一种基于自适应RNN的居民异常用电行为智能检测方法

Intelligent detection method for abnormal electricity consumption behavior of residents based on adaptive RNN

中文关键词: 配电网  数据驱动  异常用电  循环神经网络  超参数优化

英文关键词: Distribution network  Data driven  Abnormal electricity consumption  Recurrent neural network  Hyperparameter optimization

中文摘要:

      针对目前居民用电异常行为识别时存在效率低、性能差问题,提出了一种基于自适应循环神经网络（recurrent neural net-work,RNN）的异常用电行为识别模型。设计了一种合成少数过采样技术和编辑最邻近（synthetic minority oversampling tech?nique-edited nearest neighbor,SMOTE-ENN）重采样方法,增加不平衡数据集的分类性能。建立了自适应RNN检测模型,使用批量归一化RNN作为基础学习器,并结合超参数优化和缓冲区来动态调整（batch normalized recurrent neural network,BNRNN）模型。实验阶段,经过改进的SMOTE-ENN重采样后,模型分类效果大幅度提升。同时,实验验证了具有缓冲和超参数优化的所提自适应RNN模型具有最低的平均绝对误差,表明所提模型具备较优的泛化能力。试验结果验证了所提模型的实用性及优异性能,该模型可为异常用电行为检测的发展提供一定借鉴作用。

英文摘要:

      A novel model based on adaptive recurrent neural network（RNN）is proposed to address the issues of low efficiency and poor performance in identifying abnormal electricity consumption behavior among residents. Design a SMOTE-ENN resampling method to increase the classification performance of imbalanced datasets. We have established an adaptive RNN detection model, using batch normalized RNN as the basic learner, and combining hyperparameter optimization and buffer to dynamically adjust the BNRNN model. In the experimental stage, after improved SMOTE-ENN resampling, the classification performance of the model was significantly improved. At the same time, experiments have verified that the proposed adaptive RNN model with buffering and hyperparameter optimization has the lowest MAE error, indicating that the proposed model has excellent generalization ability. The experimental results validate the practicality and excellent performance of the proposed model, which can provide some reference for the development of abnormal electricity consumption behavior detection.