Attention boosted autoencoder for building energy anomaly detection

Significant energy savings can be realised from buildings if deviations the usual operating conditions are detected early, and appropriate measures taken. Building anomaly detection techniques automate identifying such instances by leveraging high dimensional data collected installed smart meters. Autoencoders allow for dimensionality reduction also model underlying distribution. However, these models treat features as independent quantities. In contrast, current work investigates an attention mechanism with autoencoder to include correlations among features. The value addition is demonstrated comparing model’s reconstruction ability ANN-based on synthetic datasets. study identifies that adding layer enables encoder–decoder architecture robust outliers in training data, thereby reducing preprocessing required. Further, tested a real-world dataset, maps generated used interpret across time dimension, establishing human-interpretable way understand model.