Unbalanced power anomaly detection model based on improved transformer and countermeasure encoder.

Area of Science:

• Electrical Engineering
• Artificial Intelligence
• Data Science

Background:

• Intelligent grid anomaly detection struggles with imbalanced data, high computational complexity, and scarce anomaly samples.
• Existing methods often exhibit low minority-class recognition and model bias.

Purpose of the Study:

• To develop a novel hybrid architecture for enhanced intelligent grid anomaly detection.
• To address challenges of imbalanced data, computational complexity, and model bias in anomaly detection.

Main Methods:

• A hybrid architecture combining an enhanced Transformer with an Adversarial Autoencoder (AAE).
• Incorporation of Locality-Sensitive Hashing (LSH) attention with Focal Loss with Temperature (FLT) for feature clustering.
• Implementation of a dynamic weighting module (Spatial-Temporal Feature Disentanglement Network - STFDN) for adaptive gradient adjustment.
• Application of Spectral Normalization to reduce memory usage for node sequences.

Main Results:

• Reduced memory usage by 52.4% (from 18.7GB to 8.9GB) using Spectral Normalization.
• Achieved a 10.4% improvement in FID score to 28.4 under Wasserstein distance constraints.
• Elevated AUPRC to 0.837 on the SGSC dataset using dynamic temperature scaling.
• Obtained an F1-score of 89.3% with 183ms inference latency on the UK-DALE dataset, suitable for edge deployment.

Conclusions:

• The proposed hybrid model effectively addresses key challenges in intelligent grid anomaly detection.
• The novel architecture demonstrates significant improvements in accuracy, efficiency, and computational resource utilization.
• This research paves the way for a new generation of automated grid operation and maintenance tools.