Combustion Dynamics Data Mining Techniques: A Way to Gain Enhanced Insight in the Combustion Processes of Fielded Gas Turbines

ABSTRACT Combustion dynamics are still an important challenge for the gas turbine operators. Modern dry low NOx combustors operate within very small tolerances of equivalence ratio, air-fuel mixing and heat release rate in order to attain low NOx emissions and combustion stability. Small changes in fuel composition, or extremes in ambient temperature can trigger combustion instabilities. Large amounts of data of real engines are available to the end user. Moreover, instead of adaptations to the hardware, the end-user is primarily interested in the actual condition of its gas turbine. Although physical insight is without any doubt an important step to enhance knowledge of the processes within the combustion chamber, these large datasets can also be exploited with data-mining techniques based on black box models, such as artificial neural networks or decision trees. In this paper, the latter approach is discussed in detail and implemented on a F-class gas turbine. The operational and combustion data, acquired over a long period on the gas turbine, have been used as the input to a commercial data-mining program in order to study the correlations between the different operational parameters and the characteristic amplitude and frequency of the combustion dynamics. Moreover, the data-mining program allows the non-linear modelling of the combustion dynamics, which in a second step has been used to carry out a parametric study. The parameters with a high influence, amongst others the gas quality, the compressor inlet temperature and the firing temperature, on the presence of combustion dynamics have been retained for modelling the behaviour of the combustion dynamics. The obtained models show good correspondence with operational experience and data gathered during gas turbine tuning operation. These models can thus be used to enhance the insight into the complex behaviour of combustion dynamics. They can be helpful for predictive maintenance and finally can be applied for the determination of tuning margins and the prevention of high combustion dynamics.