Lean premixed combustion has become state of the art technology in gas turbines for power generation because of its very low emission potential in the context of tightening pollutant emissions regulations. Lean premixed combustion is yet also prone to combustion instabilities, resulting in thermo-acoustically induced acoustic pressure oscillations (pulsations). Understanding pulsation behavior over an enginés lifetime is of interest to accurately monitor the engine status, as wear and degradation typically affect combustion behavior and result in changes of both pulsations and emissions. Such improved understanding can be exploited for optimizing both the engine operation concept and the design of relevant hardware parts. In return, pulsation and hardware optimization may lead to reduced degradation and thus inherently more robust long-term operational behavior.
The study presented here is conducted for one specific gas turbine of GE’s GT24/GT26 fleet with sequential annular combustion. Based on operational data of the examined gas turbine, a semi-empirical modeling approach is introduced to describe the pulsations measured in the first (EV) combustion chamber. The target is to reproduce measured pulsation amplitudes as well as their different behaviors with engine load. The modeling presented here has been focused on pulsations in a distinctive frequency range below 1kHz. A model based on a small set of data obtained from initial commissioning is able to represent the pulsation behavior within a normalized root mean square error of 11%. Validation with long-term engine data shows that predicted pulsation levels are reasonably matching the initial operation period but increasingly deviate with engine operating time. By using additional data from later engine commissioning and adjustments, the robustness of the model is sensibly increased. Model accuracy on the training dataset remains similar at around 11%, but validation on the long-term data shows a significant decrease of the normalized root mean square error from over 21% to below 16%. Additional model improvements to further reduce prediction errors on long-term data have been also identified.