Elsevier

Computers & Chemical Engineering

Volume 34, Issue 11, 8 November 2010, Pages 1854-1862
Computers & Chemical Engineering

Data-driven predictive control for blast furnace ironmaking process

https://doi.org/10.1016/j.compchemeng.2010.01.005Get rights and content

Abstract

High performance control of blast furnace (BF) ironmaking process is a difficult problem due to the high temperature and hostile measurement conditions for measuring devices in the process. Previous research focused on developing of accurate predictive models for silicon content in hot metal ([SI]) while control of the whole process is seldom discussed. In the present work, a data-driven predictive control method based on subspace method is presented for the blast furnace ironmaking process. The algorithm is based on input–output data and easy to implement. Simulation results show the algorithm is effective for the control application. Finally, various practical issues concerning predictive control of blast furnace ironmaking process are also addressed, such as constraint handling, control objective and output set-point selection, adaptive strategy, etc.

Introduction

Blast furnace is a kind of shaft furnace used to produce hot metal from iron ore for subsequent processing into steel. The main fuel used is coal and coke. The production process consumes a large amount of energy and emits plenty of carbon oxide which is the main cause of the greenhouse effect. Thus efficient use of energy and emission reduction becomes the primary concern. As one of the most energy intensive industrial process, the efficiency of blast furnace ironmaking has been an important problem. To increase the efficiency, modern blast furnaces are equipped with an array of supporting equipment like cowper stove to preheat the blast air, recovery systems to extract the heat from the hot gases exiting the furnace, etc. Except for technological measures, efforts have been made on thermodynamic and kinetic research and research has been made in constructing mathematical models to simulate and understand the inner phenomena in blast furnace (Bi et al., 1992, Koki et al., 2005, Nogami et al., 2005). However, due to different kinds of practical difficulties, the reported results are confronted with some unavoidable shortcomings when applied to practice, e.g. weak real time performance, excessive assumptions and simplications (Gao, Chen, Zeng, Liu, & Sun, 2009).

Recently, mathematical models have been extensively explored to deal with the control problem of blast furnace ironmaking process, most of which are concentrated on development of accurate predictive models for silicon content in hot metal. As a main indicator of thermal state in blast furnace and the most important index of pig iron quality, silicon content in hot metal must be kept at an appropriate level to facilitate the production and stable running of the ironmaking process. Thus future information about the silicon content becomes very important for blast furnace operators to judge the status of blast furnace. With such information, corrective measures can be applied to control the process. There are many research results on this issue, such as time series models (Gao et al., 2008, Waller and Saxén, 2000), statistical models (Bhattacharya, 2005), neural networks (Chen, 2001, Hao et al., 2005, Radhakrishnan and Mohamed, 2000), and others (Gao et al., 2005, Waller and Saxén, 2003). Although plenty of studies have been carried out in the last decades, experience and intuition of skilled operators are still the main method to implement blast furnace operation today. This, together with the important impact of blast furnace on the national economy, makes the research on blast furnace modeling and control still very active not only in theory but also in practice in the foreseeable future. More importantly, slight improvement of blast furnace mathematical model may result in considerable profit because of the large quantity produced. Besides, predictive models also have some shortcomings. In practice, the BF operators take control actions based on information about future silicon content obtained through predictive models, then the deviation between actual and predicted silicon content will appear and finally the operators will not trust the predictive results. To overcome this problem, it is necessary to consider both the problem of modeling and control. Model predictive control (MPC) may be a viable approach.

In this paper, a data-driven predictive controller for blast furnace ironmaking process will be constructed by combining subspace identification method and model predictive control. During the design of MPC for the ironmaking process, the solution of a quadratic programming (QP) problem is avoided through updating the input weighting matrix in the cost function. Discussion on practical issues like constraint handling, control objective and output set-point selection, adaptive strategy, etc. are also introduced.

This paper is organized as follows: in Section 2, a brief description of the blast furnace ironmaking process is given. In Section 3, the subspace identification method is presented. Section 4 gives the MPC strategy with constraint handling and Section 5 considers some practical issues in the design of MPC for the blast furnace ironmaking process. Simulation results and conclusion are presented in Sections 6 Simulation results, 7 Conclusion.

Section snippets

Process description

In a blast furnace, raw materials like iron ore, coke and limestone are dumped into the top, and preheated air is blown into the bottom. The raw material requires 6–8 h to descend to the bottom of the furnace where they become the final product of liquid hot metal and slag. These liquid products are drained from the furnace at regular intervals. The hot air that was blown into the bottom of the furnace ascends and exits from the top in 6–8 s after going through numerous chemical reactions.

During

Subspace identification method

In the last decades, subspace identification methods have obtained great development in both theory and practice and offer an attractive alternative to input–output methods due to simple and general parameterization for multi-input multi-output (MIMO) system. Three basic subspace identification methods are established, including N4SID (Overschee & De Moor, 1994) MOESP (Verhaegen & Dewilde, 1992) and CVA (Larimore, 1990). They have been applied to various industrial processes from chemical

Model predictive control

For most MPC strategies, the classical way is to identify a model of the system and then use the model to design a controller. However, it is found that the system identification step and calculation of controller parameters can be achieved by a single QR-decomposition to the input–output data based on subspace method (Favoreel et al., 1998).

The model predictive control problem can be expressed as minimization of a cost functionJ=k=1N1(rt+ky˜t+k|t)2+k=1NuΔut+k1RΔut+k1where N1 and Nu are

Practical issues

The processes of concern, BF1 with the inner volume of 2500 m3 and BF2 with the inner volume of 2000 m3, form together with their supporting equipment 2 complex industrial systems. BF1 is located in northwest China; the iron ore used for this blast furnace contains materials like rare earth elements, niobium and other metals, thus the ironmaking process is relatively unstable. BF2 is located in north China and uses better iron ores; the ironmaking process is relatively smooth. Due to the big

Simulation results

With all the practical issues considered, the MPC can then be applied. Since one step prediction of silicon content can give good accuracy while two step or three step predictions do not share equal accuracy (Gao et al., 2009), the prediction horizon N1 is set to be 1. Furthermore, the time interval of data is 1.5 h so that control horizon could not be great; here we also set it to be Nu=1. For both BF1 and BF2, the data set is divided into two parts, training set and test set. The training set

Conclusion

In this paper, a data-driven predictive controller is designed to control the blast furnace ironmaking process. The algorithm is based on the subspace method and is suited for multi-variable processes like the blast furnace ironmaking process. A constraint handling method is introduced to handle constraints for input variables. Other practical issues when applying the predictive controller to blast furnace ironmaking process are also discussed, including control objective and output set-point

Acknowledgements

The authors acknowledge the financial contribution of National Natural Science Foundation of China under Grant No. 10826100, 10901139 and 60964005, Zhejiang Provincial Natural Science Foundation of China under Grant No.Y107110, the Open Project of State Key Laboratory of Industrial Technology, Zhejiang University, under Grant No. ICT0904 and Zhejiang University Youth Seed Foundation for Interdisciplinary studies in 2009.

References (28)

  • E.F. Camacho et al.

    Model predictive control

    (1999)
  • W. Favoreel et al.

    SPC: Subspace predictive control

  • W. Favoreel et al.

    Model-free subspace based LQG design

  • G. Feng et al.

    Adaptive control systems

    (1999)
  • Cited by (0)

    View full text