Artificial intelligence for electrical steelworks modelling 

by Antonella Zaccara, Laura Laid and Stefano Dettori, Scuola Superiore Sant’Anna 

 

The role of electric steelworks is crucial in driving the steel industry transition to more sustainable and circular processes as, through the recycling of ferrous scrap, it fully reflects the concept of circular economy. Despite this, important improvements are necessary to meet the ambition of the “Green Deal Industrial Plan” that aims to accelerate the industrial transition to climate neutrality [1] [2]. To achieve the environmental goals and, at the same time, promote a socially and economically sustainable transformation, research and innovation are crucial. Advanced digital technologies, including artificial intelligence (AI) based techniques, can contribute significantly to the achievement of a climate-neutral industry [3] [4] [5]. 

The European project ALCHIMIA aims at minimizing energy and resource consumption by improving Electric Arc Furnace (EAF) charging and process management through the development of the ALCHIMIA platform, a software based on Federated Learning (FL) as a methodology for training Machine Learning (ML) models, which can support staff in decision-making processes to select the best input material mix. Modelling is the basis of the ALCHIMIA platform, as it represents the different process steps and their interconnection in affecting product quality and environmental impact.  

The modelling work covers all the stages of the EAF-based steelmaking route, starting from scrap purchasing to continuous casting, at the exit of the secondary metallurgy (Ladle Furnace, LF), as shown in Figure 1. 

Figure 1. Modelling phases 

Due to the predictive nature of the models that have to simulate non-linear systems and be integrated into the optimization framework, the need to handle complex industrial data, and the need to balance accuracy with computational cost, some ML techniques have been preferred.  

Ensemble technique 

The ensemble technique combines the predictions of several ML algorithms to obtain a more accurate one [6]. There are two main ensemble learning strategies.  

  • The first is bootstrap aggregation, also known as bagging, a meta-learning technique that trains many classifiers on different partitions of the training data and uses a combination of the predictions of all these classifiers to form the final prediction.  
  • The second is boosting, which works with weights in both the learning and prediction phases. During the learning phase, the boosting procedure trains a new model several times, each time adapting its parameters to the errors of the previous model. Once trained, the model provides a prediction based on a weighted combination of the predictions of each model [7]. 

In the Alchimia project, tests of both strategies are performed, in particular: 

  • Random Forest Regressor, a supervised learning algorithm belonging to bagging techniques, which uses an ensemble learning method for regression. The Random Forest trees are executed in parallel, and the result is aggregated into a global one [8] [9]. 
  • XGBoost, an implementation of gradient-boosting decision trees, which combine several simple regression trees, executed sequentially in a more robust model, so that each tree is trained on the error residuals of the previous sequence of trees [10] [6].  

Neural Network 

An artificial neural network, usually applied for finding non-linear relationships between inputs and output, is an interconnected network of neurons with parallel information processing. It can predict or solve unknown instances of problems by training models input-output from experimental data. Developing a model requires flexibility and accuracy of prediction. They are composed by nodes or neurons, which are connected by each other and are simple computational components used to form an input layer, one or more hidden layers and an output layer. Each node represents a specific output function, called activation function, while the connection between each two nodes represents the weight for the signal passing through the connection, which is equivalent to the memory of the artificial neural network. The output of the network varies depending on how the network is connected, the value of the weight and the incentive function. Neural networks are generally categorized in feed-forward networks, competitive networks and recurrent networks [11] [12].  

In the Alchimia project, the methodology used is based on Multi-layer Feed Forward Neural Networks, a network specialized in data regression. The feed-forward neural network architecture is the most common architecture, in which information is generated only in one direction, from input to output, and the number of inputs is equal to the number of nodes in the network. 

 

[1] EUROPEAN COMMISSION, Communication from the commission to the European parliament, the European council, the council, the European economic and social committee and the committee of the regions. A Green Deal Industrial Plan for the Net-Zero Age, Bruxelles (2023). 

[2] INTERNATIONAL ENERGY AGENCY – IEA, Achieving Net Zero Heavy Industry Sectors in G7 Members (2022). 

[3] T. A. BRANCA, V. COLLA, M. M. MURRI, A. J. SCHRÖDER., The Impact of the New Technologies and the EU Climate Objectives on the Steel Industry in Industry 4.0 and the Road to Sustainable Steelmaking in Europe: Recasting the Future, Ed. Springer International Publishing, pp. 53-75 (2024). 

[4] V. COLLA, C. PIETROSANTI, E. MALFA, K. PETERS, Environment 4.0: How digitalization and machine learning can improve the environmental footprint of the steel production processes, Materiaux et Techniques, 108: 5-6, (2020). https://doi.org/10.1051/mattech/2021007  

[5] P. GAILHOFER, A. HEROLD, J. P. SCHEMMEL, C. SCHERF, C. URRUTIA, A. R. KÖHLER, S. BRAUNGARDT, The role of Artificial Intelligence in the European Green Deal, Policy Department for Economic, Scientific and Quality of Life Policies, European Parliament (2021) 

[6] E.K. SAHIN, Assessing the predictive capability of ensemble tree methods for landslide susceptibility mapping using XGBoost, gradient boosting machine, and random forest. SN Appl. Sci. 2, 1308 (2020). https://doi.org/10.1007/s42452-020-3060-1  

[7] Ensemble Models: Bagging & Boosting, https://medium.com/analytics-vidhya/ensemble-models-bagging-boosting-c33706db0b0b, ultimo accesso 05/07/2024.   

[8] A. B. K. DIDAVI, R. G. AGBOKPANZO, M. AGBOMAHENA, Comparative study of Decision Tree, Random Forest and XGBoost performance in forecasting the power output of a photovoltaic system, 2021 4th International Conference on Bio-Engineering for Smart Technologies (BioSMART), Paris / Créteil, France, pp. 1-5, 2021. 10.1109/BioSMART54244.2021.9677566 

[9] G. LOUPPE, Understanding Random Forests: From Theory to Practice, University of Liège (2014). 

[10] T. CHEN, C. GUESTRIN, XGBoost: A Scalable Tree Boosting System, In: Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining, pp. 785-794 (2016). https://doi.org/10.1145/2939672.2939785  

[11] Y-C WU & J-W FENG, Development and Application of Artificial Neural Network, Wireless Personal Communications, Springer Nature Link, Volume 102, pp. 1645–1656, (2018). 

[12] A. ROSTAMI, M. A. ANBAZ, H. R. E. GAHROOEI, M. ARABLOO, A. BAHADORI, Accurate estimation of CO2 adsorption on activated carbon with multi-layer feed-forward neural network (MLFNN) algorithm, Egyptian Journal of Petroleum, Volume 27, Issue 1, pp. 65-73 (2018). https://doi.org/10.1016/j.ejpe.2017.01.003