Multi-granularity Robust Power Fault Prediction based on Cascaded Attention Network

Power failure has a serious impact on normal production and life. Therefore, power fault prediction is of great significance. Most existing methods of power fault prediction focus on single-granularity fault prediction based on massive data. Differently, we propose a novel cascaded attention network to predict power fault risk in multi-granularity, which is mainly composed of a client-level fault prediction transformer and a community-level fault prediction transformer. In the client-level fault prediction transformer, we introduce temporal embedding to supplement the timing information and design a mask operation to ensure the causality of feature interaction. In the community-level fault prediction transformer, we design a community fault query to aggregate the client features in a certain region (community) and predict community-level power fault risk. Moreover, there is a lot of missing and noise in power data, which leads to data sparsity. Thus, a masked self-supervised pretrain method is proposed to enhance relational modeling under sparse data, which can significantly improve the robustness of the model. Experimental results demonstrate the effectiveness of each design. In the actual data, our method achieves 99.1% precision and 95.6% recall in client-level fault prediction and achieves 99.3% precision and 96.2% recall in community-level fault prediction.