Platform 15 / 18
Business domain
Loss, Compliance & AssuranceSEIDRA
ML-Powered Theft Detection
ML-powered electricity theft detection and scoring platform for Turkish electricity distribution companies (EDAŞ). Ingests subscriber data, meter readings, OBIS events and inspection history to produce explainable theft probability scores (0-100) per subscriber, enabling field teams to prioritize non-technical loss inspections with 3x+ hit rate improvement.
15Highlights
- 013-layer ML architecture: Base Model (XGBoost), OSOS Head (XGBoost + LSTM + Rules), Boost Head
- 02100+ features across 8 categories: power triangle, per-phase, load profile, tariff integrity and more
- 03LSTM autoencoder pattern transfer — learns consumption signatures of tamper-flagged meters
- 04SHAP-based explainability with natural language Turkish explanations per suspect
- 05Configurable OBIS catalogue with 30+ entries and per-manufacturer bitmask definitions
- 06Graph analytics (PostGIS) for fraud cluster detection and geographic correlation
- 07Multi-DB adapter framework (Oracle, PostgreSQL, MSSQL, MySQL) with REST and CSV ingestion
- 083x+ inspection hit rate improvement over random targeting, auditable model lifecycle
15Modules
| Code | Module | Scope |
|---|---|---|
| ING | Data Ingestion | Multi-DB adapter framework (Oracle, PostgreSQL, MSSQL, MySQL), REST API for OSOS data, CSV bulk import and DLMS/IEC 62056-21 parser |
| OBI | OBIS Catalogue | 30+ configurable entries, per-manufacturer bitmask definitions and admin UI for editing and discovery |
| FEA | Feature Engineering | 100+ features across 8 categories: power triangle, per-phase, load profile, demand, tariff integrity, comms/time, transformer balance and peer comparison |
| ML | 3-Layer ML Architecture | Base Model (XGBoost), OSOS Head (XGBoost + LSTM + Rule Engine) and Boost Head (sparse XGBoost) with staging and promotion |
| PAT | Pattern Transfer | LSTM autoencoder learning consumption patterns of tamper-flagged meters to identify sophisticated theft in unflagged subscribers |
| EXP | Explainability | SHAP-based feature attribution with natural language Turkish explanations and confidence levels (LOW/MED/HIGH/VERY_HIGH) |
| INS | Inspection Workflow | Prioritized subscriber list, drill-down profiles, inspection order creation, field team dispatch and feedback loop |
| GRA | Graph Analytics | PostGIS-based fraud cluster detection, geographic correlation and network-level theft pattern analysis |
- INGData Ingestion
- Multi-DB adapter framework (Oracle, PostgreSQL, MSSQL, MySQL), REST API for OSOS data, CSV bulk import and DLMS/IEC 62056-21 parser
- OBIOBIS Catalogue
- 30+ configurable entries, per-manufacturer bitmask definitions and admin UI for editing and discovery
- FEAFeature Engineering
- 100+ features across 8 categories: power triangle, per-phase, load profile, demand, tariff integrity, comms/time, transformer balance and peer comparison
- ML3-Layer ML Architecture
- Base Model (XGBoost), OSOS Head (XGBoost + LSTM + Rule Engine) and Boost Head (sparse XGBoost) with staging and promotion
- PATPattern Transfer
- LSTM autoencoder learning consumption patterns of tamper-flagged meters to identify sophisticated theft in unflagged subscribers
- EXPExplainability
- SHAP-based feature attribution with natural language Turkish explanations and confidence levels (LOW/MED/HIGH/VERY_HIGH)
- INSInspection Workflow
- Prioritized subscriber list, drill-down profiles, inspection order creation, field team dispatch and feedback loop
- GRAGraph Analytics
- PostGIS-based fraud cluster detection, geographic correlation and network-level theft pattern analysis