Challenge
Aging infrastructure and manual, reactive leak-detection workflows left municipal operators battling up to 30% non-revenue water losses (“NRW”)—the volume lost between the treatment plant and the customer meter—while energy-intensive aeration processes in wastewater plants drove up utility bills. Teams spent days analyzing SCADA logs and field readings to pinpoint bursts or pressure anomalies, delaying repairs and incurring high remediation costs. Industry benchmarks show that each percentage point of NRW reduction can yield $0.75–$1.25 million in annual savings for a 100 MLD system, yet cities lacked the automated tools and analytics to realize those gains at scale.
Solution
Xylem implemented a cloud-native telemetry ingestion layer on Azure IoT Hub and Event Hubs to collect real-time pressure, flow, and acoustic data from smart meters and SCADA endpoints. Azure Functions performed normalization, geolocation tagging, and quality checks, storing cleansed streams in Cosmos DB. Anomaly detection leveraged LSTM autoencoders—trained on historical baseline patterns—to flag emerging leaks and pipe-stress events. Parallel models predicted aeration pump energy curves, identifying operating set-points causing inefficiency. A React/TypeScript dashboard, backed by FastAPI, provided interactive map overlays, trend charts, leak-heatmaps, and “what-if” scenario sliders. Role-based access controls and Slack/Webhook alerts ensured maintenance crews and operational leaders received instant notifications when thresholds were breached [1][2].
Results
- Non-revenue water losses fell by 17% on average—and up to 37% in key pipeline corridors—within six months of rollout [1].
- Wastewater aeration energy consumption dropped by 30%, saving municipalities an average of $250 k per plant annually [2].
- Preventive maintenance alerts cut unplanned pipe-failure incidents by 45%, averting approximately 1,200 failures per year.
- Self-service dashboard adoption exceeded 80% among operations teams, reducing manual analysis time by 60% and accelerating repair cycles by 70%.
Introduction & Business Context
By 2024, many municipal water authorities were grappling with escalating non-revenue water (NRW) losses—leaks, theft, and metering errors—that averaged 25–30% of treated volume, driving multi-million-dollar annual waste. Simultaneously, energy use in wastewater aeration accounted for up to 60% of a treatment plant’s electricity bill. Aging pipes, variable demand patterns, and manual diagnostics compounded these challenges, delaying repairs by days and triggering emergency pump bypasses at a premium cost. Recognizing the need for a proactive, data-driven strategy, Xylem sought to harness AI and IoT to transform its asset-management offerings into a real-time, predictive service.
Data Integration & Telemetry Architecture
The platform ingests minute-level sensor streams—pressure transients, flow rates, acoustic vibration signatures—from 5,000+ nodes across 50 networks. Azure IoT Hub secures bi-directional communication, while Event Hubs buffer data for processing. Azure Functions apply schema validation, geotagging, and anomaly scoring metadata before persisting records in Cosmos DB and archival in Azure Data Lake. This event-driven, serverless design guarantees sub-500 ms end-to-end latency and scales elastically with network growth.
Anomaly Detection & Energy Modeling
LSTM autoencoder models, trained on two years of historical baseline data, detect deviations in pressure and flow that signal hidden leaks or pipe bursts. Concurrently, a regression-LSTM hybrid forecasts aeration blower energy profiles, flagging operating points with >10% energy variance compared to expected curves. Batch retraining pipelines in Azure Machine Learning run nightly, incorporating the latest labeled incidents and ensuring model drift stays below 2% per quarter.
Dashboard & Operational Playbooks
A React front end with FastAPI delivers an interactive GIS-based dashboard: heatmap overlays visualize leak-detection confidence scores, time-series graphs track energy KPIs, and “fix-impact” simulators estimate savings from targeted interventions. Users filter by zone, time range, or asset type, then drill into raw sensor feeds and SHAP-derived anomaly explanations. Automated alerts via Slack and SMS mobilize field crews within minutes of anomaly confirmation.
Pilot Outcomes & Validation
In a six-month pilot covering three mid-sized municipalities, the platform processed over 200 million data points daily. Leak-detection precision reached 93%, reducing false-positive dispatches by 50%. Aeration energy forecasts achieved 95% accuracy with an RMSE of 2.1 kW. Operations teams reported 60% fewer emergency call-outs and a 70% faster average repair response time.
Lessons Learned & Next Steps
Rigorous data governance & training: enforcing quality checks and cross-team upskilling in AI explainability establishes the foundation for reliable anomaly detection.
Embed continuous improvement loops: integrating feedback pipelines and planning future phases (satellite moisture indices, weather forecasts, customer usage data) paves the way to a fully autonomous water-network control system.