First AI for Drinking Water Chlorination Challenge

Abstract Water distribution systems (WDS) are vital for urban societies by providing clean and safe drinking water. Water utilities have to carefully control the injection of disinfectants such as chlorine to effectively deal with natural contaminants and accidental pollution affecting the drinking water quality. This, however, constitutes a challenging task due to sparse sensor readings, complex system dynamics, and various sources of uncertainty. Despite the importance of this problem, the field lacks realistic, publicly available benchmarks to systematically evaluate and compare automated chlorine injection control strategies. To address this gap, we introduce a comprehensive benchmark specifically designed for chlorine injection control. It features realistic operational scenarios based on the Cyprus Disinfection-By-Product (CY-DBP) network, a state-of-the-art model derived from a real-world WDS. The benchmark incorporates sparse sensing conditions, realistic water demands, fluctuating source water quality, and multiple wastewater contamination events. Standardized evaluation metrics, focusing on different desiderata, enable consistent and reproducible assessment of chlorine booster station control strategies, facilitating rigorous comparison across methods. To stimulate cross-disciplinary innovation, we also hosted the benchmark as a competition at a leading artificial intelligence (AI) conference, fostering collaboration between the AI and water engineering communities. Although participants proposed diverse data-driven control approaches, none achieved fully satisfactory performance, underscoring the complexity and open challenges of the task. By providing a realistic and demanding testbed, this benchmark establishes a foundation for advancing the research on intelligent and reliable chlorine booster station control strategies.