The electric arc furnace represents the most common way of recycling steel scrap. Nevertheless, optimizing its parameters or predicting its outputs remains challenging due to the variability in scrap metal composition. While the classification of the scrap helps to some extent, there is still volatility within the same class, which can lead to significant deviations in electrical energy consumption. In addition, the scrap metal is often contaminated with non-ferrous metals, such as copper, which can lead to quality issues in the final product. Attempts to address these issues with traditional statistical methods are often unable to capture the complexity of an electric arc furnace, the process, and the reactions. Machine learning techniques to accurately predict the impact of raw material composition on energy consumption are increasingly being used to address this issue. These models significantly enhance operational efficiency and product quality. As a contribution to this area of study, this thesis applied several machine learning algorithms such as linear regression, random forest, gradient boosting, and multi-layer perceptron to predict the electrical energy consumption and copper content of the heat produced. To make predictions, the algorithms were trained with features like scrap category composition, charged scrap weight, and some time-shifted features from previous time steps utilizing different evaluation techniques. In addition, scrap categories were split into two feature sets, one of which included the scrap categories and the other the more granular storage location. The inclusion of previous time steps was done in order to capture the dynamic behavior of the furnace and its hot heel. In addition, the different cross-validation techniques, including k-fold, sliding window, and expanding window validation, were used to evaluate and understand whether the data must be treated as a time series. Finally, the models trained separately on each electric arc furnace were compared with those trained on both furnaces to understand whether it is necessary to train a model for each furnace or if a general model can be used. Given this experimental setup, the results showed that the inclusion of time-shifted features improved the accuracy of the prediction of the models. However, this improvement was not consistent across all models and tended to end up overfitting the longer the window size was chosen. Assessing the different evaluation techniques, k-fold cross-validation provided the most reliable results, while sliding and expanding window validation showed significant inconsistencies, hinting that treating the data as a time series is unnecessary. Lastly, the comparison between the models trained on each electric arc furnace separately or on both furnaces showed that the random forest and gradient boosting model benefited from this split. The random forest model, trained on the scrap categories for each electric arc furnace separately, was identified as the best-performing model across all machine learning models. It achieved a mean absolute error of 0.0400 ± 0.0011 for the copper content and 1356.00 ± 34.08 kWh for the energy consumption.