Energy consumption forecasting in non-stationary machining processes based on a physics-guided transformer

Predicting energy consumption in non-stationary machining processes is challenging due to data complexity, dynamic variations, and real-time requirements. This paper proposes a novel physics-guided Transformer model incorporating supervisory-compensatory mechanisms. Firstly, several data preprocessing and feature extraction techniques, including Lagrange interpolation, wavelet transform, and a combined approach of principal component analysis and correlation analysis, are employed to enhance data quality and identify key physical variable. Secondly, modeling accuracy and efficiency are improved by integrating physics-guided variables into the traditional Transformer model, leading to advancements in the existing de-stationary attention module. Finally, distinct training and prediction models are established, incorporating a self-supervised, self-compensation mechanism in the training phase. This mechanism utilizes ground truth for training convergence, with optimized model parameters subsequently applied to the prediction model, significantly enhancing predictive efficacy. Experimental results demonstrate that the proposed method outperforms state-of-the-art approaches, achieving improvements in energy consumption prediction accuracy of over 76 % for carbon fiber processing, 30 % for plastic, 32.7 % for aluminum, and 54.5 % for 45steel. The integration of physical principles with sequence modeling enables precise energy consumption forecasting in non-stationary machining, advancing both predictive accuracy and industrial energy management.