Modelling, Simulation and Control of Gas Turbines Using Artificial Neural Networks
Thesis DisciplineMechanical Engineering
Degree GrantorUniversity of Canterbury
Degree NameDoctor of Philosophy
This thesis investigates novel methodologies for modelling, simulation and control of gas turbines using ANNs. In the field of modelling and simulation, two different types of gas turbines are modelled and simulated using both Simulink and neural network based models. Simulated and operational data sets are employed to demonstrate the capability of neural networks in capturing complex nonlinear dynamics of gas turbines. For ANN-based modelling, the application of both static (MLP) and dynamic (NARX) networks are explored. Simulink and NARX models are set up to explore both steady-state and transient behaviours.
To develop an offline ANN-based system identification methodology for a low-power gas turbine, comprehensive computer program code including 18720 different ANN structures is generated and run in MATLAB to create and train different ANN models with feedforward multi-layer perceptron (MLP) structure. The results demonstrate that the ANN-based method can be applied accurately and reliably for the system identification of gas turbines.
In this study, Simulink and NARX models are created and validated using experimental data sets to explore transient behaviour of a heavy-duty industrial power plant gas turbine (IPGT). The results show that both Simulink and NARX models successfully capture dynamics of the system. However, NARX approach can model gas turbine behaviour with a higher accuracy compared to Simulink approach. Besides, a separate complex model of the start-up operation of the same IPGT is built and verified by using NARX models. The models are set up and verified on the basis of measured time-series data sets. It is observed that NARX models have the potential to simulate start-up operation and to predict dynamic behaviour of gas turbines.
In the area of control system design, a conventional proportional-integral-derivative (PID) controller and neural network based controllers consisting of ANN-based model predictive (MPC) and feedback linearization (NARMA-L2) controllers are designed and employed to control rotational speed of a gas turbine. The related parameters for all controllers are tuned and set up according to the requirements of the controllers design. It is demonstrated that neural network based controllers (in this case NARMA-L2) can perform even better than conventional controllers. The settling time, rise time and maximum overshoot for the response of NARMA-L2 is less than the corresponding factors for the conventional PID controller. It also follows the input changes more accurately than the PID.
Overall, it is concluded from this thesis that in spite of all the controversial issues regarding using artificial neural networks for industrial applications, they have a high and strong potential to be considered as a reliable alternative to the conventional modelling, simulation and control methodologies. The models developed in this thesis can be used offline for design and manufacturing purposes or online on sites for condition monitoring, fault detection and trouble shooting of gas turbines.