We propose a novel framework to address the problem of detecting anomalies in building electricity consumption profiles. Our method is based on two sequential steps, which combine machine learning clustering and regression methods. The first step separates weekly anomalous consumption profiles from
regular ones, for a selected timespan. This is achieved through an unsupervised machine learning clustering method applied on a representation of weekly profiles in a two-dimensional space. The results of the clustering method are used to train a regression model which predicts the future behavior of the time series. Any measured consumption which deviates from the predicted value of the regression model is flagged as anomalous, and this could potentially trigger an alarm in the system. Results are discussed and performances are compared with respect to a simple regression model. Possible applications of this method for real-time anomaly detection are briefly discussed.
Keywords Anomaly Detection, Unsupervised Machine Learning, Building Electricity Consumption