Galadrim built an explainable ML pipeline to produce daily network yield (RDR) figures from imperfect meter data, at scale and ready for operations.
12 weeks from discovery to integrated delivery
Near-100% daily RDR coverage despite missing reads
Explainable estimates with auditable data quality checks
Built to scale across ~ 60,000 meters and multiple sectors
SUEZ needed reliable daily RDR (network yield) to steer operations and support contractual commitments on a large drinking-water network. The challenge: daily consumption data were incomplete and irregular—mixing non-smart meters (manual reads) and smart meters with gaps in daily values—making direct, accurate RDR computation infeasible.
We delivered a focused, end-to-end enhancement that turns raw, imperfect inputs into trusted daily RDRs:
Data cleaning & validation: Robust ingestion, profiling, and rule-based checks to standardise time series and surface anomalies upfront.
ML imputation for missing consumption: An ensemble-based model estimates missing daily values with feature-rich inputs (time patterns, sector data, exogenous signals). We provide SHAP-based explanations for transparency and trust.
Daily RDR calculation & distribution: Deterministic RDR computation at global and sector level, packaged for SUEZ’s cloud environment with exports/APIs for downstream use.
Incomplete telemetry: Handled non-smart meters and sporadic smart-meter gaps without sacrificing reliability.
Bias & explainability: Addressed non-random missingness with a modelled approach and human-readable justifications.
Operationalisation: Delivered a repeatable, monitored pipeline with clear failure modes and fast reruns.
