Research On Anomaly Detection Method Of Building Energy Consumption Based On Deep Learning

Building energy consumption accounts for about one-third of the world’s final energy consumption,and is an important part of electric energy consumption.Due to the widespread occurrence of equipment failures,energy waste,and improper control strategies,building operations have great energy saving potential.Anomaly detection of energy consumption data is one of the most effective strategies for achieving efficient energy conservation during building operations.Traditional machine learning algorithms are widely used in anomaly detection of building energy consumption.However,with the increasing collection range and frequency of smart meters,how to rationally use building energy consumption data,improve energy consumption data utilization,and provide theoretical basis for the safety and reliability of building operations has become a new research hotspot.Traditional data mining methods have limited capabilities in processing big data,while deep learning is characterized by strong robustness and learning ability,as well as strong representation and nonlinear modeling capabilities,and is suitable for deep mining scenarios of building energy consumption data.Therefore,this paper uses the deep learning method to establish a data mining network model for anomaly detection of building energy consumption,extract the correlation features of building energy consumption data,and improve the accuracy of data mining.The main contributions and achievements of this article are as follows:(1)This paper proposes an Anomaly Detection based on Multi-Scale Temporal Convolution Network(ADMS-TCN).At present,time convolution networks have been widely used in temporal information extraction,and even in some aspects have been superior to the short and long term memory networks proposed to solve the common long-term dependency problem in general recursive neural networks.However,there is still a problem of information loss in time convolutional networks.Therefore,this paper proposes a multiscale time convolution structure.Using a multi scale structure implementation model,feature extraction is performed on energy consumption data with different expansion coefficients to achieve complementarity of temporal information to prevent information loss issues caused by exponential expansion coefficients.At the same time,a multi-scale structure is adopted to enable the model to extract long-term relationships between energy consumption data under different receptive fields,ensuring the diversity of extracted temporal information.Finally,comparative verification is performed on the smart grid dataset.It is proved that the proposed multiscale time convolution structure can achieve better detection results compared to various network structures such as time convolution networks,indicating the effectiveness of the proposed multiscale structure.Compared with traditional anomaly detection algorithms and new deep learning based detection algorithms proposed in recent years,the advantages of the model are proved.(2)This paper proposes an Anomaly Detection based on Joint Spatial Temporal Learning(ADJST)model.Currently,most building energy consumption anomaly detection methods based on deep learning focus on the temporal information of energy consumption data,ignoring the spatial information of building energy consumption data.However,for building energy consumption data,time information is important,but spatial information is also an indispensable part.Therefore,this paper proposes a spatiotemporal joint anomaly detection method.A multiscale graph convolution module is proposed to extract spatial information from energy consumption data.At the same time,two types of graphs,short-term correlation graph and long-term rule graph,are proposed to extract short-term correlation features and long-term rule features of energy consumption data,respectively.Effectively prevent feature coupling from affecting the effect of model feature extraction.Comparative validation experiments on smart grid datasets demonstrate the effectiveness of the important structures proposed in this paper.On the other hand,compared with traditional anomaly detection algorithms and related cutting-edge algorithms,the effectiveness and competitive advantages of the algorithm have been proven.(3)This paper proposes an Anomaly Detection base on Anomaly Enhancement and Saliency Analysis(ADAS).There are a large number of contextual anomalies in building energy consumption data,which are not caused by electrical failures,but rather energy waste caused by improper human operations.Such abnormal points are not significantly different from normal data and are difficult to detect.Therefore,this paper proposes an anomaly detection model based on anomaly enhancement and significance analysis to detect context anomalies in energy consumption data that are difficult to identify.The significance analysis module is used to analyze the significance of energy consumption data,and the obtained significance changes are weighted to strengthen the abnormality of building energy consumption data,increasing the difference between normal data and abnormal data.Finally,the significance analysis of the enhanced energy consumption data is conducted.Finally,the LSTM network is used to extract the features of the significant changes obtained,achieving the function of abnormal classification of energy consumption data.The effectiveness of the proposed saliency analysis module and anomaly enhancement module,as well as the advantages of the ADAS algorithm,were verified by comparing various types of building data on the Jiangnan University dataset.In summary,this paper conducts research on building energy consumption anomaly detection based on deep learning,and proposes three anomaly detection models based on deep learning: ADMS-TCN,ADJST,and ADAS.Experiments on relevant energy consumption datasets have verified the excellent performance of the algorithm proposed in this paper.