A Mechanism–Data Hybrid Approach for Predicting Energy Consumption in CNC Machine Tools

Accurate predictions of CNC machine tool energy consumption are crucial for sustainable manufacturing but remain challenging due to complex nonlinear dynamics. This paper proposes a mechanism–data hybrid framework combining physical modeling with an Attention–LSTM network. Unlike existing parallel hybrid models, this approach embeds the mechanism model’s output as a strong prior into the neural network, explicitly guiding the learning of nonlinear residuals. First, a hierarchical decoupled mechanism model is constructed to establish the physical baseline of energy consumption. Second, an Attention–LSTM network is designed to compensate for dynamic errors caused by tool wear and thermal variations. Finally, experimental validation on a three-axis CNC milling machine demonstrates that the proposed method significantly outperforms meaningful baselines, achieving a Root Mean Square Error (RMSE) of 0.0610 and an R2 of 0.9936. The framework provides a robust, physically interpretable solution for energy monitoring in intelligent manufacturing systems.