Exploring Deep Time-Series Imaging for Anomaly Detection of Building Energy Consumption

Abstract

Although deep anomaly detection (DAD) is crucial to optimize energy management in smart buildings, there is a lack of efficient research investigating DAD of energy consumption time-series. Besides, analyzing one-dimensional (1D) energy time-series does not provide many options to detect abnormal energy con-sumption. By contrast, converting 1D signals to two-dimensional (2D) representations open the doors to benefit from the success of time-series imaging in related research fields. In this paper, we explore transfer learning in anomaly detection of energy consumption. The presented method converts time-series inputs to images, which in turn serve as inputs for pre-trained CNN models. Specifically, this helps benefit from the inherent spatial invariance since the best characteristics are efficiently provided for convolutional layers. The experimental evaluation on real and simulated datasets demonstrates the promising performance of the proposed method, primarily when the support vector machine (SVM) is used to classify the features extracted using AlexNet. Accordingly, 92.24% and 99.19% weighted F1-score rates have been reached under both scenarios, respectively.