Brics Brics 1 By B. Deo, R.K. Lingamaneni, A. Dey, Prateek Mittal Strategies for Development of Optimal Process Control Models for Hot Metal Desulfurization : Conventional and AI Techniques INTRODUCTION Optimal process control models can be developed by using both conventional and AI approaches. The conventional approaches include regression models and process kinetic models and the artificial intelligence (AI) approaches are based on artificial neural nets (ANN), genetic algorithm (GA) and fuzzy rule based experts system (FRBES). Plant data on hot metal desulfurization, carried out by injecting calcium carbide, is analyzed to test and tune different types and combination of models and then evaluate their relative performance. While the control models based on process fundamentals provide a fillip to the new ideas for technological developments and improvements, the combination of conventional and AI approaches may be better option for process control on the shop floor. It is advisable to first develop and test all models and then decide about the best strategy of using them. Desulfurization of the molten carbon-saturated liquid iron, i.e. hot metal produced from the blast furnaces in an integrated steel plant, is carried out to remove sulphur to desirable levels, say from a starting sulphur content of 0.02% to a final sulphur content of < 0.005% (or <50 ppm). A powdered reagent like calcium carbide (CaC2, commercially known as calcium diamide CaD), magnesium (Mg), lime, or a mixture CaC2+Mg, CaC2+CaCO3, etc. is injected into hot metal, by using a lance, with the help of a carrier gas like nitrogen. There have been several [1-17] studies to develop control/prediction models for hot metal desulfurization on the basis of [1-9] metallurgical kinetics , statistics (regression [7] [10] models) , artificial neural nets (ANN) , genetic [11] algorithms (GA) , and fuzzy rule based expert [7] systems (FRBES) . The accuracy of prediction of different models depends to a large extent on the quality of data and the variability of process parameters. For example, in certain situations a simple regression model may perform as well or even better than the sophisticated model based on extensive application of fundamentals of metallurgical kinetics. The present work focuses on comparing the efficacy of different type of control models that can be developed for a given plant situation and the strategy that should be adopted in using a particular model. PLANT DATA The plant data used in this work relates to 400 [7, 10, 11] ton torpedoes (for details see ) in which desulfurization was carried out by injecting calcium carbide powder through a submerged 3 twin-whole lance. Nitrogen (2 m /min) was used as a carrier gas. Data for 229 heats was collected and the range of variation of each parameter was as follows: treatment time (473-1719 ± 20 seconds), hot metal weight (250-379 ±5 tons), initial sulphur content (0.009-0.044 ± 0.0005 mass %), CaD injection rate 45.0-71.0 ± 1.5 kg/min), and final sulphur content (0.003-0.018 ± 0.0002 mass %). STATISTICAL ANALYSIS OF DATA AND DEVELOPMENT OF REGRESSION MODELS By observing the magnitude of the off-diagonal correlations in the correlation matrix it was concluded that the variables (1) treatment time, (2) mass of hot metal, (3) initial sulphur content and (4) CAD injection rate, can be assumed to be independent. Multiple linear regression analysis was, therefore, carried out with final sulfur content as the dependent variable and the rest four (described above) as independent variables. The standard error of estimate was 0.0018 and multiple correlation coefficient (R) was 0.91 (R squared = 0.8229, corrected R squared "While the control models based on process fundamentals provide a fillip to the new ideas for technological developments and improvements, the combination of conventional and AI approaches may be better option for process control on the shop floor."