This thesis addresses the creation of data-driven model predictive control (MPC) schemes and optimization methods utilizing computational intelligence (CI) & machine learning (ML) tools. Both theoretical and practical aspects of CI-based MPC as well as metaheuristic optimization are taken into account, and the economic merits of the proposed algorithms are showcased over the optimization & predictive control of a diverse range of engineering applications. First, regarding metaheuristic optimization, a significant objective of this thesis is to address high-dimensional, non-convex problems with reasonable solution accuracy. For this reason, a cooperative particle swarm algorithm is devised, capable of using cooperative particle sets on grouped design variables. The grouping occurs by applying a community-detection algorithm over the sensitivity matrix of the system at hand, thus identifying design variables that are structurally or topologically interrelated. The proposed method is tested on an IEEE benchmark system, and, together with a machine-learning ensemble load prediction model that is also developed in this thesis, an effective proposition for efficient & economic smart grid dispatch is made. Second, a data-driven tracking nonlinear model predictive controller is devised based on radial basis function neural networks. Standard MPC performance heavily relies on the quality of the prediction model; if it is inaccurate, then the control actions yielded by the solution of the optimal control problem will be suboptimal for the real plant. This means that a linearized model of a high-dimensional system with significant nonlinearities will be unfit for usage within MPC, while its respective ODE-integrated form will be too computationally expensive. Such a first-principles ODE model may be extremely hard to yield for some cases, mandating a data-driven approach. Therefore, this thesis proposes complementing an MPC prediction model with radial basis function networks whenever necessary, using recorded plant data. The ability of the proposed MPC scheme in handling the two aforementioned modelling drawbacks is showcased for the case of a high-dimensional active suspension plant, as well as for the data-driven vessel trajectory inference for collision avoidance using MPC. As a natural continuation of the work on tracking MPC, the third contribution of this thesis is the creation of a data-driven economic MPC scheme for the efficient & economic control of a vessel propulsion system. This specific choice of case study is highly motivated, since it is an item of significant economic importance for the maritime sector of the Greek economy. Initially, a stabilizing EMPC control law is constructed for the vessel propulsion problem and compared to standard tracking MPC, confirming a significant difference in fuel-efficiency. Serving as proof of concept, these results inspire the development of a data-driven EMPC for vessel propulsion based on reinforcement learning. This learning scheme is able to handle structural modelling discrepancies between plant and model, therefore achieving higher closed loop performance and tangible economic benefit. Lastly, in order to leverage both the collision avoidance tracking MPC and the economic vessel propulsion EMPC results, a control law for the data-driven navigation & economic propulsion control of vessels is proposed and its theoretical foundation for further development is laid. Also, it is the author’s opinion that the work presented in this thesis is extendable to other engineering domains and practical applications.
.jpg){kind=logo}
Πανεπιστήμιο Δυτικής Αττικής (ΠΑ.Δ.Α.)
