A digital twin approach for gas turbine performance based on deep multi-model fusion

Abstract

Engine model plays a crucial role in various applications of energy system, such as health management and control optimization. However, the traditional physics-based model is challenging to address the engine-to-model mismatch while the data-driven model has difficulty in obtaining high-quality data. In this paper, a digital twin approach which innovatively integrates the machine learning, performance adaptation and component matching method is proposed to fuse physical mechanisms and big data. The bidirectional data flow between data-driven model and physics-based model is established through deep multi-model fusion. The data-driven model is constructed based on extreme learning machine for mapping adaptation factors. Whereby the physics-based model receives information to fine-tune its components. Meanwhile, the results of component matching are employed as a stable data source for predicting of data-driven model. Particularly, the parameters training the data-driven model are selected in a physics-based way based on the first-order approximation of engine. Compared with traditional adaptive method and hybrid method, the proposed approach can improve the simulation accuracy by 16.6 % and 9.9 %, respectively. The results indicate that the developed digital twin model can be used as an effective and reliable tool to predict gas turbine performance and contribute to decision-making.