Research On Energy Consumption Anomaly Detection Based On The Self-attention Mechanism Of The Time Series Autoencoder

In recent years,with the continuous development of the Internet and big data technology and the promotion of public building energy consumption systems,massive energy consumption data has accumulated in the building energy consumption platform.For the goal of building energy saving,abnormal energy consumption detection is particularly critical.The energy consumption anomaly detection task aims to automatically detect various usage anomalies of energy consumption,while saving the cost of manual anomaly detection,and also helps to find the problem causing the anomaly in time to prevent the anomaly from continuing to spread.In addition,abnormal energy consumption often has an adverse impact on energy use planning and consumption control,the detection of abnormal energy consumption is also of great significance to the overall energy distribution and control.The abnormal detection method of building energy consumption has developed rapidly in recent years,but the following problems still need to be solved in current research:(1)Data noise processing problem: the existing methods often ignore the data caused by the fluctuation of voltage and current in real applications,incorrectly classifying data noise as an abnormal sample of energy consumption,resulting in a decrease in the abnormal detection performance of the model;(2)The problem of extracting timing information: the existing methods lack sufficient consideration of the time series information present in the sensor data,which limits the detection of the model Abnormal ability;(3)Lack of abnormal labeling problem: Abnormal category labels of energy consumption data usually need to rely on expert knowledge for manual labeling.Due to the huge labor cost,there are a lot of unlabeled energy consumption data.In response to the above problems,this paper proposes an energy consumption anomaly detection algorithm based on the self-attention mechanism of the time series autoencoder.To solve the problem(1),this paper proposes to use variational autoencoders to extract hidden variable representations of data features.By introducing Gaussian noise for sampling training,the model is more robust to noise and its generalization ability.To solve the problem(2),this paper uses a multi-layer two-way long-term and short-term memory neural network and a self-attention mechanism to fully extract the time series information of the data and improve the data's representation ability.To solve the problem(3),this paper further proposes a semi-supervised energy consumption anomaly detection framework based on Kmeans clustering,which alleviates the problem of dependence on anomaly annotation.Based on the real energy consumption data set of a university and the UCR open source Italy Power Demand data set and Arrowhead data set,this paper designs and conducts an extensive comparison experiment.The experimental results show that the model proposed in this paper performs better in terms of accuracy,recall,F Score and other indicators than the current mainstream model method in the industry,which verifies the effectiveness of the algorithm.In addition,this paper also conducts a qualitative analysis on the recognition and differentiation effects of pattern anomalies through visual case analysis.The results show that the method in this paper can effectively distinguish subsequence anomalies and contextual anomalies in pattern anomalies.