When wanting to lower the atmospheric carbon dioxide concentration, it is of interest to look at the iron and steel industry, as especially the blast furnace process with its large coal consumption is a significant emittent. Electric steel mills, on the other hand, use electricity and natural gas to melt scrap metal into steel. Demand side management can thereby help to integrate renewable energy sources and increase energy efficiency. In this context, load forecasting is an important tool to know the future energy demand of such an industrial process. The purpose of this work is to develop a machine learning model that can predict the power demand of an electric steel mill’s primary aggregates as accurately as possible. Due to the high power demand of the electric arc furnace, a focus is laid on this aggregate. In the course of this thesis, literature research was conducted about machine learning algorithms, their advantages and disadvantages, and their use cases. In particular, machine learning methods used in energy system modelling were researched. A suitable method was then chosen, and based on this method, multiple models were created to forecast the aggregates’ power demand in a time-resolved manner. The method chosen for predicting the power demand was a neural network. Two types of neural networks were compared: long short-term memory networks and standard feedforward networks. Altogether six models were created, of which five are based on long short-term memory networks. The results show that long short-term memory networks can be used to predict the power demand of an electric arc furnace. By stochastically generating input parameters and realigning the predicted and actual batches of the electric arc furnace process, the model can be implemented in and used for demand side management applications.