A novel data-analytics based process for load profiling and meter-level anomaly detection in building energy consumption time series

The ongoing digitalization of building operations, combined with growing demands for energy efficiency and sustainability, has intensified the need for advanced methodologies to monitor building energy performance and promptly detect infrequent and unexpected consumption patterns. However, many existing anomaly detection approaches either depend on rigid threshold rules lacking adaptability or employ complex black-box models that offer limited interpretability and require high-resolution monitoring infrastructures, which are often unavailable in many existing buildings. In response to these challenges, this paper presents a novel data analytics-based methodology for detecting anomalies in building energy consumption time series, addressing the critical need for precise and scalable tools for energy performance monitoring. Tested on a university campus energy consumption time series, the proposed approach explicitly separates and models the shape and magnitude components of daily electrical load profiles, enabling a detailed characterization of typical consumption behaviors. Historical daily electricity data are clustered to identify recurring load shapes, serving as the groundwork for predictive models that estimate both the expected profile shape and total daily energy use based solely on readily available contextual variables, such as weather conditions and calendar information. By integrating classification and regression techniques with unsupervised learning, the framework generates synthetic daily load profile benchmarks, enabling continuous or recurrent comparisons with actual energy consumption. This supports timely, sub-daily anomaly detection while maintaining low computational cost. Through validation over a 24-week testing period, the framework successfully identified 17 days exhibiting anomalous consumption patterns, corresponding to a total overconsumption of approximately 1400 kWh. These results confirm the robustness and reliability of the approach in defining statistically informed operational baselines and detecting deviations. The method also provides practical diagnostic insight and supports scalable implementation, particularly in contexts where sensor infrastructure is limited.