基于信息熵的联邦学习异常用电识别

Information Entropy-based Federal Learning for Identifying Abnormal Electricity Consumption

为了识别出异常用电,便于电力企业进行后续的故障定位与检修,众多异常用电识别模型被提出。现有方法要求将分散在各个供电公司的数据汇集起来,统一进行模型构建。然而,这种中心化的数据汇集操作可能会在数据传输时造成用户数据泄露。为保证用户用电数据的隐私性,提出一种基于信息熵的联邦学习异常用电识别模型。该模型借助于联邦学习范式中“模型本地训练、参数在线更新”的计算策略,以存储在不同客户端的数据块作为训练数据,避免了集中式模型训练所带来的潜在数据泄露风险。此外,在联邦学习中间参数聚合过程中,各个客户端数据质量之间的差异性可能会给模型性能带来潜在影响。因此,引入Kozachenko-Leonenko k近邻评估算法获取客户端数据集的信息熵,以此表示各个客户端中间参数在聚合过程中对模型最终参数的贡献。仿真结果表明,基于信息熵的联邦学习异常用电识别方法能够在保证数据隐私的情况下,有效地识别出异常用电数据。

Numerous abnormal electricity consumption identification models have been proposed to identify abnormal electricity consumption and facilitate fault location and maintenance by electricity power companies. Existing methods require bringing together data scattered in various companies for model construction. However,this centralized data pooling operation may cause leakage of user data during transmission. In order to ensure the privacy of data,the information entropy-based federal learning for Identifying abnormal electricity consumption is proposed. This model is based on the computational strategy of "local training of model and online updating of parameters" of federal learning paradigm and uses data blocks stored in different clients as training data to avoid the potential data leakage risk caused by centralized model training. In addition,the variability of data quality among clients in the process of federal learning intermediate parameter aggregation may potentially affect the model performance. The Kozachenko-Leonenko k-nearest neighbor evaluation algorithm is used to obtain the information entropy of the client dataset as a way to characterize the contribution of each client to the final parameters. The simulation results show that the federal learning-based abnormal electricity consumption data identification can effectively identify the abnormal data in electricity consumption data while ensuring data privacy.