Towards autonomous scheduling agents for water distribution networks: Self-evolving data-centric AI via continuous...

Area of Science:

• Environmental Engineering
• Artificial Intelligence
• Water Resource Management

Background:

• Urban water distribution networks (WDNs) face challenges in adapting to fluctuating demands, impacting supply reliability and energy efficiency.
• Current data-centric AI (DCAI) for pump scheduling is limited by static models and data quality issues, hindering autonomous operation.
• Existing approaches often fail to continuously adapt to real-world network dynamics, creating a barrier for practical implementation.

Purpose of the Study:

• To develop a data-model coevolution mechanism enabling DCAI to function as an autonomous scheduling agent for WDNs.
• To address the limitations of static models and poor data quality in existing DCAI applications for WDN pump scheduling.
• To create a system capable of continuous online learning and self-evolution for adaptive WDN management.

Main Methods:

• Proposed a data-model coevolution mechanism with a dynamic data curation system for multi-criteria filtering.
• Implemented a data-coevolving model update mechanism offering retraining and fine-tuning for continuous online learning.
• Validated the approach on a real Shanghai WDN over 10-week cycles, incorporating an active learning variant.

Main Results:

• The agent-based mechanism significantly reduced Mean Absolute Error (MAE) by up to 71.03%.
• Demonstrated improvements in water supply security and energy efficiency through autonomous scheduling.
• Achieved greater than 25% MAE reduction with an active learning variant, effectively capturing minority operational patterns.

Conclusions:

• The proposed data-model coevolution mechanism transforms DCAI into an autonomous scheduling agent for WDNs.
• This synergistic paradigm enables automated adaptive updates through WDN interaction, leading to superior scheduling strategies.
• The approach enhances WDN management by ensuring reliable water supply while minimizing energy consumption and carbon emissions.