This thesis investigates how control systems based on Artificial Intelligence (AI) can be safely implemented to optimise the water-related processes in critical infrastructures across Europe, focusing on wastewater treatment plants (WWTPs) in Italy, France, and the Netherlands. The study addresses the overarching question: “How can AI-based solutions be technically effective, economically viable, safe, and acceptable for sustainable water management in Europe?”
The research follows a dual-track approach combining engineering development and policy analysis. On the engineering side, the work develops and evaluates deep reinforcement learning (DRL) controllers for aeration optimisation within activated sludge systems using real operational data provided by SUEZ Digital Solutions. Two state-of-the-art agents, Soft Actor–Critic (SAC) and Twin-Delayed DDPG (TD3), are trained interactively on a linear model for aeration, to respect operational constraints and improve the process. The agents were trained with two different configurations: with and without a buffer of historical transitions that is used as previous knowledge. After training, the agents were benchmarked across multiple disturbance scenarios, generated from real data of energy price and inflow load. Results demonstrate significant improvements compared to baseline control, achieving lower energy consumption, stable dissolved oxygen levels, and better values of redox potential in the tank. These findings confirm the technical feasibility and scalability of DRL-based aeration control for real-world deployment.
On the policy side, the research explores the institutional and governance readiness for adopting AI-based control in critical water infrastructures across Italy, France, and the Netherlands through fifteen semi-structured interviews with regulators, utility managers, and researchers. Using the Transition Model Canvas (TMC) and Multi-Level Perspective (MLP) frameworks, the analysis identifies key barriers and leverage points. In the first group, fragmented governance, infrastructural limits, and lack of AI literacy among stakeholders at all levels were identified, while the second one included regulatory sandboxes, digital-skills training, and pilot projects. Comparative insights show that France benefits from strong national coordination and incumbents (SUEZ, Veolia), Italy faces heterogeneous regional governance and uneven digitalisation, while the Netherlands provides a model of integrated and innovation-oriented regulation.
By integrating both perspectives, the thesis proposes a Transition Model Canvas for AI-based wastewater infrastructure at a European level. It maps how landscape pressures (EU AI Act, Green Deal, and Urban Wastewater Treatment Directive recast) interact with regime actors and niche innovations to shape transition pathways. The work concludes with a set of policy and design recommendations for safe and responsible AI adoption, structured into short-, medium-, and long-term phases.
Overall, this thesis demonstrates that AI-based control systems can substantially improve energy efficiency, regulatory compliance, and sustainability in wastewater management. Their successful adoption, however, requires coordinated regulatory frameworks, skills, and investment in digital infrastructures. The study illustrates how combining systems engineering with policy analysis can support the responsible digital transformation of Europe’s critical water infrastructures.