Development of a Solution for Start-up Optimization of a Thermal Power Plant

Sammanfattning: This thesis covers optimizing the first phase of the start-up of a thermal power plant using Nonlinear Model Predictive Control (NMPC) and state estimation using an Unscented Kalman Filter (UKF). The start-up has been optimized in regards to time and fuel usage. The thesis is done as a joint project between Vattenfall and Modelon. Both NMPC and UKF are nonlinear methods and require a model of the power plant. The model used in this thesis has been developed in the language Modelica in a previous master thesis and has been extended and improved upon during this thesis. The optimization and simulation of the model required by the NMPC and UKF was done within the framework of JModelica.org. Another, more detailed, model of the power plant, developed by Vattenfall, was originally planned to be used as the process to be controlled. State estimation using the UKF has been successful, with a maximum mean absolute error of 0.7 % when estimating the states of the detailed model in a reference startup. When using the NMPC to control the optimization model itself, the simulated start-up time is 70 minutes faster compared to a reference start-up using the detailed model. This is more than half the time of the first phase of the start-up. The total firing power, which relates to the fuel amount, is also considerably less, with the optimized value being about 40 % of that in the reference soft start with the detailed model. Due to difficulties in initializing the detailed model, it was not possible to run it online together with the NMPC and UKF. Running the NMPC and UKF together on the optimization model worked, but the NMPC failed to find an optimal trajectory 8 out of 10 iterations. The conclusion is that the start-up has potential for optimization, but requires more robust models to work with.