用户窃电行为是电网企业运营管理的痛点,基于数据驱动的低压用户窃电检测是当前的重要发展方向。由于窃电数据集本身高维度且样本不平衡的特点,对窃电检测模型的拟合能力和泛化能力要求极高。为此,首先利用堆栈降噪自编码器(Stacked Denoising Auto Encoder,SDAE)对低压用户日用电量数据进行特征提取工作,以避免窃电产生的极端数据对模型的影响并挖掘数据的深层特征；进而提出采用逻辑回归与深度神经网络联合训练模型进行低压用户窃电检测的方法,将逻辑回归模型的记忆能力与深度神经网络模型的泛化能力相结合,进一步提升窃电检测的精度。通过实际电网数据的实验仿真,从AUC值、准确率和召回值三个评价指标验证了所提出方法相对于传统机器学习算法具有明显的性能优势。
英文摘要:
In the low-voltage distribution network, the practices of electricity theft have been continuously causing economic losses to power grid enterprises, while the development of smart grids provides a reliable data basis for the detection of electricity theft for the data-driven low-voltage users. Due to the characteristics of high dimension and unbalanced sample in the data set of electricity theft, the fitting and generalization abilities of the detection model are highly required. On this basis, a method of detecting electricity theft by low-voltage users is proposed, which firstly uses the Stacked Denoising Auto Encoder (SDAE) to perform data mining, and then uses the Wide and Deep Model with high fitting and generalization abilities to detect electricity theft. Through the experimental simulation of actual power grid data, evaluation indicators of AUC value, accuracy rate, and recall value verify that the proposed method has obvious performance advantages over the traditional machine learning algorithm .