基于深度混合注意力网络的窃电检测

窃电检测旨在识别和检测非法或未经授权的电力使用行为。在智能电网技术高速发展的背景下，如何实现准确的窃电行为检测，是学术界和工业界广泛关注的一个重要问题。针对已有方法依赖人工特征设计以及低层特征提取能力不足的问题，提出了一种基于深度混合注意力网络的窃电检测方法，将通道注意力和自相关注意力机制相结合，在不同层次和空间范围内捕捉数据中的时间依赖性和周期性等复杂特征。所提模型在低层使用通道注意力网络来增强低层特征的表达能力，在中间层使用自相关注意力来捕捉全局上下文信息，并利用自监督方法来学习注意力参数，从而提取出更具表达力和判别力的特征表示。在中国国家电网数据集上进行实验所获得的结果表明，所提出的方法在AUC以及F1等性能指标上取得了更好的效果。

Electricity Theft Detection Based on Deep Hybrid Attention Networks

Electricity theft detection aims to identify and detect unauthorized or illegal electricity usage. In the context of rapid development of smart grid, achieving accurate electricity theft detection has become an important concern in both academia and industry. Aiming at the limitations of existing methods, which rely on manual feature design and have insufficient capability in low-level feature extraction, an electricity theft detection method based on deep hybrid attention network is proposed. The proposed model combines channel attention and self-attention mechanisms to capture complex features such as temporal dependencies and periodicity across different levels and spatial ranges in the data. Specifically, the model enhances the expression of low-level features using a channel attention network in the low layers, captures global contextual information using self-attention in the middle layers, and learns attention parameters through self-supervised learning to extract more expressive and discriminative feature representations. Experimental results on the national power grid dataset demonstrate that the proposed method can achieve better performance in terms of AUC and F1 scores.