The Use of Neural-Network-Based Technologies to Analyze the Effect of a Combination Blast on Blast-Furnace Performance Indices

Abstract

Methods of mathematical modeling ‐ not only classical methods [1, 2], but also those based on the use of artificial intelligence [3, 4] ‐ are widely used to analyze and control blast-furnace smelting. It is interesting to examine the feasibili ty of using neural networks to calculate the effect of the parameters of a combination blast (the consumption of natural gas and the oxygen content of the blast) on smelting indices ‐ furnace productivity and coke rate. We used instructional neural networks to calculate the process parameters in question. In this case, the calculation models not the smelting operation itself, but the actions that might be taken by an experienced melter foreman, metallurgical engineer, or other technologist. The aim of the modeling is to determine the control actions that should be taken and to develop recommendations. This involves also using classical mathematical models of blast-furnace smelting. The neural network, instructed by the collective experience of above-mentioned technologists, functions as a tool for controllling the blast furnace with the use of a computer. Such a network has been developed and incorporated into a program created by UGTU‐UPI: “Neuro-Robot ‐ the Melter Foreman’s Partner.” So that the program has the desired effect, it is augmented by use of the technique of calculating response functions (objective functions) and a method developed to determine the degree of competency of robots based on neural network technology (neuro-robots). The use of these additional tools significantly enhances the reliability of the results that are obtained. The program includes a file of Excel tables and files with designs of neural networks. To generate and instruct the neuro-robot, we are also using a standard software item for working with neural networks in Excel ‐ the application package Excel Neural Package developed by the company Neurok. This software includes the component Winnet 3.0 and is designed to find and model relationships hidden in large files of numerical information ‐ relationships that cannot be expressed by any known explicit analytical expression. Figure 1 shows the algorithm used to design the operation of the Neuro-Robot program. We used typical information on the current operation of a 2700-m 3 blast furnace (Table 1) to formulate the instructional data sample, which incorporates the technologists’ expert opinions on the most representative and characteristic operating situations (OSs) encountered with blast furnaces. The quantitative (number of (OSs) and temporal (the duration of the “time window”) dimensions of the sample are determined in relation to a number of factors (furnace productivity, coke rate, blast temperature, the silicon content of the pig iron, etc.), and the ultimate goal of the program’s use includes real-time control of the smelting regime, prediction of the outcomes, analysis of trial heats, and other objectives. As an example, we chose 12 weeks as the size of the time window and one week as the interval for averaging the data (see Table 1). The program was used to calculate recommendations on the optimum values for the consumption of natural gas and the oxygen content of the blast in order to be able to incorporate these two parameters into the monthly operat