Click here to flash read.
Energy management systems (EMS) rely on (non)-intrusive load monitoring
(N)ILM to monitor and manage appliances and help residents be more energy
efficient and thus more frugal. The robustness as well as the transfer
potential of the most promising machine learning solutions for (N)ILM is not
yet fully understood as they are trained and evaluated on relatively limited
data. In this paper, we propose a new approach for load monitoring in building
EMS based on dimensionality expansion of time series and transfer learning. We
perform an extensive evaluation on 5 different low-frequency datasets. The
proposed feature dimensionality expansion using video-like transformation and
resource-aware deep learning architecture achieves an average weighted F1 score
of 0.88 across the datasets with 29 appliances and is computationally more
efficient compared to the state-of-the-art imaging methods. Investigating the
proposed method for cross-dataset intra-domain transfer learning, we find that
1) our method performs with an average weighted F1 score of 0.80 while
requiring 3-times fewer epochs for model training compared to the non-transfer
approach, 2) can achieve an F1 score of 0.75 with only 230 data samples, and 3)
our transfer approach outperforms the state-of-the-art in precision drop by up
to 12 percentage points for unseen appliances.
No creative common's license