INVESTIGATING THE AUTOENCODER APPROACH TO DETECT BUILDING ENERGY FAULTS THROUGH THE DEVIATION OF REAL-TIME ENERGY USE FROM RECONSTRUCTED BASELINE

Abstract

Buildings contribute to a large proportion of global energy consumption, making maximising energy efficiency of buildings a high priority to tackle climate change. While the availability of energy data has increased and Building Automation Systems (BAS) have advanced over the years, improper configuration and faulty hardware in these systems have brought about greater energy wastage. While many studies on the effectiveness of autoencoders in fault detection have been carried out, limited research has been done for the building context. This paper investigates how the use of autoencoder can contribute in enhancing fault detection in BAS for building managers to identify and remedy these faults promptly while minimising energy wastage. This study has developed and trained a 1D Convolutional Neural Network autoencoder model to test its prediction accuracy in detecting anomalies in power meter data. The results showed that while higher threshold values set for the model can minimise the overall occurrence of false results, higher threshold values can lead to higher number of false positives than false negatives. Other findings also include the improvement of the model’s recall rate from 0.2% to 87.5% when higher values of noise (+1, +2, +3 standard deviations) are added to data with optimum threshold value set. This study concluded that autoencoders carry great potential in enhancing fault detection in buildings through analysing energy-use deviations, but more studies are required to investigate on the optimisation of autoencoder’s hyperparameters and the choice of autoencoder type that can be best applied in the building context.