Evidence before the visit.
Each account carries the technical case for inspecting it.
Detecting non-technical losses.
GridSight® NSE, two detection layers, built entirely on data the utility already had.
ZeroHardware or field integration required.
1+ yearOf rolling meter history used to build baselines.
Two layersStatistical detection + ML classification, stacked for confidence.
The challenge.
Non-technical losses were hiding in plain sight, buried in meter data that had never been analysed.
Energy theft ranks among the most stubborn revenue drains any distribution utility manages, and a major Latin American DSO running one of the region's largest networks was no exception. The World Bank puts non-technical losses at roughly 17% of all electricity generated across Latin America, money sitting in exactly the kind of untouched meter data this utility already had on hand.
Meter data had accumulated for years with no automated time-series analysis running against it, so anomalies stayed buried in the historical record until an inspection happened to expose them. Field visits followed geography and gut feel rather than statistical evidence: high cost per visit, low fraud-confirmation rate.
The signals were there the whole time: abrupt power drops, P/Q phase shifts, irregular cos φ behaviour. Nobody was watching for them systematically, so they surfaced during a scheduled spot-check, long after the revenue had
already gone.
The solution.
GridSight® NSE module, two detection layers, built entirely on data the utility already had.
Enline deployed GridSight®'s NSE module across a defined network segment, running time-series anomaly detection to build per-user and per-group consumption baselines. The modeling stack (ARIMA/SARIMA, STL decomposition and dynamic rolling standard deviations) is applied across active power, reactive power, cos φ and voltage readings. Every deviation is scored and ranked automatically.
A second pass runs Isolation Forest, alongside Random Forest and deep-learning classifiers pre-trained on Enline's broader fraud-detection dataset, against the flagged population. This separates noise from high-probability fraud candidates and ranks the output by confidence.
Where the utility held records of confirmed fraud (the same scattered evidence past spot-checks had turned up), those cases were used to fine-tune detection thresholds before any result reached a field team.
The outcome.
A ranked, evidence-backed shortlist. No hardware, no integration, no upfront cost.
Systematic non-technical-loss identification recovers revenue directly. No hardware to install, no integration project to run, no capital outlay: it runs on data the utility already owns.
Every flagged account arrives with the technical case for inspecting it. Targeted inspections built on validated, high-confidence anomalies cut wasted visits and raise the detection rate per visit, turning a reactive spot-check habit into a systematic, evidence-led programme.
Revenue, not capex.
No hardware, no integration project, no capital outlay.
Thresholds tuned on real cases.
Confirmed fraud records calibrate detection before field work.