基于对冲对抗机制与图注意力网络的异常用电检测

Abnormal Electricity Consumption Detection Based on Hedge-Antagonism Mechanism and Graph Attention Network

针对基于数据驱动的异常用电检测方法准确率不高的问题,提出一种基于配电网潮流约束的异常用电检测方法。首先,建立了基于卷积神经网络对冲反向传播网络的潮流计算模型与基于潮流方程约束的精度优化模型,实现在不完全依赖线路参数与拓扑情况下的准确潮流计算。其次,在对冲反向传播过程中加入改进对抗攻击算法形成对冲对抗机制,以电压计算值向测量值逼近为目标,实现在边界条件下功率回溯。再次,计算功率回溯值与量测值间的回溯距离并将其作为用电特征量,以加有异常累加器的图注意力网络作为分类器,实现异常回溯距离的辨识。最后,结合实际案例与仿真实验验证了所提方法的有效性。

To solve the problem that data-driven abnormal electricity consumption detection method has low accuracy, an abnormal electricity consumption detection method is proposed based on the power flow constraints of distribution networks. First, a power flow calculation model based on convolutional neural network-hedge backpropagation(CNN-HBP) network and a precision optimization model based on power flow equation constraints are established to achieve accurate power flow calculation without entirely relying on line parameters and topology. Secondly, in the process of hedge backpropagation, an improved adversarial attacks algorithm is added to form the hedge-antagonism mechanism, aiming at the voltage calculation value approaching the measured value to achieve power backtracking under boundary conditions. Thirdly, the backtracking distance between the power backtracking value and the measured value is calculated as the characteristic quantity of electricity consumption and the graph attention network(GAT) with an abnormal accumulator is used as the classifier to realize the identification of the abnormal backtracking distance. Finally, the validity of the proposed method is verified by practical cases and simulation experiments.