Machine Learning and Stochastic Control for the Optimized Automatic Piloting of Industrial Systems

My PhD thesis (French title: « Apprentissage machine et contrôle stochastique pour un pilotage automatique optimisé des systèmes industriels »), defended on 19 December 2025 at Campus de Beaulieu (Amphi Lebesgue, Université de Rennes), under a CIFRE fellowship between Purecontrol, the IRMAR mathematics lab and the IETR / ENS Rennes.

The thesis builds a single method for the energy-optimal control of wastewater treatment plant aeration, under discharge constraints and strong uncertainty on the incoming load. It is the synthesis of a line of work that runs from forecasting to estimation to uncertainty to control, with each step published on its own along the way.

• Forecasting the load. Aeration demand is driven by the incoming nitrogen load, which is uncertain. A comparison of white-, grey- and black-box forecasts, combined with data assimilation, settles which predictions are trustworthy enough to act on (Water Research, 2025).
• Identifying a model. A nonparametric local linear regression captures the reactor dynamics well, while staying light enough to control. It is cast as a stochastic state-space model and fitted with data assimilation. The grey-box estimation of a reduced ASM1 model by stochastic EM comes from the ECC 2024 work.
• Locating the uncertainty. Decomposing the sources shows that the load forecast dominates operating cost, while model error dominates constraint feasibility (ROADEF 2025). This tells the controller what to plan for.
• Controlling with it. That model feeds a stochastic dynamic programming controller that decides aeration while planning for the uncertain load, rather than betting on a single forecast (developed further in the CANUM 2026 talk). On the public BSM2 benchmark this cut the daily aeration cost by about 15% against a standard rule-based controller, with no discharge-limit violations.

The chapters follow this order, from forecasting through estimation and uncertainty analysis to the stochastic controller. The manuscript and the jury details are below.

Jury

• Jérôme Fehrenbach (Université de Toulouse), rapporteur
• Guillaume Sandou (CentraleSupélec), rapporteur
• Pierre Tandeo (IMT Atlantique), examinateur
• Georges Kariniotakis (Mines Paris – PSL), examinateur
• Valérie Monbet (Université de Rennes), directrice de thèse
• Roman Le Goff Latimier (ENS Rennes), co-encadrant
• Gautier Avril (Purecontrol), encadrant entreprise

The full manuscript is on HAL, with the official record on theses.fr.