Collective Contextual Anomaly Detection for Building Energy Consumption

Commercial and residential buildings are responsible for a substantial portion of total global energy consumption and as a result make a significant contribution to global carbon emissions. Hence, energy-saving goals that target buildings can have a major impact in reducing environmental damage. During building operation, a significant amount of energy is wasted due to equipment and human-related faults. To reduce waste, today’s smart buildings monitor energy usage with the aim of identifying abnormal consumption behaviour and notifying the building manager to implement appropriate energy-saving procedures. To this end, this research proposes the ensemble anomaly detection (EAD) framework. The EAD is a generic framework that combines several anomaly detection classifiers using majority voting. This anomaly detection classifiers are formed using existing machine learning algorithm. It is assumed that each anomaly classifier has equal weight. More importantly, to ensure diversity of anomaly classifiers, the EAD is implemented by combining pattern-based and predictionbased anomaly classifiers. For this reason, this research also proposes a new pattern-based anomaly classifier, the collective contextual anomaly detection using sliding window (CCADSW) framework. The CCAD-SW, which is also a machine leaning-based framework that identifies anomalous consumption patterns using overlapping sliding windows. The EAD framework combines the CCAD-SW, which is implemented using autoencoder, with two predictionbased anomaly classifiers that are implemented using the support vector regression and random forest machine-learning algorithms. In addition, it determines an ensemble threshold that yields an anomaly classifier with optimal anomaly detection capability and false positive minimization. Results show that the EAD performs better than the individual anomaly detection classifiers. In the EAD framework, the optimal ensemble anomaly classifier is not attained by combining the individual learners at their respective optimal performance levels. Instead, an ensemble threshold combination that yields the optimal anomaly classifier was identified by searching through the ensemble threshold space. The research was evaluated using real-world data provided by Powersmiths, located in Brampton, Ontario, Canada.