Comparative Anomaly Detection in Commercial Building Energy Consumption: Z-Score, LSTM Autoencoder, Isolation Forest, and Ensemble Approaches at Portfolio Scale

Anomaly detection in commercial building energy consumption is a critical capability for smart grid management, energy procurement optimisation, and the prevention of energy waste and meter fraud. Despite the availability of large-scale smart meter datasets and the theoretical promise of deep learning-based anomaly detection, systematic comparative evaluation of methods at portfolio scale, across hundreds of buildings simultaneously, hasas received limited attention in the literature. This paper presents a comprehensive comparison of four anomaly detection approaches applied to 36.4 million half-hourly electricity records from 1,112 commercial buildings: a statistical Z-score method, an LSTM Autoencoder trained to reconstruct normal consumption patterns, an Isolation Forest trained on temporal features, and an ensemble method combining approaches. Applied to the BDG2 dataset, the Z-score method flags 348,338 anomalies (0.955% of records) with severity breakdown LOW 211,654 / MEDIUM 62,505 / HIGH 42,319 / CRITICAL 31,860. The LSTM Autoencoder flags 1,899,045 (5.208%), trained to best validation loss 0.001596 in 23.7 minutes with thresholds at 90th percentile 0.007284, 95th percentile 0.008697, 99th percentile 0.016473. The ensemble flags 2,223,908 (6.099%). Probable cause analysis identifies unexpected weekend peaks (49.5%), unusual demand spikes (21.9%), heating/cooling failures (15.0%), equipment left on outside business hours (7.8%), and low consumption events (5.8%) as the top anomaly causes. Operational guidance for method selection is provided based on use case requirements.