Overview
Power distribution networks globally face substantial revenue loss and operational inefficiency caused by non-technical losses (NTLs) and unmetered energy consumption. Traditional detection methods rely heavily on delayed manual inspections or supervised models requiring hard-to-acquire labeled theft data.
This project implements an intelligent, unsupervised time-series anomaly detection platform paired with Explainable AI (XAI) to flag irregular consumption patterns and provide transparent, interpretable diagnostic insights for grid analysts.
Core Capabilities
- Automated Risk Scoring: Ingests high-frequency smart meter telemetry to identify irregular consumption trends and non-technical loss candidates without requiring historical labeled training data.
- Explainability Integration: Deconstructs anomaly predictions into human-interpretable risk metrics, enabling utility operators to review the specific driving factors behind each flag.
- Interactive Utility Dashboard: Provides grid operators with prioritization queues, consumption trend visualizations, and localized anomaly summaries for streamlined field auditing.
System Architecture & Engineering
- Data Processing & Feature Engineering: Cleaned, structured, and normalized multi-dimensional time-series meter telemetry. Extracted statistical, behavioral, and power-quality indicators to build resilient consumption profiles.
- Analytical Modeling: Evaluated multi-dimensional consumption deviations using unsupervised machine learning algorithms tailored for time-series outlier detection.
- Explainability Layer: Integrated post-hoc model explainability frameworks to translate model outputs into clear, audit-ready operational explanations.
- Full-Stack Application: Built a modular REST API backend integrated with a responsive React frontend for real-time monitoring and reporting.
Key Challenges & Takeaways
- Balancing anomaly sensitivity against false positives caused by benign load spikes and seasonal demand variations.
- Designing human-centered explainability interfaces that present complex statistical telemetry in an intuitive format for field inspectors.
Built With
- arduino
- dataset
- flask
- javascript
- json
- node.js
- numpy
- pandas
- react
- rest-apis
- scikit-learn
