基于深度自编码器的智能电网窃电网络攻击异常检测

Anomaly detection of smart grid stealing network attacks based on deep autoencoder

现有AMIs中的异常检测器存在浅层架构,难以捕获时间相关性以及电力消耗数据中存在的复杂模式,从而影响检测性能。提出基于长短期记忆(LSTM)的序列对序列(seq2seq)结构的深度(堆栈)自编码器。自动编码器结构的深度有助于捕获数据的复杂模式,seq2seq LSTM模型可以利用数据的时间序列特性。研究了简单自编码器、变分自编码器和注意自编码器(AEA)的性能,得出在这3种自编码器采用seq2seq结构时检测性能优于全连接结构。仿真结果表明,带有注意力机制的检测器(AEA)检出率和虚警率分别比现有性能最好的检测器高4%~21%和4%~13%。

Existing anomaly detectors in AMIs suffer from shallow architectures,which impede their ability to capture temporal correlations and complex patterns in electricity consumption data,thus impact detection performance adversely.A deep(stacked)autoencoder structure based on Long Short-Term Memory(LSTM)with a sequence-to-sequence(seq2seq)configuration is proposed.The depth of the autoencoder architecture is beneficial for capturing complex data patterns,and the seq2seq LSTM model effectively utilizes the temporal sequential characteristics of the data.The performance of simple autoencoders,variational autoencoders,and Attention Enhanced Autoencoders(AEA)was studied,revealing that using the seq2seq structure in these three types of autoencoders results in superior detection performance compared to fully connected architectures.Simulation results demonstrate that the detector with an attention mechanism(AEA)achieves a 4%~21%higher detection rate and a 4%~13%lower false alarm rate compared to the best-performing existing detectors.