Improving steel making off-gas predictions by mixing classification and regression multi-modal multivariate models

Abstract

This paper addresses the problem of real-time short-term multi-period off-gas prediction in a steel making batch process, denominated Linz-Donawitz Gas (LDG). Baselines, heuristic statistical methods, multi-modal multivariate Long Short-Term Memory (LSTM) and Ensemble Gradient Boosting Decision Tree (GBDT) strategies were proposed and compared. Proposed methods, mixing classification and regression tasks, achieved good results on recoverable LDG prediction, establishing a benchmark on subject for future works. Experiments suggest improvements from 19.4% to 15.85% on average in mean absolute percentage error (MAPE) over recent reviewed papers within a similar scenario at same steel making plant.