Abschlussarbeit

Global steel and iron production is responsible for 7.2 % of global greenhouse gases [Rit20]. To find emission-reduction strategies in this sector, researchers build global models of the steel cycle using a variety of modelling approaches and prediction techniques.

This thesis investigates inflow-driven and stock-driven dynamic stock modelling and proposes a new change-driven modelling approach. Additionally, the prediction technique by Pauliuk et al. [PWMA13] is compared to an approach proposed by Prof. Dürrwächter as well as a new machine learning-based technique using long short-term memory (LSTM) networks.

The three model approaches and three prediction techniques are implemented and combined to produce nine forecasts of the global steel production. Here, data available in 2008 is used to predict the years 2009-2022. Subsequently, this is com- pared to the actual production data in this time period to assess the performance of the methods.

Strengths and weaknesses of the six used methods are discussed. Overall, inflowdriven models and LSTM predictions performed best. The most accurate predicting model combination was the change-driven approach combined with the LSTM pro- ducing predictions with a mean absolute percentage error of 4.45 %. However, due to implausible lifetime projections in the change-driven model, its use was discouraged.

Hence, the inflow-driven approach combined with the LSTM prediction is used to derive final predictions. The resulting global production forecast suggest that the global steel demand until 2100 might be 25 % higher than previously [PWMA13] assumed, resulting in higher greenhouse gas emissions.