Power Grid Anomaly Detection Based on Transformer Architecture with Sparse Attention and Self-Supervised Learning

Power grid anomaly detection is crucial for maintaining electrical system stability and preventing potential failures. Traditional statistical methods and conventional machine learning approaches often struggle with the complex temporal dependencies and high-dimensional characteristics inherent in power consumption data. This paper proposes a novel anomaly detection framework that integrates Transformer architecture with sparse attention mechanisms, self-supervised learning, and Generative Adversarial Networks (GANs) for enhanced power grid anomaly detection. Our approach leverages the self-attention mechanism of Transformers to capture long-range temporal dependencies while employing sparse attention to reduce computational complexity and focus on the most relevant time steps. The self-supervised learning component enables the model to learn meaningful representations from unlabeled data, while the GAN framework provides additional discriminative power for anomaly identification. Extensive experiments conducted on the London Household Electricity Consumption (LCL) dataset demonstrate the superiority of our proposed method, achieving remarkable performance metrics: Precision of 0.9550, Recall of 0.8833, F1-Score of 0.9177, and AUC of 0.9525. Comprehensive comparisons with baseline models including LSTM, CNN, and AutoEncoder variants, along with detailed ablation studies, validate the effectiveness of each component in our integrated framework. The results indicate that our approach significantly outperforms traditional methods and provides a robust solution for real-world power grid monitoring applications.