Anomaly Detection of Collusion Bidding in Electricity Market Based on Deep Learning Model

Abstract

 In the electricity market, collusion bidding involves multiple entities colluding to inflate prices or conceal capacity, which disrupts marginal cost pricing. Traditional rules and single-variable thresholds struggle to detect such collusion in a timely manner. This paper addresses a real-world scenario with only six input features and highly imbalanced labels by constructing an anomaly detection framework that combines Transformer-Autoencoder (TAE) with Gaussian Mixture Model (GMM). The model uses multi-head self-attention to capture the coupling relationships among original input features, measures sample rarity using Gaussian mixture density in the space of latent variables and reconstruction errors, and automatically generates an energy threshold at the 90th percentile of the normal distribution for anomaly detection. Experimental results show that on the electricity collusion dataset, this method achieves detection performance with Precision 0.80, Recall 0.79, F10.79, and AUC 0.844. It is only ±2% sensitive to fluctuations in the energy weight λ, demonstrating robust performance. The energy distribution compresses normal sample data into spikes while pushing anomalies to the long tail, achieving efficient anomaly detection. Attention heat maps and gradient sensitivity both point to feature one as the most critical feature, validating the interpretability of the model's decision logic. TAE-GMM can be trained solely using normal samples, and the final detection threshold is determined by the proportion of normal samples, enabling simple and flexible application.