Application of data analytics processes for the detection of anomalous energy patterns in buildings

In recent years, the technological development in virtually every sector has often made it possible to consider real-world data – thanks to the ever-growing ease in collecting and storing these information – as an increasingly valuable resource to guide experts and decision-makers in a multitude of tasks. Among these, the analysis of energy consumption in large buildings is one of the areas of research that is subject to continuous innovation and refinements, as more and more data is made available through the installation of systems that ultimately aim at reducing inefficiencies by guiding the users towards a more “energetically responsible” behavior and by detecting potentially anomalous events during building operation. While collecting and storing data has seemingly become effortless, their analysis often still requires a certain degree of expert knowledge for intervention, due to the fact that it is basically impossible to define an unanimous criteria for “correct” or “incorrect” energy behavior at a whole building-level and it is even harder to investigate the individual causes of inefficiencies at a sub-meter-level starting from aggregate data. This work proposes a methodology for anomaly detection and diagnosis in large non-residential buildings that is built upon one of the newest and most promising techniques for time series analysis, the Matrix Profile (MP). Starting from an extensive review of the existing works that have contributed to the development of the Matrix Profile, its critical issues in the research field of energy data analytics are examined and a variation of the original technique, called Contextual Matrix Profile (CMP), is adopted for analysis on daily load profiles of power demand data measured by a monitoring system connected to a Medium Voltage/Low Voltage (MV/LV) transformation cabin of a university campus (i.e., Politecnico di Torino). Conventional supervised and unsupervised learning techniques, such as clustering and regression trees, are employed for the purpose of grouping together examined days with similar power demand profiles and set up the required input parameters for the CMP, while the anomaly detection step is based on the CMP output and on the combined results of two techniques – the “elbow” method and the boxplot – in order to find out the optimal number of days to be marked as “anomalous”. The root causes of unexpected behaviors in anomalous days are then investigated by defining a metric that ranks sub-loads in terms of their impact on the anomaly at a meter-level. Climatic conditions are also taken into account with the aim of providing possible explanations for the behavior of sub-loads that, during their operation, are particularly influenced by factors related to seasonality, such as external air temperature.