Comparison of ARIMA, LSTM, and XGBoost for Power Consumption Forecasting
Date:
Bachelor thesis defense on the comparison of ARIMA, LSTM, and XGBoost models for forecasting the power consumption of an industrial molding plant.
The case study was conducted using data from Fondium, a manufacturer of iron cast parts. The presentation covered the motivation, data preprocessing, feature extraction, model implementation, and evaluation. The models were trained and tested on power consumption data from January to November 2023, with features including production data, working days, downtime, and time-based variables. Evaluation metrics such as MAE, RMSE, R², and training time were used to evaluate model performance.
Results showed that XGBoost and LSTM achieved the highest prediction accuracy (R² ≈ 0.95), with XGBoost being significantly faster in training. An application example demonstrated potential cost savings of approximately €309 through optimized production scheduling based on model forecasts.
The presentation was given in German.
📥 Download the presentation