Identification And Prediction Of B.F. Ironmaking Process Based On Gaussian Processes

With the rapid development of economy, the problem of energy crisis, environment protection and global warming have received more and more attention from all over the world. It is widely recognized that cost and emission reduction is an essential way to fulfill the aim of sustainable development. As one of the most energy intensive industry in national economy, the iron and steel industry has enormous potential of cost and emission reduction. The blast furnace ironamking process is a primary sub-processes in steel industry. Hence cost and emission reduction has become an important issue in the blast furnace ironmaking process. This dissertation focuses on the identification of ironamking process via Gaussian process model using two data sets collected from No.1BF at Laiwu Steel and No.6BF at Baotou Steel respectively. Three aspects are considered, i.e., numerical prediction of silicon content, the noise model construction and tendency change prediction. This study will lay a solid foundation for the fulfillment of closed loop control of BF ironmaking process; it is helpful for cost and emission reduction in the blast furnace ironmaking process. Hence it is meaningful both in theory and practice.There are three focal points studied in this dissertation:ⅰ) Improving the accuracy of numerical prediction of silicon content to meet the requirement of the operators; ⅱ) Research on noise model of ironmaking process to facilitate the handling of outliers; ⅲ) Tendency change prediction of silicon content. All of the3points are difficult in the closed loop control of BF ironmaking process, so it is of great significance. The structure of this dissertation is as follows. After the Introduction section in Chapter1, Chapter2gave a brief summary to the Gaussian process model; Chapter3introduced the blast furnace ironmaking process and presented a brief literature review of the application of mathematical model in expert systems. Then the Gaussian process model (GP model) was adopted to build predicative model for numerical prediction of silicon content. Simulation experiments were carried out to test the GP model based on data collected from Laiwu Steel and Baotou Steel respectively. It is shown that the accuracy of prediction is relatively high for data collected from the Laiwu Steel BF, which is a relatively small blast furnace; while the accuracy of prediction model is relatively low for data collected from the Baotou Steel BF, which is a large blast furnace involving severe fluctuation during the blast furnace operation.A novel model based on Gaussian processes was proposed in Chapter4, which is named as sub-Gaussian process model in this dissertation. It is proved that this sub-Gaussian process is equivalent to GP model with a special covariance function and mean function. This sub-Gaussian process proposed a new method to construct covariance function (kernel) for GP model, which increased the application scope of GP model. Again, data collected from Laiwu Steel and Baotou Steel was used to test the proposed sub-Gaussian process. Simulation results shown that the prediction accuracy on both data sets are high, with a hit rate of85%for Laiwu Steel BF and a hit rate of79%for Baotou Steel BF.Due to the complexity of ironmaking processes, outliers arise frequently during the process. Chapter5discussed and concluded that there are three kinds of root causes which lead to outliers. To alleviate the effect of outliers, the robust Gaussian process was adopted to identify the BF ironmaking process. Robust Gaussian process is capable of handling outliers as well as analyzing the noise model. Simulation results on two BF data sets shown that the noise series for data set collected from Laiwu Steel BF approximates a Student-t distribution. It also confirmed that the noise series for both data sets did not follow a Gaussian distribution. The simulation results also shown that the robust Gaussian processes are stable even if there are outliers in data. A new model structure is then designed in the second part of Chapter5, where a basic model was adopted for modeling of ironmaking process, and then Gaussian process model was utilized to identify the errors series. To compare the modeling effects, auto regression model and Gaussian process were chosen as basic model. Simulation results on both data sets show that this kind of model structure can improve the accuracy of the basic model.Due to the importance of the thermal state change of BF hearth for operators to control ironmaking process. Chapter6focused on the prediction of thermal state change of BF. Two-classes thermal state change prediction model was built based on Gaussian processes classification model, and there are two kinds of tendency change in this model, i.e., cooling of the BFH and heating of the BFH. Simulation results show that the accuracy of this model based on Laiwu Steel BFH is satisfactory for practical use. Analysis on the results show that the difference between the volumes of the two BFs and the difference between the raw material in these two ironmaking process leads to different accuracies for the two data sets. Due to the importance of the prediction of the thermal state change of, the three state change prediction model as well as five thermal state change prediction model were further proposed using Gaussian process model. The results also confirmed that the accuracy of the model for both model were not satisfactory. The "boundary problem" is the primary problem leading to the low accuracy of the prediction, which means that the difference between different classes is not large and there are too many data samples located near the boundary. Finally, Chapter7gave the conclusion and summarized the theoretical innovations in this thesis and the direction of the further research also was investigated.