پیش بینی محتوای سیلیکون در فلز داغ با استفاده از رگرسیون بردار پشتیبانی مبتنی بر بهینه سازی ازدحام هرج و مرج ذرات

The prediction of silicon content in hot metal has been a major study subject as one of the most important means for the monitoring state in ferrous metallurgy industry. A prediction model of silicon content is established based on the support vector regression (SVR) whose optimal parameters are selected by chaos particle swarm optimization. The data of the model are collected from No. 3 BF in Panzhihua Iron and Steel Group Co. of China. The results show that the proposed prediction model has better prediction results than neural network trained by chaos particle swarm optimization and least squares support vector regression, the percentage of samples whose absolute prediction errors are less than 0.03 when predicting silicon content by the proposed model is higher than 90%, it indicates that the prediction precision can meet the requirement of practical production.

Blast furnace is the governing process of ironmaking in the ferrous metallurgy industry. The complexity of the heat and mass transfer process coupling with a large number of gas–solid, solid–solid and solid–liquid reactions, combustion processes and interphase mass transfer makes the modeling of blast furnace an extremely difficult problem. The prediction of key operational parameters, such as silicon content in hot metal, has been a major research issue. Silicon content is one of the most important indices to represent the thermal state of the blast furnace, its accurate and advance prediction can greatly help in stabilizing the blast furnace operations (Zhou, 2007). Numerous studies have focused on the accurate prediction of silicon content in hot metal by using statistical approaches and artificial intelligence approaches. However, the relationship among the operational parameters is complex and highly nonlinear, and the data collected from blast furnace are also quite noisy. Neural network modeling has been shown to reproduce nonlinear data very well while it was not very well explainable and often had the problem of overfitting leading to poor performance (Chen, 2001 and Zhang and Jin, 2007). To obtain further useful information, many predictive systems have been developed by integrating the mathematical model with expert system. Successful expert systems and predictive systems have been gradually adopted by many blast furnaces in different countries for on-line or off-line processes; however, operational results showed that some of these models did not perform well (Liu & Liu, 2003). Generally, systems based on empirical knowledge have been developed mainly by programming based on rules, while they are characterized as serious limitation to inference method based on the empirical connection between the observable findings and the metallurgy criterion (Liu & Liu, 2003). Growing efforts are made to explore innovative methods to increase prediction performance. Since the creation of the theory of support vector machines (SVM) (Vapnik, 1998 and Vapnik, 1999), rapid development of SVM in statistical learning theory encourages researchers to actively focus on applying SVM to various research fields such as document classifications and pattern recognitions. SVM possesses a great potential and has shown a superior performance as is appeared in many previous researches. This is largely due to the structural risk minimization (SRM) principle in SVM which has greater generalization ability and is superior to the empirical risk minimization (ERM) principle as adopted in the neural networks. In SVM, the results guarantee global minima whereas ERM can only locate local minima. For example, in the training process in neural networks, the results give out any number of local minima that are not promised to include global minima. Furthermore, SVM is adaptive to complex systems and is robust in dealing with corrupted data. This feature offers SVM a greater generalization ability which is the bottleneck of its predecessor, the neural network approach (Deng & Tian, 2004). On the other hand, applications of support vector regression (SVR) extended by SVM (Musicant and Alexander, 2004, Na and Upadhyaya, 2006, Quan, 2004 and Shevade and Keerthi, 2000), such as forecasting of financial market, estimation of power consumption and prediction of highway traffic flow, have been also under development and have shown many breakthroughs and excellent performances (Ding and Song, 2008, Van and Suykens, 2001, Shin et al., 2005 and Scholkopf et al., 2000). The time-varying properties of SVR applications resemble the time dependency of silicon content prediction, combined with many successful results of SVR predictions encourage our research in using SVR for silicon content modeling. Because the quality of SVR models depends on a proper setting of SVR meta-parameters, the main issue for practitioners trying to apply SVR is how to set these parameter values to ensure good generalization performance for a given dataset. Whereas existing sources on SVM regression (Deng and Tian, 2004, Kwok, 2001 and Smola and Schölkopf, 1998) give some recommendations on appropriate setting of SVM parameters, there are clearly no consensus and contradictory opinions. Hence, resampling remains the method of choice for many applications. Unfortunately, using resampling for tuning several SVR parameters is very expensive in terms of computational costs and data requirements. In this paper, we propose a new particle swarm optimization algorithm based on chaos searching (CPSO) to search the optimal parameters of SVR, then a prediction model of silicon content in hot metal (CPSO-SVR) was established by using SVR based on CPSO. The results show that this model is applicable to silicon content prediction and outperforms some previous methods. The rest of this paper is organized as follows. Section 2 provides a brief analysis of support vector regression. In Section 3, particle swarm optimization, chaos particle swarm optimization and prediction model based on CPSO-SVR are presented. Experiment and analysis are discussed in Section 4. Finally, Section 5 provides a summary and a conclusion.

In this study, we applied SVR optimized by CPSO to the prediction of silicon content in hot metal. Results show that SVR can serve as a promising alternative for existing prediction models. It can be seen from the experiment that the prediction model overcomes the main shortage of artificial neural network without defining network structure and trapping in the local optimum, so it is applicable to blast furnace process with complex relationship between input and output, and it can overcome the interference of noise data in blast furnace. Compared with SVR and neural network trained by particle swarm optimization (CPSO-NN), CPSO-SVR outperforms than SVR and CPSO-NN. The CPSO-SVR prediction model can be used on-line successfully to make a suggestion to help operators to make the right decision. In this sense, it is necessary to improve the performance of the system according to deep analysis on mechanism of the prediction model. Furthermore, efforts will be made toward a high efficient and accurate model beyond current levels. We also need to investigate the prediction performance of models on large size of training sets once enough samples are obtained.