Session: 04-18: Combustor design IV

Paper Number: 102832

102832 - Multi-Objective Experimental Combustor Development Using Surrogate Model-Based Optimization 

To reduce pollutant emissions, stationary gas turbines are operated with lean premix combustion. However, the use of lean premixed flames entails several engineering challenges as the flames are susceptible to thermoacoustic instabilities, lean blow-off and flashback. In order to ensure flame stability, a careful design of the flow field in the combustion chamber is mandatory. To increase the operational flexibility, many combustion systems use multiple fuel injection lines to allow for a variable fuel distribution. The design of such an injection system must consider the emission of various emissions and both static and dynamics flame stability. Typically, these limitations become critical under different operating conditions of the machines. Therefore, it is particularly challenging to develop a combustion chamber that ensures stable and low-emission operation over a wide operating range. In practice, the development therefore requires computationally and financially expensive simulations and experiments.

Machine learning methods are finding increasing application in a wide variety of fields. Driven by the growing availability of computing capacity, data-driven methods for model adaptation and optimization have experienced a surge in development in recent years. In this study, we show how such a data-driven modelling method can be used for burner development overcoming problems of classical experience-based development approaches.

For the present study a special pilot unit is installed into a full-scale industrial premix swirl combustor to modify the mixture field close to the flame. The SLM printed pilot features 61 different positions of fuel injection. Each of the injector lines is equipped with an individual valve, such that the distribution of fuel-air mixture upstream of the flame can be modified variously.

In atmospheric tests, we use the pilot system to adopt multiple data-driven models for different design objectives of the combustor which apply to both under full load and part load operation. The control of the fuel valves and the training of the surrogate models is fully automated. The models trained in this way can then be used for predictions for every possible injection scheme. By combining the predictions, we determine pilot configurations that result in low emissions at both full load and part load while stabilizing the flame. The results are analysed and advantages and limitations of the proposed method are discussed. The work shows the potential of the proposed approach for accelerated and efficient combustor design.

Presenting Author: Johann Moritz Reumschüssel Technische Universität Berlin

Presenting Author Biography: Moritz is a 3rd year PhD Student working on the application of Machine Learning and Data-Driven Optimization in Flow and Combustion.

Authors:

Johann Moritz Reumschüssel Technische Universität Berlin
Jakob G. R. von Saldern Technische Universität Berlin
Bernhard Ćosić MAN Energy Solutions SE
Christian Oliver Paschereit Technische Universität Berlin