WaterAdmin: Orchestrating Community Water Distribution Optimization via AI Agents

Jiaqi Wen, Pingbo Tang, Shaolei Ren, Jianyi Yang · 2026

Read WaterAdmin: Orchestrating Community Water Distribution Optimization via AI Agents by Jiaqi Wen, Pingbo Tang, Shaolei Ren, Jianyi Yang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-10343.txt.

Introduction

Wen, Tang, Ren, and Yang study community water systems, where pumps and valves must meet demand while using little energy. Standard optimization works in well-modeled settings but struggles with fast-changing local context such as human activity and weather. Language-model agents grasp such context well but cannot reliably produce live control actions. Their WaterAdmin framework pairs the two: language models abstract context above, while optimization controls operations below. Built on the EPANET simulator, it maintains pressure reliability and cuts energy use under dynamic conditions.

Catalog record

CREATORJiaqi Wen, Pingbo Tang, Shaolei Ren, Jianyi Yang
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE35102
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.10343

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

WaterAdmin: Orchestrating Community Water Distribution Optimization via AI Agents Jiaqi Wen, Pingbo Tang, Shaolei Ren, Jianyi Yang Abstract We study the operation of community water systems, where pumps and valves must be scheduled to reliably meet water demands while minimizing energy consumption. While existing optimization-based methods are effective under well-modeled environments, real-world community scenarios exhibit highly dynamic contexts—such as human activities, weather variations, etc—that significantly affect water demand patterns and operational targets across different zones. Traditional optimization approaches struggle to aggregate and adapt to such heterogeneous and rapidly evolving contextual information in real time. While Large Language Model (LLM) agents offer strong capabilities for understanding heterogeneous community context, they are not suitable for directly producing reliable real-time control actions. To address these challenges, we propose a bi-level AI-agent-based framework, WaterAdmin , which integrates LLM-based community context abstraction at the upper level with optimization-based operational control at the lower level. This design leverages the complementary strengths of both paradigms to enable adaptive and reliable operation. We implement WaterAdmin on the hydraulic simulation platform EPANET and demonstrate superior performance in maintaining pressure reliability and reducing energy consumption under highly dynamic community contexts. I Introduction Community water systems are critical infrastructure for delivering fresh water to meet demand of community consumers [ 17 , 3 , 11 ] . However, water operations are highly energy intensive, requiring substantial electricity for distribution across large-scale hydraulic networks. According to the report of U.S. Environmental Protection Agency (EPA) [ 4 ] , municipal water systems typically account for approximately 30% to 40% of total municipal electricity consumption, and the associated energy costs in the United States amount to billions of dollars annually. As urban infrastructure continues to expand, the demand for electricity and water is steadily increasing. Consequently, achieving reliable and energy-efficient water infrastructure operations has become a critical challenge for the development of sustainable and smart communities. We consider the community water distribution operations whose objectives include satisfying water demand, maintaining required pressure levels, and minimizing energy costs through the regulation of pumps and valves in the Water Distribution Network (WDN). Existing studies have developed a wide range of optimization-based control approaches [ 8 , 6 , 12 , 23 , 9 , 25 ] . However, achieving reliable and efficient water operations increasingly requires incorporating diverse and dynamic community context, such as human activities, weather conditions, and electricity grid signals, because these factors strongly influence water demand patterns, operational requirements, and electricity costs. Existing optimization methods typically cannot fully exploit this real-time community information, as it is often unstructured and obtained from various online sources. Large Language Model (LLM) agents have been utilized to enhance water system operations due to their powerful capabilities in information retrieval, multi-modal information processing, and reasoning [ 34 , 16 , 13 , 15 ] . Among them, [ 13 , 15 ] utilize LLMs to generate rule-based codes to control WDN components. [ 34 , 16 ] explore the application of LLM agents to generate system models and to provide scheduling suggestions to human operators. Despite these achievements, there still lacks an AI Agent-based optimizer that can give real-time control actions for WDN components, such as pump speed and valve positions. One may consider using LLM agents to directly generate control actions, but this introduces several limitations [ 19 , 21 ] . First, LLM agents may generate physically infeasible actions or actions that violate operational constraints, causing higher safety risks. Moreover, LLM agents often become trapped in local optima when solving optimization problems, potentially resulting in poor performance. Last but not least, the information retrieval and reasoning processes of LLM agents can incur large latency, making it infeasible for real-time water operations that require timely responses. Therefore, we argue that LLM agents are better suited for high-level planning and strategic guidance, while direct actions for pump or valve control should be output by specialized optimization-based models. These insights bring the opportunities to combine the complementary strengths of LLM agents and optimization-based approaches, enabling effective integration of community contextual information for reliable and energy-efficient water system operations. To this end, we propose a bi-level operational framework WaterAdmin that leverages the reasoning and information integration capabilities of LLM agents together with the resilience of optimization-based methods. Specifically, we make the following contributions. • We propose a bi-level framework for water operations. At the higher level, LLM agents abstract real-time community context and human instructions to generate structured operational targets, such as water demand forecasts, location-specific pressure ranges, or nominal water tank levels. To facilitate the LLM reasoning to generate informative operational targets, we perform in-context prompting which incorporate historical instances as in-context prompting and guide the LLM agent to generate target levels based on the in-context information. • At the lower level, we develop an ML-based optimizer that exploits the agent-informed structured operational targets to control pumps and valves in real time. The model is trained to explicitly utilize these agent-informed signals by simulating operational episodes in EPANET [ 1 ] with true sequences of water usage and operational targets. To address the challenges posed by non-differentiable simulators such as EPANET, we adopt a training procedure based on zeroth-order optimization to estimate gradient information. This approach enables the ML-based optimizer to capture WDN dynamics while optimizing operational objectives. • We conduct numerical evaluations on the EPANET NET3 benchmark [ 1 ] under different component configurations. The proposed WaterAdmin is compared against the rule-based control integrated in EPANET and an ML-based optimizer without exploiting LLM agent-based context abstraction. We further perform ablation studies with different LLM agents and varying windows of LLM-informed targets to assess their impact on performance. The experimental results demonstrate that integrating LLM agents with ML-based optimizers yields superior performance in both operational reliability and energy efficiency. II Water Distribution Optimization We consider the community water distribution optimization problem whose objective is to satisfy water demands and maintain water pressures with minimized energy costs by regulating the WDN components in real time. Operational states . A WDN [ 2 , 36 ] can be modeled as directed graph model \mathcal{G}=\{\mathcal{V},\mathcal{E}\} [ 38 , 7 ] . The set of network nodes \mathcal{V} includes demand nodes, water storage nodes and pump stations. For each node v\in\mathcal{V} at time step t , the water pressure is denoted as h_{v,t} (psi). For water storage nodes, the pressure is in proportional to its water level. The water usage at a demand node v at time step t is d_{v,t} ( \mathrm{m}^{3}/s ). The network nodes are interconnected by pipelines represented by the directed edges (u,v)\in\mathcal{E} . The flow rate from node u to node v at time step t is denoted as q_{(u,v),t} ( \mathrm{m^{3}/s} ) which is a function of the pressure drop from node u to node v . Meters and sensors integrated into water systems enable the real-time measurement of pressure, flow rate, water level, and other essential metrics [ 27 , 22 , 26 ] . We summarize these state variables using a state vector s_{t} . Control components . Pumps and valves are the primary components regulating operational states in water networks [ 12 , 29 , 23 ] . Pumps are deployed at pump stations

Read the complete text