Accurately predicting the energy demand of Computerized Numeric Control (CNC) machining processes before production enables the assessment of a product’s CO₂ footprint, the identification of optimization opportunities, and the implementation of energy-aware scheduling strategies. However, forecasting the energy consumption of CNC machines equipped with regenerative drives presents unique challenges, as the energy demand of a given G-command is influenced by the preceding operation. This study investigates the application of time-series Machine Learning (ML) models to better capture these temporal dependencies and improve energy consumption accuracy. A significant variance in repeated measurements was observed during the experimental phase, prompting a comparative analysis of using raw versus averaged energy values as input data. Multiple time-series model architectures, including Long Short-Term Memory (LSTM) and Temporal Convolutional Networks (TCNs), are evaluated for their ability to learn sequential patterns in a 5-axis machining process. The results reveal that while ensemble methods such as LightGBM and Random Forest achieve the highest accuracy and efficiency on the test dataset, sequence-based models demonstrate greater robustness on unseen validation data. Incorporating a small portion of validation data into training further improves ensemble performance, highlighting the trade-off between robustness and efficiency in energy demand prediction.
Schmitt et al. (Thu,) studied this question.