基于BiLSTM-CatBoost模型的电力用户异常用电检测

User Abnormal Electricity Detection Based on BiLSTM-CatBoost Model

为进一步提高电力用户异常用电检测性能,实现异常用电行为的快速准确研判,该文提出一种基于BiLSTM-CatBoost模型的用户异常用电检测方法。该模型首先采用双向长短期记忆神经网络(bidirectional long short-term memory,BiLSTM)对用户用电数据进行特征深度提取,利用用电时序数据两个方向的信息来获取特征维度;接着采用完全对称决策树为基学习器的CatBoost集成学习算法作为分类器进行检测,避免检测过拟合。BiLSTM和CatBoost均采用贝叶斯优化算法(Bayesian optimization,BO)对全局进行超参数寻优。实例对比表明,基于BiLSTM-CatBoost模型的用户异常用电检测方法在准确率、F1分数、AUC、MCC值等评价指标中均表现最优。

In order to further improve the abnormal power consumption detection performance of power users and realize the rapid and accurate judgment of abnormal power consumption behavior,a user abnormal power consumption detection method based on BiLSTM-CatBoost model is proposed.The model firstly uses Bidirectional Long Short-term Memory to extract the feature depth of user electricity data,and uses the information of two directions of electricity time series data to obtain the feature dimension.Then,the CatBoost ensemble learning algorithm based on the fully symmetric decision tree is used as the classifier to avoid overfitting.Both BiLSTM and CatBoost use Bayesian Optimization to perform global hyperparameter optimization.The comparison of examples shows that the user abnormal electricity detection method based on BiLSTM-CatBoost model performs best in evaluation indexes such as accuracy,F1 score,AUC and MCC value.