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

Abstract

Abstract The majority of premixed industrial gas turbine combustion systems feature two or more separately controlled fuel lines. Every additional fuel line improves the operational flexibility but increases the complexity of the system. When designing such a system, the goals are low emissions of various pollutants and avoiding lean blowout or extinction. Typically, these limitations become critical under different load conditions of the machines. Therefore, it is particularly challenging to develop combustors for stable and clean combustion over a wide operating range. In this study, we apply the Gaussian process regression machine learning method for application to burner development, with the aim of improving the process, which is often driven by a trial-and-error approach. To do so, a special pilot unit is installed into a full-scale industrial swirl combustor. The pilot features 61 positions of fuel injection, each of which is equipped with an individual valve, allowing to modify the fuel–air mixture close to the flame root in various degrees. In fully automatized atmospheric tests, we use the pilot system to train two surrogate models for different design objectives of the combustor, relevant for full load and part load operation, respectively. Once trained, the models allow for prediction for any possible injection scheme. In combination, they can be used to identify pilot injector configurations with an improved operation range in terms of low NOx emissions and part load stability. The adopted multimodel approach enables combustor design specifically for high operational flexibility of gas turbines, but can also be extended to other similar industrial development processes.