A data-centric decision core for AI scheduling agents in intelligent water distribution network operations

ABSTRACT A WDN scheduling agent requires four fundamental capabilities: perception, decision making, action, and learning. A data-centric AI that learns from human operational experience can serve as a decision core for the agent. Urban water distribution networks (WDNs) require real-time scheduling to balance demand, safety, and energy efficiency, requiring intelligent decision-making. Deep learning (DL) efficiently processes large-scale WDN data, but a model-centric perspective limits the optimality of its decisions, thereby restricting its application in intelligent scheduling. This study proposes a data-centric artificial intelligence (DCAI) decision core for WDN scheduling agents. The decision core is constructed through: (1) phase-dependent Gaussian kernel based multi-criteria evaluation of historical strategies; and (2) hybrid objective-subjective data sorting and filtering via heuristic optimization and Pareto-optimal selection. By learning from high-quality operational data, the decision core acquires the ability to make informed scheduling decisions. Evaluated on a large-scale WDN, the method reduced mean absolute error (MAE) by 73.75% compared with unfiltered data under high-demand conditions. Within these scenarios, hydraulic simulations indicated that scheduling decisions maintained >99% nodal pressure satisfaction, while reducing energy consumption by 3.06% and pressure fluctuation by 2.52%. SHAP-based interpretability analysis demonstrated alignment with operational principles, helping to mitigate the ‘black-box’ concern critical for deploying reliable scheduling agents. Evaluated through nearly three years of field operation, this DCAI decision core supports real-time scheduling that balances safety and efficiency in large-scale WDNs.