Time Series Anomaly Detection Based On Deep Learning Theory And Smart Grid Time Series Data

Over the past decade,social economy has experienced unprecedentedly rapid and continuous development.Production of human society is increasingly dependent on stable electricity supply,which has been made an indispensable support.In order to prevent the power system instability and even accidental large-scale power outage from happening,anomaly detection in smart grid has been attached much importance and thus plays a significant role.The first part of this thesis proposed an anomaly detection model based on encoder-decoder framework with recurrent neural network(RNN),by following which,an input time series is reconstructed and an anomaly can be detected by an unexpected high reconstruction error.This thesis validated the proposed model by using power demand data from UCR Time Series Classification Archive and IEEE 39 bus system simulation data.The result from the analysis show that the proposed encoder-decoder framework is able to successfully capture anomalies with a precision higher than 95%.The second part of this thesis proposed a model to extract image-related features of original time series.Visualize the covariance matrix of the time series.Extract the texture features(e.g.,HOG,LBP)as the characteristics of the original time series data,and perform the anomaly detection.The proposed model in this thesis has a number of inherent advantages to identify anomalies in smart grid system:First,a large number of records are being collected from intelligent monitoring devices in smart grid,only a tiny fraction of which are anomalous.The proposed encoder-decoder model is suitable for the case where distribution of positive and negative samples is unbalanced.Only normal data is needed for training the proposed model.In contrast,conventional methods for anomaly detection based on machine learning(e.g.,classification algorithm)usually require that the numbers of the two types of samples are approximately equal with each other.Oversampling or under-sampling of the samples is necessary to balance the data,which however,could result in either exaggerating insignificant features or overlooking important information.Second,conventional methods for detecting anomaly in smart grid rely to a large extent on expertise in electric power field.Excessive manpower has to be involved in analyzing the measurement data collected from intelligent motoring devices while performance of anomaly detection is still less satisfactory because of the increased data size.In contrast,the proposed method ignores the expertise in electric power field and delivers the work of anomaly detection completely to machines such that human workload can be significantly reduced.Third,inherent spatio-temporality and multi-dimensionality of the measurement data collected from intelligent monitoring devices in smart grid result in a high dimension of feature space.Visualize the time series intuitively,then extract the texture features whose dimension is far less than the original time series,that method's time complexity is reduced.It provides technical solutions for real-time anomaly detection.