A Deep-Learning Approach for Detecting Anomalies in Commercial Building Load Profiles

Smart meters have become increasingly popular for load profile monitoring and diagnostic systems. Applying data analysis methods makes obtaining information about electrical load profiles and optimizing monitoring and diagnostics easier. Diagnostics in a smart grid system can be related to smart meters that mismeasure abnormal electrical energy usage. Furthermore, cyber-attacks are also a concern that is considered an anomaly. The study introduces deep learning reconstruction models capable of detecting anomalies in daily energy consumption data for twenty-three City College buildings. The research put forward three different models for consideration: a CNN-autoencoder, an LSTM autoencoder, and a CNN-CNN Autoencoder and LSTM-LSTM Autoencoder. Based on their ability to learn the pattern of regular consumption unsupervised, these models were evaluated for detecting anomalies using reconstruction error. The LSTM-LSTM autoencoder models proposed in the study outperformed the benchmarks in detecting anomalies in the building. They also differed in distinguishing between anomalies and standard electrical load profiles.