Apprentissage machine et contrôle stochastique pour un pilotage automatique optimisé des systèmes industriels

Optimizing aeration control in wastewater treatment plants is crucial to meet both economic and environmental objectives. The development of advanced control strategies, however, is hindered by the difficulty of calibrating reliable predictive models from limited, partial, and noisy data. This thesis proposes a unified methodological framework, from modeling to control, based on state-space stochastic models and data assimilation to jointly estimate process dynamics, reconstruct latent states, and rigorously quantify uncertainties. The comparative study of different model families (mechanistic, grey-box, black-box) and real-time control strategies yields a robust control architecture combining a nonparametric model with Stochastic Dynamic Programming. This solution is validated in a realistic closed-loop simulation and tested against unmodeled disturbances such as influent variability. This framework is proven to reduce operating costs by 15–25% while ensuring constraint satisfaction, even without prior knowledge of external disturbances. Finally, the study highlights that the controller’s ability to integrate global uncertainty, including unmodeled disturbances, into its decision-making policy is ultimately more decisive for operational performance than pursuing absolute predictive accuracy.