A static prediction method for BOF steelmaking end-point based on PSO-optimised KPCA-KNNWTSVR

Abstract

The research focuses on the accurate prediction of end-point carbon content and temperature in basic oxygen furnace (BOF) steelmaking. This article proposes a static prediction method based on PSO-optimised KNN-weighted twin support vector regression (KPCA-KNNWTSVR). Firstly, key input variables, such as initial hot metal conditions, raw materials, and operational parameters, are selected through grey relational analysis to eliminate redundancy. Subsequently, Kernel Principal Component Analysis (KPCA) is utilised to extract nonlinear features and suppress noise. Simultaneously, sample weights are calculated based on the K-nearest neighbours (KNN) algorithm to enhance local fitting capability. The feature vectors and weights are then integrated into the weighted twin support vector regression (KNNWTSVR) model. Furthermore, the PSO algorithm is employed to adaptively search for the model's hyperparameters to further improve its generalisation performance. Simulation and experimental results indicate that the proposed method outperforms traditional models in both prediction accuracy and stability. With allowable errors of ±0.01 wt-% for carbon content and ±15°C for temperature. The model achieved hit rates of 80% and 88% on 100 actual production datasets. Compared with the Support Vector Machine (SVM) model, the hit rate for carbon content prediction is improved by 5 percentage points. Compared with the backpropagation (BP) neural network model, the hit rate for temperature prediction is improved by 20 percentage points. This method performs best among all compared approaches.