Deep ensemble decision network for predicting quality of converter steelmaking with small variance

Abstract

It is the key product quality requirements to simultaneously control endpoint temperature and composition of converter steelmaking process within design range. The oxidation removal reaction of carbon (C) and phosphorus (P) is a complex multiphase reaction with high-temperature, which is difficult to analyse its mechanism and realise modeling. According to the distribution of the sampled data, a deep probabilistic ensemble network is firstly designed to achieve quality prediction of the converter steelmaking process with small prediction error variance. The first network layer (level-0) consists of parallel random forest regression (RFR), extreme random forest (ERF) regression and extreme gradient boosting (XGBoost). The hyperparameters in level-0 are optimised by intelligent Bayesian optimisation algorithm. The stacked RFR structure can effectively reduce the problem of single RFR prediction deviation. The extreme randomness of ERF can enhance the performance of the stacked network to capture the diversity of data features. The parallel computing mechanism of decision tree in XGBoost can improve the training speed of stacked networks. The second network layer (level-1) uses linear regression (LR) to effectively integrate the output features to obtain the quality prediction value. Finally, by comparing with the existing networks and ablation study tests, the superiorities of the proposed network are verified by using production data from an actual steelmaking plant. The variance of the proposed network for the endpoint temperature, C and P are 73.244, 1.8 × 10⁻⁵ and 4.8 × 10⁻⁶, respectively, and others evaluation indicators have also been significantly improved by only relying on small sample data.