Energy Disaggregation via Deep Temporal Dictionary Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Signal Processing

Background:

• Energy disaggregation (ED) is crucial for understanding household electricity consumption.
• Existing methods often struggle with the complex temporal dynamics of electricity signals.

Purpose of the Study:

• To develop a novel nonlinear dictionary learning (DL) model for energy disaggregation.
• To capture deep temporal structures in electricity signals for accurate device contribution analysis.

Main Methods:

• Modeled ED as a temporal DL problem using a Long Short-Term Memory Autoencoder (LSTM-AE).
• Proposed a Deep Temporal DL (DTDL) model to learn nonlinear dictionaries in the latent space of the LSTM-AE.
• Developed a new optimization algorithm to simultaneously train the LSTM-AE and the dictionary.

Main Results:

• The DTDL model effectively learns representative temporal features of electricity signals.
• Sparse codes derived from the learned atoms indicate individual device contributions to total consumption.
• Achieved outstanding performance on the Reference ED Data Set compared to state-of-the-art methods.

Conclusions:

• The DTDL model is the first DL approach to effectively model deep temporal structures for ED.
• Demonstrated superior accuracy, precision, recall, and F-score in energy disaggregation tasks.