Prediction of acoustic pressure of the annular combustor using stacked long short-term memory network

Abstract

This paper proposes a data-driven method named stacked long short-term memory (S-LSTM) for predicting the future growth of acoustic pressure signals to detect precursors of combustion instability. The application of S-LSTM is investigated using the acoustic pressure data obtained from an annular combustor. The S-LSTM method is compared with the support vector machine (SVM) in terms of the predictive performance and also provides detailed insights into the influence of input choice by interpreting the results of S-LSTM. It is demonstrated that S-LSTM can effectively predict future pressure signals with a better error control performance compared to the SVM method. Furthermore, the feasibility of the S-LSTM in the thermoacoustic instability problem is verified using acoustic pressure data obtained from industrial combustion tests with a low-emission aero-engine. It is expected that the implementation of S-LSTM provides an early prediction solution to avoid thermoacoustic instability.