| Studying the characteristics of complex nonlinear systems and designing control
strategies necessitates the development of computationally efficient methods. In particular,
compressor systems benefit from accurate models that capture the nonlinear dynamics. Many
numerical approaches have been created using the Moore-Greitzer (MG) model. A model of a
small experimental compressor rig is developed using the Toolbox for the Modeling and
Analysis of Thermodynamic Systems (T-MATS). To verify the accuracy and capabilities of this
toolbox, experimental test data and data generated through simulation using the T-MATS model
is compared. The error is minor between the experimental data and the simulation data, which
indicates utility of such simulation models for use in further research. In this dissertation, an
overview of neural networks is given with a focus on Long-Short Term Memory (LSTM)
networks for dynamic systems. A simple test system is presented for verifying the LSTM
approach for single input single output (SISO) systems. The proposed LSTM approach is
demonstrated for the multiple input multiple output (MIMO) axial compressor model. This
dissertation investigates control systems designed using MG models. Linearization of the MG
model at various locations is executed to create segments for controller design. At each of these
linearized sections, a controller is designed and optimized using a Genetic Algorithm (GA). By
combining these, a controller network is produced. To investigate the performance, this network
is implemented with the T-MATS model and the LSTM model. It is shown that the controller
network performs well with good results on both experimentally based models. The results show
that a controller network can be designed on a mathematical model of an axial compressor that
will perform well when applied to experimentally based models.
Keywords: Compressor, Moore-Greitzer, Modeling, LSTM, Neural Network, Control |