From Alarm Sensors to Predictive Flow Intelligence

Utilities need dense, trustworthy flow information to manage sanitary sewer overflows (SSOs), assess pipe capacity, and quantify inflow and infiltration (I&I). Permanent flowmeter programs provide high-quality data but are expensive to deploy and maintain at system scale. Many utilities already operate large fleets of level sensors for overflow alarming, yet these data streams are typically treated as single-purpose alarms. This work presents a spatially constrained virtual flowmeter (VFM) framework that repurposes level sensors into quantitative dry weather flow (DWF) estimates while preserving their primary alarm function. The premise is that a level sensor can only act as a VFM when its influence can be demonstrated at a downstream reference flowmeter; without this constraint, models may look accurate in small pilots but fail when scaled across diverse hydraulics. Houston's gravity collection system provides the development testbed (about 6,100 miles of gravity pipe, 3,000+ sensor data). Objectives are to (1) identify which level sensors are hydraulically eligible to behave as VFMs based on measurable influence on a nearby reference flowmeter, (2) learn DWF behavior for eligible sensors using a repeatable, utility-scalable workflow, and (3) define acceptance criteria and safeguards so outputs are used appropriately for operations, screening, and model calibration. The methodology is organized into five steps for enterprise deployment. Step 1, spatial qualification, screens candidate sensor-flowmeter pairs using network topology, travel time, and tributary complexity. Pairs must have a continuous hydraulic path and are rejected when intermediate lateral contributions are likely to dominate the downstream signal. A practical exclusion rule removes candidates where estimated lateral inflows between the sensor and reference exceed 15% of downstream DWF, and exclusions address known backwater or controlled zones that can disconnect level response from downstream flow. Step 2, temporal qualification, confirms a stable relationship using lagged cross-correlation between level and downstream flow, estimating effective travel time and requiring an interpretable peak. Eligibility requires maximum correlation > 0.70 at a lag of 4 hours or less. Step 3, DWF pattern learning, trains sensor-specific models for qualified pairs using time-windowed level statistics (e.g., median and variability), level change rates, and time-of-day and day-type encodings to capture weekday and weekend behaviors. Training excludes wet-weather periods, post-storm recovery windows, and known sensor or operational anomalies so learned signatures represent stable DWF. Step 4, validation and grading, evaluates performance against utility-defined thresholds. Accuracy is assessed using mean absolute percentage error with a target < 15% under dry-weather conditions (stretch < 10% for high-confidence sensors). Secondary checks include volumetric agreement (target < 10% error over evaluation windows) and peak capture during diurnal maxima (target < 25% peak error). The validation protocol requires sustained data coverage (minimum 90 days of continuous records where available) and uses temporally buffered train, validation, and test splits randomized by week to reduce seasonal bias. Sensors that fail thresholds are rejected or routed for remediation such as recalibration, drift correction, or re-pairing. Step 5, operationalization, publishes VFM estimates with confidence flags that bound where estimates are trustworthy. A two-tier wet-weather strategy is conservative: during storms, the pipeline prioritizes event detection and reliability grading (Tier 1) and only produces bounded estimates (Tier 2) when conditions remain within the validated range (no reverse flow, severe backwater, or abrupt control actions). Data products include a VFM flow estimate, an eligible or ineligible designation, a rolling confidence score, and reason codes (e.g., lateral-dominated, correlation-weak, drift-suspected) to support maintainable operations. Status to date includes completion of network pairing logic and automated screening across the sensor fleet, with an initial set of qualified candidates identified for pilot calibration. Early screening suggests that about one-third of level sensors in gravity networks may satisfy the spatial and temporal constraints needed for meaningful VFM behavior, indicating potential to increase monitoring density without additional confined-space installations. The pilot will report DWF accuracy across diverse diameters, slopes, and tributary patterns and will document failure modes where VFMs are not appropriate. The significance is practical and scalable: the framework converts existing SSO monitoring investments into additional flow intelligence with clear boundaries on uncertainty. By enforcing spatial eligibility and pairing it with conservative thresholds and confidence flags, VFMs can support diagnosis of recurring surcharge, screening of basins for I&I follow-up, prioritization of field investigations, and improved calibration confidence where permanent flow coverage is limited. At utility scale, VFMs also create a consistent way to compare performance across basins and to track how maintenance actions (cleaning, repairs, or sensor service) change observed DWF signatures over time.