A big data statistical analysis platform designed to detect water and electricity theft, meter tampering, and irregular consumption patterns. Extensible to retail, transport, and logistics fraud detection.
South African municipalities lose billions of rands annually to non-revenue water (NRW) and electricity theft. Illegal connections, meter tampering, bypassed meters, and data corruption at vending points make it extremely difficult for utilities to accurately bill for consumed services. Traditional audit methods — manual meter inspections and periodic field audits — only catch a fraction of losses and are prohibitively expensive to scale.
Our client, a major South African metropolitan municipality, needed a software solution that could automatically analyse millions of daily meter readings and vending transactions to identify anomalous consumption patterns indicative of theft, tampering, or system leakage — without requiring manual inspection of every meter.
Inovosystems built a big data anomaly detection platform that ingests and analyses millions of daily meter readings, vending transactions, and network telemetry data. The system applies multiple statistical and machine learning techniques to identify consumption patterns that deviate from expected norms.
The platform detects anomalies including: sudden drops in consumption (indicating meter bypass or tampering), irregular vending patterns (suspicious token purchases), consumption profiles inconsistent with property type, zero-consumption meters with no reported faults, and abnormal flow patterns in water distribution zones. Each detected anomaly is scored by severity, cross-referenced with historical data, and queued for field investigation with prioritised dispatch recommendations. The system also generates comprehensive loss reports aligned with Department of Water and Sanitation (DWS) and NERSA reporting requirements.
Apache Kafka for real-time ingestion of meter readings and transaction data. Apache Spark for large-scale batch processing and feature extraction. Custom ETL connectors for municipality billing systems and SCADA interfaces.
TimescaleDB for time-series consumption data with built-in compression. PostgreSQL for relational metadata (customer accounts, meter registry, property records). S3-compatible object storage for raw data archives and audit trails.
Python-based ML pipeline using scikit-learn, TensorFlow, and PyOD (Python Outlier Detection) library. Ensemble of isolation forests, LSTM autoencoders for time-series anomaly detection, and statistical process control (SPC) methods.
React-based dashboard with dynamic heat maps showing anomaly density by geographic zone. Automated report generation (PDF/Excel) for regulatory submissions. REST API for integration with municipality CRM and work order management systems.
Multiple AI models work in concert to detect, classify, and prioritise anomalies across the utility network.
LSTM autoencoders learn normal consumption patterns for each meter and flag deviations. The model accounts for seasonal variations, weather patterns, and day-of-week effects, reducing false positives by 70% compared to static threshold methods.
Random forest classifier trained on known tampering signatures identifies meters showing physical tampering indicators — reverse flow, zero consumption despite active connection, and erratic reading patterns. Achieves 96% precision in tampering identification.
Graph neural network models analyse the relationship between connected meters in a distribution zone, identifying discrepancies where total supply to a zone significantly exceeds aggregated consumption — pinpointing areas likely containing illegal connections.
Time-series forecasting models project non-revenue water and electricity losses forward, enabling municipalities to quantify the financial impact of theft over time and prioritise intervention zones based on projected return on investigation effort.
The anomaly detection platform processes over 5 million meter readings daily across three South African municipalities. In the first year of deployment, the system identified ZAR 8.2 million in previously undetected revenue leakage from illegal connections and meter tampering. The deep learning models reduced false positive alerts by 70% compared to the previous rule-based system, enabling field investigation teams to focus their efforts on high-confidence leads. One municipality recovered ZAR 3.4 million in unbilled consumption within six months through targeted enforcement actions guided by the platform's recommendations. The system has since been extended to a retail client for point-of-sale transaction fraud detection.
Custom machine learning models for anomaly detection, predictive analytics, and pattern recognition across any data domain.
Learn more →Big data pipelines, real-time stream processing, and scalable storage architectures for high-volume analytics workloads.
Learn more →Data strategy, ML architecture design, and implementation guidance for organisations tackling fraud and revenue leakage.
Learn more →Bespoke analytics platforms built to your specific data sources, detection requirements, and integration landscape.
Learn more →Let's build an anomaly detection system that protects your revenue — whether in utilities, retail, transport, or logistics.
Start a Project →