Data-driven modelling, learning and stochastic predictive control for the steel industry

The steel industry involves energy-intensive processes such as combustion processes whose accurate modelling via first principles is both challenging and unlikely to lead to accurate models let alone cast time-varying dynamics and describe the inevitable wear and tear. In this paper we address the main objective which is the reduction of energy consumption and emissions along with the enhancement of the autonomy of the controlled process by online modelling and uncertainty-aware predictive control. We propose a risk-sensitive model selection procedure which makes use of the modern theory of risk measures and obtain dynamical models using process data from our experimental setting: a walking beam furnace at Swerea MEFOS. We use a scenario-based model predictive controller to track given temperature references at the three heating zones of the furnace and we train a classifier which predicts possible drops in the excess of Oxygen in each heating zone below acceptable levels. This information is then used to recalibrate the controller in order to maintain a high quality of combustion, therefore, higher thermal efficiency and lower emissions.