A physical-data-driven combined transfer learning method for gas turbine performance estimation

Abstract Gas turbine performance estimation methods rely on historical data from the same-type gas turbine. However, due to individual differences, this knowledge cannot be directly shared. In this paper, we propose a physical-data-driven combined transfer learning method. This method enables knowledge sharing between a gas turbine with complete washing cycle data and a newly commissioned gas turbine. The physical-driven models are constructed utilizing particle swarm optimization to acquire the labels used for pretraining. A data-driven model is developed through a convolutional neural network that employs labels to extract degradation features. Then, the labels for fine-tuning the network. Historical degradation knowledge is transferred to the newly commissioned gas turbine. Experiments results demonstrate that the accuracy of this method is improved by at least about 25 % compared to other methods. The computation time is reduced by at least about 98 % compared to physical-driven methods. This method enables effective performance estimation of gas turbines.