AI-Driven Framework for Adaptive Water Network Management with Proof-of-Concept Implementation: Addressing Non-Revenue Water in Jordan

Technical Summary: AI-Driven Framework for Adaptive Water Network Management in Jordan

Problem Statement
Jordan faces a critical water scarcity crisis, characterized by renewable water resources of only 90 cubic meters per capita annually. A primary driver of this inefficiency is Non-Revenue Water (NRW), where approximately 50% of produced water is lost to leakage, theft, and metering inaccuracies. Traditional management approaches in Jordan are reactive, relying on periodic leak detection surveys and manual analysis. These methods fail to provide continuous monitoring, often allowing losses to revert to previous levels after temporary interventions. Furthermore, the country's water distribution operates on intermittent supply schedules (24–48 hour cycles), creating hydraulic stresses ("hammer effect") that accelerate infrastructure deterioration, complicating the detection of anomalies against normal operational fluctuations.

Methodology
The paper proposes an intelligent framework integrating five core components to enable continuous monitoring and adaptive decision-making:

1. Data Acquisition: Utilization of SCADA systems, flow/pressure sensors, and IoT devices to capture real-time network data.
2. Big Data Infrastructure: A scalable platform for streaming, storage, and real-time analytics of high-volume sensor data.
3. Digital Twin & Hydraulic Modeling: Implementation of physics-based simulations using EPANET (via the Python toolkit EPYT) to estimate network states and calibrate models against real-time data.
4. AI Agent Intelligence: A Large Language Model (LLM) agent acting as the decision core. It employs Retrieval-Augmented Generation (RAG) to interpret operational policies and function calling to execute control logic.
5. Network Control Systems: Mechanisms for transmitting commands to actuators (valves, pumps) based on AI decisions, with safety interlocks and logging.

The framework is designed to operate in phases, starting with monitoring and alerts, progressing to AI-assisted recommendations, and eventually moving toward autonomous control as infrastructure modernizes.

Proof-of-Concept (PoC) Implementation
To validate technical feasibility, the authors developed a fully offline PoC system applied to a representative District Zone (DZ) in Amman containing 1,164 junctions and 1,310 pipes. The system architecture consists of three layers:

• Layer 1 (Hydraulic Simulation): EPYT executes hydraulic simulations and extracts time-series data for pressures and flows.
• Layer 2 (Data Processing): Python scripts parse simulation outputs, calculate statistical metrics, and identify anomalies based on thresholds (e.g., pressure < 20m or > 60m, flow deviations > 1 L/s).
• Layer 3 (AI Analysis): An offline LLM (llama3.1:8b hosted via Ollama) receives structured JSON data containing simulation results and operational context. It generates natural language health reports and prioritized recommendations without relying on cloud APIs.

Key Results
The PoC demonstrated the following technical capabilities:

• Automated Simulation & Reporting: The system successfully executed hydraulic simulations and generated structured health reports with end-to-end execution times under 2 minutes (15–30 seconds for LLM generation).
• Anomaly Detection in Intermittent Supply: In a simulated burst scenario involving a 30.1 L/s leak, the system identified the anomaly not through zone-wide pressure drops (which remained stable due to the gravity-fed, looped topology) but through local flow redistribution. The system flagged 15 pipes with flow deviations exceeding 1 L/s, converging on a 15-junction cluster to localize the burst.
• Offline Viability: The use of locally hosted open-source models eliminated cloud API costs and data exfiltration risks, demonstrating that complex network analysis can be performed on modest computational resources.
• Actionable Intelligence: The LLM successfully translated computational findings (e.g., flow deviations, pressure thresholds) into actionable operational guidance, such as isolating zones, dispatching crews, and identifying candidates for Pressure Reducing Valves (PRVs).

Significance and Claims
The paper claims to offer a scalable, cost-effective pathway for water-scarce regions to leverage intelligent automation for NRW reduction. Its primary contributions are:

1. Integration Architecture: A novel system architecture combining EPANET-based simulation, digital twin concepts, and LLM agents specifically tailored for intermittent supply conditions.
2. Offline Feasibility: A demonstration that network health analysis and structured reporting can be achieved without cloud dependencies, addressing privacy and cost concerns for utilities in resource-constrained environments.
3. Shift in Detection Paradigm: Evidence that for well-connected, gravity-fed networks, detection should shift from pressure-zone heuristics to local pipe-flow anomaly analysis, which the system successfully validated.
4. Phased Implementation: A realistic deployment pathway that accommodates Jordan's current limited automation by starting with monitoring and decision support before advancing to autonomous control.

The authors maintain a modest stance, noting that while the PoC validates the technical integration of simulation and AI, full autonomous control requires further development of safety interlocks, formal SCADA integration, and regulatory approval. The system currently serves as a decision support tool rather than a fully autonomous controller.