深度神经网络的电力数据异常检测研究

Research on Power Data Anomaly Detection Based on Deep Neural Network Techniques

提出了一种基于深度学习的检测电力数据异常方法,该方法通过两个阶段来解决问题:首先建立基于长短期记忆(LSTM)的神经网络,用于预测下一小时的样本;其次利用LSTM自编码器学习正常消费特征,第一阶段的输出被作为LSTM自编码器的输入,用于学习正常消费特征,如果输入与输出不同,表示存在异常,使用指数移动平均值作为阈值,以区分局部异常和全局异常。此外,本文还考虑了天气、时间和滞后特征,并进行了特征选择以找到最佳组合。通过比较异常和正常消耗的验证方法,结果显示异常期间电力消耗显著增加,时间和滞后特征能够提高所提出方法的效率和性能。

In this paper,a deep learning-based method for detecting anomalies in electricity data is proposed,which solves the problem in two stages:firstly,a neural network based on long short-term memory(LSTM)is built for predicting the next hour's samples.Secondly,the output of the first stage is used as the input to the LSTM self-encoder for learning normal consumption features using the LSTM self-encoder.If the input differs from the output,it indicates the presence of anomalies,and an exponential moving average is used as a threshold to distinguish between local anomalies and global anomalies.In addition,weather features,time and lag features are considered in this paper,and feature selection is performed to find the best combination.By comparing the validation methods of anomalous and normal consumption,the results show a significant increase in power consumption during anomalies,and the time and lag features can improve the efficiency and performance of the proposed method.