军事知识服务平台
基于特征选择与AE-TCN-LA的线损率异常检测方法
An Anomaly Detection Method for Line Loss Rate Based on Feature Selection and AE-TCN-LA
摘要 针对配电网线损数据非线性强、时序耦合度高及异常模式隐蔽导致的检测难题,提出一种融合特征选择、自编码器(AE)、时间卷积网络(TCN)与局部注意力机制(LA)的线损率异常检测模型.首先,采用互信息与随机森林相结合的两阶段特征选择策略,剔除冗余特征,筛选出与线损异常关联度高的关键变量;其次,构建自编码器与时间卷积网络并行的双分支架构,其中自编码器通过重构输入序列挖掘数据深层分布特征以增强异常辨识能力,时间卷积网络则引入深度可分离卷积以降低参数量并精准捕获长短期时序依赖;随后,在双分支输出端嵌入多头局部注意力机制,强化模型对关键特征维度及重要时间片段的聚焦能力;最后,引入自适应权重系数动态协调重构误差与预测误差的损失贡献.基于真实配电网线损数据集的仿真实验表明,所提模型的准确率、精确率、召回率及F1值分别达到91.17%、90.48%、93.78%和92.10%,较对比模型具有更优的综合检测性能.
Abstract:
To address the detection challenges posed by strong nonlinearity,high temporal coupling,and concealed anomaly patterns in distribution network line loss data,this paper proposes a line loss rate anomaly detection model that integrates feature selection,an autoencoder(AE),a temporal convolutional network(TCN),and a local attention mechanism(LA).First,a two-stage feature selection strategy combining mutual information and random forest is adopted to eliminate redundant features and screen out key variables highly correlated with line loss anomalies.Second,a dual-branch architecture is constructed with an autoencoder and a temporal convolutional network running in parallel.The AE reconstructs the input sequence to mine deep distributional features of the data,thereby enhancing anomaly discriminability,while the TCN incorporates depthwise separable convolutions to reduce parameter count and accurately capture both short-term and long-term temporal dependencies.Subsequently,a multi-head LA is embedded at the output of the dual branches to strengthen the model's focus on critical feature dimensions and important temporal segments.Finally,adaptive weight coefficients are introduced to dynamically coordinate the loss contributions of reconstruction error and prediction error.Simulation experiments based on real-world distribution network line loss datasets demonstrate that the proposed model achieves accuracy,precision,recall,and F1-score of 91.17%,90.48%,93.78%,and 92.10%,respectively,exhibiting superior overall detection performance over comparison models.
刊 名 智慧电力 ISTICEIPKU
年,卷(期): 2026, 54(9)
分类号: TM
机标分类号: TP393.08 TN929.1 TP18
在线出版日期: 2026年11月9日