alirezaaminzadeh/smart-meter-anomaly-detector

Smart Meter Anomaly Detector

Anomaly detection on smart meter (energy consumption) data. Flags unusual consumption patterns using an autoencoder or a classical method (e.g. Isolation Forest). Optional minimal Gradio dashboard for uploading series and viewing scores.

Overview

Designed for utilities or building management: input a time series of consumption readings; output anomaly scores or binary labels. Useful for fraud detection, fault detection, or demand response analysis.

Model

• Autoencoder (PyTorch): reconstructs a window of consumption; high reconstruction error indicates anomaly. Model in model.py, training in train.py, inference in inference.py; checkpoint ./checkpoints/anomaly_ae.pt. Optional Gradio app in app.py.

Usage

Training:

pip install -r requirements.txt
python train.py

Demo:

python app.py

Upload a CSV with a consumption column or use synthetic data; the app shows anomaly score or label.

Data

• CSV with a consumption or value column (and optional timestamp). Training uses normal data; evaluation can include labeled anomalies.

Limitations / future work

• Threshold for "anomaly" is data-dependent; may need tuning per deployment.
• Could add explanation (which time steps contribute most to the score).

Author

Alireza Aminzadeh

• Email: alireza.aminzadeh@hotmail.com
• Hugging Face: syeedalireza
• LinkedIn: alirezaaminzadeh