Gas turbine multi-working conditions identification and performance prediction based on deep learning and knowledge

Abstract

Performance prediction is crucial for monitoring, controlling and optimizing gas turbine (GT) operations. Due to significant performance variation under different working conditions, a single model cannot adequately represent all scenarios. In this paper, we propose a novel multi-working conditions performance prediction framework for GT, leveraging deep learning and professional knowledge. The GT is used in gas-steam combined cycle power plants that utilize low calorific value gases. A unique multi-working condition identification model has been established, enabling accurate identification of the current operating status of GT. Additionally, a dynamic model has been developed to fully utilize the temporally varying data. The entire GT model, spanning from unstart to steady-state, is constructed through model fusion using a mathematical mechanism. Compared to the single working condition model, our approach demonstrates superior performance prediction results on the actual GT operating dataset. The mean square error (MSE), mean absolute error (MAE) and correlation coefficient (CORR) are 0.7501, 0.5872, and 0.9973, respectively. These results highlight a substantial improvement in prediction accuracy and robustness outperforming the single working condition model in same contexts, effectively capturing operational characteristics and offering valuable insights for optimizing GT operations. The findings may also contribute to advancing GT research.