Electricity Theft Detection using XAI

Detecting Electricity theft using smart meter data

Comment

TRACK 1: SOCIAL GOOD

Project Description

AI-Based Reliable and Explainable Energy Theft Detection

Electricity theft is a major challenge worldwide, leading to billions in losses, grid instability, unfair billing, and operational inefficiencies for utilities. Traditional theft detection relies on manual audits or supervised AI models that require labeled theft data which is rarely available.

Our project uses unsupervised anomaly detection on smart meter data to automatically flag unusual consumption behavior that may indicate energy theft, while also explaining why a pattern looks suspicious. This balances detection accuracy with transparency so operators can review findings rather than trusting a black-box model.

We built the solution using:

Python for data processing & modeling

Pandas / NumPy for feature engineering

scikit-learn (Isolation Forest) for anomaly detection

SHAP for explainable AI

Flask API + Node.js backend for integration

React for the dashboard frontend

The pipeline ingests smart meter readings, engineers behavioral and power-quality features, detects anomalies, and presents interpretable results to utility operators.

Purpose

Electricity theft affects millions of people. Losses are often passed onto honest consumers through higher tariffs, and unstable load conditions can damage infrastructure and appliances. At the same time, falsely accusing users is unacceptable and creates distrust.

The problem statement was choosen for the following reasons:

It has a real-world social and economic impact

Smart meters are becoming widespread

Our key motivation was to create a system that is: data-driven explainable practical for field use

Instead of claiming perfect theft detection, the system aims to assist analysts by highlighting suspicious usage patterns and showing the reasoning behind each flag.

If developed further, utilities could:

prioritize field inspections more efficiently

reduce non-technical losses

minimize billing discrepancies

improve grid reliability

All while maintaining fairness and transparency.

How it Works

Users (such as utility analysts) can upload or connect smart-meter data to the system. The application then:

Preprocesses & aggregates meter readings

Smart-meter energy, voltage, current, and frequency readings are resampled into 15-minute intervals and cleaned.

Engineers meaningful behavioral features

Including:

consumption trends and daily patterns

peer-comparison scores

rolling averages and variability

voltage and load stability metrics

peak vs off-peak behavior

Detects anomalies using Isolation Forest

The ML model identifies the ~1% most unusual readings across all meters — these become investigation candidates.

Explains each anomaly using SHAP

So operators can see:

which variables influenced detection

whether behavior looks abnormal vs peers

whether supply quality influenced readings

Displays results in a web dashboard

Key features include:

list of flagged meters/events

anomaly scores

trend visualization

human-readable explanations

Dataset

We tested using a real-world smart meter dataset from Mathura (India) containing residential & commercial consumption logs. No personally identifiable information is used.

This ensures realistic behavior patterns.

The Demo (2–5 minute walkthrough)

The demo video includes:

00:00 Problem Overview
00:21 Data Processing
00:39 Anomaly Detection (Isolation Forest)
00:46 Explainability (SHAP)
00:58 Deployment & Interface
01:38 System Integration
01:51 Conclusion

Built With

Share this project:

Updates