Deep Learning-Based Anomaly Detection in Energy Consumption Data of Buildings: A Comparative Study

In this study, we evaluate the performance of four different encoder-decoder models for anomaly detection in building smart meter data using a reconstruction-based approach. The results are compared against an ARIMA-based model, which serves as the baseline. Model training and evaluation are carried out using Nested Cross-Validation for time series, a robust model assessment technique that includes an inner loop for hyperparameter tuning and an outer loop for performance assessment. Our findings indicate that the Variational Autoencoder consistently outperforms the other models in terms of anomaly detection. However, the model performance is highly sensitive to the choice of z-score threshold, as well as the quality of the input data and of the anomaly labels.