| Summary: | Separating valuable minerals from waste rock is an important step in the production of metals. This is for copper ore done through a process called flotation. A flotation series consists of tank cells in series where the minerals are collected in a froth on top of the cells. The level control of the flotation cells is important in order to be able to collect the froth. The first four flotation cells and the buffer tank before the series is the process considered in this thesis. This process is found in the concentrator connected to Boliden’s copper mine Aitik located near Gällivare in the north of Sweden. A simulation model of the process was developed using both physical modeling and experimental data from the real process. When the simulation model of the process had been developed, different control structures were tested and evaluated. The control structures that were tested were coupled PI-controllers, an LQ-controller, an MPC-controller and a state feedback controller where the state feedback was determined using reinforcement learning. The reference-tracking properties of the different controllers were similar while a bigger difference could be seen when it came to disturbance rejection. The PI-controllers gave a stable performance but their disturbance rejection was not as good as for the other controllers. One advantage with the PI-structure is its simplicity. Unlike the LQ- and the MPC-controllers, it does not need a model of the process to control it. The MPC-controller outperformed the other controllers when it came to disturbance rejection, but it was a bit more sensitive to model errors than the LQ-controller which also performed well. The reinforcement-learning-based controller did not give a better performance than the LQ-controller and it had issues with robustness in the tuning process, making it less reliable than the other controllers. The tuning process for it also required experiments that are unreasonable to perform on the real process. There is potential in reinforcement learning ...
|
|---|