E. Donskoi 1 , J. R. Manuel 1 , J. M. F. Clout 2 , Y. Zhang 3
1 CSIRO Minerals, Pullenvale, Kenmore QLD, Australia 2 Fortescue Metals Group Limited, West Perth, Australia 3 Hamersley Iron Rio Tinto Ltd., Shanghai, CHINA
ABSTRACT. The quality of iron ore sinter is a critical factor determining the productivity of blast furnaces for ironmaking. Csiro has therefore been developing capabilities for predicting sinter characteristics, which enables sinter quality to be improved/optimized and preliminary assessments to be made of the suitability of specific ores or ore blends for sinter production. An extensive database of pilot-scale sintering experimental results has been used to create mathematical models for predicting different sinter properties. In addition to size distribution and other physical and chemical characteristics which are usually used for sinter quality prediction, the mineralogical and textural characteristics of iron ores intended for sintering have also been taken into account. This approach has been quite successful, the variation of sinter reduction degradation index (RDI) that is accounted for by explanatory variables (R-sq) being 87%, for example. Optimisation criteria have been developed that take into account several sinter characteristics simultaneously and optimisation of different iron ore blends to produce target sinter characteristics has been carried out. Modelling results and their preliminary validation are discussed.
Introduction In spite of the development and introduction to industrial production of a large number of new iron making processes (1), the major source of iron production (more than 90%) is still the blast furnace. The purpose of a blast furnace is to remove oxygen from iron oxides and to convert iron oxides into liquid iron which can be cooled down to produce pig iron for further processing into steel. Iron oxide can be fed to the blast furnace in the form of raw ore (lump), pellets or sinter. In current ironmaking practice, sinter comprises up to 70-85% of the total ferrous burden. Sinter is produced by agglomerating fine ore, utilising coke or another carbonaceous material as fuel. Specific amounts of different additives, such as fine- sized limestone, magnesite, dolomite, quartzite, serpentine, hydrated lime, etc, are also added to balance sinter chemistry for the blast furnace and flux matrix melt- forming reactions. The main aim of iron ore sintering is to produce a strong and reducible agglomerate (2). To produce sinter at a plant, the granulated sinter mixture is fed onto the sinter strand and ignited from the top. Air is drawn down through the bed with suction fans and a flame front propagates through the bed. During high temperature treatment, the sinter mix is partially melted, and when the sinter is cooled the melt solidifies into a bonding phase for the unreacted materials (see Fig 1).
2-40
Figure 1. Reflected light optical photomicrograph of hematite sinter structure, showing remnant hematite nuclei, the melt-derived matrix and porosity.
Several different approaches have been used to develop mathematical models simulating the iron ore sintering process (3-5). Usually such models describe the heat and mass transfer, drying and condensation of water, gas flow, coke combustion, and charge melting and solidification phenomena that occur during the process. They can also include modelling of phenomena like shrinkage of the bed, variation of granule diameter, channelling factors, void fraction (5), etc. Such modelling enables calculation of different parameters inside the bed like composition, solid and gas temperature, and porosity. These models are very good for understanding the underlying mechanisms of the process and in prediction of the bed behaviour. Unfortunately, no previous mathematical phenomenological models have been developed that connect the parameters described above and some of the parameters of the sinter which are extremely important for the industry, including Tumble Index (TI), Reduction Degradation Index (RDI) and Reducibility Index (RI) (for definition of these parameters see the Appendix). Parameters such as Productivity and Fuel Rate can be obtained from such modelling. However, due to the extreme complexity of the process, results obtained from the resulting models can be significantly different from experimental results. In the work described here, a different approach based on empirical modelling was employed. CSIRO (Commonwealth Scientific and Industrial Research Organisation) has performed a large number of sintering experiments over recent years. These experiments were performed with many different iron ores and blends, covering a wide range of sintering behaviour. From this work, a large database has been built up, which contains data on the resulting sinter, including physical, chemical and mineralogical parameters of the blends and specific conditions under which the experiments were performed. The most significant variables for modelling sintering outputs were identified and regression models developed. Further optimisation criteria have been developed and then modelling was performed to find optimal iron ore blends within realistic constraints. The effect of different variables on sinter properties has been studied by different authors. Loo et al (6) and Hsieh (7) showed that the properties of the raw materials used for sintering significantly affect the melt formation and assimilation, nucleus matrix pore pore pore matrix 1mm hematite nucleus 2-41 sintering reactions and production of bonding phases. Hosotani et al (8) showed the influence of alumina on melt formation in the sintering process and its further effect on sinter properties. Higuchi and Heerema (9) showed the influence of sintering conditions on the reduction behaviour of pure hematite compacts. No articles have been found in the open literature where all the dependencies of raw material properties and sintering conditions were combined in separate models for predicting sinter parameters, although a number of attempts are known to have been made within industry using an empirical approach to modelling sinter properties. It is believed that none of these considered the textural characteristics of iron ore used for sintering, while CSIRO research has shown that textural characteristics significantly affect the quality of the sinter and should be taken into account during modelling of sinter properties.
Initial task
The initial task was to find the best sinter mix composition for nine different ores (further referred to as ores A, B, C, G, H, I). The ore blends were selected on the basis of different criteria, such as lowest cost, highest TI, highest productivity and lowest fuel rate. Simultaneously, the blends had to contain minimum percentages of particular ores and have chemical compositions acceptable for specific blast furnace requirements. Two sets of iron ores were developed, ie, high quality and correspondingly higher cost blends, and lower cost blends with acceptable quality. The constraints for the high quality blends were:
(1) Final sinter Total Iron 57-58%, SiO 2 < 5%, basicity = 1.8-2.0, MgO = 1.6- 2.5%. (2) Initial blend not less than 30%, but not more than 60%, in total of ores A, B, C and D.
The constraints for the low cost blends were:
(1) Final sinter Total Iron < 57%, 4.3% < SiO 2 < 5.5%, basicity = 1.8-2.0, MgO = 1.6-2.5% (2) Initial blend with 60% or more in total of ores A, B, C and D. (3) Acceptable sinter TI >65% (pot-grate equivalent).
Approach
CSIRO has been conducting pot-grate sintering (pilot scale) on different ores and blends for a number of years. The technique used is consistent for all sintering runs. Information about the initial sinter mixture and sintering results were put into a sintering database. The modelling was based on 95 carefully selected optimum runs from CSIRO pot-grate testing in which the return fines were in balance and there were no abnormalities in physical or process variables. Variables which had the greatest influence on sintering performance and which were mathematically most significant were used. `Major variables which were used for modelling were total iron content, the amount of alumina in the -1mm size fraction, blend moisture content, coke level, bulk density of the iron ore mixture without coke or fluxes, basicity (ratio of CaO to SiO 2 by mass), loss of mass during ignition (LOI) up to 371 C (LOI 371), LOI 900, ore mass percentages in the -0.63 2-42 mm, -1 mm and +4 mm size fractions, and the proportion of dense hematite ore texture. Details on ore textural classifications have been provided by Box et al (10). All variables characterising sintering are interrelated, so different sets of variables can be chosen for modelling. However, the set of variables mentioned above was found to be the most efficient.
Table 1. List of objective functions used in the modelling (Prod=Productivity). Optimisation criteria 1 Cost 6 TI/Cost 2 Prod/cost 6a TI-Cost*20.43 2a 100+prod-cost*52.1 7 TI 3 Productivity 8 Prod*TI 4 Fuel Rate 8a Prod+2.55*TI 5 Prod/Cost/Fuel Rate 9 Prod+2.55*TI-cost*52.1 5a 100+prod-cost*52.1-FuelRate*0.91 9a Prod*TI/cost
After modelling the sintering process parameters, ie, productivity, RDI, TI and fuel rate, optimisation objective functions (optimisation criteria) were developed. These objective functions included separate sintering process parameters and their combinations (see Table 1). For the same combination of sintering parameters, two different objective functions were developed. One objective function was just a simple combination of products and ratios of sintering parameters, eg, (Prod*TI/FuelRate/Cost) which aims (if maximised) to maximise productivity and TI while minimising cost and fuel rate. Another objective function with the same goal as the previous example is the sum of sintering parameters with weighting coefficients (eg, Prod+2.55*TI-Fuel Rate*0.91-cost*52.1). These coefficients correspond to the respective variability of the parameters involved (standard deviations) obtained from the database of experimental results. If special preference needs to be given to maximisation or minimisation of certain parameters, the weighting coefficients can be changed correspondingly. The optimisation spreadsheet was created in Microsoft Excel and contained information about the nine individual ores from which blends were created. This information includes chemical and textural composition, cost, bulk density and size distribution information. The spreadsheet also contains information about coke and fluxes that have to be added to meet sinter composition requirements. Among these fluxes were limestone, magnesite, dolomite, quartzite, serpentine and hydrated lime. To find optimal blends, the objective functions described above and the Excel Solver function were used for iterative optimisation. Concurrently with optimisation 2-43 and optimal blend determination, all parameters for the resulting blends were calculated and tabulated.
Results
As an example of sintering parameter modelling, regression analysis details for RDI are given below. Regression Analysis: RDI as a function of Fe T (total iron content of the initial iron ore blend), SG (bulk density of the initial blend calculated as the average of the bulk densities of the component ores), -1 mm Al 2 O 3 (percentage of alumina in the -1 mm ore size fraction), H1M (percentage of dense hematite ore texture in the blend), -1 mm percentage in the blend), coke (percentage of coke in the initial mixture) and LOI 900 (loss on ignition at 900 C). The regression equation is:
47 cases used, 47 cases cont ai n mi ssi ng val ues
Pr edi ct or Coef T P VI F Const ant K1 6. 30 0. 000 FeT K2 - 6. 08 0. 000 3. 0 ( FeT- Mean( FeT) ) ^2 K3 - 5. 35 0. 000 1. 4 - 1mmAl 2O3 K4 4. 32 0. 000 2. 3 - 1mm K5 - 2. 40 0. 021 1. 2 LOI 900*H1M +4 ( Wt %) K6 - 2. 60 0. 013 1. 9 Coke*SG K7 3. 97 0. 000 1. 9
S = 3. 40715 R- Sq = 87. 0% R- Sq( adj ) = 85. 0%
Anal ysi s of Var i ance
Sour ce DF SS MS F P Regr essi on 6 3100. 02 516. 67 44. 51 0. 000 Resi dual Er r or 40 464. 35 11. 61 Tot al 46 3564. 36
Sour ce DF Seq SS FeT 1 1150. 66 ( FeT- Mean( FeT) ) ^2 1 947. 40 - 1mmAl 2O3 1 684. 66 - 1mm 1 23. 76 LOI 900*H1M 1 110. 14 Coke *SG 1 183. 39
As can be seen from the modelling, all t statistics for the regression coefficients are very large (>2.4) and the P-values are very low (<0.021), which means that all regression coefficients are significantly different from zero. From the analysis of variance it can also be seen that the F statistic is very large (44.5), which provides good evidence against the null hypothesis K1=K2=K3=K7=0. Variance Inflation Factors (VIF) are quite small (<3.0), which means that the explanatory variables have small colinearity and the estimates of regression coefficients are quite stable (small changes in explanatory variables do not result in large changes to the estimated regression coefficients). The R-Sq value is equal to 87.0%, which means that 87% of total variation is explained by this model and only 13% is explained by other factors not included in 2-44 the model and/or is a result of random measurement errors. For other modelling parameters, the proportions of the variation of the response variable (modelling parameters) that were explained by explanatory variables (initial characteristics of the mixture) (R-Sq) were 82% for productivity, 88% for TI and 90% for fuel rate. More than 100 theoretical blends corresponding to different optimisation criteria and constraints were investigated. From this initial list, results for one of the exercises are shown in Figure 2, the aim being to find the blend with the best productivity corresponding to a given modelled Tumble Index. The trade-off between productivity and tumble index can be clearly seen. Tumble Index P r o d u c t i v i t y 72 71 70 69 68 50 49 48 47 46 45 44 43 1 8 7 6 5 4 3 2 Scatterplot of Productivity vs Tumble I ndex
Figure 2. Relationship between maximum sinter productivity and Tumble Index produced by the optimisation process.
Some other results for the blend optimisation procedure are presented in Table 2 (these blends are different from blends in Figure 2). The optimisation criteria were: 1 Minimum cost 2 Maximum Productivity 3 Minimum Fuel Rate 4 Maximum TI 5 Maximum (Productivity+2.55*TI-cost*52.1) 6 Maximum (Productivity+2.55TI+2.64Fe total)
Table 2. Examples of blend calculations for various optimisation criteria. Sinter parameters % of iron ore in the blend
Fe Total sinter TI Pro- duc- tivity Fuel Rate Cost Blend RDI Basi- city B C D E F G H I 1 57.00 67.72 41.88 51.43 0.745 32.0 2.00 48.87 0.00 28.50 0.00 0.00 0.00 0.00 0.00 2 59.38 68.37 49.30 54.09 0.867 30.0 1.80 0.00 24.25 0.00 31.12 8.08 0.00 0.00 17.37 3 58.00 72.60 44.15 44.24 0.867 29.4 1.89 0.00 0.00 24.46 0.00 0.00 2.06 52.48 0.00 4 58.00 72.69 43.24 44.37 0.865 30.2 1.89 0.00 0.00 30.34 0.00 0.00 0.57 48.90 0.00 5 58.00 72.66 43.78 44.32 0.860 30.1 1.89 0.00 0.00 28.77 0.00 0.00 1.71 48.61 0.00 6 59.66 72.28 43.91 44.29 0.906 29.6 1.80 0.00 0.00 24.66 0.00 0.00 14.52 43.01 0.00 2-45
Table 2 clearly demonstrates the optimisation capability of the modelling technique presented above. The trade-off between different sintering parameters is obvious. The first blend has significantly lower cost than all others, but total iron, TI and productivity are also lower than for the other blends. The modelled productivity for the second blend is the highest, as expected. However, TI was the second lowest of the blends listed. Blend 4 has the largest TI of the blends (72.69), although this is very close to the TI of Blend 3 (72.6). At the same time, the productivity of Blend 4 (43.24) is significantly lower than that predicted for Blend 3 (44,15), so Blend 3 would be preferred in this case. Blends 5 and 6 have combined optimisation criteria, ie, to maximise TI and productivity and to minimise cost for Blend 5, and to maximize iron content for Blend 6. It can be seen that their TI and productivity are quite close, while iron content (but also cost) for Blend 6 is significantly higher than for Blend 5. Verification of modelling results at pot-grate sintering scale (pilot scale) is very expensive. For this reason, initial verification was carried out using laboratory- scale and small-scale sintering. The laboratory-scale sintering technique (compact sintering) used is the same as that described by Clout and Manuel (11). A small fluxed ore sample (~5 g) is heated in a tube furnace using a controlled heating profile in a N 2 /O 2 atmosphere simulating sintering conditions. After firing, rapid cooling is carried out in the same atmosphere. This technique is designed to simulate the matrix fraction of the sinter and give a relative measurement of matrix melting characteristics, including strength, via a tumble/ abrasion test. By calibrating compact sintering melt temperature results against pot-grate sintering results on the same series of blends, it is possible to predict blend fuel rate. Compact sintering also gives fundamental information about sinter mineralogy. In Figure 3, the compact-scale sintering results are shown (the blends are different from those shown in Table 2). As can be seen, correlations between experimentally determined TI, compact TI and theoretically determined TI are quite strong, which demonstrates the predictive capabilities of both methods, ie, mathematical modelling and compact sintering.
2-46 Relationship Between Experimental TI and Theoretical TI y =0.7707x +31.863 R 2 =0.8689 y =0.6719x +37.791 R 2 =0.6802 82.0 82.5 83.0 83.5 84.0 84.5 85.0 85.5 86.0 86.5 87.0 66.0 67.0 68.0 69.0 70.0 71.0 72.0 Model Predi cted TI (%+6.3mm) C o m p a c t
T I
( %
+ 2 m m ) Low Cost Blends High Quality Blends
Figure 3. Relationship between experimental (measured) TI and theoretical (predicted) TI.
Figure 4 compares experimental and modelled results for Tumble Index and RDI for several single (100%) ore blends. As can be seen, the experimental results are quite close to the modelling results. These single ore blends were included in the database on which modelling was performed. Strictly speaking, such a comparison does not demonstrate the predictability of models for new blends that are not part of the modelling database. As already pointed out, pot-grate sintering experiments are very expensive to perform and require large amounts of raw materials, which limits the number of blends and conditions which can be feasibly tested. Unfortunately, at this stage there is not enough data to demonstrate the predictability of the model for new blends, but this will be the subject of further research and validation of the approach outlined above. 2-47 Tumble Index Modelling 60 62 64 66 68 70 72 74 D G H I A Experiment Prediction
RDI Modelling 0 10 20 30 40 50 60 D G H I A Experi ment Predi cti on
Figure 4. Comparison of experimental results and modelling results for various single ore blends for a) Tumble Index and b) RDI.
Conclusions and Future Development
A mathematical modelling method for predicting sinter properties has been presented. The ability of the models to reflect the variation in sinter properties via explanatory variables was quite high (R-Sq values were 82% for productivity, 88% for TI, 86% for RDI and 90% for Fuel Rate), which demonstrates the potential for reliable prediction of major sinter characteristics. Such modelling enables determination of major parameters affecting various sinter properties and a better understanding of the effect of different parameters such as chemical composition, density and ore textural characteristics, on sinter quality. The modelling showed that not only should physical, chemical and mineralogical characteristics of iron ores and fluxes be taken into account, but the textural characteristics of the iron ores comprising the blend should also be included. Comparison of modelling results with results from compact scale sintering demonstrates the capability of both methods for predicting a range of sinter characteristics. Optimisation criteria and the method described above open up the possibility of developing blends with specified characteristics and/or testing the feasibility of 2-48 creating sinter with the required properties from a specified set of iron ores. The effect of substitution of one iron ore for another in a given blend can also be studied. To further validate the method, the results of more recent pot-grate sintering runs need to be analysed. As part of further work, the database will be significantly extended not only by including more recent results, but also by including more chemical and textural parameters characterising the initial sinter mix. To take into account the effect of the larger number of parameters on sinter properties, the principal component analysis method will be used.
Acknowledgments
The financial support of Hamersley China for this work and permission to publish the results is gratefully acknowledged. The contributions from Mr Jasbir Khosa and Dr Liming Lu of CSIRO Minerals, who provided pot-grate sintering data and information for this paper, are greatly appreciated. Dr Ralph Holmes is thanked for reviewing the final manuscript.
APPENDIX
Fuel Rate - weight (in kilograms) of dry coke required to produce one tonne of product sinter. Productivity - amount of sinter product expressed as tonnes per square metre of grate area of sintering machine per day. Reducibility Index (RI) - the reducibility test is carried out to evaluate the reduction behaviour of various iron oxide burden materials in the middle zone of the blast furnace shaft and is a measure of the amount of oxygen removed from the sample under certain standard conditions. Reduction Degradation Index (RDI) - is a measure of the breakdown of the blast furnace burden materials under conditions chosen to resemble those in the upper part (low temperature zone) of the blast furnace shaft. In the test, a 500 g sample of - 19 + 16 mm sinter is reduced isothermally at 550 C for 30 minutes in a fixed bed with a 30% CO/70% N 2 gas mixture flowing at 15 L/min. After cooling under N 2 , the sample is tumbled for 30 minutes at a speed of 30 rpm and then screened at 3.15 mm. The RDI is the weight % of tumbled product passing 3.15 mm. Tumble Index (TI- Sinter Strength) - weight % +6.3 mm material remaining after the tumble treatment in which 15 Kg of -40+10 mm sinter is tumbled in a 1 m diameter drum for 200 revolutions at a speed of 25 rpm.
REFERENCES
1. Girija, P. and Gejierstam, J.(Ed.). Tradition and Innovation in the History of Iron Making. Nainital, PAHAR, 2002. 2. German, R.M. Sintering theory and practice. John Wiley & Sons, 1996. 3. Nath, N.K. and Mitra, K. Mathematical modeling and optimization of two- layer sintering process for sinter quality and fuel efficiency using genetic algorithm. MATERIALS AND MANUFACTURING PROCESSES 2005, 20(3), 335-349. 4. Patisson, F., Bellot, J.P., Ablitzer, D., Marliere, E., Dulcy, C. and Steiler, J.M. Mathematical-Modelling of Iron-Ore Sintering Process. IRONMAKING & STEELMAKING 1991 18(2), 89-95. 2-49 5. Cumming, M.J. and Thurlby, J.A. Developments in Modeling and Simulation of Iron-Ore Sintering. IRONMAKING & STEELMAKING 1990 17(4), 245- 254. 6. Loo, C.E., Williams, R.P., Matthews, L.T. Influence of Material Properties on High-Temperature Zone Reactions in Sintering of Iron Ore. TRANSACTIONS OF THE INSTITUTION OF MINING AND METALLURGY SECTION C- MINERAL PROCESSING AND EXTRACTIVE METALLURGY 1992, 101, C7-C16. 7. Hsieh, L.H.. Effect of Raw Material Composition on the Sintering Properties. ISIJ INTERNATIONAL 2005, 45(4), 551-559. 8. Hosotani, Y., Konno, N., Kabuto, S., Kitamura M. and Abe T. Influence of Alumina Component on Melt Formation in Sintering Process and Sinter Quality Improving Technology. TETSU TO HAGANE- JOURNAL OF THE IRON AND STEEL INSTITUTE OF JAPAN, 83 (6), 347-352. 9. Higuchi, K. and Heerema, R.H. Influence of Sintering Conditions on the Reduction Behaviour of Pure Hematite Compacts. MINERALS ENGINEERING 2003, 16 (5), 463-477. 10. Box, J., Phillips, J. and Clout J.M.F., 2002. Use of Geological Material Types for Predicting Iron Ore Product Characteristics. In: Proceedings of the 1st Japan-Australia Symposium on Iron and Steelmaking, Fuwa-Ward Symposium, ISIJ, Kyoto University, April 4-5, 2002. 11. Clout, J.M.F. and Manuel, J.R., 2003. Fundamental investigations of differences in bonding mechanisms in iron ore sinter formed from magnetite concentrates and hematite ores. Powder Technology, 130, 393-399.