Data based predictive controlApplication to water distribution networks

Resumen

In this thesis, the main goal is to propose novel data based predictive controllers to cope with complex industrial infrastructures such as water distribution networks. This sort of systems have several inputs and out- puts, complicate nonlinear dynamics, binary actuators and they are usually perturbed by disturbances and noise and require real-time control implemen- tation. The proposed controllers have to deal successfully with these issues while using the available information, such as past operation data of the process, or system properties as fading dynamics. To this end, the control strategies presented in this work follow a predic- tive control approach. The control action computed by the proposed data- driven strategies are obtained as the solution of an optimization problem that is similar in essence to those used in model predictive control (MPC) based on a cost function that determines the performance to be optimized. In the proposed approach however, the prediction model is substituted by an inference data based strategy, either to identify a model, an unknown control law or estimate the future cost of a given decision. As in MPC, the proposed strategies are based on a receding horizon implementation, which implies that the optimization problems considered have to be solved online. In order to obtain problems that can be solved e ciently, most of the strategies proposed in this thesis are based on direct weight optimization for ease of implementation and computational complexity reasons. Linear convex combination is a simple and strong tool in continuous domain and computational load associated with the constrained optimization problems generated by linear convex combination are relatively soft. This fact makes the proposed data based predictive approaches suitable to be used in real time applications. The proposed approaches selects the most adequate information (similar to the current situation according to output, state, input, disturbances,etc.), in particular, data which is close to the current state or situation of the system. Using local data can be interpreted as an implicit local linearisation of the system every time we solve the model-free data driven optimization problem. This implies that even though, model free data driven approaches presented in this thesis are based on linear theory, they can successfully deal with nonlinear systems because of the implicit information available in the database. Finally, a learning-based approach for robust predictive control design for multi-input multi-output (MIMO) linear systems is also presented, in which the effect of the estimation and measuring errors or the effect of unknown perturbations in large scale complex system is considered.