Abstract
The intermittent nature of operation and unpredictable availability of renewable sources of energy (e.g., wind and solar) would require the combustors in fossil-fuel power plants, sharing the same grid, to operate with large turn-down ratios. This brings in new challenges of suppressing high-amplitude pressure oscillations (e.g., thermoacoustic instabilities (TAI)) in combustors. These pressure oscillations are usually self-sustained, as they occur within a feedback loop, and may induce severe thermomechanical stresses in structural components of combustors, which often lead to performance degradation and even system failures. Thus, prediction of thermoacoustic instabilities is a critical issue for both design and operation of combustion systems. From this perspective, it is important to identify operating conditions which can potentially lead to thermoacoustic instabilities. In this regard, data-driven approaches have shown considerable success in predicting the instability map as a function of operating conditions. However, often the available data are limited to learn such a relationship efficiently in a data-driven approach for a practical combustion system. In this work, a proof-of-concept demonstration of transfer learning is provided, whereby a deep neural network trained on relatively inexpensive experiments in an electrically heated Rijke tube has been adapted to predict the unstable operating conditions for a swirl-stabilized lean-premixed laboratory scaled combustor, for which data are expensive to obtain. The operating spaces and underlying flow physics of these two combustion systems are different, and hence this work presents a strong case of using transfer learning as a potential data-driven solution for transferring knowledge across domains. The results show that the knowledge transfer from the electrically heated Rijke tube apparatus helps in formulating an accurate data-driven surrogate model for predicting the unstable operating conditions in the swirl-stabilized combustor, even though the available data are significantly less for the latter.