Research On Endpoint Prediction Model Of Basic Oxygen Furnace Steelmaking Based On Relevance Vector Machine

The endpoint carbon content and temperature prediction of basic oxygen furnace (BOF) steelmaking is an important component part of BOF steelmaking process control. The performance of endpoint prediction model has an important affect on whether the required molten steel can be obtained or not. BOF steelmaking is a complex nonlinear process, and it contains many physical-chemical reactions, so the process is difficult to be described by just using classical mechanistic analysis. In this study, on account of the features of practical BOF steelmaking data, modified relevance vector machine (RVM) and particle swarm optimization (PSO) algorithm are utilized to construct the endpoint prediction models.The environment of BOF steelmaking production workshop is complex and disturbance often emerges, therefore, the production data is vulnerable to be contaminated by noise or outliers. In addition, classical RVM is sensitive to outliers and has weak robustness. To overcome this drawback, a robust relevance vector machine is proposed. The individual noise variance coefficient is introduced for each training sample and the iteration formulas of hyperparameters and noise variance coefficients are derived according to Bayesian evidence procedure. In the process of training, the noise variance coefficients of outliers will be reduced with the increase of the prediction error, which can detect and remove outliers as well as improve the robustness of the model. In addition, an adaptive kernel relevance vector machine based on PSO is presented to deal with the problem that the regression performance of classical RVM is often influenced by kernel parameters. In the adaptive kernel RVM, different kernel parameters are set for different input features, and PSO is utilized to optimize the kernel parameters at an interval of several iterations so as to improve the regression accuracy of RVM. According to the optimization result of kernel parameters, the correlations between input and output variables can be analyzed. Based on the proposed robust relevance vector machine and kernel parameter optimization algorithm, endpoint carbon content and temperature prediction models of BOF steelmaking are constructed, and the practical data of a steel plant is used to take simulations. The results indicate that the proposed models are able to achieve good hit ratio and accuracy, which is promising in the practical utilization.