In this master thesis, the recently introduced Mobile Model Predictive Control (MoMPC) approach for open water systems with uncertain dynamics is discussed, where there are no sensors or actuators installed in the system that would allow for a fully automatic operation. MoMPC is a configuration of Model Predictive Control (MPC) that explicitly incorporates the role of a mobile operator travelling between the points of interest, i.e., nodes, of the system as instructed by a remote centralised controller. The operator provides the controller with up-to-date measurements from the locations visited and acts as the actuator as required by the remote controller.
In this research, four areas of improvement of MoMPC from literature are explored and some possible solutions are proposed, resulting in a new method, called Multiple-Action Mobile Model Predictive Control (MaMoMPC).
First, the MoMPC approach is generalised to open water systems described as a network, wherein for each node a unique set of actions is possible, e.g., at some nodes only actuation is possible, while at others both measuring and actuation is possible.
Secondly, in MoMPC, controlling the system is only allowed until a predefined control horizon, after which there is often still some setpoint water level error present, which is penalised until the end of the prediction. Cyclic control is proposed to include some simplified estimate of future control in the MPC optimisation problem past the control horizon, without introducing extra computational burden. By including cyclic control the future effort to drive the water levels to the setpoints is better represented in the prediction, improving system performance.
The third area of improvement consists of the consideration of the limitations of the mobile operators in the optimisation problem. Until now, the human operators were assumed to be able to work continuously without requiring breaks. An extension that keeps track of the energy levels of the human operators is proposed, which can be used by the controller to schedule breaks for the human operators.
Finally, another shortcoming in the MoMPC approaches from literature is the discrepancy between the predicted state of the system and the actual state. Depending on the number of operators available the measuring and actuating actions will be sparse in time. Furthermore, the system is subjected to external disturbances and will always have some modelling errors. As a result, there is some uncertainty on the predicted state of the system. This uncertainty about the system can become large when some measurement locations are not visited regularly. Moreover, the uncertainty about the predicted state of the system may result in reduced system performance and constraint violations. To ensure the predicted system state does not drift too far from the actual state, the information gathering capabilities of the system have to be augmented. To that end, three methods to weigh the measurement frequency are proposed.
To evaluate, a case study is performed on a realistic numerical model of the Dez main irrigation canal in Iran. In the first part of the case study, the system performance when adding cyclic control to the Time Instant Optimisation Mobile Model Predictive Control (TIO-MoMPC) approach from literature is evaluated. Including cyclic control improved the reference tracking performance during a scenario without noise with statistical significance. In the second part of the case study, noise is added to the numerical model and the MaMoMPC approach with uncertainty weighing methods and cyclic control is evaluated. The results show that the addition of the uncertainty weighing methods yields enhanced disturbance rejection and reference tracking performance.