Thermoacoustic Oscillations: Modeling, Identification and Control

In this multidisciplinary collaboration with our industrial partner ALSTOM, a test rig of a combustor has been set up in our lab, see Fig. 1. The air is electrically preheated and premixed with natural gas. The flame is stabilized by an EV burner, and the exhaust gases are subsequently discharged. Microphones distributed along the combustor measure the acoustic field. Loudspeakers as well as a secondary high-bandwidth fuel injector are used as actuators to decrease the pressure oscillations, which exceed 160 dB.

In order to design modern model-based controllers, a model of the combustor has to be developed. Our approach is based on a network model shown in Fig. 2. It consists of different blocks such as ducts, burner and flame. Each block is then described based on physical principles and also identified experimentally. The model is then assembled and various transfer functions can be extracted. For instance, the transfer function from a loudspeaker as actuator to a microphone pressure signal has been modeled and found to agree well with measurements, see Fig. 3.

Passive control approaches using Helmholtz resonators have been evaluated, which can have a decisive effect in certain frequency ranges. However, changing this frequency range would involve hardware changes. Therefore, active control strategies are more promising. A robust H_infinity-controller has been designed with a loudspeaker as actuator, and found to be superior to a simple Gain-Delay algorithm.

In addition, different controllers are improved online while the combustor is running, by using an evolutionary algorithm called CMA-ES. This allows to find an optimal controller in a changing environment. Figure 6 shows the cost function landscape for a Gain-Delay controller. The algorithm successfully finds the minimum value, and continues to track it despite changing combustor behavior, because of heat-up effect.