Abstract
Machine learning (ML) models are increasingly deployed in urban water systems to optimize operations, enhance efficiency, and curb resource consumption amid growing sustainability demands. Yet, their transferability across plants is hampered by scenario differences-variations in environmental factors, protocols, and data distributions-that erode performance and necessitate energy-intensive retraining. While existing strategies focus on minimizing these differences via domain adaptation or fine-tuning, none exploit them as inherent prior knowledge for improved generalization. Here we show an environmental information adaptive transfer network (EIATN) framework that can leverage scenario differences to enable effective generalization across distinct prediction tasks within the same water plant. By evaluating EIATN across four scenario categories and 16 diverse ML architectures-yielding 64 models in total-we demonstrate its feasibility, with bidirectional long short-term memory emerging as the top performer, achieving a mean absolute percentage error of just 3.8 % while requiring only 32.8 % of the typical data volume. In a case study of Shenzhen's urban water system, it reduced carbon emissions by 40.8 % compared to fine-tuning and 66.8 % relative to direct modeling from scratch. EIATN unlocks the reuse of vast existing ML models in water systems, yielding substantial energy savings and fostering equitable, low-carbon intelligent management.