Parallel prediction method for damage factors and life loss of turbine high-pressure rotors based on soft parameter sharing mechanism of multi-task learning

In order to ensure the safe operation of steam turbines and extend rotor life, this study aims to develop a high-precision rolling prediction method for turbine rotor temperature, thermal stress and life loss at dangerous points. To address the limitations of traditional prediction methods, a parallel prediction method based on the soft parameter sharing of multi-task learning (MTL-SPS-PPM) is proposed. The model employs a soft parameter sharing mechanism to enhance knowledge transfer across tasks, a multi-model fusion strategy to improve feature extraction, and a rolling multi-step prediction framework for long-term trend forecasting. Furthermore, future power scheduling commands are incorporated as input features to improve adaptability to dynamic operational conditions. Experimental results demonstrate that MTL-SPS-PPM significantly outperforms traditional single-task learning models. Compared to the best-performing baseline models, MSE decreased by 0.305, 0.01, and 0.112, MAE decreased by 0.176, 0.135, and 0.089, and R 2 increased by 0.14, 0.067, and 0.051 for temperature, thermal stress, and life loss predictions, respectively. The proposed approach achieves higher prediction accuracy and efficiency, providing a reliable tool for real-time health monitoring and operational optimization of turbine units. This work lays the foundation for integrating physics-based modeling with deep learning to enhance predictive performance under complex operating conditions. • MTL-SPS-PPM was proposed for parallel prediction of rotor damage factors and life loss. • Prediction accuracy was enhanced with future power scheduling commands as input features. • Soft parameter sharing mechanism improves multi-task learning efficiency. • Multi-model fusion strategy boosts feature extraction and time-series learning. • Rolling multi-step prediction framework ensures long-term trend forecasting.