基于深度孪生自回归网络的无监督异常用电检测

异常用电检测旨在识别出不符合正常用电规律或者违反用电合约的用电行为。针对现有基于重构的检测方法依赖标记的正常样本和难以捕捉复杂时间依赖性的问题，提出一种基于深度孪生自回归网络的无监督异常用电行为检测模型（DSAD）。所提模型通过两个孪生自回归子网络来分别独立地对无标记的输入数据进行重构，再将两个子网络的重构误差相结合来预测数据中的正常样本，并利用多头自注意力机制来有效地捕捉时间依赖性、周期性和随机性等复杂特征。在大规模时序数据集和国家电网真实用电数据集上进行实验所获得的结果表明，所提模型在AUC以及AP等性能指标上取得了更好的检测效果。

Unsupervised abnormal power consumption detection via deep siamese autoregressive network

Abnormal electricity consumption detection aims to identify electricity consumption behaviors that do not conform to normal electricity consumption patterns or violate electricity consumption contracts. To address the issues of existing reconstruction-based detection methods relying on labeled normal samples and failing to capture complex time dependencies, this paper proposed an unsupervised abnormal electricity consumption detection model based on deep siamese autoregressive networks(DSAD), which used two siamese autoregressive subnetworks to independently reconstruct the unlabeled input data, and then combined the reconstruction errors of the two subnetworks to predict the normal samples in the data, and utilized multi-head self-attention mechanism to effectively capture complex features such as time dependency, periodicity and randomness. The results obtained from experiments on large-scale time series datasets and real electricity consumption datasets from state grid show that the proposed method achieves better detection performance in terms of AUC and AP.