François Gauthier-Clerc Thesis defense

Abstract : Climate change and the energy transition demand greater efficiency and flexibility in energy management solutions for industry and utilities.
However, the facilities involved do not always have the financial resources to implement the advanced control systems required. This situation fosters the development of accessible control methods that can be deployed at low cost across a wide range of systems.
This thesis focuses on facilities that leverage data representative of their behavior under operational conditions (sampled data directly collected from the system in operation). It draws on identification methods capable of inferring dynamic models of a physical process solely from experimental data, thus avoiding costly physical modeling. The work examines traditional and neural models and learning methods suitable for this context. It delves into the challenge of learning ad hoc models for predictive control and establishes two benchmarks tailored to the target industrial context. Ultimately, it proposes a new methodological building block aimed at reducing operational energy costs : a novel black-box model that can represent nonlinear behavior while leveraging, for the targeted application class, the extrapolation capabilities of linear models. This model is directly usable for predictive control without requiring a complex state estimation process.