Power Consumption Forecasting for Industrial Systems
Published:
Overview
In my Bachelor’s thesis at HTWG Konstanz, I developed and compared machine learning models for forecasting industrial power consumption. The work involved preprocessing production-related time series data, training multiple models, and evaluating them using standard error metrics.
The project was conducted using real-world production data provided by the industrial company Fondium in Singen, Germany, and focused on improving prediction accuracy for energy management.
Methods & Tools
- Data Processing: Python (Pandas, NumPy, Scikit-learn, statsmodels)
- Models: ARIMA, LSTM (PyTorch), XGBoost
- Evaluation Metrics: MAE, RMSE, R²
- Visualization: Matplotlib, Seaborn
- Environment: PyCharm, Anaconda
Results
- XGBoost and LSTM achieved the highest prediction accuracy (R² ≈ 0.95), with XGBoost being significantly faster in training.
- An application scenario demonstrated potential weekly cost savings of around €309 through optimized production scheduling based on model forecasts.
Model performance comparison on validation (10 splits) and test datasets:
| Model | MAE (Val.) | RMSE (Val.) | R² (Val.) | MAE (Test) | RMSE (Test) | R² (Test) | Training Time |
|---|---|---|---|---|---|---|---|
| ARIMA | 33.442 | 42.625 | 0.698 | 32.932 | 41.144 | 0.741 | 69.679 s |
| LSTM | 18.073 | 23.424 | 0.915 | 12.481 | 16.353 | 0.959 | 35.371 s |
| XGBoost | 16.801 | 21.657 | 0.927 | 13.436 | 18.032 | 0.950 | 0.520 s |
These results demonstrate that XGBoost offers an optimal balance between prediction performance and computational efficiency.
Notes: Due to confidentiality, the full thesis document is not publicly available.