Development of a data-driven methodology for the detection and diagnosis of meter-level energy anomalies in buildings

In 2023, electricity consumption in Italy reached 287.4 TWh, with the building sector alone accounting for approximately 40% of this total. This significant share underscores the pressing need for effective energy monitoring and management strategies within buildings to support both operational efficiency and long-term sustainability goals. In this context, Energy Management and Information Systems (EMIS) have emerged as essential tools, offering integrated platforms for real-time data collection, performance benchmarking, and decision support. Among the various functionalities provided by EMIS, Anomaly Detection and Diagnosis (ADD) is particularly valuable. It enables the identification of abnormal consumption patterns and potential system inefficiencies in energy consumption, serving as a critical leverage for energy optimization and supporting decision making. However, traditional data-driven ADD approaches, often based solely on statistical or machine learning methods, which tend to oversimplify or ignore the physical-operational context of building systems. This limitation restricts diagnostic capabilities and the ability to identify root causes of energy anomalies. To address the identified gap, this thesis presents a hybrid Anomaly Detection and Diagnosis methodology integrated into an Energy Information System (EIS). The approach combines unsupervised detection, supervised learning, and statistical methods to improve the reliability of both detection and diagnosis. A key contribution is a novel meter-level ADD process capable of detecting anomalies in electrical energy consumption time series and performing diagnoses using sub-load level data. The framework supports multi-scale analysis by operating at both aggregated and sub-load levels. A notable innovation is the integration of thermal sensitivity analysis, which enables automatic identification and characterization of temperature-dependent electrical loads. This prevents misclassification of infrequent patterns caused by extreme weather conditions. Additionally, a probabilistic diagnostic layer, based on Bayesian Networks, is developed to localize anomalies within the electrical load hierarchy, track their propagation, and quantify their impact relative to a baseline of expected normal operation. The methodology is validated using real energy data from the Politecnico di Torino campus, confirming its robustness and applicability in a complex real-world environment. The result is a comprehensive and interpretable ADD framework that effectively integrates data-driven methods with domain expertise to enhance energy management in buildings.