Méthodes de commande extrémale avec action anticipative : application à des systèmes d’énergie renouvelable soumis à des perturbations mesurables

Abstract

Real-time optimization aims to bring and maintain a system at its optimum point of operation, regardless of external disturbances that may affect this optimal point. In many situations, these systems are non-linear and have an unknown dynamic. Also, the use of a real-time optimization method which does not require a fundamental dynamic model and takes into account the response time of the system is necessary. Extremum-seeking control is one approach able to optimize in real time such a type of system. It only requires an empirical model to estimate the gradient of the objective function prior to control it at zero, the optimum of the objective function. However, this optimization method does not converge quickly and/or precisely to the desired optimum operating point when the system is regularly exposed to external disturbances. Our research project aims to develop new approaches to make the extremum-seeking control scheme able to bring a system with unknown dynamics to converge quickly and accurately to its optimal operating point, despite the presence of external disturbances. In this thesis, three approaches are developed. The basic idea of these approaches is to integrate in the extremumseeking control method an anticipative action based on a multilayer neural network model which gives an estimate of the position of the optimal operating point of the system with regards to the external disturbances to which it is submitted. This estimate is used to adapt in real time the parameters of the extremum seeking control in order to improve its accuracy and reduce its time of convergence under external disturbances. The three approaches offer three different ways to control the parameters of the extremum-seeking control scheme. The first two approaches improve the performance of the extremum seeking control loop when the system faces measurable disturbances. The first approach uses the estimation of the system’s optimum provided by the anticipative action as an initial condition for the extremum-seeking controller in order to quickly push the system close to its optimum and then fine tune this position using the extremum seeking control loop. The second approach adapts the amplitude of the excitation signal of the extremum-seeking control loop according to the system optimum estimated by the anticipative action in order to ensure the maximum speed of convergence while guaranteeing a better precision around the optimum. The third approach improves the performance of the extremum-seeking control scheme despite the presence of measurable and/or unmeasurable disturbances by adapting online the neural network model used in the anticipative. A detailed theoretical study of the proposed approaches is provided. This study establishes the stability of two approaches. Moreover, it formulates the gain obtained with the proposed methods in terms of time of convergence compared with the perturbation method. The improvement of the performance of the approaches is validated through simulations of a microbial fuel cell. It is also validated experimentally using a photovoltaic system. The results of this experimental study show that the optimum tracking efficiency of each of the three proposed methods is improved up to 30%, 20% and 15% respectively compared to the perturbation method.

Date	19 Oct 2018
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Lyne Woodward (Supervisor) & Ouassima Akhrif (Co-supervisor)