the value that should be determined by the model. However, it was not used because the furnace stops for material load, whose duration is quite irregular, making its use tough. This paper is focused in one of these variables, the steel tap temperature. A primary criterion is to reach a valid
1005
temperature value (besides the carbon content) to stop the heat, but is arduous to obtain an accurate measurement because:
There is not a continuous sensor inside the furnace. Temperatures are measured with a pyrometer with a limited precision. There is not a uniform temperature in the furnace because its size, the materials flow and convection heat transmission. There is a lack of accuracy because of the own measurement method or operators errors.
The target temperature to melt materials, in the analyzed DC electrical arc furnace, is around 1620 1C although that value could be modified due subsequent treatments. To overcome this temperature value implies a waste of energy. Nevertheless, energy lost in cooling during the measurement delay is significant when the temperature is lower than the target and furnace must be switched on again. In both cases, there are energy losses, the tap-to-tap times are longer and therefore the process yield is reduced. A better temperature prediction model based in parameters of the process could improve a lot/could derive to an important improvement of the process efficiency. Fig. 3 shows the increase of the average electric consumption (E) when it is needed a second or third temperature measurement (a second or third stopping of furnace therefore).
Fig. 3. Electric consumption comparative for different number of temperature measurements.
ARTICLE IN PRESS 1006
J.M.M. Mesa Ferna´ndez et al. / Engineering Applications of Artificial Intelligence 21 (2008) 1001–1012
Fig. 4. Consecutive temperature difference.
The first temperature measurements were considered not valid to go through refining stage and these cases represent around 45% of the heats. Only a reduction of this rate can be a significant process improvement. Just a reduction of this rate can be a significant process improvement. A model capable to predict the tap temperature during the heat would be essential to introduce corrective actions immediately, with the consequent energy saving and optimization of the process. Also, the operator would be provided with a less subjective reason of process evolution based in acquired data to stop the heating and to measure tap temperature.
adjusted for its use in the studied furnace and therefore preliminary relevant variables for prediction were selected. Neural networks are successfully used for modeling, classifying and controlling complex systems, and a large number of training algorithms have been developed. Good performance to data noise, the possibility of training again to fit properly in changeable conditions and its capability to detect non-lineal relations become neural networks in an adequate choice to develop a temperature prediction model. Projects on EAF control have focused mainly on electrode control, where neural networks were used alone or in combination with other techniques. In a first stage, the method to develop a predictive temperature model was based in Perceptron multilayer neural networks (Rosenblatt, 1958) and can be summarized as follows:
2.3. Temperature data treatment As it was previously mentioned, in this research the tap temperature was selected as target variable of the prediction model to develop and it requires a special attention. The study of collected temperature data showed that variable contains very low values (down 1550 1C) and very high ones (up 1700 1C). Low values are due to stop the heat process before reaching the target value, and high ones, when the required temperature value is widely exceeded, are due to delay unnecessary process stopping. In both cases, the energy consumption is increased. Also, many temperature measurement errors were detected in data cleaning due to the difficult measurement conditions. In some cases, the temperature difference between two consecutive measurements was too high in a small period of time (a 50 1C temperature change in 1 or 2 min as shown in Fig. 4) or temperature increased when the furnace was off. Temperature values were also contrasted too with data from ladle furnace where another temperature measurement is realized at the starting point of the process. A number of cases with unexpected temperature values higher in ladle than EAF were detected and rejected. 3. Tap temperature modeling A statistical regression model developed by S. Ko¨hle and his team in BFI (Betriebsforschungsinstitu, Du¨sseldorf, Germany) was used as starting point (Ko¨hle, 2002). This model estimates the specific energy required in EAFs and was based in data collected from 50 EAFs. This regression model is applicable to any DC or AC furnaces and it was
As starting point and taking the Ko¨hle formulae as a reference model, the main process parameters were considered as initial model inputs. Then several neural networks topologies were trained with BackPropMomentum algorithm (Rumelhart et al., 1994), using different training parameters in order to reach the best model accuracy. The results were evaluated by using the quadratic medium error to test patterns. The contribution of each variable into the model was evaluated using a sensitivity analysis. The next step was to add new inputs to improve model accuracy. A SOM network (self organizing map) (Kohonen, 2006) was trained verifying that the new set of variables is capable to categorize the output variable.
Neural networks must be trained with a representative process dataset before they can generalize to new data and can securely take action within a controller. Several nets, with a typical multilayer Perceptron topology but changing the number of input and hidden neurons, were trained with a dataset corresponding to more than 2500 heats.
Table 3 Neural model variables t tfu tv tpaf Ch Cal Dol HBI O2 O2eaf C RO2 C E Eeaf fa fb
Total time Melting time Idle time Refining time Scrap Lime Dolomite Hot briquetted iron Total injected oxygen Specific oxygen injected at refining Injected carbon Injected oxygen and carbon rate Electric power Specific electric power at refining Mean amplitude (frequencies 2116–2248 Hz) Mean amplitude (frequencies 10,252–10,436 Hz)
ARTICLE IN PRESS J.M.M. Mesa Ferna´ndez et al. / Engineering Applications of Artificial Intelligence 21 (2008) 1001–1012
The input neurons number in the best model trained was 16 (Table 3), including time variables (tfu, tv, t and tpaf), raw materials (Ch, Cal, Dol, HBI), oxygen and coal injection (O2, O2eaf , C, RO2 C ), electric consumption and sound amplitude in two frequencies (fa, fb). The success percentage in training were over 80% of with an absolute error 25 1C and over 75% with a tolerance of 20 1C. In the temperature, SOM map shown in Fig. 5 (corresponding to the inputs of the selected model) four groups can be identified corresponding to heats with a different behavior. Number 1 group (highest temperature values) is related with high oxygen–coal rates and the sound levels (fb frequency) in those heats are higher than usually. It could be a sign of wrong coal injection. Number 2 group (part of lowest temperature heats) is classified by the two sound variables (fa and fb) and the oxygen injection in the last process phase (O2eaf ). A change in raw materials (low scrap and dolomite charge and high cal and HBI values) seems to be the cause of number 3 group. The last group, number 4, (middle-high temperature values) could be related to a high oxygen injection, especially during refining, and a high electrical consumption. Fig. 6 presents the behavior of the neural model for 100 cases and it is possible to observe very good predictions, but considerable errors in some cases too. However, the neural model accuracy is much better than the results achieved by using statistical regression model, which was taken as reference. In order to get a better accuracy and to reduce the bigger errors which occurred mainly in the extremes of work interval, a new approach was tried.
1007
To take advantage of the good behavior of neural networks as classifier, a model with different structure was tried, which is showed in Fig. 7. A first block, constituted by a neuronal network multilayer, classifies the value of tap temperature. Each one of the output neurons corresponds with a certain category that covers a range of temperature. It could be considered that a greater number of neurons in the exit layer would increase model accuracy. The obtained classification usually will not have a clear winning neuron, but the solution will be a set of membership values. The use of many output neurons would make more difficult the creation of inference functions to get a temperature prediction. On the other hand, a greater number of output neurons increase the complexity of the neuronal network and can reduce its ability to generalize, mainly when the number of training patterns is limited as in this case (Zhang, 2000). Another aspect to consider for the selection of output categories are temperature ranges usually used in plant. The proposed model have five output neurons, labeled as MA, A, N, B and MB. They correspond, respectively, to very high, high, medium, low and very low temperature. That classification will assign a value between 0 and 1 to each one of the outputs indicating membership degree of an input vector to each category. As it was mentioned before, the second block transforms these membership degrees in a value of prediction of temperature. The work range of temperature extends from 1550 to 1750 1C. Outside this interval it would be enough if
Fig. 5. SOM map of the selected temperature model.
ARTICLE IN PRESS 1008
J.M.M. Mesa Ferna´ndez et al. / Engineering Applications of Artificial Intelligence 21 (2008) 1001–1012
Fig. 6. Neural model predictions.
Fig. 7. Neural classifier model scheme.
the temperature in each was classified as ‘‘very low’’ or ‘‘very high’’ because they would be far to target temperature (around 1620 1C).
Several membership functions, with triangular and trapezoidal shapes and with different intervals and slopes, were tried to calculate the degree of membership to each
ARTICLE IN PRESS J.M.M. Mesa Ferna´ndez et al. / Engineering Applications of Artificial Intelligence 21 (2008) 1001–1012
1009
Fig. 8. Selected membership function.
temperature category. Some of the studied alternatives were:
distribute uniformly the number of patterns along work temperature interval; extend the work interval, diminishing the slope of triangles that define the membership to each category; use a trapezoidal function around the objective value of temperature (1620 1C); define different triangle slopes, higher for the central zone and lower in the ends where the number of data is smaller.
Fig. 8 shows the selected function, which fits better to the criteria used in EAF by technicians. Selected membership function is composed of triangular functions according to the expression that is in Fig. 8. Outside the ends of work interval the output value would be 1 in ‘‘very low’’ or ‘‘very high’’ groups according to it corresponds. Using that selected membership function, patterns for neural network training (75%) and model testing (25%). were generated. New neural network has the same input variables as the previous predictive model but hidden neurons number was modified. This allows evaluating if the neuronal network works better as classifier or as predictive function. A very important aspect in model structure, because it determines its results enormously, is how to convert these
Fig. 9. Inference function selected.
five classification values in a temperature prediction. For the selection of the inference function in order to turn the five values of the output neurons to a temperature prediction were tried several rules like select only the winner neuron value, the two best neuron values (only if they are corresponding to contiguous temperature groups), a weigh of all the output values, etc. The best results were achieved by the winner neuron value when it is greater than 80% of sum or, in the opposite case, weighing up the two best values but only if they are contiguous (Fig. 9). Trying to use more complex rules to weigh up the influence of three or more neurons give worse results.
ARTICLE IN PRESS 1010
J.M.M. Mesa Ferna´ndez et al. / Engineering Applications of Artificial Intelligence 21 (2008) 1001–1012
Fig. 10. Actual versus estimated temperatures.
The results achieved with the prediction model were around 80% of accuracy for an absolute error under 20 1C and a 90% of accuracy for a relative tolerance of a 25 1C. Real measurements versus model predictions are shown in Fig. 10. When modeling the temperature, prediction should not be unbiased. As it was already discussed, the accuracy of the temperature measurements is low due to the difficult operation in that hard environment and the lack of representativeness of a single measure in a unique point. In this condition, if the predicted value is under the actual temperature, efficiency gets worse but process is safe and it is guaranteed that all the materials are melted. However, an estimation of the temperature over the actual value when temperature is under target could cause part of the material to solidify. Unmelted steel scraps could cause different problems in the installation as, for instance, to obstruct melted steel flow or to clog up the tapping hole but also human hazards. Considering the risk, model has been intentionally shifted to keep mean in a safer region. Standard error deviation is around 14 1C and, as it happened with the previous neuronal models, the greater errors are located at the ends of the work interval. These errors are due to the small number of data existing in these zones and they correspond with the taps that have had more difficulties during the process. The precision reached by artificial intelligent modeling techniques are very high as regards the previous statistical regression model. The improvement in temperature prediction with the neural working as a classifier is significant and could be raised by introducing some improve inference
rules or modifying the membership functions in plant testing (Fig. 11). Continuous prediction during the process would allow obtaining a better estimation of the temperature at the moment of the decision to stop the process and to pour melting steel in ladle. The obtained measurement would have to validate model prediction. It is not required to obtain the exact value of temperature measurement, but an approximated value would be enough, because this measurement has its own error. With the objective to validate the results of the model, a simulation of operation was made as it would take place in the plant. Off-line simulation was made by introducing continuous data to neural classifier model. The output temperature values were compared with the measurements from EAF and some samples are showed in Fig. 12. It is possible to observe in these samples the good behavior of prediction model, even when several measurements were taken and, therefore, the furnace was stopped for some time. Neural network model will be installed next in plant control for the accomplishment of an online tests stage and, after the required adjustments, to be incorporated later in EAF control. 4. Conclusions Process modeling in electrical arc furnaces makes by means of classic methods with a suitable precision it is very difficult. In this work, techniques that extract the knowledge of empirical data gathered of plants in operation have
ARTICLE IN PRESS J.M.M. Mesa Ferna´ndez et al. / Engineering Applications of Artificial Intelligence 21 (2008) 1001–1012
Fig. 11. Model results comparison.
Fig. 12. Simulation off-line samples.
1011
ARTICLE IN PRESS 1012
J.M.M. Mesa Ferna´ndez et al. / Engineering Applications of Artificial Intelligence 21 (2008) 1001–1012
been used. The measurement or storage errors and incomplete cases found in the gathered data required an intense treatment and filtering of data. The high tolerance to the noise of neuronal networks causes that this artificial intelligence technique is suitable for modeling this process. Developed neuronal networks used as predictive functions improve significantly the results obtained with previous models based on statistical regression, like the model of Ko¨hle. Finally, neural networks are used as classifier in combination with membership functions and inference rules to develop a hybrid temperature prediction model. As a consequence of this structure was obtained an improvement in precision reached with regards to the previous predictive neural network model and the decrease of the errors at the ends of the temperature prediction interval. On-line estimation of tap temperature will help plant operator to monitor process and to know when target tap temperature is reached with the consequent energy saving and optimization of the process. References Astigarraga, J., 1998. Hornos De Arco Para Fusio´n De Acero. McGrawHill. Birat, J.P., 2000. A futures study analysis of the technological evolution of the EAF by 2010. Revue de Metallurgie-Cahiers D Informations Techniques 97 (11), 1347–1363. Buydens, J.M., Nyssen, P., Marique, C., Salamone, P., 1998. The dynamic control of the slag foaming operation in the electric arc furnace. Revue de Metallurgie-Cahiers D Informations Techniques 95 (4), 501–509. Energetics, Inc., 2003. Steel industry technology roadmap. Barriers and pathways for yield improvements American iron and steel institute.
Fruehan, R.J., Fortini, O., Paxton, H.W., and Brindle, R., 2000. Theoretical Minimum Energies to Produce Steel for Selected Conditions. Energetics, Inc., Columbia, MD, US Department of Energy Office of Industrial Technologies Washington, DC. Karr, C.L., Wilson, E.L., 2005. Improved electric arc furnace operation via implementation of a geno-fuzzy control system. Materials and Manufacturing Processes 20 (3), 381–405. Ko¨hle, S., 2002. Recent improvements in modeling energy consumption of electric arc furnaces. In: Proceedings of the 7th European Electric Steelmaking Conference. Kohonen, T., 2006. Self-organizing neural projections. Neural Networks 19 (6–7), 723–733. Mapelli, C., Baragiola, S., 2006. Evaluation of energy and exergy performances in EAF during melting and refining period. Ironmaking & Steelmaking 33 (5), 379–388. Montequin, V.R., Fernandez, F.O., Martinez, N.D., Rodriguez, J.A.G., 2002. Using artificial intelligence to model heavy plate mill rolling. JOM—Journal of the Minerals Metals & Materials Society 54 (7), 46–50. Ortega, F., 2003. Data based modelling for prediction and control of internal cleanliness in steel strips. Thermec’2003, parts 1–5, 426-4, 3897–3902. Rosenblatt, F., 1958. The Perceptron—a probabilistic model for information-storage and organization in the brain. Psychological Review 65 (6), 386–408. Rumelhart, D.E., Widrow, B., Lehr, M.A., 1994. The basic ideas in neural networks. Communications of the ACM 37 (3), 87–92. Vidacak, B., Arvanitidis, I., Jonsson, P.G., Sjoberg, P., 2002. Observation on foaming of EAF slags in the production of stainless steel. Scandinavian Journal of Metallurgy 31 (5), 321–327. Wilson, E.L., Karr, C.L., Bennett, J.P., 2004. An adaptive, intelligent control system for slag foaming. Applied Intelligence 20 (2), 165–177. Yi, S.H., Kim, S.M., 2002. Effect of surface-active species on the foam stability in the EAF slag system. Scandinavian Journal of Metallurgy 31 (2), 148–152. Zhang, G.Q.P., 2000. Neural networks for classification: a survey. IEEE Transactions on Systems Man and Cybernetics Part C—Applications and Reviews 30 (4), 451–462.