Towards predictive control of ferroalloy furnaces: combining thermochemistry, inventory modelling and systems engineering

Ferroalloy furnaces are well known to be extremely difficult to control from a metallurgical perspective. Mostly metallurgical control is still reactive in nature, causing a significant amount of reworking and product blending to achieve a proper product specification. Despite numerous efforts towards the model based control of primary smelting operations, actual metallurgical feedforward control seems to remain elusive. One of the key reasons in the failure to model these metallurgical reactors well, is that fundamentally orientated pyrometallurgists, statisticians, data miners and control engineers very often have diverging ideas how to model these systems. Furthermore, metallurgists have often come to rely on their ability to develop empirical regression models of furnace behaviour, as they have lost their faith in the ability of thermochemical and metallurgical engineering fundamentals to accurately predict the outcomes from furnaces. This paper shows how good thermochemical modelling can be combined with proper technological and systems analysis, as well as a proper inventory model, to predict the alloy and slag chemistry with good accuracy one tap into the future, at a time horizon of typically around 4 to 6 hours. For this modelling approach, a number of tools will be used, such as data reconciliation, thermodynamic modelling using FactSage®, inventory and heel estimation, neural networks and statistical data exploration. Furthermore, systems identification theory will be combined with the approaches given above to develop a truly dynamic predictive model. This paper seeks to demonstrate the need for a hybrid approach, which incorporates plant data and readily accessible thermodynamic data to develop a proper dynamic model.