Detection of Power System Anomalies Using a Fusion of Machine Learning & Deep Learning

This repository contains the implementation and research artifacts for the paper:

Detection of Power System Anomalies Using a Fusion of Machine Learning & Deep Learning
By Jishnu Teja Dandamudi and Rupa Kandula
Amrita School of Artificial Intelligence, Amrita Vishwa Vidyapeetham, Coimbatore, Tamil Nadu, India

---

🧠 Overview

Anomaly detection in Phasor Measurement Unit (PMU) data is critical to maintaining stability and security in modern power grids. This project introduces a hybrid anomaly detection framework that combines:

• Isolation Forest (IF) for statistical anomaly detection
• LSTM Autoencoder for temporal anomaly detection

These techniques are fused using a weighted strategy to improve detection accuracy, particularly in the presence of missing data, noise, or cyber-induced anomalies.

---

---

🧪 Methodology

LSTM Autoencoder

• Captures temporal patterns from sequential PMU data
• Uses reconstruction error to detect deviations from normal behavior

Isolation Forest (IF)

• Identifies statistical outliers based on recursive partitioning
• Trained on the reconstruction errors from the LSTM

Fusion Strategy

• Final anomaly score: S_f(x) = α × S_LSTM(x) + β × S_IF(x) where (α + β = 1)
• Classification is done using a 95th percentile threshold on the final anomaly score

---

📊 Evaluation

Metrics used:

• Precision: 88.21%
• Recall: 88.29%
• F1 Score: 88.25%

Compared against state-of-the-art techniques like CyResGrid and GC-LSTM+ResNet.

---

📚 Dataset

Used the publicly available dataset:
Realistic Labelled PMU Data for Cyber-Power Anomaly Detection Using Real-Time Synchrophasor Testbed
Available on IEEE DataPort

---

🧩 Future Enhancements

• Adaptive thresholding using Bayesian Optimization or RL
• Self-supervised models using contrastive learning
• Federated learning for decentralized anomaly detection
• Integration of Explainable AI (XAI) for model interpretability

---